Update version and HISTORY.md for 11.5.2 release

Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/14887

This reverts two changes to the range-lock endpoint comparator and restores the prior, direction-agnostic behavior:
- https://github.com/facebook/rocksdb/pull/14831 ("Fix range lock manager crash with reverse comparator"), which made RangeTreeLockManager::CompareDbtEndpoints reverse-aware, and
- https://github.com/facebook/rocksdb/pull/14880 (a follow-up that only corrected the equal-length / point-range branch).

Why: #14831 taught the endpoint comparator about reverse-ordered column families by flipping the suffix tie-break based on comparator direction. However, callers that use range locking on a reverse-ordered CF already flip the start/end endpoints themselves before calling lock_range() -- for example, MyRocks does this in ha_rocksdb::set_range_lock() (see its "RangeFlagsShouldBeFlippedForRevCF" comment). The two corrections stack and produce left > right again, which:
(a) trips paranoid_invariant(compare(left, right) <= 0) in toku::locktree::try_acquire_lock() and aborts on equality/point locks, and
(b) for range scans, mis-orders the endpoints and produces an incorrect lock range, leading to silent correctness issues under concurrency.
#14880 only addressed (a) (the equal-length / point-range branch), not (b), so it was an incomplete fix at the wrong layer.

Decision: keep CompareDbtEndpoints direction-agnostic and make it an explicit caller requirement to pass a flipped (swapped start/end, with negated infimum/supremum flags) range for range locking to work on a reverse-ordered column family. A function-level comment documenting this contract is added so reverse-awareness is not reintroduced here.

Removes the RangeLockWithReverseComparator unit test added by #14831 (it asserts the reverted direction-flipped ordering). Keeps a regression test that a point range lock [infimum(K), supremum(K)] works on a reverse-comparator column family.

Reviewed By: xingbowang

Differential Revision: D109713989

fbshipit-source-id: 4ea0999c92645e1adc9cd1114d924cad142f2001

.github/actions/pre-steps/action.yml

@@ -2,6 +2,13 @@ name: pre-steps
 runs:
   using: composite
   steps:
+  - name: Dump ulimits for diagnostics
+    run: |
+      echo "=== ulimits (soft) ==="
+      ulimit -a -S || true
+      echo "=== ulimits (hard) ==="
+      ulimit -a -H || true
+    shell: bash
   - name: Install lld linker for faster builds
     run: apt-get update -y && apt-get install -y lld 2>/dev/null || true
     shell: bash

.github/actions/windows-build-steps/action.yml

@@ -3,7 +3,7 @@ inputs:
   suite-run:
     description: Comma-separated list of test suites to run (empty to skip C++ tests)
     required: false
-    default: arena_test,db_basic_test,db_test,db_test2,db_merge_operand_test,bloom_test,c_test,coding_test,crc32c_test,dynamic_bloom_test,env_basic_test,env_test,hash_test,random_test
+    default: arena_test,db_basic_test,db_test,db_etc2_test,db_merge_operand_test,bloom_test,c_test,coding_test,crc32c_test,dynamic_bloom_test,env_basic_test,env_test,hash_test,random_test
   run-java:
     description: Whether to run Java tests
     required: false

.github/workflows/pr-jobs.yml

@@ -373,14 +373,17 @@ jobs:
           crash_duration: 240
     container:
       image: ghcr.io/facebook/rocksdb_ubuntu:22.1
-      options: --shm-size=16gb
+      options: --shm-size=16gb --ulimit nofile=1048576:1048576
     steps:
     - uses: actions/checkout@v4.1.0
     - uses: "./.github/actions/pre-steps"
     - uses: "./.github/actions/setup-ccache"
       with:
         cache-key-prefix: mini-crashtest
-    - run: ulimit -S -n `ulimit -H -n` && make V=1 -j8 CRASH_TEST_EXT_ARGS='--duration=${{ matrix.crash_duration }} --max_key=2500000' ${{ matrix.crash_test_target }}
+    - run: |
+        ulimit -S -n $(ulimit -Hn) # Raise open files soft limit to hard limit
+        echo "Updated ulimit -Hn: $(ulimit -Hn) -Sn: $(ulimit -Sn)"
+        make V=1 -j8 CRASH_TEST_EXT_ARGS='--duration=${{ matrix.crash_duration }} --max_key=2500000' ${{ matrix.crash_test_target }}
     - uses: "./.github/actions/teardown-ccache"
       if: always()
     - uses: "./.github/actions/post-steps"
@@ -534,7 +537,7 @@ jobs:
           suite_run: db_test
           run_java: "false"
         - test_shard: other
-          suite_run: arena_test,db_basic_test,db_test2,db_merge_operand_test,bloom_test,c_test,coding_test,crc32c_test,dynamic_bloom_test,env_basic_test,env_test,hash_test,random_test
+          suite_run: arena_test,db_basic_test,db_etc2_test,db_merge_operand_test,bloom_test,c_test,coding_test,crc32c_test,dynamic_bloom_test,env_basic_test,env_test,hash_test,random_test
           run_java: "false"
         - test_shard: java
           suite_run: ""

.gitignore

@@ -52,7 +52,7 @@ GRTAGS
 GTAGS
 rocksdb_dump
 rocksdb_undump
-db_test2
+db_etc2_test
 trace_analyzer
 block_cache_trace_analyzer
 io_tracer_parser

BUCK

@@ -4868,6 +4868,12 @@ cpp_unittest_wrapper(name="db_encryption_test",
             extra_compiler_flags=[])
 
 
+cpp_unittest_wrapper(name="db_etc2_test",
+            srcs=["db/db_etc2_test.cc"],
+            deps=[":rocksdb_test_lib"],
+            extra_compiler_flags=[])
+
+
 cpp_unittest_wrapper(name="db_etc3_test",
             srcs=["db/db_etc3_test.cc"],
             deps=[":rocksdb_test_lib"],
@@ -5024,12 +5030,6 @@ cpp_unittest_wrapper(name="db_test",
             extra_compiler_flags=[])
 
 
-cpp_unittest_wrapper(name="db_test2",
-            srcs=["db/db_test2.cc"],
-            deps=[":rocksdb_test_lib"],
-            extra_compiler_flags=[])
-
-
 cpp_unittest_wrapper(name="db_universal_compaction_test",
             srcs=["db/db_universal_compaction_test.cc"],
             deps=[":rocksdb_test_lib"],

CLAUDE.md

@@ -149,6 +149,26 @@ Cache management is critical for RocksDB's performance.
 
 When reviewing RocksDB code (or preparing code for review), use this checklist:
 
+### Contract Boundaries
+- [ ] Is each behavior owned by the right layer? High-level policy (for example,
+  "compaction wants this I/O mode") should live at the caller/policy layer, while
+  lower layers should expose generic mechanisms (for example, "open a fresh
+  reader", "skip shared cache insertion", or "use these FileOptions").
+- [ ] Do comments and names describe local contracts rather than leaking a
+  specific caller's rationale into reusable APIs? Generic code should not need to
+  know about one current use case unless the API itself is intentionally
+  use-case-specific.
+- [ ] Does each flag or parameter control one coherent behavior? If one boolean
+  starts implying ownership, cache policy, I/O mode, prefetching, and caller
+  identity, split it into explicit flags or an options struct.
+- [ ] Could a future caller use this lower-level API without accidentally
+  inheriting assumptions from compaction, backup, user reads, or a particular
+  table format? If not, tighten the contract with assertions, clearer names, or
+  a narrower API.
+- [ ] Are implementation details not being used as policy signals? Prefer an
+  explicit contract over inferring behavior from incidental fields such as file
+  options, cache handles, or current table-reader state.
+
 ### Correctness
 - [ ] Does the change preserve database semantics (e.g., snapshot isolation, key ordering)?
 - [ ] Are all error cases handled appropriately?
@@ -204,19 +224,84 @@ The following patterns emerged as frequent sources of review feedback:
 
 10. **Platform Compatibility:** Ensure changes work correctly on all supported platforms (Linux, Windows, macOS) and with all supported compilers (GCC, Clang, MSVC).
 
+11. **Contract Boundary Leaks:** When a change plumbs a new option or use-case
+specific behavior through multiple subsystems, review the call chain for
+contract leaks. Caller-specific rationale belongs at the call site or public API
+documentation; reusable layers should expose precise, layer-local capabilities.
+Watch especially for comments mentioning one caller in generic code, booleans
+that silently bundle several behaviors, and downstream code inferring policy
+from an implementation detail instead of an explicit option.
+
 ---
 
 ## Important tips
 
 ### Build system
-* There are 3 build system. Make, CMake, BUCK(meta internal).
+* There are 3 build system. Make for git clones, BUCK (meta internal) for hg
+  clones, and CMake for some special cases.
 * When a new .cc file is added, update Makefile, CMakeLists.txt, src.mk, BUCK.
 * Don't manually edit BUCK file, after updating src.mk, run
     /usr/local/bin/python3 buckifier/buckify_rocksdb.py to update it
-* Use make to build and run the test. CMake and BUCK are not used locally.
-* Use `make dbg` command to build all of the unit test in debug mode.
 * For -j in make command, use the number of CPU cores to decide it.
 
+### When to run `make clean` (avoid mixing build modes)
+
+The Makefile does **not** track build mode, so object files from a prior
+build are silently reused even when compiled with different flags, leading
+to confusing linker errors, sanitizer false negatives, ODR violations, or
+"phantom" bugs.
+
+Run `make clean` before switching any of these:
+* **`ASSERT_STATUS_CHECKED=1` ↔ unset** — changes the `Status` class layout (ABI break).
+* **Sanitizer builds** — toggling any of `COMPILE_WITH_ASAN=1`,
+    `COMPILE_WITH_UBSAN=1`, `COMPILE_WITH_TSAN=1` on/off.
+* `DEBUG_LEVEL=0` (release) ↔ `DEBUG_LEVEL=1` (debug, default for `make dbg`).
+* Different compilers, `OPT` levels, or other flags affecting codegen/ABI.
+
+**Notable exception:** `DEBUG_LEVEL=2` can be safely mixed with
+`DEBUG_LEVEL=1` — rebuild a subset of files with `DEBUG_LEVEL=2` to get
+extra/more accurate runtime checks for those files without a full clean.
+
+When in doubt, `make clean` is cheap insurance compared to chasing a
+phantom bug.
+
+### Source checks
+* Run `make check-sources` before committing. This catches non-ASCII
+    characters in source files and other source-level issues that CI will
+    reject. In particular, **do not use Unicode characters** (em dashes,
+    smart quotes, etc.) in comments or strings -- use ASCII equivalents
+    (`--` instead of em dash, `'` instead of smart quote, etc.).
+
+### RTTI and dynamic_cast
+* Production code and `db_stress` must build in **release mode
+    (`-fno-rtti`)**. Do not use `dynamic_cast` anywhere except unit tests.
+    Use `static_cast_with_check` from `util/cast_util.h` (validates with
+    `dynamic_cast` in debug builds, plain `static_cast` in release).
+* Unit tests (`*_test.cc`) are built in debug mode with RTTI enabled.
+
+### Cross-platform / portability
+Local `make` only exercises Linux with GCC/Clang, but CI
+(`.github/workflows/pr-jobs.yml` and `nightly.yml`) gates on a much wider
+matrix, so portability breaks are invisible locally until CI fails. Code must
+build (and where noted, run tests) across:
+
+| Axis | Must support |
+|------|--------------|
+| OS | Linux (x86_64 + ARM), macOS, Windows |
+| Compiler | GCC, Clang (libstdc++ **and** libc++), AppleClang, **MSVC (VS2022)**, MinGW (Linux cross-compile, build-only, no gflags) |
+| Build system | Make, CMake, and BUCK (internal) -- keep all in sync (see "Build system" above) |
+| Config | release (`-fno-rtti`), `ASSERT_STATUS_CHECKED`, ASAN/UBSAN/TSAN, folly, unity build, JNI/Java |
+
+Treat these as constraints to satisfy and infer the specifics from them before
+adding any system header, libc call, or compiler-specific construct. The most
+common trap: anything that compiles under GCC/Clang on Linux but not under
+**MSVC/MinGW** -- e.g. unguarded POSIX-only headers/functions (`<unistd.h>`,
+`<sys/*.h>`, `getpid`, `_exit`, ...) or GCC/Clang extensions
+(`__attribute__`, `__builtin_*`, VLAs, `alloca`). Prefer the `port::`/`Env`
+abstractions; otherwise guard with `#ifdef OS_WIN` (POSIX `<unistd.h>` ->
+Windows `<process.h>`). Because libc++ is also tested, include what you use
+rather than relying on libstdc++ transitive includes.
+
 ### Unit Test
 * After all of the unit tests are added, review them and try to extract common
     reusable utility functions to reduce code duplication due to copy past between
@@ -232,6 +317,9 @@ The following patterns emerged as frequent sources of review feedback:
     COERCE_CONTEXT_SWITCH=1 make {test_binary}
     ./{test_binary} --gtest_filter="*YourTestName*" --gtest_repeat=5
     ```
+* For CI-style flaky tests that do not reproduce with `gtest_parallel.py`,
+    `--gtest_repeat`, or normal coerce-mode runs, inspect
+    `tools/gtest_parallel_repro.py --help`.
 
 ### Unit test dedup guidelines
 * Extract helper functions for repeated patterns such as object

CMakeLists.txt

@@ -654,6 +654,8 @@ if(USE_FOLLY)
     FMT_INST_PATH)
     exec_program(ls ARGS -d ${FOLLY_INST_PATH}/../gflags* OUTPUT_VARIABLE
     GFLAGS_INST_PATH)
+    exec_program(ls ARGS -d ${FOLLY_INST_PATH}/../glog* OUTPUT_VARIABLE
+    GLOG_INST_PATH)
     set(Boost_DIR ${BOOST_INST_PATH}/lib/cmake/Boost-1.83.0)
     if(EXISTS ${FMT_INST_PATH}/lib64)
       set(fmt_DIR ${FMT_INST_PATH}/lib64/cmake/fmt)
@@ -661,6 +663,42 @@ if(USE_FOLLY)
       set(fmt_DIR ${FMT_INST_PATH}/lib/cmake/fmt)
     endif()
     set(gflags_DIR ${GFLAGS_INST_PATH}/lib/cmake/gflags)
+    if(EXISTS ${GLOG_INST_PATH}/lib64/cmake/glog)
+      set(GLOG_CMAKE_DIR ${GLOG_INST_PATH}/lib64/cmake/glog)
+    elseif(EXISTS ${GLOG_INST_PATH}/lib/cmake/glog)
+      set(GLOG_CMAKE_DIR ${GLOG_INST_PATH}/lib/cmake/glog)
+    endif()
+    if(NOT GLOG_CMAKE_DIR)
+      find_library(GETDEPS_GLOG_LIBRARY NAMES glogd glog
+        PATHS ${GLOG_INST_PATH}/lib64 ${GLOG_INST_PATH}/lib
+        NO_DEFAULT_PATH)
+      if(NOT GETDEPS_GLOG_LIBRARY OR
+         NOT EXISTS ${GLOG_INST_PATH}/include/glog/logging.h)
+        message(FATAL_ERROR "Could not find getdeps glog under "
+          "${GLOG_INST_PATH}")
+      endif()
+      set(GLOG_CMAKE_DIR ${CMAKE_CURRENT_BINARY_DIR}/getdeps-glog-cmake)
+      file(MAKE_DIRECTORY ${GLOG_CMAKE_DIR})
+      file(WRITE ${GLOG_CMAKE_DIR}/glog-config.cmake
+        "if(NOT TARGET glog::glog)\n"
+        "  add_library(glog::glog UNKNOWN IMPORTED)\n"
+        "  set_target_properties(glog::glog PROPERTIES\n"
+        "    IMPORTED_LOCATION \"${GETDEPS_GLOG_LIBRARY}\"\n"
+        "    INTERFACE_INCLUDE_DIRECTORIES \"${GLOG_INST_PATH}/include\")\n"
+        "endif()\n"
+        "set(glog_FOUND TRUE)\n"
+        "set(Glog_FOUND TRUE)\n"
+        "set(GLOG_FOUND TRUE)\n"
+        "set(GLOG_LIBRARY \"${GETDEPS_GLOG_LIBRARY}\")\n"
+        "set(GLOG_LIBRARIES \"${GETDEPS_GLOG_LIBRARY}\")\n"
+        "set(GLOG_INCLUDE_DIR \"${GLOG_INST_PATH}/include\")\n")
+    endif()
+    set(glog_DIR ${GLOG_CMAKE_DIR})
+    set(Glog_DIR ${GLOG_CMAKE_DIR})
+    # Unlike gflags, glog may also be installed system-wide. Pre-resolve the
+    # getdeps copy so folly-config.cmake cannot fall back to another version.
+    find_package(glog CONFIG REQUIRED PATHS ${GLOG_CMAKE_DIR}
+                 NO_DEFAULT_PATH)
 
     exec_program(sed ARGS -i 's/gflags_shared//g'
     ${FOLLY_INST_PATH}/lib/cmake/folly/folly-targets.cmake)
@@ -1453,6 +1491,7 @@ if(WITH_TESTS)
         db/db_clip_test.cc
         db/db_dynamic_level_test.cc
         db/db_encryption_test.cc
+        db/db_etc2_test.cc
         db/db_etc3_test.cc
         db/db_flush_test.cc
         db/db_inplace_update_test.cc
@@ -1476,7 +1515,6 @@ if(WITH_TESTS)
         db/db_table_properties_test.cc
         db/db_tailing_iter_test.cc
         db/db_test.cc
-        db/db_test2.cc
         db/db_logical_block_size_cache_test.cc
         db/db_universal_compaction_test.cc
         db/db_wal_test.cc

HISTORY.md

@@ -1,6 +1,86 @@
 # Rocksdb Change Log
 > NOTE: Entries for next release do not go here. Follow instructions in `unreleased_history/README.txt`
 
+## 11.5.2 (06/25/2026)
+### Bug Fixes
+* Reverted PR14831 that made range_lock_manager aware of reverse-order CF
+
+## 11.5.1 (06/24/2026)
+### Bug Fixes
+* Fixed a bug where closing a read-only DB instance could delete live SST files created by a concurrent read-write DB sharing the same directory.
+
+## 11.5.0 (06/16/2026)
+### New Features
+* External table readers that open files through `ExternalTableOptions::fs` now update RocksDB SST/file-read statistics and file IO listener callbacks, making external file IO activity visible in existing read metrics.
+* Added `DB::PrepareFileIngestion()`, a two-phase form of `IngestExternalFile`/`IngestExternalFiles`. `PrepareFileIngestion()` performs all of the work that does not require the DB mutex (validating the arguments, reading each external file's metadata, and linking/copying the files into the DB) and returns an opaque `FileIngestionHandle`. `DB::CommitFileIngestionHandle()` (or `DB::CommitFileIngestionHandles()` for several handles at once) then makes the prepared files visible; multiple handles are committed atomically in a single MANIFEST write, or none are. `FileIngestionHandle::Abort()` (or simply destroying the handle) cancels a prepared ingestion and rolls it back. This gives flexibility to the application to Prepare a file separately from when its committed, and may shorten an applications critical path on `IngestExternalFile`.
+* Added two stats histograms reporting the per-call latency (in microseconds) of each successful `IngestExternalFile`/`IngestExternalFiles` call, split by phase: `rocksdb.ingest.external.file.prepare.micros` (argument validation and reading/validating/linking the external files, which does not block live writes) and `rocksdb.ingest.external.file.run.micros` (the ingestion performed under the DB mutex while live writes are blocked). Failed ingestions are not recorded.
+* Added `DBOptions::use_direct_io_for_compaction_reads` (default false). When enabled, compaction-input SST reads use `O_DIRECT` while user reads remain buffered, avoiding page-cache eviction of hot user-read data by sequential compaction scans. Pair with `use_direct_io_for_flush_and_compaction = true` on write-heavy workloads for direct I/O on both compaction inputs and outputs. Rejected at Open when combined with `allow_mmap_reads = true`. No-op when `use_direct_reads = true` is already set (since all reads are already direct).
+
+### Public API Changes
+* `DB::GetPreparedFileInfoForExternalSstIngestion()` can now prepare metadata for a live DB-generated SST file so it can be passed to `IngestExternalFileArg::file_infos`. The input path must exactly match a live table file owned by the source DB, and the returned metadata handle must outlive the ingestion.
+* `ExternalSstFileInfo` gained a `prepared_file_info` member produced by `SstFileWriter::Finish`, and `IngestExternalFileArg` gained a `file_infos` field: a vector of borrowed `const PreparedFileInfo*` pointers, parallel to `external_files`. The owning `ExternalSstFileInfo::prepared_file_info` handle must outlive the ingestion. When set, `IngestExternalFiles()` reuses the supplied per-file metadata instead of re-opening and scanning each file to recompute it, avoiding that extra I/O.
+* Corrected the public C++ option spelling from `memtable_veirfy_per_key_checksum_on_seek` to `memtable_verify_per_key_checksum_on_seek`. RocksDB continues to accept the old misspelled OPTIONS-file key for compatibility, while newly serialized OPTIONS files use the corrected key.
+* Added experimental read-scoped block buffer provider API, configured through `ReadOptions::read_scoped_block_buffer_provider`, for supported block-based table iterator scans and MultiScan reads to use caller-provided read-scoped storage for final data-block contents. When configured, supported provider-backed scan data-block reads bypass the data-block cache. The provider is ignored when mmap reads are enabled. RocksDB may still use ordinary temporary scratch for serialized block bytes, such as when a block may be compressed. Get and MultiGet do not currently provide API guarantees for this provider.
+* Added `FileSystem::SyncFile()` and `Env::SyncFile()` public APIs for syncing or fsyncing a file by name without requiring callers to reopen it as writable.
+
+### Behavior Changes
+* The default `compression` for column families is now `kLZ4Compression` rather than `kSnappyCompression`. This affects only column families that do not explicitly set `compression`, and only newly-written SST files. The change is fully compatible: existing data remains readable with no migration needed, as RocksDB selects the decompressor per block. LZ4 offers slightly better compression ratios and decompression CPU efficiency vs. Snappy and similar compression CPU, tested across various server CPUs. When support is not compiled in, the fallback path for default compression is LZ4 -> Snappy -> NoCompression.
+* Disable parallel compression (ignore `CompressionOptions::parallel_threads`) for fast built-in compressors: Snappy, LZ4 (accelerated, not LZ4HC), and accelerated levels of ZSTD (level < 0). These do not generally benefit from parallel compression. This behavior can be overridden with a custom Compressor from a custom CompressionManager.
+* `PosixFileSystem::OptimizeForCompactionTableRead` now delegates to the base `FileSystem::OptimizeForCompactionTableRead` implementation before applying its Linux-only compaction-readahead clamp (added for #12038). The returned `FileOptions::use_direct_reads` therefore reflects `DBOptions::use_direct_reads`, consistent with the base class and with other `FileSystem` implementations. Previously the override ignored the base implementation and keyed both the returned options and the Linux readahead clamp off the incoming `FileOptions` rather than `DBOptions`. At the in-tree call site the incoming `FileOptions::use_direct_reads` already matches `DBOptions::use_direct_reads`, so this is effectively a no-op for existing configurations; it removes a latent inconsistency where a caller passing `FileOptions` whose `use_direct_reads` disagreed with `DBOptions` would have had the global flag silently ignored for compaction-input reads on Linux. (Note: the `#12038` readahead clamp still keys off the global `use_direct_reads`; it is not skipped for the compaction-only `use_direct_io_for_compaction_reads` path, since that flag enables O_DIRECT after this hook runs.)
+* Brought more unity to `kLZ4Compression` and `kLZ4HCCompression` by giving each access to the other's compression levels. Negative (fast) values previously only available to LZ4 and positive (slow) values previously only available to LZ4HC are now available to both, so configuring a non-default `compression_opts.level` now selects the LZ4 compressor variant. This is a behavior change for previously tolerated but dubious configurations such as positive compression level with `kLZ4Compression`. `kLZ4Compression` and `kLZ4HCCompression` keep their effective default levels, equivalent to -1 and 9 respectively.
+* Removed a discontinuity in `kZSTD` compression levels at `compression_opts.level == 0` by mapping that setting to `level = -1` rather than to `level = 3` as the ZSTD library does internally. This improves the compression auto-tuning landscape.
+
+### Bug Fixes
+* Fixed a bug where the range lock manager would crash with an assertion failure when using a reverse comparator column family.
+
+### Performance Improvements
+* Reduced commit latency when ingesting many external files by allowing `IngestExternalFileOptions::file_opening_threads` to open table readers for committed ingested files using multiple threads.
+* Reduced commit latency for large external file ingestions into the last level by adding `IngestExternalFileOptions::prefetch_lmax_index_and_filter_blocks`, which can skip commit-time index and filter block prefetching for cache-backed table metadata.
+
+## 11.4.0 (06/02/2026)
+### Public API Changes
+* Added `rocksdb_options_set_memtable_batch_lookup_optimization()` and `rocksdb_options_get_memtable_batch_lookup_optimization()` to the C API, exposing the existing `AdvancedColumnFamilyOptions::memtable_batch_lookup_optimization` field. This allows C API users (and downstream language bindings) to enable the skip-list memtable's batch-lookup optimization for `MultiGet`, which caches the search path between consecutive keys and reduces per-key cost from O(log N) to O(log d) where d is the distance between consecutive keys.
+* Added `rocksdb_readoptions_set_optimize_multiget_for_io()` and `rocksdb_readoptions_get_optimize_multiget_for_io()` to the C API, exposing the existing `ReadOptions::optimize_multiget_for_io` field. This allows C API users (and downstream language bindings) to opt out of the multi-level parallel MultiGet path, which only takes effect when the library is built with `USE_COROUTINES`.
+* Added `rocksdb_block_based_table_index_block_search_type_auto` enum constant and the `rocksdb_block_based_options_set_uniform_cv_threshold()` setter to the C API. The constant exposes `BlockSearchType::kAuto`, and the setter exposes `BlockBasedTableOptions::uniform_cv_threshold`. Both are required for `kAuto` index-block search to take effect from the C API: without setting `uniform_cv_threshold >= 0`, the per-block `is_uniform` footer bit is never written, and `kAuto` falls back to binary search at read time.
+* Added `EventListener::OnDBShutdownBegin`, a callback that fires once when a DB begins shutdown, including when RocksDB cleans up a failed `DB::Open()` attempt.
+* Added a new PerfContext counter `blob_cache_read_byte` for blob cache bytes read.
+
+### Behavior Changes
+* Remote compaction now falls back to local compaction when the primary cannot deserialize a successful `CompactionServiceResult` before installing any remote output files.
+* StringToMap (rocksdb/convenience.h) now preserves the outer braces of nested values so each map entry is in self-contained form and can be embedded directly into another `key=value;` string (e.g. SetOptions) without losing inner ';' delimiters. A new symmetric MapToString utility is provided.
+
+### Bug Fixes
+* Fixed a bug where `AbortAllCompactions()` could leave automatic compaction work unscheduled when aborting before the background worker picked it, causing compaction and write stalls until DB restart.
+* Fix a bug where db had a false positive compaction corruption error due to remote compaction service result with `has_accurate_num_input_records=false` not being serialized.
+* Fixed WritePrepared TransactionDB cleanup after retryable commit or rollback write failures so prepared transactions remain rollbackable instead of leaving unresolved prepared state.
+* Fix a rare corruption bug for tiered compaction that incorrectly moved last level range tombstones into proximal level. This corruption error is surfaced when force_consistency_checks is enabled.
+
+### Performance Improvements
+* Reduce the likelihood of user-facing metadata operations blocking on MANIFEST rotation when file creation is slow, such as on remote storage. MANIFEST write batches containing foreground edits from `CreateColumnFamily()`, `DropColumnFamily()`, `CreateColumnFamilyWithImport()`, `IngestExternalFile()`, and `DeleteFilesInRanges()` now get a relaxed file size threshold for triggering MANIFEST rotation; background-only MANIFEST write batches, such as compaction and flush, continue to use the normal threshold. This change should make auto-tuning manifest file size more attractive (see `max_manifest_file_size` and `max_manifest_space_amp_pct` options).
+
+
+## 11.3.0 (05/15/2026)
+### New Features
+* Add experimental DB option `async_wal_precreate` to precreate the next WAL file in a background thread and reduce foreground WAL rotation latency. The option is sanitized to false when WAL recycling is enabled.
+* Added a new `EventListener::OnCompactionPreCommit` callback that fires after a compaction job finishes but before its input files are released (i.e. while `FileMetaData::being_compacted` is still true). Listeners that maintain bookkeeping of which files are currently being compacted can clean up such state in this new callback to avoid races with concurrent compaction picking, where another thread might pick up the same files for a new compaction immediately after `being_compacted` is flipped back to false but before `OnCompactionCompleted` fires. The default implementation is a no-op so this is not a breaking change.
+* Add mutable DBOption `optimize_manifest_for_recovery` (default false). When enabled, RocksDB can reduce recovery work after a clean shutdown, which may lower DB::Open latency on warm reopens.
+* Added public utility APIs `ParseCompressionNameForDisplay()` to convert `TableProperties::compression_name` into a human-readable compression name for both legacy and format_version 7+ SST metadata, including custom `CompressionManager`-provided display names for custom compression types.
+* Add `reuse_manifest_on_open` DBOption (default false). When enabled, DB::Open reuses the existing MANIFEST file for append instead of creating a fresh one, avoiding the cost of serializing the entire database state into a new MANIFEST on the first post-open write. To prevent this feature from interfering with manifest file size auto-tuning, an extra forward-compatible field is now always added to the MANIFEST (to track the last "compacted" size).
+
+### Behavior Changes
+* Read-only open with `error_if_wal_file_exists=true` now tolerates empty WAL files so empty precreated WALs do not prevent inspection.
+* WriteCommitted TransactionDB now matches WritePrepared and WriteUnprepared compaction filtering in both single- and two-write-queue modes: a compaction filter's FilterMergeOperand will not be invoked on merge operands at or below the latest published sequence number.
+
+### Bug Fixes
+* Fixed blob-backed wide-column merge reads to preserve correct status
+propagation and resolution across memtable, read-only, and secondary DB
+paths.
+* Fixed a bug where `DB::GetCreationTimeOfOldestFile()` could return inaccurate results instead of the real creation time when called shortly after opening a legacy DB (one whose manifest lacks `file_creation_time`) with `open_files_async = true`. The API now waits for background SST file loading to complete only when needed; modern DBs are unaffected.
+* Fixed merge reads against wide-column/blob-backed base values to preserve precise failure statuses, including `GetMergeOperands()` and direct-write memtable reads.
+* Fix bug in range tombstone synthesis that covers live keys added during an IngestExternalFile
+* Reject the empty string as a column family name in `DB::CreateColumnFamily` / `DB::CreateColumnFamilies`. Previously such calls returned OK and a usable handle, but the column family was not persisted in the manifest, so any data written to it was silently lost on DB reopen.
+* Fixed a bug where a WriteCommitted TransactionDB using commit-bypass WBWI ingestion could drop an entry that is still visible at the published sequence boundary.
+
 ## 11.2.0 (04/18/2026)
 ### New Features
 * Added experimental `DBOptions::fast_sst_open` option. When enabled, RocksDB retrieves opaque file system metadata for SST files after flush, compaction, and external file ingestion, persists it in the MANIFEST, and passes it back to the file system on subsequent file opens to accelerate DB open time.

Makefile

@@ -327,6 +327,9 @@ missing_make_config_paths := $(shell				\
 $(foreach path, $(missing_make_config_paths), \
 	$(warning Warning: $(path) does not exist))
 
+# This (the first rule) must depend on "all".
+default: all
+
 ifeq ($(PLATFORM), OS_AIX)
 # no debug info
 else ifneq ($(PLATFORM), IOS)
@@ -485,9 +488,6 @@ ifdef ROCKSDB_MODIFY_NPHASH
   PLATFORM_CXXFLAGS += -DROCKSDB_MODIFY_NPHASH=1
 endif
 
-# This (the first rule) must depend on "all".
-default: all
-
 WARNING_FLAGS = -W -Wextra -Wall -Wsign-compare -Wshadow \
   -Wunused-parameter
 
@@ -887,23 +887,43 @@ coverage: clean
 parallel_tests = $(patsubst %,parallel_%,$(PARALLEL_TEST))
 .PHONY: gen_parallel_tests $(parallel_tests)
 # Shard size controls how many test cases run per process. The actual number
-# of shards per binary is: min(ceil(test_count / GTEST_SHARD_SIZE), NCORES * 8)
+# of shards per binary is: min(ceil(test_count / shard_size), NCORES * 8)
 # This adapts to machine size: many small shards on beefy machines, fewer
 # larger shards on CI (typically 2-4 cores).
 GTEST_SHARD_SIZE ?= 10
 NCORES ?= $(shell nproc 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null || echo 4)
 
+# Per-binary overrides for GTEST_SHARD_SIZE, used to chop up binaries whose
+# individual test cases are slow enough that the default of 10 leaves one
+# shard on the critical path of "make check". Format: whitespace-separated
+# tokens "BINARY:SIZE". Sizes were tuned from `make suggest-slow-tests`
+# output to keep max shard time under ~20s on a beefy box. Reducing too far
+# adds per-shard fixed overhead, so don't use 1 unless individual tests are
+# >5s.
+SHARD_SIZE_OVERRIDES ?= \
+  point_lock_manager_stress_test:1 \
+  external_sst_file_test:1 \
+  external_sst_file_basic_test:2 \
+  db_test:3 \
+  compaction_service_test:1
+
 $(parallel_tests):
 	$(AM_V_at)TEST_BINARY=$(patsubst parallel_%,%,$@); \
   TEST_COUNT=` \
     (./$$TEST_BINARY --gtest_list_tests 2>/dev/null || echo "  list_failure") \
     | grep -c '^ '`; \
 	if [ "$$TEST_COUNT" -le 0 ]; then TEST_COUNT=1; fi; \
+	SHARD_SIZE=$(GTEST_SHARD_SIZE); \
+	for o in $(SHARD_SIZE_OVERRIDES); do \
+		case "$$o" in \
+			"$$TEST_BINARY":*) SHARD_SIZE=$${o#*:} ;; \
+		esac; \
+	done; \
 	MAX_SHARDS=$$(( $(NCORES) * 8 )); \
-	NUM_SHARDS=$$(( (TEST_COUNT + $(GTEST_SHARD_SIZE) - 1) / $(GTEST_SHARD_SIZE) )); \
+	NUM_SHARDS=$$(( (TEST_COUNT + SHARD_SIZE - 1) / SHARD_SIZE )); \
 	if [ "$$NUM_SHARDS" -gt "$$MAX_SHARDS" ]; then NUM_SHARDS=$$MAX_SHARDS; fi; \
 	if [ "$$NUM_SHARDS" -le 0 ]; then NUM_SHARDS=1; fi; \
-	echo "  Generating $$NUM_SHARDS shards for $$TEST_BINARY ($$TEST_COUNT tests)"; \
+	echo "  Generating $$NUM_SHARDS shards for $$TEST_BINARY ($$TEST_COUNT tests, shard_size=$$SHARD_SIZE)"; \
 	SHARD_IDX=0; \
 	while [ "$$SHARD_IDX" -lt "$$NUM_SHARDS" ]; do \
 		if [ -n "$(CI_TOTAL_SHARDS)" ] && [ $$(( $$SHARD_IDX % $(CI_TOTAL_SHARDS) )) -ne $(CI_SHARD_INDEX) ]; then \
@@ -929,31 +949,64 @@ gen_parallel_tests:
 	$(AM_V_at)$(FIND) t -type f -name 'run-*' -exec rm -f {} \;
 	$(MAKE) $(parallel_tests)
 
-# Reorder input lines (which are one per test) so that the
-# longest-running tests appear first in the output.
-# Do this by prefixing each selected name with its duration,
-# sort the resulting names, and remove the leading numbers.
-# FIXME: the "100" we prepend is a fake time, for now.
-# FIXME: squirrel away timings from each run and use them
-# (when present) on subsequent runs to order these tests.
+# Reorder input lines (one per test or per shard) so the longest-running
+# tests/shards start first under "make check" parallel scheduling. The trick
+# is to prefix matched names with a fake duration (100), sort numerically
+# descending, then strip the prefix. Slow-matched tests therefore come ahead
+# of all unmatched tests; ordering within each group is unspecified.
 #
-# Without this reordering, these two tests would happen to start only
-# after almost all other tests had completed, thus adding 100 seconds
-# to the duration of parallel "make check".  That's the difference
-# between 4 minutes (old) and 2m20s (new).
+# Why this matters: gnu_parallel hands out jobs in input order. If a slow
+# test starts only after most others have finished, the run is bottlenecked
+# on its tail. Front-loading slow tests overlaps them with the bulk of
+# faster ones for much better wall-clock time.
 #
-# 152.120 PASS t/DBTest.FileCreationRandomFailure
-# 107.816 PASS t/DBTest.EncodeDecompressedBlockSizeTest
+# Maintenance:
+#   1) Run `make check` so LOG is up-to-date.
+#   2) Run `make suggest-slow-tests` to see candidate slow binaries.
+#   3) Add binaries with high per-shard time (slow critical path) or high
+#      total time across many shards (need early queueing for fan-out).
+# Each alternative is wrapped in `^.*NAME.*$$` so the *whole input line* is
+# captured into $$1; then `s,(...),100 $$1,` prepends "100 " to the line.
 #
-# With sharded test execution, prioritize binaries known to be slow.
-# These generate many shards and should start early for good load balancing.
+# Tiers below are based on observed timings (see suggest-slow-tests).
+# Tier 1: max single-shard time >= 30s (critical-path bottlenecks).
+# Tier 2: max single-shard time 15-30s.
+# Tier 3: huge total time across many tiny shards; front-loading them keeps
+#         the tail of the run busy while big shards finish.
 slow_test_regexp = \
-	^.*block_based_table_reader_test.*$$|^.*table_test.*$$|^.*block_test.*$$|^.*write_prepared_transaction_test.*$$|^.*transaction_test.*$$|^.*external_sst_file_test.*$$|^.*db_wal_test.*$$|^.*db_with_timestamp_basic_test.*$$|^.*db_test-.*$$
+	^.*point_lock_manager_stress_test.*$$|^.*db_test.*$$|^.*external_sst_file_test.*$$|^.*compaction_service_test.*$$|^.*corruption_test.*$$|^.*comparator_db_test.*$$|^.*external_sst_file_basic_test.*$$|^.*rate_limiter_test.*$$|^.*db_compaction_test.*$$|^.*write_prepared_transaction_test.*$$|^.*db_merge_operator_test.*$$|\
+	^.*db_dynamic_level_test.*$$|^.*db_bloom_filter_test.*$$|^.*error_handler_fs_test.*$$|^.*merge_helper_test.*$$|^.*transaction_test.*$$|^.*db_kv_checksum_test.*$$|^.*inlineskiplist_test.*$$|\
+	^.*db_with_timestamp_basic_test.*$$|^.*table_test.*$$|^.*db_wal_test.*$$|^.*block_based_table_reader_test.*$$|^.*block_test.*$$
 prioritize_long_running_tests =						\
   perl -pe 's,($(slow_test_regexp)),100 $$1,'				\
     | sort -k1,1gr							\
     | sed 's/^[.0-9]* //'
 
+# Helper: print binaries observed to be slow in the last `make check` run.
+# Use this to decide whether `slow_test_regexp` above needs updating.
+# Aggregates per-binary across shards; flags any binary whose max single-shard
+# time is >= 20s OR whose total time is >= 200s.
+.PHONY: suggest-slow-tests
+suggest-slow-tests:
+	@if [ ! -s LOG ]; then \
+	  echo "No (or empty) LOG file. Run 'make check' first." >&2; exit 1; \
+	fi
+	@bash -c '$(quoted_perl_command)' < LOG | awk '\
+	  { \
+	    t=$$1; name=$$3; \
+	    sub(/^t\/run-/, "", name); \
+	    sub(/-shard-[0-9]+$$/, "", name); \
+	    total[name] += t; \
+	    if (t > max[name]) max[name] = t; \
+	    count[name]++; \
+	  } \
+	  END { \
+	    printf "%8s %8s %5s  %s\n", "TOTAL_s", "MAX_s", "SHRDS", "BINARY"; \
+	    for (n in total) \
+	      if (max[n] >= 20 || total[n] >= 200) \
+	        printf "%8.1f %8.1f %5d  %s\n", total[n], max[n], count[n], n; \
+	  }' | (read header; echo "$$header"; sort -k2,2gr)
+
 # "make check" uses
 # Run with "make J=1 check" to disable parallelism in "make check".
 # Run with "make J=200% check" to run two parallel jobs per core.
@@ -990,7 +1043,7 @@ check_0:
 		        awk -v s=$(CI_SHARD_INDEX) -v n=$(CI_TOTAL_SHARDS) '(NR-1)%n==s'; \
 		      else cat; fi; \
 		fi; \
-		find t -name 'run-*' -print; \
+		$(FIND) t -name 'run-*' -print; \
 	} \
 	  | $(prioritize_long_running_tests)				\
 	  | grep -E '$(tests-regexp)'					\
@@ -1012,7 +1065,7 @@ valgrind_check_0:
 	  '  run "make watch-log" in a separate window' '';		\
 	{								\
 	  printf './%s\n' $(filter-out $(PARALLEL_TEST) %skiplist_test options_settable_test, $(TESTS));		\
-	  find t -name 'run-*' -print; \
+	  $(FIND) t -name 'run-*' -print; \
 	}								\
 	  | $(prioritize_long_running_tests)				\
 	  | grep -E '$(tests-regexp)'					\
@@ -1230,11 +1283,17 @@ clean-not-downloaded: clean-ext-libraries-bin clean-rocks clean-not-downloaded-r
 
 clean-rocks:
 # Not practical to exactly match all versions/variants in naming (e.g. debug or not)
-	rm -f ${LIBNAME}*.so* ${LIBNAME}*.a
-	rm -f $(BENCHMARKS) $(TOOLS) $(TESTS) $(PARALLEL_TEST) $(MICROBENCHS)
-	rm -rf $(CLEAN_FILES) ios-x86 ios-arm scan_build_report
-	$(FIND) . -name "*.[oda]" -exec rm -f {} \;
-	$(FIND) . -type f \( -name "*.gcda" -o -name "*.gcno" \) -exec rm -f {} \;
+	@echo Removing link targets
+	@rm -f ${LIBNAME}*.so* ${LIBNAME}*.a $(BENCHMARKS) $(TOOLS) $(TESTS) $(PARALLEL_TEST) $(MICROBENCHS)
+	@echo Finding and cleaning other files
+	@rm -rf $(CLEAN_FILES) ios-x86 ios-arm scan_build_report
+	@if $(FIND) Makefile -regextype awk &> /dev/null; then \
+		$(FIND) . -regextype awk \
+			-regex '.*/([.].*|third-party)' -prune -o \
+			-type f -regex '.*[.]([oda]|gcda|gcno)' -exec rm -f {} \; ; \
+	else \
+		$(FIND) . -name "*.[oda]" -exec rm -f {} \; ; \
+	fi
 
 clean-rocksjava: clean-rocks
 	rm -rf jl jls
@@ -1247,7 +1306,7 @@ clean-ext-libraries-all:
 	rm -rf bzip2* snappy* zlib* lz4* zstd*
 
 clean-ext-libraries-bin:
-	find . -maxdepth 1 -type d \( -name bzip2\* -or -name snappy\* -or -name zlib\* -or -name lz4\* -or -name zstd\* \) -prune -exec rm -rf {} \;
+	$(FIND) . -maxdepth 1 -type d \( -name bzip2\* -or -name snappy\* -or -name zlib\* -or -name lz4\* -or -name zstd\* \) -prune -exec rm -rf {} \;
 
 tags:
 	ctags -R .
@@ -1525,7 +1584,7 @@ db_open_with_config_test: $(OBJ_DIR)/db/db_open_with_config_test.o $(TEST_LIBRAR
 db_test: $(OBJ_DIR)/db/db_test.o $(TEST_LIBRARY) $(LIBRARY)
 	$(AM_LINK)
 
-db_test2: $(OBJ_DIR)/db/db_test2.o $(TEST_LIBRARY) $(LIBRARY)
+db_etc2_test: $(OBJ_DIR)/db/db_etc2_test.o $(TEST_LIBRARY) $(LIBRARY)
 	$(AM_LINK)
 
 db_etc3_test: $(OBJ_DIR)/db/db_etc3_test.o $(TEST_LIBRARY) $(LIBRARY)

build_tools/getdeps_fallback_mirror.py

@@ -1,62 +1,188 @@
 #!/usr/bin/env python3
+#  Copyright (c) Meta Platforms, Inc. and affiliates.
+#  This source code is licensed under both the GPLv2 (found in the COPYING file in the root directory)
+#  and the Apache 2.0 License (found in the LICENSE.Apache file in the root directory).
 """
 Pre-download packages with unreliable mirrors using fallback mirrors.
 Reads package info from folly's getdeps manifest files.
 """
-import sys
-import os
-import hashlib
-import subprocess
 import configparser
+import hashlib
+import os
+import shutil
+import sys
+import urllib.request
+
+DOWNLOAD_TIMEOUT_SECONDS = 120
+DOWNLOAD_CHUNK_BYTES = 64 * 1024
+MAX_DOWNLOAD_BYTES = 50 * 1024 * 1024
+
+MIRROR_FALLBACKS = {
+    "ftpmirror.gnu.org/gnu/": [
+        "https://mirrors.kernel.org/gnu/",
+        "https://ftpmirror.gnu.org/gnu/",
+        "https://ftp.gnu.org/gnu/",
+    ],
+    "ftp.gnu.org/gnu/": [
+        "https://mirrors.kernel.org/gnu/",
+        "https://ftpmirror.gnu.org/gnu/",
+        "https://ftp.gnu.org/gnu/",
+    ],
+}
+
+# These packages must have URLs matching MIRROR_FALLBACKS; other packages are
+# left for getdeps.py's normal download path.
+PACKAGES_TO_CHECK = ("autoconf", "automake", "libtool", "libiberty")
+
 
 def sha256_file(path):
     """Calculate SHA256 hash of a file."""
     h = hashlib.sha256()
     try:
-        with open(path, 'rb') as f:
-            for chunk in iter(lambda: f.read(65536), b''):
+        with open(path, "rb") as f:
+            for chunk in iter(lambda: f.read(65536), b""):
                 h.update(chunk)
         return h.hexdigest()
     except Exception:
         return None
 
+
 def parse_manifest(manifest_path):
     """Parse a getdeps manifest file to extract download info."""
-    config = configparser.ConfigParser()
+    # folly manifests can contain bare keys in sections unrelated to downloads.
+    config = configparser.ConfigParser(allow_no_value=True, interpolation=None)
     try:
-        config.read(manifest_path)
-        if 'download' in config:
-            return {
-                'url': config['download'].get('url', ''),
-                'sha256': config['download'].get('sha256', ''),
-            }
-    except Exception:
-        pass
+        with open(manifest_path, encoding="utf-8") as manifest_file:
+            config.read_file(manifest_file)
+    except Exception as ex:
+        print(f"  {os.path.basename(manifest_path)}: WARNING - parse failed: {ex}")
+        return None
+
+    if "download" in config:
+        return {
+            "url": config["download"].get("url", ""),
+            "sha256": config["download"].get("sha256", ""),
+        }
     return None
 
+
+def file_size(path):
+    try:
+        return os.path.getsize(path)
+    except Exception:
+        return None
+
+
 def get_fallback_mirrors(url):
     """Get fallback mirror URLs for a given URL."""
-    # Fallback mirror patterns for known unreliable hosts
-    mirror_fallbacks = {
-        "ftp.gnu.org/gnu/": [
-            "https://mirrors.kernel.org/gnu/",
-            "https://ftpmirror.gnu.org/gnu/",
-            "https://ftp.gnu.org/gnu/",
-        ],
-        "ftpmirror.gnu.org/gnu/": [
-            "https://mirrors.kernel.org/gnu/",
-            "https://ftpmirror.gnu.org/gnu/",
-            "https://ftp.gnu.org/gnu/",
-        ],
-    }
-
-    for pattern, mirrors in mirror_fallbacks.items():
+    for pattern, mirrors in MIRROR_FALLBACKS.items():
         if pattern in url:
             # Extract the path after the pattern
             path_start = url.find(pattern) + len(pattern)
             path = url[path_start:]
             return [mirror + path for mirror in mirrors]
-    return [url]  # No fallback, use original
+    return []
+
+
+def download_url(url, filepath):
+    """Download URL to filepath without leaving partial files behind."""
+    tmp_filepath = filepath + ".tmp"
+    if os.path.exists(tmp_filepath):
+        os.remove(tmp_filepath)
+
+    request = urllib.request.Request(
+        url, headers={"User-Agent": "rocksdb-getdeps-fallback/1.0"}
+    )
+    try:
+        with urllib.request.urlopen(
+            request, timeout=DOWNLOAD_TIMEOUT_SECONDS
+        ) as response, open(tmp_filepath, "wb") as output:
+            copied = 0
+            while True:
+                chunk = response.read(DOWNLOAD_CHUNK_BYTES)
+                if not chunk:
+                    break
+
+                copied += len(chunk)
+                if copied > MAX_DOWNLOAD_BYTES:
+                    raise Exception(
+                        f"download exceeds {MAX_DOWNLOAD_BYTES} bytes"
+                    )
+                output.write(chunk)
+        os.replace(tmp_filepath, filepath)
+    finally:
+        if os.path.exists(tmp_filepath):
+            os.remove(tmp_filepath)
+
+
+def prepare_download(package, info, download_dir, cache_dir):
+    url = info["url"]
+    expected_sha256 = info["sha256"]
+    mirrors = get_fallback_mirrors(url)
+    if not mirrors:
+        return False
+
+    if not expected_sha256:
+        print(f"  {package}: WARNING - skipped fallback without sha256")
+        return False
+
+    # getdeps uses format: {package}-{filename}
+    filename = f"{package}-{os.path.basename(url)}"
+    filepath = os.path.join(download_dir, filename)
+    cache_path = os.path.join(cache_dir, filename)
+
+    # Check if already valid.
+    actual_sha256 = sha256_file(filepath) if os.path.exists(filepath) else None
+    if actual_sha256 == expected_sha256:
+        print(f"  {filename}: OK (already downloaded)")
+        return True
+    if actual_sha256 is not None:
+        print(
+            f"  {filename}: WARNING - removing invalid download "
+            f"sha256={actual_sha256}"
+        )
+        os.remove(filepath)
+
+    # The cache is only an opportunistic single-build accelerator; callers
+    # should not share it across concurrent builds without external locking.
+    actual_sha256 = sha256_file(cache_path) if os.path.exists(cache_path) else None
+    if actual_sha256 == expected_sha256:
+        print(f"  {filename}: OK (from cache)")
+        shutil.copy2(cache_path, filepath)
+        return True
+    if actual_sha256 is not None:
+        print(
+            f"  {filename}: WARNING - removing invalid cache "
+            f"sha256={actual_sha256}"
+        )
+        os.remove(cache_path)
+
+    # Try fallback mirrors.
+    for mirror_url in mirrors:
+        print(f"  {filename}: trying {mirror_url}...")
+        try:
+            download_url(mirror_url, filepath)
+        except Exception as ex:
+            print(f"  {filename}: WARNING - download failed: {ex}")
+            continue
+
+        actual_sha256 = sha256_file(filepath)
+        if actual_sha256 == expected_sha256:
+            size = file_size(filepath)
+            print(f"  {filename}: OK (downloaded, {size} bytes)")
+            shutil.copy2(filepath, cache_path)
+            return True
+
+        size = file_size(filepath)
+        print(
+            f"  {filename}: WARNING - sha256 mismatch from {mirror_url}: "
+            f"expected={expected_sha256} actual={actual_sha256} size={size}"
+        )
+        os.remove(filepath)
+
+    print(f"  {filename}: WARNING - all mirrors failed")
+    return False
+
 
 def main():
     if len(sys.argv) != 4:
@@ -64,60 +190,39 @@ def main():
         sys.exit(1)
 
     download_dir, cache_dir, manifests_dir = sys.argv[1], sys.argv[2], sys.argv[3]
+    os.makedirs(download_dir, exist_ok=True)
+    os.makedirs(cache_dir, exist_ok=True)
 
-    # Packages known to have unreliable mirrors
-    packages_to_check = ["autoconf", "automake", "libtool"]
-
-    for package in packages_to_check:
+    checked = 0
+    ready = 0
+    for package in PACKAGES_TO_CHECK:
         manifest_path = os.path.join(manifests_dir, package)
-        if not os.path.exists(manifest_path):
+        if not os.path.isfile(manifest_path):
             continue
 
         info = parse_manifest(manifest_path)
-        if not info or not info['url'] or not info['sha256']:
+        if not info or not info["url"]:
             continue
 
-        # Determine filename from URL
-        url = info['url']
-        expected_sha256 = info['sha256']
-        url_filename = os.path.basename(url)
-
-        # getdeps uses format: {package}-{filename}
-        filename = f"{package}-{url_filename}"
-        filepath = os.path.join(download_dir, filename)
-        cache_path = os.path.join(cache_dir, filename)
-
-        # Check if already valid
-        if os.path.exists(filepath) and sha256_file(filepath) == expected_sha256:
-            print(f"  {filename}: OK (already downloaded)")
+        if not info["sha256"]:
+            print(f"  {package}: WARNING - skipped fallback without sha256")
             continue
 
-        # Check cache
-        if os.path.exists(cache_path) and sha256_file(cache_path) == expected_sha256:
-            print(f"  {filename}: OK (from cache)")
-            subprocess.run(['cp', cache_path, filepath], check=True)
+        if not get_fallback_mirrors(info["url"]):
+            print(
+                f"  {package}: WARNING - skipped fallback without known mirror "
+                f"for {info['url']}"
+            )
             continue
 
-        # Try fallback mirrors
-        mirrors = get_fallback_mirrors(url)
-        downloaded = False
-        for mirror_url in mirrors:
-            print(f"  {filename}: trying {mirror_url}...")
-            try:
-                subprocess.run(['wget', '-q', '-O', filepath, mirror_url], check=True, timeout=120)
-                if sha256_file(filepath) == expected_sha256:
-                    print(f"  {filename}: OK (downloaded)")
-                    subprocess.run(['cp', filepath, cache_path], check=False)
-                    downloaded = True
-                    break
-                else:
-                    os.remove(filepath)
-            except Exception:
-                if os.path.exists(filepath):
-                    os.remove(filepath)
+        checked += 1
+        try:
+            if prepare_download(package, info, download_dir, cache_dir):
+                ready += 1
+        except Exception as ex:
+            print(f"  {package}: WARNING - fallback preparation failed: {ex}")
 
-        if not downloaded:
-            print(f"  {filename}: WARNING - all mirrors failed")
+    print(f"  fallback mirror downloads ready: {ready}/{checked}")
 
 if __name__ == "__main__":
     main()

claude_md/add_option.md

@@ -46,7 +46,7 @@ This document provides guidance on how to add new options to RocksDB's public AP
 
 ## Pattern 1: Adding a Standard Column Family Option
 
-Example reference: commit `94e65a2e0b4f817aa4bfa4c96cdf867e7980d7bc` (memtable_veirfy_per_key_checksum_on_seek)
+Example reference: commit `94e65a2e0b4f817aa4bfa4c96cdf867e7980d7bc` (memtable_verify_per_key_checksum_on_seek)
 
 ### Step 1: Define the Option in Public Header
 
@@ -63,7 +63,7 @@ Add the option with documentation in `AdvancedColumnFamilyOptions` struct:
 // operation.
 // This option depends on memtable_protection_bytes_per_key to be non zero.
 // If memtable_protection_bytes_per_key is zero, no validation is performed.
-bool memtable_veirfy_per_key_checksum_on_seek = false;
+bool memtable_verify_per_key_checksum_on_seek = false;
 ```
 
 ### Step 2: Add to Internal Options Structs
@@ -74,14 +74,14 @@ Add to `MutableCFOptions` struct (or `ImmutableCFOptions` for immutable options)
 
 ```cpp
 // In MutableCFOptions constructor from Options:
-memtable_veirfy_per_key_checksum_on_seek(
-    options.memtable_veirfy_per_key_checksum_on_seek),
+memtable_verify_per_key_checksum_on_seek(
+    options.memtable_verify_per_key_checksum_on_seek),
 
 // In MutableCFOptions default constructor:
-memtable_veirfy_per_key_checksum_on_seek(false),
+memtable_verify_per_key_checksum_on_seek(false),
 
 // In MutableCFOptions struct member declarations:
-bool memtable_veirfy_per_key_checksum_on_seek;
+bool memtable_verify_per_key_checksum_on_seek;
 ```
 
 ### Step 3: Register for Serialization/Deserialization
@@ -91,9 +91,9 @@ bool memtable_veirfy_per_key_checksum_on_seek;
 Add to the options type info map for serialization:
 
 ```cpp
-{"memtable_veirfy_per_key_checksum_on_seek",
+{"memtable_verify_per_key_checksum_on_seek",
  {offsetof(struct MutableCFOptions,
-           memtable_veirfy_per_key_checksum_on_seek),
+           memtable_verify_per_key_checksum_on_seek),
   OptionType::kBoolean, OptionVerificationType::kNormal,
   OptionTypeFlags::kMutable}},
 ```
@@ -101,8 +101,8 @@ Add to the options type info map for serialization:
 Add logging in `MutableCFOptions::Dump()`:
 
 ```cpp
-ROCKS_LOG_INFO(log, "memtable_veirfy_per_key_checksum_on_seek: %d",
-               memtable_veirfy_per_key_checksum_on_seek);
+ROCKS_LOG_INFO(log, "memtable_verify_per_key_checksum_on_seek: %d",
+               memtable_verify_per_key_checksum_on_seek);
 ```
 
 ### Step 4: Update Options Helper
@@ -112,8 +112,8 @@ ROCKS_LOG_INFO(log, "memtable_veirfy_per_key_checksum_on_seek: %d",
 Add to `UpdateColumnFamilyOptions()`:
 
 ```cpp
-cf_opts->memtable_veirfy_per_key_checksum_on_seek =
-    moptions.memtable_veirfy_per_key_checksum_on_seek;
+cf_opts->memtable_verify_per_key_checksum_on_seek =
+    moptions.memtable_verify_per_key_checksum_on_seek;
 ```
 
 ### Step 5: Add to Options Settable Test
@@ -123,7 +123,7 @@ cf_opts->memtable_veirfy_per_key_checksum_on_seek =
 Add to the test string in `ColumnFamilyOptionsAllFieldsSettable`:
 
 ```cpp
-"memtable_veirfy_per_key_checksum_on_seek=1;"
+"memtable_verify_per_key_checksum_on_seek=1;"
 ```
 
 ### Step 6: Add db_stress Support
@@ -131,23 +131,23 @@ Add to the test string in `ColumnFamilyOptionsAllFieldsSettable`:
 **File: `db_stress_tool/db_stress_common.h`**
 
 ```cpp
-DECLARE_bool(memtable_veirfy_per_key_checksum_on_seek);
+DECLARE_bool(memtable_verify_per_key_checksum_on_seek);
 ```
 
 **File: `db_stress_tool/db_stress_gflags.cc`**
 
 ```cpp
 DEFINE_bool(
-    memtable_veirfy_per_key_checksum_on_seek,
-    ROCKSDB_NAMESPACE::Options().memtable_veirfy_per_key_checksum_on_seek,
-    "Sets CF option memtable_veirfy_per_key_checksum_on_seek.");
+    memtable_verify_per_key_checksum_on_seek,
+    ROCKSDB_NAMESPACE::Options().memtable_verify_per_key_checksum_on_seek,
+    "Sets CF option memtable_verify_per_key_checksum_on_seek.");
 ```
 
 **File: `db_stress_tool/db_stress_test_base.cc`**
 
 ```cpp
-options.memtable_veirfy_per_key_checksum_on_seek =
-    FLAGS_memtable_veirfy_per_key_checksum_on_seek;
+options.memtable_verify_per_key_checksum_on_seek =
+    FLAGS_memtable_verify_per_key_checksum_on_seek;
 ```
 
 ### Step 7: Add db_bench Support
@@ -156,12 +156,12 @@ options.memtable_veirfy_per_key_checksum_on_seek =
 
 ```cpp
 // Flag definition (near related flags):
-DEFINE_bool(memtable_veirfy_per_key_checksum_on_seek, false,
-            "Sets CF option memtable_veirfy_per_key_checksum_on_seek");
+DEFINE_bool(memtable_verify_per_key_checksum_on_seek, false,
+            "Sets CF option memtable_verify_per_key_checksum_on_seek");
 
 // Apply flag to options (in InitializeOptionsFromFlags or similar):
-options.memtable_veirfy_per_key_checksum_on_seek =
-    FLAGS_memtable_veirfy_per_key_checksum_on_seek;
+options.memtable_verify_per_key_checksum_on_seek =
+    FLAGS_memtable_verify_per_key_checksum_on_seek;
 ```
 
 ### Step 8: Add Crash Test Support
@@ -169,7 +169,7 @@ options.memtable_veirfy_per_key_checksum_on_seek =
 **File: `tools/db_crashtest.py`**
 
 ```python
-"memtable_veirfy_per_key_checksum_on_seek": lambda: random.choice([0] * 7 + [1]),
+"memtable_verify_per_key_checksum_on_seek": lambda: random.choice([0] * 7 + [1]),
 ```
 
 Also add constraint handling in `finalize_and_sanitize()` if needed:
@@ -177,7 +177,7 @@ Also add constraint handling in `finalize_and_sanitize()` if needed:
 ```python
 # only skip list memtable representation supports paranoid memory checks
 if dest_params.get("memtablerep") != "skip_list":
-    dest_params["memtable_veirfy_per_key_checksum_on_seek"] = 0
+    dest_params["memtable_verify_per_key_checksum_on_seek"] = 0
 ```
 
 ### Step 9: Add Release Note
@@ -185,7 +185,7 @@ if dest_params.get("memtablerep") != "skip_list":
 **File: `unreleased_history/new_features/<descriptive_name>.md`**
 
 ```markdown
-A new flag memtable_veirfy_per_key_checksum_on_seek is added to AdvancedColumnFamilyOptions. When it is enabled, it will validate key checksum along the binary search path on skiplist based memtable during seek operation.
+A new flag memtable_verify_per_key_checksum_on_seek is added to AdvancedColumnFamilyOptions. When it is enabled, it will validate key checksum along the binary search path on skiplist based memtable during seek operation.
 ```
 
 ---

db/arena_wrapped_db_iter.cc

@@ -30,13 +30,78 @@ inline static SequenceNumber GetSeqNum(const DBImpl* db, const Snapshot* s) {
 Status ArenaWrappedDBIter::GetProperty(std::string prop_name,
                                        std::string* prop) {
   if (prop_name == "rocksdb.iterator.super-version-number") {
+    if (!internal_iter_initialized_) {
+      *prop = std::to_string(sv_number_);
+      return Status::OK();
+    }
     // First try to pass the value returned from inner iterator.
     if (!db_iter_->GetProperty(prop_name, prop).ok()) {
       *prop = std::to_string(sv_number_);
     }
     return Status::OK();
   }
-  return db_iter_->GetProperty(prop_name, prop);
+  if (!EnsureInternalIteratorInitialized(nullptr).ok()) {
+    return db_iter_->status();
+  }
+  return db_iter_->GetProperty(std::move(prop_name), prop);
+}
+
+void ArenaWrappedDBIter::CleanupDeferredSuperVersion() {
+  if (deferred_sv_ != nullptr) {
+    assert(db_impl_ != nullptr);
+    db_impl_->CleanupIteratorSuperVersion(
+        deferred_sv_, read_options_.background_purge_on_iterator_cleanup);
+    deferred_sv_ = nullptr;
+  }
+}
+
+void ArenaWrappedDBIter::DestroyDBIter() {
+  db_iter_->~DBIter();
+  CleanupDeferredSuperVersion();
+}
+
+void ArenaWrappedDBIter::ColumnFamilyDataUnrefDeleter::operator()(
+    ColumnFamilyData* cfd) const {
+  if (cfd == nullptr) {
+    return;
+  }
+  assert(db_impl != nullptr);
+
+  InstrumentedMutexLock lock(db_impl->mutex());
+  cfd->UnrefAndTryDelete();
+}
+
+void ArenaWrappedDBIter::DestroyDBIterAndArena() {
+  DestroyDBIter();
+  arena_.~Arena();
+}
+
+Status ArenaWrappedDBIter::EnsureInternalIteratorInitialized(
+    const MultiScanArgs* scan_opts) {
+  if (internal_iter_initialized_) {
+    return Status::OK();
+  }
+  if (db_impl_ == nullptr || deferred_cfd_ == nullptr ||
+      deferred_sv_ == nullptr) {
+    Status s = Status::InvalidArgument(
+        "Internal iterator cannot be initialized without deferred DB state");
+    db_iter_->set_status(s);
+    db_iter_->set_valid(false);
+    return s;
+  }
+
+  const MultiScanArgs* pruning_scan_opts =
+      scan_opts != nullptr && scan_opts->HasBoundedScanRanges() ? scan_opts
+                                                                : nullptr;
+  child_read_options_ = read_options_;
+  child_read_options_.snapshot = nullptr;
+  InternalIterator* internal_iter = db_impl_->NewInternalIterator(
+      child_read_options_, deferred_cfd_, deferred_sv_, &arena_, sequence_,
+      /*allow_unprepared_value=*/true, this, pruning_scan_opts);
+  deferred_cfd_ = nullptr;
+  deferred_sv_ = nullptr;
+  SetIterUnderDBIterImpl(internal_iter);
+  return Status::OK();
 }
 
 void ArenaWrappedDBIter::Init(
@@ -44,18 +109,26 @@ void ArenaWrappedDBIter::Init(
     const MutableCFOptions& mutable_cf_options, const Version* version,
     const SequenceNumber& sequence, uint64_t version_number,
     ReadCallback* read_callback, ColumnFamilyHandleImpl* cfh,
-    bool expose_blob_index, bool allow_refresh, ReadOnlyMemTable* active_mem) {
+    bool expose_blob_index, bool allow_refresh, ReadOnlyMemTable* active_mem,
+    DBImpl* db_impl, ColumnFamilyData* cfd) {
   read_options_ = read_options;
+  child_read_options_ = read_options;
   if (!CheckFSFeatureSupport(env->GetFileSystem().get(),
                              FSSupportedOps::kAsyncIO)) {
     read_options_.async_io = false;
   }
   read_options_.total_order_seek |= ioptions.prefix_seek_opt_in_only;
 
-  db_iter_ = DBIter::NewIter(
-      env, read_options_, ioptions, mutable_cf_options,
-      ioptions.user_comparator, /*internal_iter=*/nullptr, version, sequence,
-      read_callback, active_mem, cfh, expose_blob_index, &arena_);
+  if (cfh != nullptr) {
+    db_impl = cfh->db();
+    cfd = cfh->cfd();
+  }
+
+  db_iter_ = DBIter::NewIter(env, read_options_, ioptions, mutable_cf_options,
+                             ioptions.user_comparator,
+                             /*internal_iter=*/nullptr, version, sequence,
+                             read_callback, active_mem, /*cfh=*/nullptr,
+                             expose_blob_index, &arena_, db_impl, cfd);
 
   sv_number_ = version_number;
   allow_refresh_ = allow_refresh;
@@ -65,13 +138,13 @@ void ArenaWrappedDBIter::Init(
 
 void ArenaWrappedDBIter::MaybeAutoRefresh(bool is_seek,
                                           DBIter::Direction direction) {
-  if (cfh_ != nullptr && read_options_.snapshot != nullptr && allow_refresh_ &&
-      read_options_.auto_refresh_iterator_with_snapshot) {
+  if (cfd_ref_ != nullptr && read_options_.snapshot != nullptr &&
+      allow_refresh_ && read_options_.auto_refresh_iterator_with_snapshot) {
     // The intent here is to capture the superversion number change
     // reasonably soon from the time it actually happened. As such,
     // we're fine with weaker synchronization / ordering guarantees
     // provided by relaxed atomic (in favor of less CPU / mem overhead).
-    uint64_t cur_sv_number = cfh_->cfd()->GetSuperVersionNumberRelaxed();
+    uint64_t cur_sv_number = cfd_ref_->GetSuperVersionNumberRelaxed();
     if ((sv_number_ != cur_sv_number) && status().ok()) {
       // Changing iterators' direction is pretty heavy-weight operation and
       // could have unintended consequences when it comes to prefix seek.
@@ -150,13 +223,13 @@ void ArenaWrappedDBIter::DoRefresh(const Snapshot* snapshot,
   // present in the error log, but won't be reflected in the iterator status.
   // This is by design as we expect compaction to clean up those obsolete files
   // eventually.
-  db_iter_->~DBIter();
-
-  arena_.~Arena();
+  DestroyDBIterAndArena();
   new (&arena_) Arena();
 
-  auto cfd = cfh_->cfd();
-  auto db_impl = cfh_->db();
+  auto cfd = cfd_ref_.get();
+  auto db_impl = db_impl_;
+  assert(cfd != nullptr);
+  assert(db_impl != nullptr);
 
   SuperVersion* sv = cfd->GetReferencedSuperVersion(db_impl);
   assert(sv->version_number >= sv_number);
@@ -164,23 +237,27 @@ void ArenaWrappedDBIter::DoRefresh(const Snapshot* snapshot,
   if (read_callback_) {
     read_callback_->Refresh(read_seq);
   }
+  // TODO: Preserve Prepare() scan options across Refresh() so a refreshed
+  // MultiScan iterator can rebuild the same pruned tree.
   Init(env, read_options_, cfd->ioptions(), sv->mutable_cf_options, sv->current,
-       read_seq, sv->version_number, read_callback_, cfh_, expose_blob_index_,
-       allow_refresh_, allow_mark_memtable_for_flush_ ? sv->mem : nullptr);
+       read_seq, sv->version_number, read_callback_, nullptr,
+       expose_blob_index_, allow_refresh_,
+       allow_mark_memtable_for_flush_ ? sv->mem : nullptr, db_impl, cfd);
 
   InternalIterator* internal_iter = db_impl->NewInternalIterator(
       read_options_, cfd, sv, &arena_, read_seq,
       /* allow_unprepared_value */ true, /* db_iter */ this);
   SetIterUnderDBIter(internal_iter);
+  internal_iter_initialized_ = true;
 }
 
 Status ArenaWrappedDBIter::Refresh(const Snapshot* snapshot) {
-  if (cfh_ == nullptr || !allow_refresh_) {
+  if (cfd_ref_ == nullptr || db_impl_ == nullptr || !allow_refresh_) {
     return Status::NotSupported("Creating renew iterator is not allowed.");
   }
   assert(db_iter_ != nullptr);
-  auto cfd = cfh_->cfd();
-  auto db_impl = cfh_->db();
+  auto cfd = cfd_ref_.get();
+  auto db_impl = db_impl_;
 
   // TODO(yiwu): For last_seq_same_as_publish_seq_==false, this is not the
   // correct behavior. Will be corrected automatically when we take a snapshot
@@ -193,6 +270,13 @@ Status ArenaWrappedDBIter::Refresh(const Snapshot* snapshot) {
   TEST_SYNC_POINT("ArenaWrappedDBIter::Refresh:1");
   TEST_SYNC_POINT("ArenaWrappedDBIter::Refresh:2");
 
+  if (!internal_iter_initialized_) {
+    Status s = EnsureInternalIteratorInitialized(nullptr);
+    if (!s.ok()) {
+      return s;
+    }
+  }
+
   while (true) {
     if (sv_number_ != cur_sv_number) {
       DoRefresh(snapshot, cur_sv_number);
@@ -250,25 +334,47 @@ Status ArenaWrappedDBIter::Refresh(const Snapshot* snapshot) {
   return Status::OK();
 }
 
+void ArenaWrappedDBIter::Prepare(const MultiScanArgs& scan_opts) {
+  if (prepare_called_) {
+    db_iter_->set_status(Status::InvalidArgument(
+        "Prepare called more than once on the same iterator"));
+    db_iter_->set_valid(false);
+    return;
+  }
+  prepare_called_ = true;
+
+  Status s = db_iter_->SetScanOptionsForPrepare(scan_opts);
+  if (!s.ok()) {
+    return;
+  }
+
+  const MultiScanArgs* pruning_scan_opts =
+      scan_opts.HasBoundedScanRanges() ? &scan_opts : nullptr;
+  s = EnsureInternalIteratorInitialized(pruning_scan_opts);
+  if (!s.ok()) {
+    return;
+  }
+
+  db_iter_->PrepareInternalChildren();
+}
+
 ArenaWrappedDBIter* NewArenaWrappedDbIterator(
     Env* env, const ReadOptions& read_options, ColumnFamilyHandleImpl* cfh,
     SuperVersion* sv, const SequenceNumber& sequence,
     ReadCallback* read_callback, DBImpl* db_impl, bool expose_blob_index,
     bool allow_refresh, bool allow_mark_memtable_for_flush) {
   ArenaWrappedDBIter* db_iter = new ArenaWrappedDBIter();
+  ColumnFamilyData* cfd = cfh->cfd();
   db_iter->Init(env, read_options, cfh->cfd()->ioptions(),
                 sv->mutable_cf_options, sv->current, sequence,
                 sv->version_number, read_callback, cfh, expose_blob_index,
                 allow_refresh,
                 allow_mark_memtable_for_flush ? sv->mem : nullptr);
-  if (cfh != nullptr && allow_refresh) {
-    db_iter->StoreRefreshInfo(cfh, read_callback, expose_blob_index);
+  if (allow_refresh && cfd != nullptr) {
+    db_iter->StoreRefreshInfo(read_callback, expose_blob_index);
   }
-
-  InternalIterator* internal_iter = db_impl->NewInternalIterator(
-      db_iter->GetReadOptions(), cfh->cfd(), sv, db_iter->GetArena(), sequence,
-      /*allow_unprepared_value=*/true, db_iter);
-  db_iter->SetIterUnderDBIter(internal_iter);
+  db_iter->StoreDeferredInitInfo(db_impl, cfd, sv, sequence,
+                                 allow_mark_memtable_for_flush);
 
   return db_iter;
 }

db/arena_wrapped_db_iter.h

@@ -10,6 +10,7 @@
 #pragma once
 #include <stdint.h>
 
+#include <memory>
 #include <string>
 
 #include "db/db_impl/db_impl.h"
@@ -33,10 +34,17 @@ class Version;
 // When using the class's Iterator interface, the behavior is exactly
 // the same as the inner DBIter.
 class ArenaWrappedDBIter : public Iterator {
+  struct ColumnFamilyDataUnrefDeleter {
+    DBImpl* db_impl = nullptr;
+    void operator()(ColumnFamilyData* cfd) const;
+  };
+  using ColumnFamilyDataRef =
+      std::unique_ptr<ColumnFamilyData, ColumnFamilyDataUnrefDeleter>;
+
  public:
   ~ArenaWrappedDBIter() override {
     if (db_iter_ != nullptr) {
-      db_iter_->~DBIter();
+      DestroyDBIter();
     } else {
       assert(false);
     }
@@ -51,7 +59,7 @@ class ArenaWrappedDBIter : public Iterator {
   // Set the internal iterator wrapped inside the DB Iterator. Usually it is
   // a merging iterator.
   virtual void SetIterUnderDBIter(InternalIterator* iter) {
-    db_iter_->SetIter(iter);
+    SetIterUnderDBIterImpl(iter);
   }
 
   void SetMemtableRangetombstoneIter(
@@ -60,26 +68,48 @@ class ArenaWrappedDBIter : public Iterator {
   }
 
   bool Valid() const override { return db_iter_->Valid(); }
-  void SeekToFirst() override { db_iter_->SeekToFirst(); }
-  void SeekToLast() override { db_iter_->SeekToLast(); }
+  void SeekToFirst() override {
+    if (!EnsureInternalIteratorInitialized(nullptr).ok()) {
+      return;
+    }
+    db_iter_->SeekToFirst();
+  }
+  void SeekToLast() override {
+    if (!EnsureInternalIteratorInitialized(nullptr).ok()) {
+      return;
+    }
+    db_iter_->SeekToLast();
+  }
   // 'target' does not contain timestamp, even if user timestamp feature is
   // enabled.
   void Seek(const Slice& target) override {
+    if (!EnsureInternalIteratorInitialized(nullptr).ok()) {
+      return;
+    }
     MaybeAutoRefresh(true /* is_seek */, DBIter::kForward);
     db_iter_->Seek(target);
   }
 
   void SeekForPrev(const Slice& target) override {
+    if (!EnsureInternalIteratorInitialized(nullptr).ok()) {
+      return;
+    }
     MaybeAutoRefresh(true /* is_seek */, DBIter::kReverse);
     db_iter_->SeekForPrev(target);
   }
 
   void Next() override {
+    if (!EnsureInternalIteratorInitialized(nullptr).ok()) {
+      return;
+    }
     db_iter_->Next();
     MaybeAutoRefresh(false /* is_seek */, DBIter::kForward);
   }
 
   void Prev() override {
+    if (!EnsureInternalIteratorInitialized(nullptr).ok()) {
+      return;
+    }
     db_iter_->Prev();
     MaybeAutoRefresh(false /* is_seek */, DBIter::kReverse);
   }
@@ -96,12 +126,13 @@ class ArenaWrappedDBIter : public Iterator {
   Status Refresh() override;
   Status Refresh(const Snapshot*) override;
 
-  bool PrepareValue() override { return db_iter_->PrepareValue(); }
-
-  void Prepare(const MultiScanArgs& scan_opts) override {
-    db_iter_->Prepare(scan_opts);
+  bool PrepareValue() override {
+    return EnsureInternalIteratorInitialized(nullptr).ok() &&
+           db_iter_->PrepareValue();
   }
 
+  void Prepare(const MultiScanArgs& scan_opts) override;
+
   // FIXME: we could just pass SV in for mutable cf option, version and version
   // number, but this is used by SstFileReader which does not have a SV.
   void Init(Env* env, const ReadOptions& read_options,
@@ -110,27 +141,53 @@ class ArenaWrappedDBIter : public Iterator {
             const SequenceNumber& sequence, uint64_t version_number,
             ReadCallback* read_callback, ColumnFamilyHandleImpl* cfh,
             bool expose_blob_index, bool allow_refresh,
-            ReadOnlyMemTable* active_mem);
+            ReadOnlyMemTable* active_mem, DBImpl* db_impl = nullptr,
+            ColumnFamilyData* cfd = nullptr);
 
-  // Store some parameters so we can refresh the iterator at a later point
-  // with these same params
-  void StoreRefreshInfo(ColumnFamilyHandleImpl* cfh,
-                        ReadCallback* read_callback, bool expose_blob_index) {
-    cfh_ = cfh;
+  // Store parameters used only by explicit/auto-refresh.
+  void StoreRefreshInfo(ReadCallback* read_callback, bool expose_blob_index) {
     read_callback_ = read_callback;
     expose_blob_index_ = expose_blob_index;
   }
 
+  void StoreDeferredInitInfo(DBImpl* db_impl, ColumnFamilyData* cfd,
+                             SuperVersion* sv, const SequenceNumber& sequence,
+                             bool allow_mark_memtable_for_flush) {
+    assert(cfd != nullptr);
+    db_impl_ = db_impl;
+    cfd->Ref();
+    cfd_ref_ = ColumnFamilyDataRef(cfd, ColumnFamilyDataUnrefDeleter{db_impl});
+    deferred_cfd_ = cfd;
+    deferred_sv_ = sv;
+    sequence_ = sequence;
+    allow_mark_memtable_for_flush_ = allow_mark_memtable_for_flush;
+  }
+
  private:
+  Status EnsureInternalIteratorInitialized(const MultiScanArgs* scan_opts);
+  void SetIterUnderDBIterImpl(InternalIterator* iter) {
+    db_iter_->SetIter(iter);
+    internal_iter_initialized_ = true;
+  }
+  void CleanupDeferredSuperVersion();
+  void DestroyDBIter();
+  void DestroyDBIterAndArena();
   void DoRefresh(const Snapshot* snapshot, uint64_t sv_number);
   void MaybeAutoRefresh(bool is_seek, DBIter::Direction direction);
 
   DBIter* db_iter_ = nullptr;
   Arena arena_;
-  uint64_t sv_number_;
-  ColumnFamilyHandleImpl* cfh_ = nullptr;
+  uint64_t sv_number_ = 0;
+  DBImpl* db_impl_ = nullptr;
+  ColumnFamilyDataRef cfd_ref_{nullptr, ColumnFamilyDataUnrefDeleter{}};
+  ColumnFamilyData* deferred_cfd_ = nullptr;
+  SuperVersion* deferred_sv_ = nullptr;
+  SequenceNumber sequence_ = kMaxSequenceNumber;
+  bool internal_iter_initialized_ = false;
+  bool prepare_called_ = false;
   ReadOptions read_options_;
-  ReadCallback* read_callback_;
+  ReadOptions child_read_options_;
+  ReadCallback* read_callback_ = nullptr;
   bool expose_blob_index_ = false;
   bool allow_refresh_ = true;
   bool allow_mark_memtable_for_flush_ = true;

db/blob/blob_file_builder.cc

@@ -22,6 +22,7 @@
 #include "logging/logging.h"
 #include "options/cf_options.h"
 #include "options/options_helper.h"
+#include "rocksdb/file_system.h"
 #include "rocksdb/slice.h"
 #include "rocksdb/status.h"
 #include "test_util/sync_point.h"
@@ -200,6 +201,8 @@ Status BlobFileBuilder::OpenBlobFileIfNeeded() {
   {
     assert(file_options_);
     fo_copy = *file_options_;
+    fo_copy.open_contract = FileOpenContract::kNoReopenForWrite |
+                            FileOpenContract::kNoReadersWhileOpenForWrite;
     fo_copy.write_hint = write_hint_;
     Status s = NewWritableFile(fs_, blob_file_path, &file, fo_copy);
 

db/blob/blob_file_cache.cc

@@ -123,6 +123,7 @@ Status BlobFileCache::OpenBlobFileReaderUncached(
   Statistics* const statistics = immutable_options_->stats;
   RecordTick(statistics, NO_FILE_OPENS);
 
+  assert(file_options_);
   Status s = BlobFileReader::Create(
       *immutable_options_, read_options, *file_options_, column_family_id_,
       blob_file_read_hist_, blob_file_number, io_tracer_,

db/blob/blob_file_partition_manager.cc

@@ -24,6 +24,7 @@
 #include "file/read_write_util.h"
 #include "file/writable_file_writer.h"
 #include "logging/logging.h"
+#include "rocksdb/file_system.h"
 #include "util/aligned_buffer.h"
 #include "util/compression.h"
 #include "util/mutexlock.h"
@@ -199,7 +200,9 @@ Status BlobFilePartitionManager::OpenNewBlobFile(Partition* partition,
   }
 
   std::unique_ptr<FSWritableFile> file;
-  Status s = NewWritableFile(fs_, blob_file_path, &file, file_options_);
+  FileOptions writer_file_options = file_options_;
+  writer_file_options.open_contract = FileOpenContract::kNoReopenForWrite;
+  Status s = NewWritableFile(fs_, blob_file_path, &file, writer_file_options);
   if (!s.ok()) {
     RemoveFilePartitionMapping(blob_file_number);
     return s;
@@ -208,7 +211,7 @@ Status BlobFilePartitionManager::OpenNewBlobFile(Partition* partition,
   const bool perform_data_verification =
       checksum_handoff_file_types_.Contains(FileType::kBlobFile);
   auto file_writer = std::make_unique<WritableFileWriter>(
-      std::move(file), blob_file_path, file_options_, clock_, io_tracer_,
+      std::move(file), blob_file_path, writer_file_options, clock_, io_tracer_,
       statistics_, Histograms::BLOB_DB_BLOB_FILE_WRITE_MICROS, listeners_,
       file_checksum_gen_factory_, perform_data_verification);
 

db/blob/blob_file_reader.cc

@@ -12,6 +12,7 @@
 #include "db/blob/blob_log_format.h"
 #include "file/file_prefetch_buffer.h"
 #include "file/filename.h"
+#include "file/read_write_util.h"
 #include "monitoring/statistics_impl.h"
 #include "options/cf_options.h"
 #include "rocksdb/file_system.h"
@@ -20,6 +21,7 @@
 #include "table/format.h"
 #include "table/multiget_context.h"
 #include "test_util/sync_point.h"
+#include "util/aligned_buffer.h"
 #include "util/compression.h"
 #include "util/stop_watch.h"
 
@@ -105,24 +107,6 @@ Status BlobFileReader::OpenFile(
 
   constexpr IODebugContext* dbg = nullptr;
 
-  {
-    TEST_SYNC_POINT("BlobFileReader::OpenFile:GetFileSize");
-
-    const Status s =
-        fs->GetFileSize(blob_file_path, IOOptions(), file_size, dbg);
-    if (!s.ok()) {
-      return s;
-    }
-  }
-
-  if (!skip_footer_size_check &&
-      *file_size < BlobLogHeader::kSize + BlobLogFooter::kSize) {
-    return Status::Corruption("Malformed blob file");
-  }
-  if (skip_footer_size_check && *file_size < BlobLogHeader::kSize) {
-    return Status::Corruption("Malformed blob file");
-  }
-
   std::unique_ptr<FSRandomAccessFile> file;
   FileOptions reader_file_opts = file_opts;
 
@@ -142,6 +126,24 @@ Status BlobFileReader::OpenFile(
 
   assert(file);
 
+  {
+    Status s = GetFileSizeFromOpenFileOrPath(
+        file.get(), fs, blob_file_path, file_size, dbg,
+        FileSizeFallback::kNotSupportedOnly,
+        []() { TEST_SYNC_POINT("BlobFileReader::OpenFile:GetFileSize"); });
+    if (!s.ok()) {
+      return s;
+    }
+  }
+
+  if (!skip_footer_size_check &&
+      *file_size < BlobLogHeader::kSize + BlobLogFooter::kSize) {
+    return Status::Corruption("Malformed blob file");
+  }
+  if (skip_footer_size_check && *file_size < BlobLogHeader::kSize) {
+    return Status::Corruption("Malformed blob file");
+  }
+
   if (immutable_options.advise_random_on_open) {
     file->Hint(FSRandomAccessFile::kRandom);
   }
@@ -165,7 +167,7 @@ Status BlobFileReader::ReadHeader(const RandomAccessFileReader* file_reader,
 
   Slice header_slice;
   Buffer buf;
-  AlignedBuf aligned_buf;
+  AlignedBuffer direct_io_buffer;
 
   {
     TEST_SYNC_POINT("BlobFileReader::ReadHeader:ReadFromFile");
@@ -175,7 +177,7 @@ Status BlobFileReader::ReadHeader(const RandomAccessFileReader* file_reader,
 
     const Status s =
         ReadFromFile(file_reader, read_options, read_offset, read_size,
-                     statistics, &header_slice, &buf, &aligned_buf);
+                     statistics, &header_slice, &buf, &direct_io_buffer);
     if (!s.ok()) {
       return s;
     }
@@ -216,7 +218,7 @@ Status BlobFileReader::ReadFooter(const RandomAccessFileReader* file_reader,
 
   Slice footer_slice;
   Buffer buf;
-  AlignedBuf aligned_buf;
+  AlignedBuffer direct_io_buffer;
 
   {
     TEST_SYNC_POINT("BlobFileReader::ReadFooter:ReadFromFile");
@@ -226,7 +228,7 @@ Status BlobFileReader::ReadFooter(const RandomAccessFileReader* file_reader,
 
     const Status s =
         ReadFromFile(file_reader, read_options, read_offset, read_size,
-                     statistics, &footer_slice, &buf, &aligned_buf);
+                     statistics, &footer_slice, &buf, &direct_io_buffer);
     if (!s.ok()) {
       return s;
     }
@@ -257,10 +259,11 @@ Status BlobFileReader::ReadFromFile(const RandomAccessFileReader* file_reader,
                                     const ReadOptions& read_options,
                                     uint64_t read_offset, size_t read_size,
                                     Statistics* statistics, Slice* slice,
-                                    Buffer* buf, AlignedBuf* aligned_buf) {
+                                    Buffer* buf,
+                                    AlignedBuffer* direct_io_buffer) {
   assert(slice);
   assert(buf);
-  assert(aligned_buf);
+  assert(direct_io_buffer);
 
   assert(file_reader);
 
@@ -277,15 +280,15 @@ Status BlobFileReader::ReadFromFile(const RandomAccessFileReader* file_reader,
 
   if (file_reader->use_direct_io()) {
     constexpr char* scratch = nullptr;
+    AlignedBufferAllocationContext direct_io_context{direct_io_buffer};
 
     s = file_reader->Read(io_options, read_offset, read_size, slice, scratch,
-                          aligned_buf, &dbg);
+                          &direct_io_context, &dbg);
   } else {
     buf->reset(new char[read_size]);
-    constexpr AlignedBuf* aligned_scratch = nullptr;
 
     s = file_reader->Read(io_options, read_offset, read_size, slice, buf->get(),
-                          aligned_scratch, &dbg);
+                          nullptr, &dbg);
   }
 
   if (!s.ok()) {
@@ -349,7 +352,7 @@ Status BlobFileReader::GetBlob(
 
   Slice record_slice;
   Buffer buf;
-  AlignedBuf aligned_buf;
+  AlignedBuffer direct_io_buffer;
 
   bool prefetched = false;
 
@@ -379,7 +382,7 @@ Status BlobFileReader::GetBlob(
     const Status s =
         ReadFromFile(file_reader_.get(), read_options, record_offset,
                      static_cast<size_t>(record_size), statistics_,
-                     &record_slice, &buf, &aligned_buf);
+                     &record_slice, &buf, &direct_io_buffer);
     if (!s.ok()) {
       return s;
     }
@@ -477,7 +480,7 @@ void BlobFileReader::MultiGetBlob(
   }
 
   Buffer buf;
-  AlignedBuf aligned_buf;
+  AlignedBuffer direct_io_buffer;
 
   Status s;
   bool direct_io = file_reader_->use_direct_io();
@@ -500,8 +503,9 @@ void BlobFileReader::MultiGetBlob(
   IODebugContext dbg;
   s = file_reader_->PrepareIOOptions(read_options, opts, &dbg);
   if (s.ok()) {
+    AlignedBufferAllocationContext direct_io_context{&direct_io_buffer};
     s = file_reader_->MultiRead(opts, read_reqs.data(), read_reqs.size(),
-                                direct_io ? &aligned_buf : nullptr, &dbg);
+                                &direct_io_context, &dbg);
   }
   if (!s.ok()) {
     for (auto& req : read_reqs) {

db/blob/blob_file_reader.h

@@ -108,7 +108,7 @@ class BlobFileReader {
                              const ReadOptions& read_options,
                              uint64_t read_offset, size_t read_size,
                              Statistics* statistics, Slice* slice, Buffer* buf,
-                             AlignedBuf* aligned_buf);
+                             AlignedBuffer* direct_io_buffer);
 
   static Status VerifyBlob(const Slice& record_slice, const Slice& user_key,
                            uint64_t value_size);

db/blob/blob_file_reader_test.cc

@@ -7,10 +7,12 @@
 
 #include <cassert>
 #include <string>
+#include <utility>
 
 #include "db/blob/blob_contents.h"
 #include "db/blob/blob_log_format.h"
 #include "db/blob/blob_log_writer.h"
+#include "env/composite_env_wrapper.h"
 #include "env/mock_env.h"
 #include "file/filename.h"
 #include "file/read_write_util.h"
@@ -136,6 +138,167 @@ class BlobFileReaderTest : public testing::Test {
   std::unique_ptr<Env> mock_env_;
 };
 
+namespace {
+
+// Returns a stale path-level size for one target blob file while leaving the
+// opened file handle untouched. This lets the test verify
+// BlobFileReader::Create prefers the opened handle's size when the path stat
+// lags behind.
+class StalePathSizeFileSystem : public FileSystemWrapper {
+ public:
+  static const char* kClassName() { return "StalePathSizeFileSystem"; }
+
+  StalePathSizeFileSystem(const std::shared_ptr<FileSystem>& target,
+                          std::string stale_path)
+      : FileSystemWrapper(target), stale_path_(std::move(stale_path)) {}
+
+  const char* Name() const override { return kClassName(); }
+
+  IOStatus GetFileSize(const std::string& fname, const IOOptions& options,
+                       uint64_t* file_size, IODebugContext* dbg) override {
+    if (fname == stale_path_) {
+      *file_size = 0;
+      return IOStatus::OK();
+    }
+    return FileSystemWrapper::GetFileSize(fname, options, file_size, dbg);
+  }
+
+ private:
+  const std::string stale_path_;
+};
+
+// Makes opened blob file handles fail GetFileSize() with a hard error so the
+// reader path can verify that such errors are propagated directly.
+class FailingOpenFileSizeRandomAccessFile
+    : public FSRandomAccessFileOwnerWrapper {
+ public:
+  explicit FailingOpenFileSizeRandomAccessFile(
+      std::unique_ptr<FSRandomAccessFile>&& target)
+      : FSRandomAccessFileOwnerWrapper(std::move(target)) {}
+
+  IOStatus GetFileSize(uint64_t* /*result*/) override {
+    return IOStatus::IOError("open file size failed");
+  }
+};
+
+// Makes opened blob file handles report GetFileSize() as unsupported so the
+// reader path exercises its path-level fallback logic.
+class NotSupportedOpenFileSizeRandomAccessFile
+    : public FSRandomAccessFileOwnerWrapper {
+ public:
+  explicit NotSupportedOpenFileSizeRandomAccessFile(
+      std::unique_ptr<FSRandomAccessFile>&& target)
+      : FSRandomAccessFileOwnerWrapper(std::move(target)) {}
+
+  IOStatus GetFileSize(uint64_t* /*result*/) override {
+    return IOStatus::NotSupported("open file size not supported");
+  }
+};
+
+class OpenFileSizeFileSystem : public FileSystemWrapper {
+ public:
+  OpenFileSizeFileSystem(const std::shared_ptr<FileSystem>& target,
+                         std::string target_path)
+      : FileSystemWrapper(target), target_path_(std::move(target_path)) {}
+
+  IOStatus NewRandomAccessFile(const std::string& fname,
+                               const FileOptions& options,
+                               std::unique_ptr<FSRandomAccessFile>* result,
+                               IODebugContext* dbg) override {
+    std::unique_ptr<FSRandomAccessFile> file;
+    IOStatus s =
+        FileSystemWrapper::NewRandomAccessFile(fname, options, &file, dbg);
+    if (!s.ok()) {
+      return IOStatus(std::move(s));
+    }
+    if (fname == target_path_) {
+      *result = WrapTargetFile(std::move(file));
+    } else {
+      *result = std::move(file);
+    }
+    return IOStatus::OK();
+  }
+
+ protected:
+  virtual std::unique_ptr<FSRandomAccessFile> WrapTargetFile(
+      std::unique_ptr<FSRandomAccessFile>&& file) = 0;
+
+ private:
+  const std::string target_path_;
+};
+
+// Wraps the target blob file in FailingOpenFileSizeRandomAccessFile while
+// leaving all other files alone.
+class FailingOpenFileSizeFileSystem : public OpenFileSizeFileSystem {
+ public:
+  static const char* kClassName() { return "FailingOpenFileSizeFileSystem"; }
+
+  FailingOpenFileSizeFileSystem(const std::shared_ptr<FileSystem>& target,
+                                std::string target_path)
+      : OpenFileSizeFileSystem(target, std::move(target_path)) {}
+
+  const char* Name() const override { return kClassName(); }
+
+ private:
+  std::unique_ptr<FSRandomAccessFile> WrapTargetFile(
+      std::unique_ptr<FSRandomAccessFile>&& file) override {
+    return std::unique_ptr<FSRandomAccessFile>(
+        new FailingOpenFileSizeRandomAccessFile(std::move(file)));
+  }
+};
+
+// Wraps the target blob file in NotSupportedOpenFileSizeRandomAccessFile while
+// leaving all other files alone.
+class NotSupportedOpenFileSizeFileSystem : public OpenFileSizeFileSystem {
+ public:
+  static const char* kClassName() {
+    return "NotSupportedOpenFileSizeFileSystem";
+  }
+
+  NotSupportedOpenFileSizeFileSystem(const std::shared_ptr<FileSystem>& target,
+                                     std::string target_path)
+      : OpenFileSizeFileSystem(target, std::move(target_path)) {}
+
+  const char* Name() const override { return kClassName(); }
+
+ private:
+  std::unique_ptr<FSRandomAccessFile> WrapTargetFile(
+      std::unique_ptr<FSRandomAccessFile>&& file) override {
+    return std::unique_ptr<FSRandomAccessFile>(
+        new NotSupportedOpenFileSizeRandomAccessFile(std::move(file)));
+  }
+};
+
+// Forces the target blob file onto the NotSupported open-handle-size branch and
+// then fails the fallback path-level GetFileSize() call.
+class FailingFallbackPathSizeFileSystem
+    : public NotSupportedOpenFileSizeFileSystem {
+ public:
+  static const char* kClassName() {
+    return "FailingFallbackPathSizeFileSystem";
+  }
+
+  FailingFallbackPathSizeFileSystem(const std::shared_ptr<FileSystem>& target,
+                                    std::string target_path)
+      : NotSupportedOpenFileSizeFileSystem(target, target_path),
+        target_path_(std::move(target_path)) {}
+
+  const char* Name() const override { return kClassName(); }
+
+  IOStatus GetFileSize(const std::string& fname, const IOOptions& options,
+                       uint64_t* file_size, IODebugContext* dbg) override {
+    if (fname == target_path_) {
+      return IOStatus::IOError("fallback path size failed");
+    }
+    return FileSystemWrapper::GetFileSize(fname, options, file_size, dbg);
+  }
+
+ private:
+  const std::string target_path_;
+};
+
+}  // namespace
+
 TEST_F(BlobFileReaderTest, CreateReaderAndGetBlob) {
   Options options;
   options.env = mock_env_.get();
@@ -428,6 +591,211 @@ TEST_F(BlobFileReaderTest, CreateReaderAndGetBlob) {
   }
 }
 
+// Verifies Create() prefers the file size reported by the opened handle over a
+// stale path-level size query by returning a readable reader even when the path
+// stat intentionally reports 0 for the blob file.
+TEST_F(BlobFileReaderTest, CreateReaderUsesOpenedFileSizeWhenPathSizeIsStale) {
+  const std::string db_path = test::PerThreadDBPath(
+      mock_env_.get(),
+      "BlobFileReaderTest_CreateReaderUsesOpenedFileSizeWhenPathSizeIsStale");
+
+  constexpr uint64_t blob_file_number = 1;
+  const std::string blob_file_path = BlobFileName(db_path, blob_file_number);
+
+  auto stale_size_fs = std::make_shared<StalePathSizeFileSystem>(
+      mock_env_->GetFileSystem(), blob_file_path);
+  CompositeEnvWrapper stale_size_env(mock_env_.get(), stale_size_fs);
+
+  Options options;
+  options.env = &stale_size_env;
+  options.cf_paths.emplace_back(db_path, 0);
+  options.enable_blob_files = true;
+
+  ImmutableOptions immutable_options(options);
+
+  constexpr uint32_t column_family_id = 1;
+  constexpr bool has_ttl = false;
+  constexpr ExpirationRange expiration_range;
+  constexpr char key[] = "key";
+  constexpr char blob[] = "blob";
+
+  uint64_t blob_offset = 0;
+  uint64_t blob_size = 0;
+  WriteBlobFile(immutable_options, column_family_id, has_ttl, expiration_range,
+                expiration_range, blob_file_number, key, blob, kNoCompression,
+                &blob_offset, &blob_size);
+
+  std::unique_ptr<BlobFileReader> reader;
+  ReadOptions read_options;
+  read_options.verify_checksums = false;
+  ASSERT_OK(BlobFileReader::Create(
+      immutable_options, read_options, FileOptions(), column_family_id,
+      /*blob_file_read_hist=*/nullptr, blob_file_number, nullptr /*IOTracer*/,
+      &reader));
+  ASSERT_NE(reader, nullptr);
+
+  std::unique_ptr<BlobContents> value;
+  uint64_t bytes_read = 0;
+  ASSERT_OK(reader->GetBlob(read_options, key, blob_offset, blob_size,
+                            kNoCompression,
+                            /*prefetch_buffer=*/nullptr,
+                            /*allocator=*/nullptr, &value, &bytes_read));
+  ASSERT_NE(value, nullptr);
+  ASSERT_EQ(value->data(), Slice(blob));
+  ASSERT_EQ(bytes_read, blob_size);
+}
+
+// Verifies Create() falls back to the path-level GetFileSize() call when the
+// opened file handle reports that size queries are unsupported.
+TEST_F(BlobFileReaderTest,
+       CreateReaderFallsBackToPathSizeWhenOpenedFileSizeNotSupported) {
+  const std::string db_path = test::PerThreadDBPath(
+      mock_env_.get(),
+      "BlobFileReaderTest_"
+      "CreateReaderFallsBackToPathSizeWhenOpenedFileSizeNotSupported");
+
+  constexpr uint64_t blob_file_number = 1;
+  const std::string blob_file_path = BlobFileName(db_path, blob_file_number);
+
+  auto not_supported_open_file_size_fs =
+      std::make_shared<NotSupportedOpenFileSizeFileSystem>(
+          mock_env_->GetFileSystem(), blob_file_path);
+  CompositeEnvWrapper not_supported_open_file_size_env(
+      mock_env_.get(), not_supported_open_file_size_fs);
+
+  Options options;
+  options.env = &not_supported_open_file_size_env;
+  options.cf_paths.emplace_back(db_path, 0);
+  options.enable_blob_files = true;
+
+  ImmutableOptions immutable_options(options);
+
+  constexpr uint32_t column_family_id = 1;
+  constexpr bool has_ttl = false;
+  constexpr ExpirationRange expiration_range;
+  constexpr char key[] = "key";
+  constexpr char blob[] = "blob";
+
+  uint64_t blob_offset = 0;
+  uint64_t blob_size = 0;
+  WriteBlobFile(immutable_options, column_family_id, has_ttl, expiration_range,
+                expiration_range, blob_file_number, key, blob, kNoCompression,
+                &blob_offset, &blob_size);
+
+  std::unique_ptr<BlobFileReader> reader;
+  ReadOptions read_options;
+  read_options.verify_checksums = false;
+  ASSERT_OK(BlobFileReader::Create(
+      immutable_options, read_options, FileOptions(), column_family_id,
+      /*blob_file_read_hist=*/nullptr, blob_file_number, nullptr /*IOTracer*/,
+      &reader));
+  ASSERT_NE(reader, nullptr);
+
+  std::unique_ptr<BlobContents> value;
+  uint64_t bytes_read = 0;
+  ASSERT_OK(reader->GetBlob(read_options, key, blob_offset, blob_size,
+                            kNoCompression,
+                            /*prefetch_buffer=*/nullptr,
+                            /*allocator=*/nullptr, &value, &bytes_read));
+  ASSERT_NE(value, nullptr);
+  ASSERT_EQ(value->data(), Slice(blob));
+  ASSERT_EQ(bytes_read, blob_size);
+}
+
+// Verifies Create() propagates a real error from the opened file handle's
+// GetFileSize() rather than masking it with any path-level fallback.
+TEST_F(BlobFileReaderTest, CreateReaderPropagatesOpenedFileSizeError) {
+  const std::string db_path = test::PerThreadDBPath(
+      mock_env_.get(),
+      "BlobFileReaderTest_CreateReaderPropagatesOpenedFileSizeError");
+
+  constexpr uint64_t blob_file_number = 1;
+  const std::string blob_file_path = BlobFileName(db_path, blob_file_number);
+
+  auto failing_open_file_size_fs =
+      std::make_shared<FailingOpenFileSizeFileSystem>(
+          mock_env_->GetFileSystem(), blob_file_path);
+  CompositeEnvWrapper failing_open_file_size_env(mock_env_.get(),
+                                                 failing_open_file_size_fs);
+
+  Options options;
+  options.env = &failing_open_file_size_env;
+  options.cf_paths.emplace_back(db_path, 0);
+  options.enable_blob_files = true;
+
+  ImmutableOptions immutable_options(options);
+
+  constexpr uint32_t column_family_id = 1;
+  constexpr bool has_ttl = false;
+  constexpr ExpirationRange expiration_range;
+  constexpr char key[] = "key";
+  constexpr char blob[] = "blob";
+
+  uint64_t blob_offset = 0;
+  uint64_t blob_size = 0;
+  WriteBlobFile(immutable_options, column_family_id, has_ttl, expiration_range,
+                expiration_range, blob_file_number, key, blob, kNoCompression,
+                &blob_offset, &blob_size);
+
+  std::unique_ptr<BlobFileReader> reader;
+  ReadOptions read_options;
+  Status s = BlobFileReader::Create(
+      immutable_options, read_options, FileOptions(), column_family_id,
+      /*blob_file_read_hist=*/nullptr, blob_file_number, nullptr /*IOTracer*/,
+      &reader);
+  ASSERT_TRUE(s.IsIOError());
+  ASSERT_EQ(reader, nullptr);
+}
+
+// Verifies Create() also propagates errors from the path-level fallback size
+// query when the opened file handle only reports NotSupported for GetFileSize.
+TEST_F(
+    BlobFileReaderTest,
+    CreateReaderPropagatesFallbackPathSizeErrorWhenOpenedFileSizeNotSupported) {
+  const std::string db_path =
+      test::PerThreadDBPath(mock_env_.get(),
+                            "BlobFileReaderTest_"
+                            "CreateReaderPropagatesFallbackPathSizeErrorWhenOpe"
+                            "nedFileSizeNotSupported");
+
+  constexpr uint64_t blob_file_number = 1;
+  const std::string blob_file_path = BlobFileName(db_path, blob_file_number);
+
+  auto failing_fallback_path_size_fs =
+      std::make_shared<FailingFallbackPathSizeFileSystem>(
+          mock_env_->GetFileSystem(), blob_file_path);
+  CompositeEnvWrapper failing_fallback_path_size_env(
+      mock_env_.get(), failing_fallback_path_size_fs);
+
+  Options options;
+  options.env = &failing_fallback_path_size_env;
+  options.cf_paths.emplace_back(db_path, 0);
+  options.enable_blob_files = true;
+
+  ImmutableOptions immutable_options(options);
+
+  constexpr uint32_t column_family_id = 1;
+  constexpr bool has_ttl = false;
+  constexpr ExpirationRange expiration_range;
+  constexpr char key[] = "key";
+  constexpr char blob[] = "blob";
+
+  uint64_t blob_offset = 0;
+  uint64_t blob_size = 0;
+  WriteBlobFile(immutable_options, column_family_id, has_ttl, expiration_range,
+                expiration_range, blob_file_number, key, blob, kNoCompression,
+                &blob_offset, &blob_size);
+
+  std::unique_ptr<BlobFileReader> reader;
+  ReadOptions read_options;
+  Status s = BlobFileReader::Create(
+      immutable_options, read_options, FileOptions(), column_family_id,
+      /*blob_file_read_hist=*/nullptr, blob_file_number, nullptr /*IOTracer*/,
+      &reader);
+  ASSERT_TRUE(s.IsIOError());
+  ASSERT_EQ(reader, nullptr);
+}
+
 TEST_F(BlobFileReaderTest, Malformed) {
   // Write a blob file consisting of nothing but a header, and make sure we
   // detect the error when we open it for reading
@@ -849,7 +1217,6 @@ class BlobFileReaderIOErrorTest
 
 INSTANTIATE_TEST_CASE_P(BlobFileReaderTest, BlobFileReaderIOErrorTest,
                         ::testing::ValuesIn(std::vector<std::string>{
-                            "BlobFileReader::OpenFile:GetFileSize",
                             "BlobFileReader::OpenFile:NewRandomAccessFile",
                             "BlobFileReader::ReadHeader:ReadFromFile",
                             "BlobFileReader::ReadFooter:ReadFromFile",

db/blob/blob_source.cc

@@ -74,6 +74,7 @@ Status BlobSource::GetBlobFromCache(
     assert(cached_blob->GetValue());
 
     PERF_COUNTER_ADD(blob_cache_hit_count, 1);
+    PERF_COUNTER_ADD(blob_cache_read_byte, cached_blob->GetValue()->size());
     RecordTick(statistics_, BLOB_DB_CACHE_HIT);
     RecordTick(statistics_, BLOB_DB_CACHE_BYTES_READ,
                cached_blob->GetValue()->size());

db/blob/blob_source_test.cc

@@ -229,6 +229,7 @@ TEST_F(BlobSourceTest, GetBlobsFromCache) {
     // Retrieved the blob cache num_blobs * 3 times via TEST_BlobInCache,
     // GetBlob, and TEST_BlobInCache.
     ASSERT_EQ((int)get_perf_context()->blob_cache_hit_count, 0);
+    ASSERT_EQ((int)get_perf_context()->blob_cache_read_byte, 0);
     ASSERT_EQ((int)get_perf_context()->blob_read_count, num_blobs);
     ASSERT_EQ((int)get_perf_context()->blob_read_byte, total_bytes);
     ASSERT_GE((int)get_perf_context()->blob_checksum_time, 0);
@@ -262,6 +263,8 @@ TEST_F(BlobSourceTest, GetBlobsFromCache) {
       blob_bytes += blob_sizes[i];
       total_bytes += bytes_read;
       ASSERT_EQ((int)get_perf_context()->blob_cache_hit_count, i);
+      ASSERT_EQ((int)get_perf_context()->blob_cache_read_byte,
+                blob_bytes - blob_sizes[i]);
       ASSERT_EQ((int)get_perf_context()->blob_read_count, i + 1);
       ASSERT_EQ((int)get_perf_context()->blob_read_byte, total_bytes);
 
@@ -269,11 +272,13 @@ TEST_F(BlobSourceTest, GetBlobsFromCache) {
                                                blob_offsets[i]));
 
       ASSERT_EQ((int)get_perf_context()->blob_cache_hit_count, i + 1);
+      ASSERT_EQ((int)get_perf_context()->blob_cache_read_byte, blob_bytes);
       ASSERT_EQ((int)get_perf_context()->blob_read_count, i + 1);
       ASSERT_EQ((int)get_perf_context()->blob_read_byte, total_bytes);
     }
 
     ASSERT_EQ((int)get_perf_context()->blob_cache_hit_count, num_blobs);
+    ASSERT_EQ((int)get_perf_context()->blob_cache_read_byte, blob_bytes);
     ASSERT_EQ((int)get_perf_context()->blob_read_count, num_blobs);
     ASSERT_EQ((int)get_perf_context()->blob_read_byte, total_bytes);
 
@@ -312,6 +317,7 @@ TEST_F(BlobSourceTest, GetBlobsFromCache) {
     // Retrieved the blob cache num_blobs * 3 times via TEST_BlobInCache,
     // GetBlob, and TEST_BlobInCache.
     ASSERT_EQ((int)get_perf_context()->blob_cache_hit_count, num_blobs * 3);
+    ASSERT_EQ((int)get_perf_context()->blob_cache_read_byte, blob_bytes * 3);
     ASSERT_EQ((int)get_perf_context()->blob_read_count, 0);  // without i/o
     ASSERT_EQ((int)get_perf_context()->blob_read_byte, 0);   // without i/o
 
@@ -690,6 +696,8 @@ TEST_F(BlobSourceTest, MultiGetBlobsFromMultiFiles) {
     // TEST_BlobInCache.
     ASSERT_EQ((int)get_perf_context()->blob_cache_hit_count,
               num_blobs * blob_files);
+    ASSERT_EQ((int)get_perf_context()->blob_cache_read_byte,
+              blob_value_bytes * blob_files);
     ASSERT_EQ((int)get_perf_context()->blob_read_count,
               num_blobs * blob_files);  // blocking i/o
     ASSERT_EQ((int)get_perf_context()->blob_read_byte,
@@ -750,6 +758,8 @@ TEST_F(BlobSourceTest, MultiGetBlobsFromMultiFiles) {
     // via MultiGetBlob and TEST_BlobInCache.
     ASSERT_EQ((int)get_perf_context()->blob_cache_hit_count,
               num_blobs * blob_files * 2);
+    ASSERT_EQ((int)get_perf_context()->blob_cache_read_byte,
+              blob_value_bytes * blob_files * 2);
     ASSERT_EQ((int)get_perf_context()->blob_read_count,
               0);  // blocking i/o
     ASSERT_EQ((int)get_perf_context()->blob_read_byte,

db/blob/db_blob_direct_write_test.cc

@@ -10,11 +10,14 @@
 
 #include <algorithm>
 #include <array>
+#include <functional>
 #include <limits>
 #include <mutex>
 #include <set>
 #include <sstream>
 #include <string>
+#include <unordered_set>
+#include <utility>
 #include <vector>
 
 #include "cache/compressed_secondary_cache.h"
@@ -26,10 +29,12 @@
 #include "db/db_test_util.h"
 #include "db/db_with_timestamp_test_util.h"
 #include "db/wide/wide_column_test_util.h"
+#include "env/composite_env_wrapper.h"
 #include "file/filename.h"
 #include "file/random_access_file_reader.h"
 #include "port/stack_trace.h"
 #include "rocksdb/convenience.h"
+#include "rocksdb/file_system.h"
 #include "rocksdb/trace_reader_writer.h"
 #include "rocksdb/trace_record.h"
 #include "rocksdb/utilities/replayer.h"
@@ -101,6 +106,206 @@ class RecordingBlobDirectWritePartitionStrategy
   std::vector<Call> calls_;
 };
 
+namespace {
+
+// Matches BlobDB file names so the test filesystem can special-case only active
+// blob files without changing unrelated file opens.
+bool IsBlobFilePath(const std::string& fname) {
+  const size_t basename_pos = fname.find_last_of("/\\");
+  const std::string basename = basename_pos == std::string::npos
+                                   ? fname
+                                   : fname.substr(basename_pos + 1);
+  uint64_t file_number = 0;
+  FileType file_type = kWalFile;
+  return ParseFileName(basename, &file_number, &file_type) &&
+         file_type == kBlobFile;
+}
+
+// Keeps track of when an active blob file stops being writer-visible so the
+// test filesystem can model current-size visibility only while the file remains
+// active.
+class ActiveBlobVisibilityWritableFile : public FSWritableFileOwnerWrapper {
+ public:
+  ActiveBlobVisibilityWritableFile(std::unique_ptr<FSWritableFile>&& target,
+                                   std::function<void()> on_close)
+      : FSWritableFileOwnerWrapper(std::move(target)),
+        on_close_(std::move(on_close)) {}
+
+  ~ActiveBlobVisibilityWritableFile() override { DeactivateOnce(); }
+
+  IOStatus Close(const IOOptions& options, IODebugContext* dbg) override {
+    IOStatus s = target()->Close(options, dbg);
+    DeactivateOnce();
+    return s;
+  }
+
+ private:
+  void DeactivateOnce() {
+    if (on_close_) {
+      on_close_();
+      on_close_ = nullptr;
+    }
+  }
+
+  std::function<void()> on_close_;
+};
+
+// Simulates a remote readable file whose GetFileSize() either exposes the live
+// on-disk size for default-contract active blobs or hides it for
+// append-only-no-readers files.
+class ActiveBlobVisibilityRandomAccessFile
+    : public FSRandomAccessFileOwnerWrapper {
+ public:
+  ActiveBlobVisibilityRandomAccessFile(
+      std::unique_ptr<FSRandomAccessFile>&& target, bool expose_current_size,
+      std::shared_ptr<FileSystem> underlying_fs, std::string fname)
+      : FSRandomAccessFileOwnerWrapper(std::move(target)),
+        expose_current_size_(expose_current_size),
+        underlying_fs_(std::move(underlying_fs)),
+        fname_(std::move(fname)) {}
+
+  IOStatus GetFileSize(uint64_t* result) override {
+    if (!expose_current_size_) {
+      *result = 0;
+      return IOStatus::OK();
+    }
+    // Delegate to the underlying FS path-level GetFileSize (bypassing the
+    // wrapper that reports 0 for active blobs) to simulate a remote FS that
+    // exposes the current file size through the open handle.
+    return underlying_fs_->GetFileSize(fname_, IOOptions(), result,
+                                       /*dbg=*/nullptr);
+  }
+
+ private:
+  const bool expose_current_size_;
+  std::shared_ptr<FileSystem> underlying_fs_;
+  std::string fname_;
+};
+
+// Models a remote filesystem where active direct-write blobs are only
+// reader-visible when the writer does not promise no concurrent readers.
+class RemoteBlobVisibilityFileSystem : public FileSystemWrapper {
+ public:
+  explicit RemoteBlobVisibilityFileSystem(const std::shared_ptr<FileSystem>& fs)
+      : FileSystemWrapper(fs) {}
+
+  static const char* kClassName() { return "RemoteBlobVisibilityFileSystem"; }
+  const char* Name() const override { return kClassName(); }
+
+  IOStatus NewWritableFile(const std::string& fname, const FileOptions& opts,
+                           std::unique_ptr<FSWritableFile>* result,
+                           IODebugContext* dbg) override {
+    std::unique_ptr<FSWritableFile> file;
+    IOStatus s = target()->NewWritableFile(fname, opts, &file, dbg);
+    if (!s.ok() || !IsBlobFilePath(fname)) {
+      *result = std::move(file);
+      return IOStatus(std::move(s));
+    }
+
+    const bool has_no_reopen_for_write_contract = HasFileOpenContract(
+        opts.open_contract, FileOpenContract::kNoReopenForWrite);
+    const bool has_no_readers_while_open_for_write_contract =
+        HasFileOpenContract(opts.open_contract,
+                            FileOpenContract::kNoReadersWhileOpenForWrite);
+    {
+      std::lock_guard<std::mutex> lock(mu_);
+      saw_no_reopen_for_write_writer_contract_ |=
+          has_no_reopen_for_write_contract;
+      saw_no_readers_while_open_for_write_writer_contract_ |=
+          has_no_readers_while_open_for_write_contract;
+      if (!has_no_readers_while_open_for_write_contract) {
+        active_blob_paths_.insert(fname);
+      }
+    }
+
+    if (has_no_readers_while_open_for_write_contract) {
+      *result = std::move(file);
+      return IOStatus::OK();
+    }
+
+    result->reset(new ActiveBlobVisibilityWritableFile(
+        std::move(file), [this, tracked_path = std::string(fname)]() {
+          DeactivateBlobPath(tracked_path);
+        }));
+    return IOStatus::OK();
+  }
+
+  IOStatus NewRandomAccessFile(const std::string& fname,
+                               const FileOptions& opts,
+                               std::unique_ptr<FSRandomAccessFile>* result,
+                               IODebugContext* dbg) override {
+    std::unique_ptr<FSRandomAccessFile> file;
+    IOStatus s = target()->NewRandomAccessFile(fname, opts, &file, dbg);
+    if (!s.ok() || !IsBlobFilePath(fname)) {
+      *result = std::move(file);
+      return IOStatus(std::move(s));
+    }
+
+    const bool has_no_readers_while_open_for_write_contract =
+        HasFileOpenContract(opts.open_contract,
+                            FileOpenContract::kNoReadersWhileOpenForWrite);
+    const bool active_blob = IsActiveBlobPath(fname);
+    {
+      std::lock_guard<std::mutex> lock(mu_);
+      saw_no_readers_while_open_for_write_reader_contract_ |=
+          has_no_readers_while_open_for_write_contract;
+    }
+
+    if (!active_blob) {
+      *result = std::move(file);
+      return IOStatus::OK();
+    }
+
+    result->reset(new ActiveBlobVisibilityRandomAccessFile(
+        std::move(file), !has_no_readers_while_open_for_write_contract, target_,
+        fname));
+    return IOStatus::OK();
+  }
+
+  IOStatus GetFileSize(const std::string& fname, const IOOptions& options,
+                       uint64_t* file_size, IODebugContext* dbg) override {
+    if (IsActiveBlobPath(fname)) {
+      *file_size = 0;
+      return IOStatus::OK();
+    }
+    return target()->GetFileSize(fname, options, file_size, dbg);
+  }
+
+  bool SawNoReopenForWriteWriterContract() const {
+    std::lock_guard<std::mutex> lock(mu_);
+    return saw_no_reopen_for_write_writer_contract_;
+  }
+
+  bool SawNoReadersWhileOpenForWriteWriterContract() const {
+    std::lock_guard<std::mutex> lock(mu_);
+    return saw_no_readers_while_open_for_write_writer_contract_;
+  }
+
+  bool SawNoReadersWhileOpenForWriteReaderContract() const {
+    std::lock_guard<std::mutex> lock(mu_);
+    return saw_no_readers_while_open_for_write_reader_contract_;
+  }
+
+ private:
+  void DeactivateBlobPath(const std::string& fname) {
+    std::lock_guard<std::mutex> lock(mu_);
+    active_blob_paths_.erase(fname);
+  }
+
+  bool IsActiveBlobPath(const std::string& fname) const {
+    std::lock_guard<std::mutex> lock(mu_);
+    return active_blob_paths_.find(fname) != active_blob_paths_.end();
+  }
+
+  mutable std::mutex mu_;
+  bool saw_no_reopen_for_write_writer_contract_ = false;
+  bool saw_no_readers_while_open_for_write_writer_contract_ = false;
+  bool saw_no_readers_while_open_for_write_reader_contract_ = false;
+  std::unordered_set<std::string> active_blob_paths_;
+};
+
+}  // namespace
+
 class DBBlobDirectWriteTest : public DBTestBase {
  protected:
   DBBlobDirectWriteTest()
@@ -816,6 +1021,30 @@ TEST_F(DBBlobDirectWriteTest, DirectWriteImmutableMemtableRead) {
   verify_reads();
 }
 
+// Verifies active blob direct-write files promise no write reopen without also
+// promising no concurrent readers.
+TEST_F(DBBlobDirectWriteTest, DirectWriteUsesReadableContractForRemoteFile) {
+  auto remote_fs =
+      std::make_shared<RemoteBlobVisibilityFileSystem>(env_->GetFileSystem());
+  std::unique_ptr<Env> remote_env(new CompositeEnvWrapper(env_, remote_fs));
+
+  Options options = GetDirectWriteOptions();
+  options.env = remote_env.get();
+  Reopen(options);
+
+  const std::string key = "remote_blob_key";
+  const std::string value(128, 'w');
+
+  ASSERT_OK(Put(key, value));
+  ASSERT_TRUE(remote_fs->SawNoReopenForWriteWriterContract());
+  ASSERT_FALSE(remote_fs->SawNoReadersWhileOpenForWriteWriterContract());
+
+  ASSERT_EQ(Get(key), value);
+  ASSERT_FALSE(remote_fs->SawNoReadersWhileOpenForWriteReaderContract());
+
+  Close();
+}
+
 TEST_F(DBBlobDirectWriteTest, DirectWriteRefreshesReaderAfterFlush) {
   Options options = GetDirectWriteOptions();
   options.blob_direct_write_partitions = 1;

db/builder.cc

@@ -179,6 +179,8 @@ Status BuildTable(
       TEST_SYNC_POINT_CALLBACK("BuildTable:create_file", &use_direct_writes);
 #endif  // !NDEBUG
       FileOptions fo_copy = file_options;
+      fo_copy.open_contract = FileOpenContract::kNoReopenForWrite |
+                              FileOpenContract::kNoReadersWhileOpenForWrite;
       fo_copy.write_hint = write_hint;
       IOStatus io_s = NewWritableFile(fs, fname, &file, fo_copy);
       assert(s.ok());

db/c.cc

@@ -148,6 +148,7 @@ using ROCKSDB_NAMESPACE::TransactionDB;
 using ROCKSDB_NAMESPACE::TransactionDBOptions;
 using ROCKSDB_NAMESPACE::TransactionLogIterator;
 using ROCKSDB_NAMESPACE::TransactionOptions;
+using ROCKSDB_NAMESPACE::TxnDBWritePolicy;
 using ROCKSDB_NAMESPACE::WaitForCompactOptions;
 using ROCKSDB_NAMESPACE::WALRecoveryMode;
 using ROCKSDB_NAMESPACE::WritableFile;
@@ -1384,6 +1385,10 @@ void rocksdb_backup_engine_info_destroy(
   delete info;
 }
 
+void rocksdb_backup_engine_stop_backup(rocksdb_backup_engine_t* be) {
+  be->rep->StopBackup();
+}
+
 void rocksdb_backup_engine_close(rocksdb_backup_engine_t* be) {
   delete be->rep;
   delete be;
@@ -1465,6 +1470,20 @@ uint64_t rocksdb_backup_engine_options_get_restore_rate_limit(
   return options->rep.restore_rate_limit;
 }
 
+void rocksdb_backup_engine_options_set_backup_rate_limiter(
+    rocksdb_backup_engine_options_t* options, rocksdb_ratelimiter_t* limiter) {
+  if (limiter) {
+    options->rep.backup_rate_limiter = limiter->rep;
+  }
+}
+
+void rocksdb_backup_engine_options_set_restore_rate_limiter(
+    rocksdb_backup_engine_options_t* options, rocksdb_ratelimiter_t* limiter) {
+  if (limiter) {
+    options->rep.restore_rate_limiter = limiter->rep;
+  }
+}
+
 void rocksdb_backup_engine_options_set_max_background_operations(
     rocksdb_backup_engine_options_t* options, int val) {
   options->rep.max_background_operations = val;
@@ -3771,6 +3790,11 @@ void rocksdb_block_based_options_set_index_block_search_type(
       static_cast<BlockBasedTableOptions::BlockSearchType>(v);
 }
 
+void rocksdb_block_based_options_set_uniform_cv_threshold(
+    rocksdb_block_based_table_options_t* options, double v) {
+  options->rep.uniform_cv_threshold = v;
+}
+
 void rocksdb_block_based_options_set_data_block_hash_ratio(
     rocksdb_block_based_table_options_t* options, double v) {
   options->rep.data_block_hash_table_util_ratio = v;
@@ -4242,6 +4266,15 @@ void rocksdb_set_options(rocksdb_t* db, int count, const char* const keys[],
   SaveError(errptr, db->rep->SetOptions(options_map));
 }
 
+void rocksdb_set_db_options(rocksdb_t* db, int count, const char* const keys[],
+                            const char* const values[], char** errptr) {
+  std::unordered_map<std::string, std::string> options_map;
+  for (int i = 0; i < count; i++) {
+    options_map[keys[i]] = values[i];
+  }
+  SaveError(errptr, db->rep->SetDBOptions(options_map));
+}
+
 void rocksdb_set_options_cf(rocksdb_t* db,
                             rocksdb_column_family_handle_t* handle, int count,
                             const char* const keys[],
@@ -5038,6 +5071,16 @@ unsigned char rocksdb_options_get_use_direct_io_for_flush_and_compaction(
   return opt->rep.use_direct_io_for_flush_and_compaction;
 }
 
+void rocksdb_options_set_use_direct_io_for_compaction_reads(
+    rocksdb_options_t* opt, unsigned char v) {
+  opt->rep.use_direct_io_for_compaction_reads = v;
+}
+
+unsigned char rocksdb_options_get_use_direct_io_for_compaction_reads(
+    rocksdb_options_t* opt) {
+  return opt->rep.use_direct_io_for_compaction_reads;
+}
+
 void rocksdb_options_set_allow_mmap_reads(rocksdb_options_t* opt,
                                           unsigned char v) {
   opt->rep.allow_mmap_reads = v;
@@ -5286,6 +5329,15 @@ size_t rocksdb_options_get_recycle_log_file_num(rocksdb_options_t* opt) {
   return opt->rep.recycle_log_file_num;
 }
 
+void rocksdb_options_set_async_wal_precreate(rocksdb_options_t* opt,
+                                             unsigned char v) {
+  opt->rep.async_wal_precreate = v;
+}
+
+unsigned char rocksdb_options_get_async_wal_precreate(rocksdb_options_t* opt) {
+  return opt->rep.async_wal_precreate;
+}
+
 void rocksdb_options_set_soft_pending_compaction_bytes_limit(
     rocksdb_options_t* opt, size_t v) {
   opt->rep.soft_pending_compaction_bytes_limit = v;
@@ -5389,6 +5441,16 @@ size_t rocksdb_options_get_memtable_huge_page_size(rocksdb_options_t* opt) {
   return opt->rep.memtable_huge_page_size;
 }
 
+void rocksdb_options_set_memtable_batch_lookup_optimization(
+    rocksdb_options_t* opt, unsigned char v) {
+  opt->rep.memtable_batch_lookup_optimization = v;
+}
+
+unsigned char rocksdb_options_get_memtable_batch_lookup_optimization(
+    rocksdb_options_t* opt) {
+  return opt->rep.memtable_batch_lookup_optimization;
+}
+
 void rocksdb_options_set_hash_skip_list_rep(rocksdb_options_t* opt,
                                             size_t bucket_count,
                                             int32_t skiplist_height,
@@ -5820,6 +5882,8 @@ uint64_t rocksdb_perfcontext_metric(rocksdb_perfcontext_t* context,
       return rep->number_async_seek;
     case rocksdb_blob_cache_hit_count:
       return rep->blob_cache_hit_count;
+    case rocksdb_blob_cache_read_byte:
+      return rep->blob_cache_read_byte;
     case rocksdb_blob_read_count:
       return rep->blob_read_count;
     case rocksdb_blob_read_byte:
@@ -6238,6 +6302,16 @@ unsigned char rocksdb_readoptions_get_async_io(rocksdb_readoptions_t* opt) {
   return opt->rep.async_io;
 }
 
+void rocksdb_readoptions_set_optimize_multiget_for_io(
+    rocksdb_readoptions_t* opt, unsigned char v) {
+  opt->rep.optimize_multiget_for_io = v;
+}
+
+unsigned char rocksdb_readoptions_get_optimize_multiget_for_io(
+    rocksdb_readoptions_t* opt) {
+  return opt->rep.optimize_multiget_for_io;
+}
+
 void rocksdb_readoptions_set_timestamp(rocksdb_readoptions_t* opt,
                                        const char* ts, size_t tslen) {
   if (ts == nullptr) {
@@ -7491,6 +7565,11 @@ void rocksdb_transactiondb_options_set_default_lock_timeout(
   opt->rep.default_lock_timeout = default_lock_timeout;
 }
 
+void rocksdb_transactiondb_options_set_write_policy(
+    rocksdb_transactiondb_options_t* opt, int write_policy) {
+  opt->rep.write_policy = static_cast<TxnDBWritePolicy>(write_policy);
+}
+
 void rocksdb_transactiondb_options_set_use_per_key_point_lock_mgr(
     rocksdb_transactiondb_options_t* opt, int use_per_key_point_lock_mgr) {
   opt->rep.use_per_key_point_lock_mgr = use_per_key_point_lock_mgr;

db/c_test.c

@@ -6,6 +6,7 @@
 #include "rocksdb/c.h"
 
 #include <assert.h>
+#include <inttypes.h>
 #include <stddef.h>
 #include <stdio.h>
 #include <stdlib.h>
@@ -14,7 +15,6 @@
 #ifndef OS_WIN
 #include <unistd.h>
 #endif
-#include <inttypes.h>
 
 // Can not use port/port.h macros as this is a c file
 #ifdef OS_WIN
@@ -32,6 +32,14 @@ int geteuid() {
 
 #endif
 
+static uint64_t GetTestId(void) {
+#ifdef OS_WIN
+  return (uint64_t)geteuid();
+#else
+  return ((uint64_t)geteuid() << 32) ^ (uint64_t)getpid();
+#endif
+}
+
 const char* phase = "";
 static char dbname[200];
 static char sstfilename[200];
@@ -848,20 +856,21 @@ int main(int argc, char** argv) {
   char* err = NULL;
   int run = -1;
 
-  snprintf(dbname, sizeof(dbname), "%s/rocksdb_c_test-%d", GetTempDir(),
-           ((int)geteuid()));
+  snprintf(dbname, sizeof(dbname), "%s/rocksdb_c_test-%" PRIu64, GetTempDir(),
+           GetTestId());
 
-  snprintf(dbbackupname, sizeof(dbbackupname), "%s/rocksdb_c_test-%d-backup",
-           GetTempDir(), ((int)geteuid()));
+  snprintf(dbbackupname, sizeof(dbbackupname),
+           "%s/rocksdb_c_test-%" PRIu64 "-backup", GetTempDir(), GetTestId());
 
   snprintf(dbcheckpointname, sizeof(dbcheckpointname),
-           "%s/rocksdb_c_test-%d-checkpoint", GetTempDir(), ((int)geteuid()));
+           "%s/rocksdb_c_test-%" PRIu64 "-checkpoint", GetTempDir(),
+           GetTestId());
 
-  snprintf(sstfilename, sizeof(sstfilename), "%s/rocksdb_c_test-%d-sst",
-           GetTempDir(), ((int)geteuid()));
+  snprintf(sstfilename, sizeof(sstfilename),
+           "%s/rocksdb_c_test-%" PRIu64 "-sst", GetTempDir(), GetTestId());
 
-  snprintf(dbpathname, sizeof(dbpathname), "%s/rocksdb_c_test-%d-dbpath",
-           GetTempDir(), ((int)geteuid()));
+  snprintf(dbpathname, sizeof(dbpathname),
+           "%s/rocksdb_c_test-%" PRIu64 "-dbpath", GetTempDir(), GetTestId());
 
   StartPhase("create_objects");
   cmp = rocksdb_comparator_create(NULL, CmpDestroy, CmpCompare, CmpName);
@@ -888,6 +897,9 @@ int main(int argc, char** argv) {
   rocksdb_block_based_options_set_block_cache(table_options, cache);
   rocksdb_block_based_options_set_data_block_index_type(table_options, 1);
   rocksdb_block_based_options_set_data_block_hash_ratio(table_options, 0.75);
+  rocksdb_block_based_options_set_index_block_search_type(
+      table_options, rocksdb_block_based_table_index_block_search_type_auto);
+  rocksdb_block_based_options_set_uniform_cv_threshold(table_options, 0.2);
   rocksdb_block_based_options_set_top_level_index_pinning_tier(table_options,
                                                                1);
   rocksdb_block_based_options_set_partition_pinning_tier(table_options, 2);
@@ -948,6 +960,19 @@ int main(int argc, char** argv) {
   CheckNoError(err);
   CheckGet(db, roptions, "foo", "hello");
 
+  StartPhase("set_db_options");
+  {
+    const char* keys[] = {"stats_dump_period_sec"};
+    const char* values[] = {"10"};
+    rocksdb_set_db_options(db, 1, keys, values, &err);
+    CheckNoError(err);
+
+    keys[0] = "not_a_db_option";
+    rocksdb_set_db_options(db, 1, keys, values, &err);
+    CheckCondition(err != NULL);
+    Free(&err);
+  }
+
   StartPhase("backup_and_restore");
   {
     rocksdb_destroy_db(options, dbbackupname, &err);
@@ -1015,6 +1040,12 @@ int main(int argc, char** argv) {
     Free(&before_db_id);
     Free(&after_db_id);
 
+    rocksdb_backup_engine_stop_backup(be);
+    rocksdb_backup_engine_create_new_backup(be, db, &err);
+    CheckCondition(err != NULL);
+    free(err);
+    err = NULL;
+
     rocksdb_backup_engine_close(be);
   }
 
@@ -1144,9 +1175,11 @@ int main(int argc, char** argv) {
     static char cf_export_path[200];
     static char db_import_path[200];
     snprintf(cf_export_path, sizeof(cf_export_path),
-             "%s/rocksdb_c_test-%d-cf_export", GetTempDir(), ((int)geteuid()));
+             "%s/rocksdb_c_test-%" PRIu64 "-cf_export", GetTempDir(),
+             GetTestId());
     snprintf(db_import_path, sizeof(db_import_path),
-             "%s/rocksdb_c_test-%d-db_import", GetTempDir(), ((int)geteuid()));
+             "%s/rocksdb_c_test-%" PRIu64 "-db_import", GetTempDir(),
+             GetTestId());
 
     rocksdb_options_t* db_options = rocksdb_options_create();
     rocksdb_column_family_handle_t* cf_export =
@@ -2055,8 +2088,9 @@ int main(int argc, char** argv) {
     // use dbpathname2 as the cf_path for "cf1"
     rocksdb_dbpath_t* dbpath2;
     char dbpathname2[200];
-    snprintf(dbpathname2, sizeof(dbpathname2), "%s/rocksdb_c_test-%d-dbpath2",
-             GetTempDir(), ((int)geteuid()));
+    snprintf(dbpathname2, sizeof(dbpathname2),
+             "%s/rocksdb_c_test-%" PRIu64 "-dbpath2", GetTempDir(),
+             GetTestId());
     dbpath2 = rocksdb_dbpath_create(dbpathname2, 1024 * 1024);
     const rocksdb_dbpath_t* cf_paths[1] = {dbpath2};
     rocksdb_options_set_cf_paths(cf_options_2, cf_paths, 1);
@@ -2757,6 +2791,10 @@ int main(int argc, char** argv) {
     CheckCondition(
         1 == rocksdb_options_get_use_direct_io_for_flush_and_compaction(o));
 
+    rocksdb_options_set_use_direct_io_for_compaction_reads(o, 1);
+    CheckCondition(1 ==
+                   rocksdb_options_get_use_direct_io_for_compaction_reads(o));
+
     rocksdb_options_set_is_fd_close_on_exec(o, 1);
     CheckCondition(1 == rocksdb_options_get_is_fd_close_on_exec(o));
 
@@ -2812,6 +2850,15 @@ int main(int argc, char** argv) {
     rocksdb_options_set_memtable_huge_page_size(o, 25);
     CheckCondition(25 == rocksdb_options_get_memtable_huge_page_size(o));
 
+    // memtable_batch_lookup_optimization defaults to 0; flip to 1 to verify
+    // the setter is wired through to the underlying C++ ColumnFamilyOptions
+    // field.
+    CheckCondition(0 ==
+                   rocksdb_options_get_memtable_batch_lookup_optimization(o));
+    rocksdb_options_set_memtable_batch_lookup_optimization(o, 1);
+    CheckCondition(1 ==
+                   rocksdb_options_get_memtable_batch_lookup_optimization(o));
+
     rocksdb_options_set_max_successive_merges(o, 26);
     CheckCondition(26 == rocksdb_options_get_max_successive_merges(o));
 
@@ -2869,6 +2916,9 @@ int main(int argc, char** argv) {
     rocksdb_options_set_track_and_verify_wals_in_manifest(o, 42);
     CheckCondition(1 ==
                    rocksdb_options_get_track_and_verify_wals_in_manifest(o));
+    CheckCondition(0 == rocksdb_options_get_async_wal_precreate(o));
+    rocksdb_options_set_async_wal_precreate(o, 1);
+    CheckCondition(1 == rocksdb_options_get_async_wal_precreate(o));
 
     /* Blob Options */
     rocksdb_options_set_enable_blob_files(o, 1);
@@ -2972,6 +3022,8 @@ int main(int argc, char** argv) {
     CheckCondition(1 == rocksdb_options_get_use_direct_reads(copy));
     CheckCondition(
         1 == rocksdb_options_get_use_direct_io_for_flush_and_compaction(copy));
+    CheckCondition(
+        1 == rocksdb_options_get_use_direct_io_for_compaction_reads(copy));
     CheckCondition(1 == rocksdb_options_get_is_fd_close_on_exec(copy));
     CheckCondition(18 == rocksdb_options_get_stats_dump_period_sec(copy));
     CheckCondition(5 == rocksdb_options_get_stats_persist_period_sec(copy));
@@ -2995,6 +3047,8 @@ int main(int argc, char** argv) {
                    rocksdb_options_get_memtable_prefix_bloom_size_ratio(copy));
     CheckCondition(24 == rocksdb_options_get_max_compaction_bytes(copy));
     CheckCondition(25 == rocksdb_options_get_memtable_huge_page_size(copy));
+    CheckCondition(
+        1 == rocksdb_options_get_memtable_batch_lookup_optimization(copy));
     CheckCondition(26 == rocksdb_options_get_max_successive_merges(copy));
     CheckCondition(27 == rocksdb_options_get_bloom_locality(copy));
     CheckCondition(1 == rocksdb_options_get_inplace_update_support(copy));
@@ -3248,6 +3302,12 @@ int main(int argc, char** argv) {
     CheckCondition(
         1 == rocksdb_options_get_use_direct_io_for_flush_and_compaction(o));
 
+    rocksdb_options_set_use_direct_io_for_compaction_reads(copy, 0);
+    CheckCondition(
+        0 == rocksdb_options_get_use_direct_io_for_compaction_reads(copy));
+    CheckCondition(1 ==
+                   rocksdb_options_get_use_direct_io_for_compaction_reads(o));
+
     rocksdb_options_set_is_fd_close_on_exec(copy, 0);
     CheckCondition(0 == rocksdb_options_get_is_fd_close_on_exec(copy));
     CheckCondition(1 == rocksdb_options_get_is_fd_close_on_exec(o));
@@ -3429,6 +3489,12 @@ int main(int argc, char** argv) {
     rocksdb_readoptions_set_async_io(ro, 1);
     CheckCondition(1 == rocksdb_readoptions_get_async_io(ro));
 
+    // optimize_multiget_for_io defaults to true (1); flip to 0 to verify the
+    // setter is wired through to the underlying C++ ReadOptions field.
+    CheckCondition(1 == rocksdb_readoptions_get_optimize_multiget_for_io(ro));
+    rocksdb_readoptions_set_optimize_multiget_for_io(ro, 0);
+    CheckCondition(0 == rocksdb_readoptions_get_optimize_multiget_for_io(ro));
+
     rocksdb_readoptions_destroy(ro);
   }
 
@@ -3674,6 +3740,52 @@ int main(int argc, char** argv) {
     CheckCondition(37 ==
                    rocksdb_backup_engine_options_get_restore_rate_limit(bdo));
 
+    {
+      rocksdb_ratelimiter_t* limiter = rocksdb_ratelimiter_create_with_mode(
+          1024 * 1024, 100 * 1000, 10, 2, 0);
+      rocksdb_backup_engine_options_set_backup_rate_limiter(bdo, limiter);
+      rocksdb_backup_engine_options_set_restore_rate_limiter(bdo, limiter);
+      rocksdb_ratelimiter_destroy(limiter);
+
+      rocksdb_backup_engine_options_t* rate_beo =
+          rocksdb_backup_engine_options_create(dbbackupname);
+      rocksdb_ratelimiter_t* backup_limiter =
+          rocksdb_ratelimiter_create_with_mode(1024 * 1024, 100 * 1000, 10, 2,
+                                               0);
+      rocksdb_ratelimiter_t* restore_limiter =
+          rocksdb_ratelimiter_create_with_mode(1024 * 1024, 100 * 1000, 10, 2,
+                                               0);
+      rocksdb_backup_engine_options_set_backup_rate_limiter(rate_beo,
+                                                            backup_limiter);
+      rocksdb_backup_engine_options_set_restore_rate_limiter(rate_beo,
+                                                             restore_limiter);
+      rocksdb_ratelimiter_destroy(backup_limiter);
+      rocksdb_ratelimiter_destroy(restore_limiter);
+      rocksdb_env_t* rate_benv = rocksdb_create_default_env();
+      rocksdb_backup_engine_t* rate_be =
+          rocksdb_backup_engine_open_opts(rate_beo, rate_benv, &err);
+      rocksdb_backup_engine_options_destroy(rate_beo);
+      rocksdb_env_destroy(rate_benv);
+      CheckNoError(err);
+      rocksdb_backup_engine_create_new_backup(rate_be, db, &err);
+      CheckNoError(err);
+      {
+        char rate_restore_path[220];
+        snprintf(rate_restore_path, sizeof(rate_restore_path),
+                 "%s.rate_restore", dbname);
+        rocksdb_restore_options_t* rate_restore_opts =
+            rocksdb_restore_options_create();
+        rocksdb_backup_engine_restore_db_from_latest_backup(
+            rate_be, rate_restore_path, rate_restore_path, rate_restore_opts,
+            &err);
+        CheckNoError(err);
+        rocksdb_restore_options_destroy(rate_restore_opts);
+        rocksdb_destroy_db(options, rate_restore_path, &err);
+        err = NULL;
+      }
+      rocksdb_backup_engine_close(rate_be);
+    }
+
     rocksdb_backup_engine_options_set_max_background_operations(bdo, 20);
     CheckCondition(
         20 == rocksdb_backup_engine_options_get_max_background_operations(bdo));
@@ -3785,6 +3897,43 @@ int main(int argc, char** argv) {
     }
   }
 
+  StartPhase("transactiondb_set_write_policy_prepared");
+  {
+    char wp_dbname[200];
+    rocksdb_transactiondb_t* wp_txn_db;
+    rocksdb_transactiondb_options_t* wp_txn_opts;
+    snprintf(wp_dbname, sizeof(wp_dbname), "%s/rocksdb_c_test-txnwp-%" PRIu64,
+             GetTempDir(), GetTestId());
+    rocksdb_destroy_db(options, wp_dbname, &err);
+    Free(&err);
+    wp_txn_opts = rocksdb_transactiondb_options_create();
+    rocksdb_transactiondb_options_set_write_policy(
+        wp_txn_opts, rocksdb_txndb_write_policy_write_prepared);
+    rocksdb_options_set_create_if_missing(options, 1);
+    // create new TransactionDB with write policy WRITE_PREPARED
+    wp_txn_db =
+        rocksdb_transactiondb_open(options, wp_txn_opts, wp_dbname, &err);
+    CheckNoError(err);
+    {
+      rocksdb_transaction_options_t* wp_txn_options =
+          rocksdb_transaction_options_create();
+      rocksdb_transaction_t* wp_txn =
+          rocksdb_transaction_begin(wp_txn_db, woptions, wp_txn_options, NULL);
+      // write one record within a transaction
+      rocksdb_transaction_put(wp_txn, "k", 1, "v", 1, &err);
+      CheckNoError(err);
+      rocksdb_transaction_commit(wp_txn, &err);
+      CheckNoError(err);
+      rocksdb_transaction_destroy(wp_txn);
+      rocksdb_transaction_options_destroy(wp_txn_options);
+    }
+    CheckTxnDBGet(wp_txn_db, roptions, "k", "v");
+    rocksdb_transactiondb_close(wp_txn_db);
+    rocksdb_transactiondb_options_destroy(wp_txn_opts);
+    rocksdb_destroy_db(options, wp_dbname, &err);
+    CheckNoError(err);
+  }
+
   StartPhase("transactions");
   {
     rocksdb_close(db);
@@ -4422,7 +4571,7 @@ int main(int argc, char** argv) {
     rocksdb_options_set_max_open_files(opts, -1);
     rocksdb_options_set_create_if_missing(opts, 1);
     snprintf(secondary_path, sizeof(secondary_path),
-             "%s/rocksdb_c_test_secondary-%d", GetTempDir(), ((int)geteuid()));
+             "%s/rocksdb_c_test_secondary-%" PRIu64, GetTempDir(), GetTestId());
     db1 = rocksdb_open_as_secondary(opts, dbname, secondary_path, &err);
     CheckNoError(err);
 
@@ -4881,7 +5030,13 @@ int main(int argc, char** argv) {
     rocksdb_options_set_create_if_missing(null_opts, 1);
     rocksdb_options_set_compaction_service(null_opts, null_service);
 
-    const char* null_db = "rocksdb_c_test_null_service";
+    char null_db[200];
+    int null_db_len = snprintf(null_db, sizeof(null_db),
+                               "%s/rocksdb_c_test_null_service-%" PRIu64,
+                               GetTempDir(), GetTestId());
+    if (null_db_len <= 0 || (size_t)null_db_len >= sizeof(null_db)) {
+      abort();
+    }
 
     rocksdb_t* null_db_handle = rocksdb_open(null_opts, null_db, &err);
     CheckNoError(err);

db/compaction/compaction.h

@@ -454,6 +454,18 @@ class Compaction {
     return notify_on_compaction_completion_;
   }
 
+  // Called by DBImpl::NotifyOnCompactionPreCommit to ensure the
+  // OnCompactionPreCommit listener callback fires at most once per
+  // compaction lifecycle, even though the notify helper is invoked at multiple
+  // potential ReleaseCompactionFiles() sites for safety.
+  void SetNotifyOnCompactionPreCommitCalled() {
+    notify_on_compaction_pre_commit_called_ = true;
+  }
+
+  bool WasNotifyOnCompactionPreCommitCalled() const {
+    return notify_on_compaction_pre_commit_called_;
+  }
+
   static constexpr int kInvalidLevel = -1;
 
   // Evaluate proximal output level. If the compaction supports
@@ -626,6 +638,12 @@ class Compaction {
   // begin.
   bool notify_on_compaction_completion_;
 
+  // Tracks whether OnCompactionPreCommit has already fired for this
+  // compaction. The notify helper is called from multiple potential
+  // ReleaseCompactionFiles() call sites; this flag guarantees that listeners
+  // observe the callback exactly once per compaction.
+  bool notify_on_compaction_pre_commit_called_ = false;
+
   // Enable/disable GC collection for blobs during compaction.
   bool enable_blob_garbage_collection_;
 

db/compaction/compaction_job.cc

@@ -173,6 +173,7 @@ CompactionJob::CompactionJob(
       fs_(db_options.fs, io_tracer),
       file_options_for_read_(
           fs_->OptimizeForCompactionTableRead(file_options, db_options_)),
+      file_options_for_compaction_input_read_(file_options_for_read_),
       versions_(versions),
       shutting_down_(shutting_down),
       manual_compaction_canceled_(manual_compaction_canceled),
@@ -203,6 +204,11 @@ CompactionJob::CompactionJob(
   assert(job_context);
   assert(job_context->snapshot_context_initialized);
 
+  if (db_options_.use_direct_io_for_compaction_reads &&
+      !db_options_.use_direct_reads) {
+    file_options_for_compaction_input_read_.use_direct_reads = true;
+  }
+
   const auto* cfd = compact_->compaction->column_family_data();
   ThreadStatusUtil::SetEnableTracking(db_options_.enable_thread_tracking);
   ThreadStatusUtil::SetColumnFamily(cfd);
@@ -1536,10 +1542,20 @@ InternalIterator* CompactionJob::CreateInputIterator(
 
   // Although the v2 aggregator is what the level iterator(s) know about,
   // the AddTombstones calls will be propagated down to the v1 aggregator.
+  const bool open_ephemeral_table_reader =
+      db_options_.use_direct_io_for_compaction_reads &&
+      !db_options_.use_direct_reads;
+  FileOptions& input_file_options =
+      open_ephemeral_table_reader ? file_options_for_compaction_input_read_
+                                  : file_options_for_read_;
+  TEST_SYNC_POINT_CALLBACK(
+      "CompactionJob::CreateInputIterator:InputFileOptions",
+      &input_file_options);
   iterators.raw_input =
       std::unique_ptr<InternalIterator>(versions_->MakeInputIterator(
           read_options, sub_compact->compaction, sub_compact->RangeDelAgg(),
-          file_options_for_read_, boundaries.start, boundaries.end));
+          input_file_options, boundaries.start, boundaries.end,
+          open_ephemeral_table_reader));
   InternalIterator* input = iterators.raw_input.get();
 
   if (boundaries.start.has_value() || boundaries.end.has_value()) {
@@ -2065,10 +2081,18 @@ Status CompactionJob::FinishCompactionOutputFile(
     std::pair<SequenceNumber, SequenceNumber> keep_seqno_range{
         0, kMaxSequenceNumber};
     if (sub_compact->compaction->SupportsPerKeyPlacement()) {
+      // Point entries are routed to proximal output only when their seqno is
+      // strictly greater than `proximal_after_seqno_`. Range tombstones use a
+      // [lower, upper) filter, so split them at the next seqno to preserve the
+      // same boundary.
+      SequenceNumber range_del_split_seqno = proximal_after_seqno_;
+      if (range_del_split_seqno < kMaxSequenceNumber) {
+        range_del_split_seqno++;
+      }
       if (outputs.IsProximalLevel()) {
-        keep_seqno_range.first = proximal_after_seqno_;
+        keep_seqno_range.first = range_del_split_seqno;
       } else {
-        keep_seqno_range.second = proximal_after_seqno_;
+        keep_seqno_range.second = range_del_split_seqno;
       }
     }
     CompactionIterationStats range_del_out_stats;
@@ -2462,6 +2486,8 @@ Status CompactionJob::OpenCompactionOutputFile(SubcompactionState* sub_compact,
   auto temperature =
       sub_compact->compaction->GetOutputTemperature(outputs.IsProximalLevel());
   fo_copy.temperature = temperature;
+  fo_copy.open_contract = FileOpenContract::kNoReopenForWrite |
+                          FileOpenContract::kNoReadersWhileOpenForWrite;
   fo_copy.write_hint = write_hint_;
 
   Status s;

db/compaction/compaction_job.h

@@ -462,6 +462,8 @@ class CompactionJob {
   FileSystemPtr fs_;
   // env_option optimized for compaction table reads
   FileOptions file_options_for_read_;
+  // file_options_for_read_ plus compaction-input-only overrides.
+  FileOptions file_options_for_compaction_input_read_;
   VersionSet* versions_;
   const std::atomic<bool>* shutting_down_;
   const std::atomic<bool>& manual_compaction_canceled_;

db/compaction/compaction_service_job.cc

@@ -189,14 +189,24 @@ CompactionJob::ProcessKeyValueCompactionWithCompactionService(
     return compaction_status;
   }
 
-  // CompactionServiceJobStatus::kSuccess was returned, but somehow we failed to
-  // read the result. Consider this as an installation failure
   if (!s.ok()) {
-    sub_compact->status = s;
+    // Wait() returned kSuccess, but the primary host could not deserialize the
+    // remote result before importing any output files into this DB. The remote
+    // output is still isolated in the service-managed staging directory, so it
+    // is safe to fall back to local compaction for the same job and notify the
+    // service via OnInstallation(kUseLocal).
+    assert(sub_compact->status.ok());
+    assert(sub_compact->GetMutableCompactionOutputs().empty());
+    assert(sub_compact->GetMutableProximalOutputs().empty());
+    ROCKS_LOG_WARN(db_options_.info_log,
+                   "[%s] [JOB %d] Failed to parse remote compaction result "
+                   "(%s), falling back to local compaction",
+                   compaction->column_family_data()->GetName().c_str(), job_id_,
+                   s.ToString().c_str());
     compaction_result.status.PermitUncheckedError();
     db_options_.compaction_service->OnInstallation(
-        response.scheduled_job_id, CompactionServiceJobStatus::kFailure);
-    return CompactionServiceJobStatus::kFailure;
+        response.scheduled_job_id, CompactionServiceJobStatus::kUseLocal);
+    return CompactionServiceJobStatus::kUseLocal;
   }
   sub_compact->status = compaction_result.status;
 
@@ -421,6 +431,14 @@ Status CompactionServiceCompactionJob::Run() {
   // Please note that input stats will be updated by primary host when all
   // subcompactions are finished
   UpdateCompactionJobOutputStatsFromInternalStats(status, internal_stats_);
+  // Whole-file filtering changes which keys are fed to the iterator, so the
+  // iterator-based input record count should not be treated as exact.
+  for (size_t level = 1; level < c->num_input_levels(); ++level) {
+    if (!c->filtered_input_levels(level).empty()) {
+      compaction_result_->stats.has_accurate_num_input_records = false;
+      break;
+    }
+  }
   // and set fields that are not propagated as part of the update
   compaction_result_->stats.is_manual_compaction = c->is_manual_compaction();
   compaction_result_->stats.is_full_compaction = c->is_full_compaction();
@@ -641,6 +659,10 @@ static std::unordered_map<std::string, OptionTypeInfo>
         {"cpu_micros",
          {offsetof(struct CompactionJobStats, cpu_micros), OptionType::kUInt64T,
           OptionVerificationType::kNormal, OptionTypeFlags::kNone}},
+        {"has_accurate_num_input_records",
+         {offsetof(struct CompactionJobStats, has_accurate_num_input_records),
+          OptionType::kBoolean, OptionVerificationType::kNormal,
+          OptionTypeFlags::kNone}},
         {"num_input_records",
          {offsetof(struct CompactionJobStats, num_input_records),
           OptionType::kUInt64T, OptionVerificationType::kNormal,

db/compaction/compaction_service_test.cc

@@ -127,6 +127,7 @@ class MyTestCompactionService : public CompactionService {
 
   void OnInstallation(const std::string& /*scheduled_job_id*/,
                       CompactionServiceJobStatus status) override {
+    installation_callback_count_.fetch_add(1);
     final_updated_status_ = status;
   }
 
@@ -167,6 +168,8 @@ class MyTestCompactionService : public CompactionService {
     return final_updated_status_.load();
   }
 
+  int GetOnInstallationCount() { return installation_callback_count_.load(); }
+
  protected:
   InstrumentedMutex mutex_;
   const std::string db_path_;
@@ -195,6 +198,7 @@ class MyTestCompactionService : public CompactionService {
   std::vector<std::shared_ptr<EventListener>> listeners_;
   std::vector<std::shared_ptr<TablePropertiesCollectorFactory>>
       table_properties_collector_factories_;
+  std::atomic_int installation_callback_count_{0};
   std::atomic<CompactionServiceJobStatus> final_updated_status_{
       CompactionServiceJobStatus::kUseLocal};
 
@@ -607,6 +611,101 @@ TEST_F(CompactionServiceTest, StandaloneDeleteRangeTombstoneOptimization) {
   SyncPoint::GetInstance()->DisableProcessing();
 }
 
+TEST_F(CompactionServiceTest,
+       StandaloneDeleteRangeTombstoneMakesInputCountInaccurate) {
+  Options options = CurrentOptions();
+  options.compaction_style = CompactionStyle::kCompactionStyleUniversal;
+  options.compaction_verify_record_count = true;
+  ReopenWithCompactionService(&options);
+
+  std::vector<std::string> files;
+  {
+    SstFileWriter sst_file_writer(EnvOptions(), options);
+    std::string file1 = dbname_ + "file1.sst";
+    ASSERT_OK(sst_file_writer.Open(file1));
+    ASSERT_OK(sst_file_writer.Put("a", "a1"));
+    ASSERT_OK(sst_file_writer.Put("b", "b1"));
+    ExternalSstFileInfo file1_info;
+    ASSERT_OK(sst_file_writer.Finish(&file1_info));
+    files.push_back(std::move(file1));
+
+    std::string file2 = dbname_ + "file2.sst";
+    ASSERT_OK(sst_file_writer.Open(file2));
+    ASSERT_OK(sst_file_writer.Put("x", "x1"));
+    ASSERT_OK(sst_file_writer.Put("y", "y1"));
+    ExternalSstFileInfo file2_info;
+    ASSERT_OK(sst_file_writer.Finish(&file2_info));
+    files.push_back(std::move(file2));
+  }
+
+  IngestExternalFileOptions ifo;
+  ASSERT_OK(db_->IngestExternalFile(files, ifo));
+  ASSERT_EQ(Get("a"), "a1");
+  ASSERT_EQ(Get("b"), "b1");
+  ASSERT_EQ(Get("x"), "x1");
+  ASSERT_EQ(Get("y"), "y1");
+  ASSERT_EQ(2, NumTableFilesAtLevel(6));
+
+  auto my_cs = GetCompactionService();
+  uint64_t comp_num = my_cs->GetCompactionNum();
+
+  size_t num_files_after_filtered = 0;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "VersionSet::MakeInputIterator:NewCompactionMergingIterator",
+      [&](void* arg) {
+        num_files_after_filtered = *static_cast<size_t*>(arg);
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  {
+    // The standalone range tombstone can fully cover the old versioned files,
+    // so universal compaction may filter those files before iteration.
+    files.clear();
+    SstFileWriter sst_file_writer(EnvOptions(), options);
+    std::string file2 = dbname_ + "file2.sst";
+    ASSERT_OK(sst_file_writer.Open(file2));
+    ASSERT_OK(sst_file_writer.DeleteRange("a", "z"));
+    ExternalSstFileInfo file2_info;
+    ASSERT_OK(sst_file_writer.Finish(&file2_info));
+    files.push_back(std::move(file2));
+
+    std::string file3 = dbname_ + "file3.sst";
+    ASSERT_OK(sst_file_writer.Open(file3));
+    ASSERT_OK(sst_file_writer.Put("a", "a2"));
+    ASSERT_OK(sst_file_writer.Put("b", "b2"));
+    ExternalSstFileInfo file3_info;
+    ASSERT_OK(sst_file_writer.Finish(&file3_info));
+    files.push_back(std::move(file3));
+
+    std::string file4 = dbname_ + "file4.sst";
+    ASSERT_OK(sst_file_writer.Open(file4));
+    ASSERT_OK(sst_file_writer.Put("x", "x2"));
+    ASSERT_OK(sst_file_writer.Put("y", "y2"));
+    ExternalSstFileInfo file4_info;
+    ASSERT_OK(sst_file_writer.Finish(&file4_info));
+    files.push_back(std::move(file4));
+  }
+
+  ASSERT_OK(db_->IngestExternalFile(files, ifo));
+  ASSERT_OK(db_->WaitForCompact(WaitForCompactOptions()));
+  ASSERT_GE(my_cs->GetCompactionNum(), comp_num + 1);
+
+  CompactionServiceResult result;
+  my_cs->GetResult(&result);
+  ASSERT_OK(result.status);
+  ASSERT_TRUE(result.stats.is_remote_compaction);
+  ASSERT_FALSE(result.stats.has_accurate_num_input_records);
+  ASSERT_EQ(1, num_files_after_filtered);
+
+  ASSERT_EQ(Get("a"), "a2");
+  ASSERT_EQ(Get("b"), "b2");
+  ASSERT_EQ(Get("x"), "x2");
+  ASSERT_EQ(Get("y"), "y2");
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
 TEST_F(CompactionServiceTest, CompactionOutputFileIOError) {
   Options options = CurrentOptions();
   options.disable_auto_compactions = true;
@@ -939,6 +1038,37 @@ TEST_F(CompactionServiceTest, VerifyInputRecordCount) {
   SyncPoint::GetInstance()->ClearAllCallBacks();
 }
 
+TEST_F(CompactionServiceTest, InaccurateInputRecordCount) {
+  Options options = CurrentOptions();
+  options.disable_auto_compactions = true;
+  ReopenWithCompactionService(&options);
+  GenerateTestData();
+
+  auto my_cs = GetCompactionService();
+
+  std::string start_str = Key(15);
+  std::string end_str = Key(45);
+  Slice start(start_str);
+  Slice end(end_str);
+  uint64_t comp_num = my_cs->GetCompactionNum();
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "CompactionServiceCompactionJob::Run:0", [&](void* arg) {
+        CompactionServiceResult* compaction_result =
+            *(static_cast<CompactionServiceResult**>(arg));
+        ASSERT_TRUE(compaction_result != nullptr);
+        compaction_result->stats.has_accurate_num_input_records = false;
+        compaction_result->stats.num_input_records = 0;
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), &start, &end));
+  ASSERT_GE(my_cs->GetCompactionNum(), comp_num + 1);
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
 TEST_F(CompactionServiceTest, EmptyResult) {
   Options options = CurrentOptions();
   options.disable_auto_compactions = true;
@@ -1281,6 +1411,10 @@ TEST_F(CompactionServiceTest, TruncatedOutput) {
 
 TEST_F(CompactionServiceTest, CustomFileChecksum) {
   Options options = CurrentOptions();
+  // Pin compression so the auto-compacted LSM shape (and thus whether the
+  // manual CompactRange below schedules a remote compaction) doesn't depend on
+  // the default compression type. kNoCompression is always available.
+  options.compression = kNoCompression;
   options.file_checksum_gen_factory = GetFileChecksumGenCrc32cFactory();
   ReopenWithCompactionService(&options);
   GenerateTestData();
@@ -1448,7 +1582,7 @@ TEST_F(CompactionServiceTest, FailedToStart) {
   ASSERT_TRUE(s.IsIncomplete());
 }
 
-TEST_F(CompactionServiceTest, InvalidResult) {
+TEST_F(CompactionServiceTest, InvalidResultFallsBackToLocal) {
   Options options = CurrentOptions();
   options.disable_auto_compactions = true;
   ReopenWithCompactionService(&options);
@@ -1456,15 +1590,22 @@ TEST_F(CompactionServiceTest, InvalidResult) {
   GenerateTestData();
 
   auto my_cs = GetCompactionService();
+  ASSERT_EQ(0, my_cs->GetOnInstallationCount());
   my_cs->OverrideWaitResult("Invalid Str");
+  int compaction_num_before = my_cs->GetCompactionNum();
 
   std::string start_str = Key(15);
   std::string end_str = Key(45);
   Slice start(start_str);
   Slice end(end_str);
+  // The remote result is malformed before any installation starts, so the
+  // primary should recover by rerunning the compaction locally for this job.
   Status s = db_->CompactRange(CompactRangeOptions(), &start, &end);
-  ASSERT_FALSE(s.ok());
-  ASSERT_EQ(CompactionServiceJobStatus::kFailure,
+  ASSERT_OK(s);
+  ASSERT_GE(my_cs->GetCompactionNum(), compaction_num_before + 1);
+  VerifyTestData();
+  ASSERT_EQ(1, my_cs->GetOnInstallationCount());
+  ASSERT_EQ(CompactionServiceJobStatus::kUseLocal,
             my_cs->GetFinalCompactionServiceJobStatus());
 }
 

db/compaction/tiered_compaction_test.cc

@@ -2897,6 +2897,61 @@ TEST_P(PrecludeLastLevelTest, RangeDelsCauseFileEndpointsToOverlap) {
   Close();
 }
 
+TEST_F(PrecludeLastLevelTestBase,
+       RangeDelAtProximalSeqnoBoundaryStaysInLastLevel) {
+  constexpr int kNumLevels = 7;
+
+  Options options = CurrentOptions();
+  options.env = mock_env_.get();
+  options.num_levels = kNumLevels;
+  options.disable_auto_compactions = true;
+  DestroyAndReopen(options);
+  Defer close_db([&] { Close(); });
+
+  ASSERT_OK(Put(Key(2), "old"));
+  ASSERT_OK(Put(Key(12), "old"));
+  ManagedSnapshot snapshot(db_.get());
+  ASSERT_OK(db_->DeleteRange(WriteOptions(), db_->DefaultColumnFamily(), Key(2),
+                             Key(12)));
+  ASSERT_OK(Flush());
+  MoveFilesToLevel(6);
+
+  const int l5_file_keys[][2] = {{0, 4}, {5, 9}};
+  for (const auto& l5_file : l5_file_keys) {
+    ASSERT_OK(Put(Key(l5_file[0]), "hot"));
+    ASSERT_OK(Put(Key(l5_file[1]), "hot"));
+    ASSERT_OK(Flush());
+    MoveFilesToLevel(5);
+  }
+  ASSERT_EQ("0,0,0,0,0,2,1", FilesPerLevel());
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "CompactionJob::PrepareTimes():preclude_last_level_min_seqno",
+      [](void* arg) { *static_cast<SequenceNumber*>(arg) = 0; });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(db_->SetOptions({{"preclude_last_level_data_seconds", "10000"}}));
+
+  auto l5_files = GetLevelFileMetadatas(5);
+  auto l6_files = GetLevelFileMetadatas(6);
+  ASSERT_EQ(2, l5_files.size());
+  ASSERT_EQ(1, l6_files.size());
+
+  ASSERT_OK(db_->CompactFiles(CompactionOptions(),
+                              {MakeTableFileName(l5_files[1]->fd.GetNumber()),
+                               MakeTableFileName(l6_files[0]->fd.GetNumber())},
+                              6));
+
+  uint64_t l6_range_deletions = 0;
+  for (auto* f : GetLevelFileMetadatas(5)) {
+    ASSERT_EQ(0, f->num_range_deletions);
+  }
+  for (auto* f : GetLevelFileMetadatas(6)) {
+    l6_range_deletions += f->num_range_deletions;
+  }
+  ASSERT_GT(l6_range_deletions, 0);
+}
+
 // Tests DBIter::GetProperty("rocksdb.iterator.write-time") return a data's
 // approximate write unix time.
 class IteratorWriteTimeTest

db/db_basic_test.cc

@@ -761,6 +761,111 @@ TEST_F(DBBasicTest, ReuseManifestOnOpenFullStackComposition) {
   EXPECT_EQ("v", Get("k"));
 }
 
+// Regression test for WAL recovery while publishing a fresh MANIFEST. The test
+// stores SSTs in a separate DB path, injects failure after CURRENT points at
+// the new MANIFEST, and simulates crash cleanup; the recovered SST must survive
+// because the synced MANIFEST references it.
+TEST_F(DBBasicTest, RecoverySstDirSyncedBeforeFreshManifestPublish) {
+  auto fault_fs = std::make_shared<FaultInjectionTestFS>(env_->GetFileSystem());
+  std::unique_ptr<Env> fault_env(NewCompositeEnv(fault_fs));
+
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.disable_auto_compactions = true;
+  options.env = fault_env.get();
+  options.reuse_manifest_on_open = false;
+  options.db_paths.emplace_back(dbname_ + "_2", 1ULL << 30);
+
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("base", "value"));
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(Put("recovered", "value"));
+  ASSERT_OK(db_->FlushWAL(true));
+  Close();
+
+  fault_fs->ResetState();
+
+  std::atomic<bool> fail_after_current_publish{true};
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "VersionSet::ProcessManifestWrites:AfterSetCurrentFile", [&](void* arg) {
+        if (fail_after_current_publish.exchange(false)) {
+          ASSERT_NE(nullptr, arg);
+          IOStatus* io_s = static_cast<IOStatus*>(arg);
+          ASSERT_OK(*io_s);
+          *io_s = IOStatus::IOError("injected current publish aftermath");
+        }
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  Status s = TryReopen(options);
+  ASSERT_TRUE(s.IsIOError()) << s.ToString();
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  ASSERT_OK(fault_fs->DeleteFilesCreatedAfterLastDirSync(IOOptions(), nullptr));
+
+  s = TryReopen(options);
+  ASSERT_OK(s);
+  ASSERT_EQ("value", Get("base"));
+  ASSERT_EQ("value", Get("recovered"));
+  Close();
+}
+
+// Regression test for WAL recovery while appending to a reused MANIFEST. The
+// first reopen forces recovery to create an SST and then fail after MANIFEST
+// sync. The simulated crash cleanup deletes files without a prior directory
+// sync; the recovered SST must survive because the synced MANIFEST references
+// it.
+TEST_F(DBBasicTest,
+       ReuseManifestOnOpenSyncsRecoverySstDirBeforeManifestAppend) {
+  auto fault_fs = std::make_shared<FaultInjectionTestFS>(env_->GetFileSystem());
+  std::unique_ptr<Env> fault_env(NewCompositeEnv(fault_fs));
+
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.disable_auto_compactions = true;
+  options.env = fault_env.get();
+  options.reuse_manifest_on_open = true;
+
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("base", "value"));
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(Put("recovered", "value"));
+  ASSERT_OK(db_->FlushWAL(true));
+  Close();
+
+  fault_fs->ResetState();
+
+  std::atomic<bool> fail_after_manifest_sync{true};
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "VersionSet::ProcessManifestWrites:AfterSyncManifest", [&](void* arg) {
+        if (fail_after_manifest_sync.exchange(false)) {
+          ASSERT_NE(nullptr, arg);
+          IOStatus* io_s = static_cast<IOStatus*>(arg);
+          ASSERT_OK(*io_s);
+          *io_s = IOStatus::IOError("injected manifest sync aftermath");
+        }
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  Status s = TryReopen(options);
+  ASSERT_TRUE(s.IsIOError()) << s.ToString();
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  ASSERT_OK(fault_fs->DeleteFilesCreatedAfterLastDirSync(IOOptions(), nullptr));
+
+  s = TryReopen(options);
+  ASSERT_OK(s);
+  ASSERT_EQ("value", Get("base"));
+  ASSERT_EQ("value", Get("recovered"));
+  Close();
+}
+
 // Direct unit test for WritableFileWriter's initial_file_size parameter:
 // verifies the visible size accessors report the existing on-disk size
 // immediately, rather than the constructor's zero default.
@@ -6226,7 +6331,7 @@ TEST_F(DBBasicTest, DisallowMemtableWrite) {
   Options options_disallow = options_allow;
   options_disallow.disallow_memtable_writes = true;
   options_disallow.paranoid_memory_checks = true;
-  options_disallow.memtable_veirfy_per_key_checksum_on_seek = true;
+  options_disallow.memtable_verify_per_key_checksum_on_seek = true;
 
   DestroyAndReopen(options_allow);
   // CFs allowing and disallowing memtable write

db/db_block_cache_test.cc

@@ -1482,10 +1482,13 @@ TEST_P(DBBlockCacheTypeTest, AddRedundantStats) {
   // Access just data, forcing redundant load+insert
   ReadOptions read_options;
   std::unique_ptr<Iterator> iter{db_->NewIterator(read_options)};
+  ASSERT_OK(iter->Refresh());
+
   cache->SetNthLookupNotFound(1);
   iter->SeekToFirst();
   ASSERT_TRUE(iter->Valid());
   ASSERT_EQ(iter->key(), "bar");
+  ASSERT_EQ(iter->value(), "value");
 
   EXPECT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD));
   EXPECT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD));

db/db_bloom_filter_test.cc

@@ -2022,7 +2022,11 @@ TEST_F(DBBloomFilterTest, MutableFilterPolicy) {
   double expected_bpk = 10.0;
   // Other configs to try
   std::vector<std::pair<std::string, double>> configs = {
-      {"ribbonfilter:10:-1", 7.0}, {"bloomfilter:5", 5.0}, {"nullptr", 0.0}};
+      {"ribbonfilter:10:-1", 7.0},
+      {"bloomfilter:5", 5.0},
+      {"nullptr", 0.0},
+      // As serialized in OPTIONS file
+      {"{id=ribbonfilter:12:-1;bloom_before_level=-1;}", 8.4}};
 
   table_options.cache_index_and_filter_blocks = true;
   options.table_factory.reset(NewBlockBasedTableFactory(table_options));

db/db_compaction_abort_test.cc

@@ -427,6 +427,38 @@ TEST_F(DBCompactionAbortTest, AbortAutomaticCompaction) {
   }
 }
 
+TEST_F(DBCompactionAbortTest, AbortScheduledAutomaticCompactionBeforePick) {
+  // The goal is to cover an automatic compaction that has been scheduled, but
+  // aborts before the worker picks and removes a column family from the
+  // compaction queue. This pins the scheduler bookkeeping invariant that
+  // ResumeAllCompactions() must be able to schedule that still-queued work.
+  constexpr int kCompactionTrigger = 4;
+  constexpr int kNumL0Files = kCompactionTrigger + 1;
+  Options options = GetOptionsWithStats();
+  options.level0_file_num_compaction_trigger = kCompactionTrigger;
+  options.max_background_compactions = 1;
+  options.max_subcompactions = 1;
+  options.disable_auto_compactions = false;
+  Reopen(options);
+
+  SyncPointAbortHelper helper("BackgroundCallCompaction:0");
+  helper.Setup(dbfull());
+
+  PopulateData(/*num_files=*/kNumL0Files, /*keys_per_file=*/100,
+               /*value_size=*/1000);
+  helper.CleanupAndWait();
+
+  const uint64_t compact_write_bytes_before_resume =
+      stats_->getTickerCount(COMPACT_WRITE_BYTES);
+  dbfull()->ResumeAllCompactions();
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  ASSERT_GT(stats_->getTickerCount(COMPACT_WRITE_BYTES),
+            compact_write_bytes_before_resume);
+  ASSERT_LT(NumTableFilesAtLevel(0), kNumL0Files);
+  VerifyDataIntegrity(/*num_keys=*/100);
+}
+
 TEST_F(DBCompactionAbortTest, AbortAndVerifyNoOutputFiles) {
   Options options = CurrentOptions();
   options.level0_file_num_compaction_trigger = 4;

db/db_compaction_test.cc

@@ -8,11 +8,13 @@
 // found in the LICENSE file. See the AUTHORS file for names of contributors.
 
 #include <tuple>
+#include <utility>
 
 #include "compaction/compaction_picker_universal.h"
 #include "db/blob/blob_index.h"
 #include "db/db_test_util.h"
 #include "db/dbformat.h"
+#include "db/table_cache.h"
 #include "env/mock_env.h"
 #include "file/filename.h"
 #include "port/port.h"
@@ -6651,6 +6653,572 @@ TEST_P(DBCompactionDirectIOTest, DirectIO) {
 INSTANTIATE_TEST_CASE_P(DBCompactionDirectIOTest, DBCompactionDirectIOTest,
                         testing::Bool());
 
+// With use_direct_io_for_compaction_reads OFF, compaction reads must stay
+// buffered: neither the compaction-input FileOptions nor a kernel O_DIRECT
+// open should fire. Runs on every platform (the sync points just don't fire
+// where O_DIRECT isn't reachable). Pairs with
+// UseDirectIoForCompactionReadsEndToEnd for the on case.
+TEST_F(DBCompactionTest, UseDirectIoForCompactionReadsOffStaysBuffered) {
+  Options options = CurrentOptions();
+  Destroy(options);
+  options.create_if_missing = true;
+  options.disable_auto_compactions = true;
+  options.use_direct_reads = false;
+  options.use_direct_io_for_compaction_reads = false;
+  options.use_direct_io_for_flush_and_compaction = false;
+
+  std::atomic<bool> observed_direct_compaction_read{false};
+  std::atomic<int> observed_callbacks{0};
+  std::atomic<int> observed_odirect_opens{0};
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "CompactionJob::CreateInputIterator:InputFileOptions", [&](void* arg) {
+        const auto* fo = static_cast<const FileOptions*>(arg);
+        observed_callbacks.fetch_add(1, std::memory_order_relaxed);
+        if (fo->use_direct_reads) {
+          observed_direct_compaction_read.store(true,
+                                                std::memory_order_relaxed);
+        }
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "NewRandomAccessFile:O_DIRECT", [&](void* /*arg*/) {
+        observed_odirect_opens.fetch_add(1, std::memory_order_relaxed);
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(TryReopen(options));
+
+  const std::string value(4096, 'v');
+  for (int i = 0; i < 64; ++i) {
+    ASSERT_OK(Put(Key(i), value));
+  }
+  ASSERT_OK(Flush());
+  for (int i = 0; i < 64; ++i) {
+    ASSERT_OK(Put(Key(i), value));
+  }
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(dbfull()->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  ASSERT_GT(observed_callbacks.load(), 0);
+  ASSERT_FALSE(observed_direct_compaction_read.load());
+  ASSERT_EQ(0, observed_odirect_opens.load());
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+  Destroy(options);
+}
+
+TEST_F(DBCompactionTest,
+       UseDirectIoForCompactionReadsUsesBoundedEphemeralReaders) {
+  auto fs = std::make_shared<MockFileSystem>(Env::Default()->GetSystemClock(),
+                                             /*supports_direct_io=*/true);
+  std::unique_ptr<Env> mock_env = NewCompositeEnv(fs);
+
+  Options options = CurrentOptions();
+  options.env = mock_env.get();
+  Destroy(options);
+  options.create_if_missing = true;
+  options.disable_auto_compactions = true;
+  options.use_direct_reads = false;
+  options.use_direct_io_for_compaction_reads = true;
+  options.use_direct_io_for_flush_and_compaction = false;
+  options.max_subcompactions = 3;
+  options.statistics = CreateDBStatistics();
+
+  BlockBasedTableOptions table_options;
+  table_options.block_cache = NewLRUCache(64 << 20);
+  table_options.cache_index_and_filter_blocks = true;
+  table_options.pin_l0_filter_and_index_blocks_in_cache = true;
+  table_options.filter_policy.reset(NewBloomFilterPolicy(10));
+  options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+
+  std::atomic<int> fresh_reader_opens{0};
+  std::atomic<int> malformed_fresh_reader_options{0};
+  std::atomic<int> avoid_shared_metadata_cache_opens{0};
+  std::atomic<int> shared_metadata_cache_uses{0};
+  std::atomic<size_t> input_iterator_file_bound{0};
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "TableCache::FindTable:FreshTableReader", [&](void* arg) {
+        const auto* open_options = static_cast<TableCacheOpenOptions*>(arg);
+        fresh_reader_opens.fetch_add(1, std::memory_order_relaxed);
+        if (!open_options->open_ephemeral_table_reader ||
+            !open_options->avoid_shared_metadata_cache ||
+            !open_options->skip_filters) {
+          malformed_fresh_reader_options.fetch_add(1,
+                                                   std::memory_order_relaxed);
+        }
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "BlockBasedTable::PrefetchIndexAndFilterBlocks:SharedMetadataCache",
+      [&](void* arg) {
+        const auto* context = static_cast<std::pair<bool, bool>*>(arg);
+        const bool avoid_shared_metadata_cache = context->first;
+        const bool use_cache = context->second;
+        if (avoid_shared_metadata_cache) {
+          avoid_shared_metadata_cache_opens.fetch_add(
+              1, std::memory_order_relaxed);
+          if (use_cache) {
+            shared_metadata_cache_uses.fetch_add(1, std::memory_order_relaxed);
+          }
+        }
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "VersionSet::MakeInputIterator:NewCompactionMergingIterator",
+      [&](void* arg) {
+        input_iterator_file_bound.fetch_add(*static_cast<size_t*>(arg),
+                                            std::memory_order_relaxed);
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(TryReopen(options));
+
+  const std::string value(4096, 'v');
+  for (int i = 0; i < 64; ++i) {
+    ASSERT_OK(Put(Key(i), value));
+  }
+  ASSERT_OK(Flush());
+  for (int i = 0; i < 64; ++i) {
+    ASSERT_OK(Put(Key(i), value));
+  }
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(dbfull()->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  ASSERT_GT(fresh_reader_opens.load(), 0);
+  EXPECT_EQ(0, malformed_fresh_reader_options.load());
+  ASSERT_GT(avoid_shared_metadata_cache_opens.load(), 0);
+  EXPECT_EQ(0, shared_metadata_cache_uses.load());
+  ASSERT_GT(input_iterator_file_bound.load(), 0);
+  EXPECT_LE(static_cast<size_t>(fresh_reader_opens.load()),
+            input_iterator_file_bound.load());
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+  Destroy(options);
+}
+
+// End-to-end check that use_direct_io_for_compaction_reads opens compaction
+// inputs with O_DIRECT while use_direct_reads stays off (user reads buffered).
+// The NewRandomAccessFile:O_DIRECT sync point in env/fs_posix.cc fires once
+// per fresh open with the O_DIRECT flag, so this proves the kernel path, not
+// just the FileOptions. Only runs on platforms that take the O_DIRECT path.
+#if !defined(OS_MACOSX) && !defined(OS_OPENBSD) && !defined(OS_SOLARIS) && \
+    !defined(OS_WIN)
+TEST_F(DBCompactionTest, UseDirectIoForCompactionReadsEndToEnd) {
+  if (!IsDirectIOSupported()) {
+    ROCKSDB_GTEST_BYPASS("Direct IO not supported");
+    return;
+  }
+
+  Options options = CurrentOptions();
+  Destroy(options);
+  options.create_if_missing = true;
+  options.disable_auto_compactions = true;
+  // User reads stay buffered, compaction reads should switch to O_DIRECT.
+  options.use_direct_reads = false;
+  options.use_direct_io_for_compaction_reads = true;
+  // Isolate the read-side change; leave the compaction write path buffered.
+  options.use_direct_io_for_flush_and_compaction = false;
+
+  // Sync-point callbacks fire on compaction threads while the test thread
+  // reads these counters, so use atomics to avoid a data race.
+  std::atomic<int> observed_run_starts{0};
+  std::atomic<int> observed_odirect_opens{0};
+  std::atomic<bool> observed_direct_compaction_read{false};
+  std::atomic<int> observed_callbacks{0};
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({});
+  // Plumbing-level probe: the compaction-input FileOptions should carry
+  // use_direct_reads = true when the new flag is enabled.
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "CompactionJob::CreateInputIterator:InputFileOptions", [&](void* arg) {
+        const auto* fo = static_cast<const FileOptions*>(arg);
+        observed_callbacks.fetch_add(1, std::memory_order_relaxed);
+        if (fo != nullptr && fo->use_direct_reads) {
+          observed_direct_compaction_read.store(true,
+                                                std::memory_order_relaxed);
+        }
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "CompactionJob::Run():Start", [&](void* /*arg*/) {
+        observed_run_starts.fetch_add(1, std::memory_order_relaxed);
+      });
+  // Kernel-level probe: this fires only when open() is issued with O_DIRECT,
+  // proving we change the actual cache mode, not just the FileOptions struct.
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "NewRandomAccessFile:O_DIRECT", [&](void* /*arg*/) {
+        observed_odirect_opens.fetch_add(1, std::memory_order_relaxed);
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  Status s = TryReopen(options);
+  if (s.IsNotSupported() || s.IsInvalidArgument()) {
+    ROCKSDB_GTEST_BYPASS(
+        "Direct IO reads not supported in this test environment");
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+    return;
+  }
+  ASSERT_OK(s);
+
+  // Produce two L0 files with OVERLAPPING key ranges so that CompactRange has
+  // actual merge work to do (otherwise RocksDB performs a trivial file move
+  // and never constructs a CompactionJob).
+  const std::string value(4096, 'v');
+  for (int i = 0; i < 64; ++i) {
+    ASSERT_OK(Put(Key(i), value));
+  }
+  ASSERT_OK(Flush());
+  for (int i = 0; i < 64; ++i) {
+    ASSERT_OK(Put(Key(i), value));
+  }
+  ASSERT_OK(Flush());
+
+  // User reads should still go through the buffered path. Confirm that the
+  // option does not silently flip use_direct_reads for user reads.
+  for (int i = 0; i < 8; ++i) {
+    std::string actual;
+    ASSERT_OK(db_->Get(ReadOptions(), Key(i), &actual));
+    ASSERT_EQ(value, actual);
+  }
+
+  ASSERT_OK(dbfull()->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  // Wait for compaction to complete and CompactionJob to be constructed.
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  // Confirm the compaction actually ran; otherwise the missing sync-point hits
+  // would be a test-setup problem, not an option regression.
+  ASSERT_GT(observed_run_starts.load(), 0)
+      << "CompactionJob::Run():Start never fired; CompactRange did not "
+         "schedule a compaction.";
+  ASSERT_GT(observed_callbacks.load(), 0);
+  ASSERT_TRUE(observed_direct_compaction_read.load());
+  // At least one compaction-input open went through O_DIRECT. Without the
+  // TableCache bypass this would be zero, since compaction would reuse the
+  // buffered handles cached for user reads.
+  EXPECT_GT(observed_odirect_opens.load(), 0)
+      << "no compaction-input opens went through O_DIRECT; "
+         "observed_odirect_opens="
+      << observed_odirect_opens.load();
+
+  // Quick sanity sweep after compaction to confirm data is intact.
+  for (int i = 0; i < 64; ++i) {
+    std::string actual;
+    ASSERT_OK(db_->Get(ReadOptions(), Key(i), &actual));
+    ASSERT_EQ(value, actual);
+  }
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+  Destroy(options);
+}
+
+// Exercise the LevelIterator bypass path (L1+ compactions) with range
+// tombstones, where the ephemeral TableReader's lifetime is coupled to the
+// range_tombstone_iter the file iterator returns. The end-to-end test above
+// only makes two L0 files, which take the direct NewIterator path and never
+// hit LevelIterator. Here we build L1/L2 with tombstones and compact L1->L2 so
+// LevelIterator::NewFileIterator drives the bypass; a wrong reader/iterator
+// lifetime would crash or trip sanitizers as LevelIterator switches files.
+// Correctness is checked by computing the expected state of every key and
+// asserting Get() matches: wave-2 puts beat wave-1, and each key's most-recent
+// covering tombstone shadows older puts.
+TEST_F(DBCompactionTest,
+       UseDirectIoForCompactionReadsLevelIteratorWithTombstones) {
+  if (!IsDirectIOSupported()) {
+    ROCKSDB_GTEST_BYPASS("Direct IO not supported");
+    return;
+  }
+
+  Options options = CurrentOptions();
+  Destroy(options);
+  options.create_if_missing = true;
+  options.disable_auto_compactions = true;
+  options.use_direct_reads = false;
+  options.use_direct_io_for_compaction_reads = true;
+  options.use_direct_io_for_flush_and_compaction = false;
+  // Small files / small level base so we can pack data into L1 and L2 with
+  // a few flushes and CompactRange calls instead of needing millions of keys.
+  options.write_buffer_size = 64 * 1024;
+  options.target_file_size_base = 64 * 1024;
+  options.max_bytes_for_level_base = 256 * 1024;
+  options.level0_file_num_compaction_trigger = 100;  // never auto-trigger
+
+  std::atomic<int> observed_odirect_opens{0};
+  std::atomic<int> observed_run_starts{0};
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "NewRandomAccessFile:O_DIRECT", [&](void* /*arg*/) {
+        observed_odirect_opens.fetch_add(1, std::memory_order_relaxed);
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "CompactionJob::Run():Start", [&](void* /*arg*/) {
+        observed_run_starts.fetch_add(1, std::memory_order_relaxed);
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  Status s = TryReopen(options);
+  if (s.IsNotSupported() || s.IsInvalidArgument()) {
+    ROCKSDB_GTEST_BYPASS(
+        "Direct IO reads not supported in this test environment");
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+    return;
+  }
+  ASSERT_OK(s);
+
+  // Use distinguishable values per wave so we can verify which wave's put
+  // won the merge for each key, not just "some put won".
+  const std::string wave1_value(1024, '1');
+  const std::string wave2_value(1024, '2');
+
+  auto write_batch = [&](int begin, int end, const std::string& value,
+                         bool with_range_tombstone) {
+    for (int i = begin; i < end; ++i) {
+      ASSERT_OK(Put(Key(i), value));
+    }
+    if (with_range_tombstone) {
+      // Drop a slice in the middle of the just-written range. The
+      // DeleteRange follows the puts in this batch, so its sequence number
+      // is higher and it shadows the puts on [del_lo, del_hi) within this
+      // SST.
+      ASSERT_OK(db_->DeleteRange(WriteOptions(), db_->DefaultColumnFamily(),
+                                 Key(begin + (end - begin) / 4),
+                                 Key(begin + 3 * (end - begin) / 4)));
+    }
+    ASSERT_OK(Flush());
+  };
+
+  // Wave 1: four flushed SSTs covering [0, 800), each with a tombstone
+  // covering the middle half of its own range. Then compact down so the
+  // next phase exercises LevelIterator over L1+ files.
+  constexpr int kWave1Begin = 0;
+  constexpr int kWave1End = 800;
+  constexpr int kWave1BatchSize = 200;
+  for (int batch_begin = kWave1Begin; batch_begin < kWave1End;
+       batch_begin += kWave1BatchSize) {
+    write_batch(batch_begin, batch_begin + kWave1BatchSize, wave1_value,
+                /*with_range_tombstone=*/true);
+  }
+  ASSERT_OK(dbfull()->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  // Wave 2: two more flushed SSTs at L0 that overlap wave 1, each with their
+  // own range tombstone. The puts in wave 2 have higher seqno than wave 1,
+  // so they win when not tombstoned.
+  struct Wave2Range {
+    int begin;
+    int end;
+  };
+  const Wave2Range wave2_ranges[] = {{50, 250}, {350, 550}};
+  for (const auto& r : wave2_ranges) {
+    write_batch(r.begin, r.end, wave2_value, /*with_range_tombstone=*/true);
+  }
+
+  const int run_starts_before = observed_run_starts.load();
+  const int odirect_before = observed_odirect_opens.load();
+
+  // Compact everything together, forcing a LevelIterator over the lower-level
+  // files on the bypass path. Wrong ephemeral reader / tombstone-iter
+  // lifetimes should trip sanitizers here.
+  ASSERT_OK(dbfull()->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  ASSERT_GT(observed_run_starts.load(), run_starts_before)
+      << "expected at least one compaction to run during the L1+ phase";
+  EXPECT_GT(observed_odirect_opens.load(), odirect_before)
+      << "no compaction-input opens went through O_DIRECT during L1+ "
+         "compaction; LevelIterator bypass path may be broken";
+
+  // Compute the precise expected state per key from the test parameters.
+  // For each k in the wave-1 range:
+  //   - If k is inside a wave-2 batch: wave-2 wins. NotFound iff k is in
+  //     that batch's tombstone range; otherwise present with wave2_value.
+  //   - Else: wave 1 wins. NotFound iff k is in its batch's tombstone
+  //     range; otherwise present with wave1_value.
+  enum class Expectation { kAbsent, kWave1Value, kWave2Value };
+  auto classify = [&](int k) -> Expectation {
+    // Wave 2 first (it shadows wave 1 wherever it covers).
+    for (const auto& r : wave2_ranges) {
+      if (k >= r.begin && k < r.end) {
+        const int width = r.end - r.begin;
+        const int del_lo = r.begin + width / 4;
+        const int del_hi = r.begin + 3 * width / 4;
+        if (k >= del_lo && k < del_hi) {
+          return Expectation::kAbsent;
+        }
+        return Expectation::kWave2Value;
+      }
+    }
+    // Fall back to wave 1.
+    const int batch = (k - kWave1Begin) / kWave1BatchSize;
+    const int batch_begin = kWave1Begin + batch * kWave1BatchSize;
+    const int del_lo = batch_begin + kWave1BatchSize / 4;
+    const int del_hi = batch_begin + 3 * kWave1BatchSize / 4;
+    if (k >= del_lo && k < del_hi) {
+      return Expectation::kAbsent;
+    }
+    return Expectation::kWave1Value;
+  };
+
+  int present_w1 = 0;
+  int present_w2 = 0;
+  int absent = 0;
+  std::string actual;
+  for (int k = kWave1Begin; k < kWave1End; ++k) {
+    const Status get_s = db_->Get(ReadOptions(), Key(k), &actual);
+    const Expectation exp = classify(k);
+    switch (exp) {
+      case Expectation::kAbsent:
+        EXPECT_TRUE(get_s.IsNotFound())
+            << "key " << k << " expected NotFound (covered by tombstone); "
+            << "got status=" << get_s.ToString()
+            << " value_len=" << (get_s.ok() ? actual.size() : 0);
+        ++absent;
+        break;
+      case Expectation::kWave1Value:
+        ASSERT_OK(get_s) << "key " << k << " expected wave-1 value";
+        EXPECT_EQ(wave1_value, actual)
+            << "key " << k << " expected wave-1 value but got a different one";
+        ++present_w1;
+        break;
+      case Expectation::kWave2Value:
+        ASSERT_OK(get_s) << "key " << k << " expected wave-2 value";
+        EXPECT_EQ(wave2_value, actual)
+            << "key " << k << " expected wave-2 value but got a different one";
+        ++present_w2;
+        break;
+    }
+  }
+  // All three buckets should be non-empty, so the test really exercises both
+  // tombstone paths and the wave-2-wins path. A zero here means the setup
+  // drifted and the setup (not these checks) should be fixed.
+  EXPECT_GT(present_w1, 0);
+  EXPECT_GT(present_w2, 0);
+  EXPECT_GT(absent, 0);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+  Destroy(options);
+}
+
+// With the option enabled, run reader threads against the live DB while manual
+// compactions run in the background. Each compaction-input file is open through
+// an ephemeral O_DIRECT handle while the shared TableCache still serves user
+// reads through a buffered handle, so the same SST is open in two cache modes
+// at once. This stresses that coexistence under TSAN/ASAN/UBSAN; it asserts
+// reads stay consistent and final values match the last writes, not anything
+// about timing.
+TEST_F(DBCompactionTest, UseDirectIoForCompactionReadsConcurrentReadStress) {
+  if (!IsDirectIOSupported()) {
+    ROCKSDB_GTEST_BYPASS("Direct IO not supported");
+    return;
+  }
+
+  Options options = CurrentOptions();
+  Destroy(options);
+  options.create_if_missing = true;
+  options.disable_auto_compactions = true;
+  options.use_direct_reads = false;
+  options.use_direct_io_for_compaction_reads = true;
+  options.use_direct_io_for_flush_and_compaction = false;
+  options.write_buffer_size = 64 * 1024;
+  options.target_file_size_base = 64 * 1024;
+  options.max_bytes_for_level_base = 256 * 1024;
+  options.level0_file_num_compaction_trigger = 100;  // never auto-trigger
+
+  Status s = TryReopen(options);
+  if (s.IsNotSupported() || s.IsInvalidArgument()) {
+    ROCKSDB_GTEST_BYPASS(
+        "Direct IO reads not supported in this test environment");
+    return;
+  }
+  ASSERT_OK(s);
+
+  constexpr int kNumKeys = 512;
+  constexpr int kValueSize = 256;
+  const std::string base_value(kValueSize, 'b');
+
+  // Initial population, flushed into a few L0 SSTs.
+  for (int i = 0; i < kNumKeys; ++i) {
+    ASSERT_OK(Put(Key(i), base_value));
+    if (i > 0 && i % 128 == 0) {
+      ASSERT_OK(Flush());
+    }
+  }
+  ASSERT_OK(Flush());
+
+  std::atomic<bool> stop{false};
+  std::atomic<int64_t> reads_observed{0};
+  std::atomic<int64_t> read_errors{0};
+
+  // Reader threads loop until stop is set, recording any status that isn't OK
+  // or NotFound. The point is to surface Corruption/IOError/use-after-free, so
+  // the main thread can fail the test.
+  constexpr int kNumReaders = 4;
+  std::vector<port::Thread> readers;
+  readers.reserve(kNumReaders);
+  for (int t = 0; t < kNumReaders; ++t) {
+    readers.emplace_back([&, t]() {
+      std::string value;
+      while (!stop.load(std::memory_order_acquire)) {
+        const int k =
+            (t * 7919 +
+             static_cast<int>(reads_observed.load(std::memory_order_relaxed))) %
+            kNumKeys;
+        Status read_s = db_->Get(ReadOptions(), Key(k), &value);
+        if (!read_s.ok() && !read_s.IsNotFound()) {
+          read_errors.fetch_add(1, std::memory_order_relaxed);
+        }
+        reads_observed.fetch_add(1, std::memory_order_relaxed);
+      }
+    });
+  }
+
+  // Drive a few rounds of writes and compactions on the main thread while
+  // the readers hammer Get(). Each round overwrites every key with a
+  // round-tagged value and then compacts.
+  constexpr int kRounds = 4;
+  std::string last_value = base_value;
+  for (int round = 0; round < kRounds; ++round) {
+    const std::string round_value(kValueSize,
+                                  static_cast<char>('A' + (round % 26)));
+    for (int i = 0; i < kNumKeys; ++i) {
+      ASSERT_OK(Put(Key(i), round_value));
+      if (i > 0 && i % 64 == 0) {
+        ASSERT_OK(Flush());
+      }
+    }
+    ASSERT_OK(Flush());
+    ASSERT_OK(dbfull()->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+    ASSERT_OK(dbfull()->TEST_WaitForCompact());
+    last_value = round_value;
+  }
+
+  stop.store(true, std::memory_order_release);
+  for (auto& reader : readers) {
+    reader.join();
+  }
+
+  // Every key must now hold the final round's value. Catches wholesale
+  // corruption the sampling readers might miss, and confirms compaction
+  // finished under the bypass path.
+  std::string value;
+  for (int i = 0; i < kNumKeys; ++i) {
+    ASSERT_OK(db_->Get(ReadOptions(), Key(i), &value));
+    EXPECT_EQ(last_value, value);
+  }
+  EXPECT_GT(reads_observed.load(), 0)
+      << "reader threads never observed a single read; test was a no-op";
+  EXPECT_EQ(0, read_errors.load())
+      << "concurrent reads against compaction with bypass-path saw "
+      << read_errors.load() << " non-OK/non-NotFound status returns";
+
+  Destroy(options);
+}
+#endif  // !defined(OS_MACOSX) && !defined(OS_OPENBSD) && ...
+
 class CompactionPriTest : public DBTestBase,
                           public testing::WithParamInterface<uint32_t> {
  public:

db/db_etc2_test.cc

@@ -35,7 +35,7 @@ namespace ROCKSDB_NAMESPACE {
 
 class DBTest2 : public DBTestBase {
  public:
-  DBTest2() : DBTestBase("db_test2", /*env_do_fsync=*/true) {}
+  DBTest2() : DBTestBase("db_etc2_test", /*env_do_fsync=*/true) {}
 };
 
 TEST_F(DBTest2, OpenForReadOnly) {

db/db_etc3_test.cc

@@ -1,9 +1,12 @@
-//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  Copyright (c) Meta Platforms, Inc. and affiliates.
 //  This source code is licensed under both the GPLv2 (found in the
 //  COPYING file in the root directory) and Apache 2.0 License
 //  (found in the LICENSE.Apache file in the root directory).
 
 #include "db/db_test_util.h"
+#include "rocksdb/convenience.h"
+#include "rocksdb/metadata.h"
+#include "rocksdb/sst_file_writer.h"
 
 namespace ROCKSDB_NAMESPACE {
 
@@ -43,50 +46,98 @@ TEST_F(DBEtc3Test, AutoTuneManifestSize) {
   ASSERT_EQ(DBOptions{}.max_manifest_space_amp_pct, 500);
 
   Options options = CurrentOptions();
-  ASSERT_OK(db_->SetOptions({{"level0_file_num_compaction_trigger", "20"}}));
+  // Test setup: create many flushed files. Keep level compaction semantics so
+  // DeleteFilesInRanges can remove non-L0 files, but prevent automatic
+  // compactions and write stalls from adding unrelated behavior.
+  options.disable_auto_compactions = true;
+  options.level0_slowdown_writes_trigger = 100;
+  options.level0_stop_writes_trigger = 200;
 
-  // Use large column family names to essentially control the amount of payload
-  // data needed for the manifest file. Drop manifest entries don't include the
-  // CF name so are small.
+  // Test strategy: use large column family names to control the rough amount
+  // of payload added to the MANIFEST. Drop manifest entries do not include the
+  // CF name, so they are small.
+  //
+  // Most CF helper calls piggy-back a background manifest write so the main
+  // auto-tuning phases continue to test the unrelaxed background threshold
+  // even though CF manipulation itself is foreground. Phase-specific foreground
+  // checks disable that piggy-backed background write.
   uint64_t prev_manifest_num = 0, cur_manifest_num = 0;
   std::deque<ColumnFamilyHandle*> handles;
   int counter = 5;
-  auto AddCfFn = [&]() {
+  auto UpdateManifestNumsFrom = [&](uint64_t before_manifest_num) {
+    prev_manifest_num = before_manifest_num;
+    cur_manifest_num = dbfull()->TEST_Current_Manifest_FileNo();
+  };
+  auto BackgroundManifestWriteFn = [&]() {
+    uint64_t before_manifest_num = cur_manifest_num;
+    ASSERT_OK(Put("x", std::to_string(counter++)));
+    ASSERT_OK(Flush());
+    UpdateManifestNumsFrom(before_manifest_num);
+  };
+  auto AddCfFn = [&](bool include_background_manifest_write = true) {
+    uint64_t before_manifest_num = cur_manifest_num;
     std::string name = "cf" + std::to_string(counter++);
     name.resize(1000, 'a');
     ASSERT_OK(db_->CreateColumnFamily(options, name, &handles.emplace_back()));
-    prev_manifest_num = cur_manifest_num;
     cur_manifest_num = dbfull()->TEST_Current_Manifest_FileNo();
+    if (include_background_manifest_write) {
+      BackgroundManifestWriteFn();
+    }
+    UpdateManifestNumsFrom(before_manifest_num);
   };
-  auto DropCfFn = [&]() {
+  auto DropCfFn = [&](bool include_background_manifest_write = true) {
+    uint64_t before_manifest_num = cur_manifest_num;
     ASSERT_OK(db_->DropColumnFamily(handles.front()));
     ASSERT_OK(db_->DestroyColumnFamilyHandle(handles.front()));
     handles.pop_front();
-    prev_manifest_num = cur_manifest_num;
-    cur_manifest_num = dbfull()->TEST_Current_Manifest_FileNo();
-  };
-  auto TrivialManifestWriteFn = [&]() {
-    ASSERT_OK(Put("x", std::to_string(counter++)));
-    ASSERT_OK(Flush());
-    prev_manifest_num = cur_manifest_num;
     cur_manifest_num = dbfull()->TEST_Current_Manifest_FileNo();
+    if (include_background_manifest_write) {
+      BackgroundManifestWriteFn();
+    }
+    UpdateManifestNumsFrom(before_manifest_num);
   };
 
+  // ---- Phase 1: foreground threshold relaxation is bounded ----
+  //
+  // Foreground operations should only get about 25% extra headroom, not an
+  // unbounded threshold. With 1000-byte CF names and a 3000-byte normal limit,
+  // the relaxed limit should allow the first four foreground-only CF additions
+  // but require rotation on the fifth.
+  options.max_manifest_file_size = 3000;
+  options.max_manifest_space_amp_pct = 0;
+  DestroyAndReopen(options);
+  cur_manifest_num = dbfull()->TEST_Current_Manifest_FileNo();
+  prev_manifest_num = cur_manifest_num;
+  for (int i = 1; i <= 4; ++i) {
+    AddCfFn(/*include_background_manifest_write=*/false);
+    ASSERT_EQ(prev_manifest_num, cur_manifest_num);
+  }
+  AddCfFn(/*include_background_manifest_write=*/false);
+  ASSERT_LT(prev_manifest_num, cur_manifest_num);
+  while (!handles.empty()) {
+    ASSERT_OK(db_->DestroyColumnFamilyHandle(handles.front()));
+    handles.pop_front();
+  }
+
+  // ---- Phase 2: no auto-tuning means frequent rotation ----
+  //
   options.max_manifest_file_size = 1000000;
   options.max_manifest_space_amp_pct = 0;  // no auto-tuning yet
   DestroyAndReopen(options);
+  cur_manifest_num = dbfull()->TEST_Current_Manifest_FileNo();
+  prev_manifest_num = cur_manifest_num;
 
-  // With the generous (minimum) maximum manifest size, should not be rotated
+  // With the generous minimum manifest size, should not be rotated.
   AddCfFn();
   AddCfFn();
   AddCfFn();
   ASSERT_EQ(prev_manifest_num, cur_manifest_num);
 
-  // Change options for small max and (still) no auto-tuning
+  // Lower the minimum while still disabling auto-tuning.
   ASSERT_OK(db_->SetDBOptions({{"max_manifest_file_size", "3000"}}));
 
   // Takes effect on the next manifest write
-  TrivialManifestWriteFn();
+  BackgroundManifestWriteFn();
   ASSERT_LT(prev_manifest_num, cur_manifest_num);
 
   // Now we have to rewrite the whole manifest on each write because the
@@ -97,28 +148,31 @@ TEST_F(DBEtc3Test, AutoTuneManifestSize) {
   ASSERT_LT(prev_manifest_num, cur_manifest_num);
   AddCfFn();
   ASSERT_LT(prev_manifest_num, cur_manifest_num);
-  TrivialManifestWriteFn();
+  BackgroundManifestWriteFn();
   ASSERT_LT(prev_manifest_num, cur_manifest_num);
 
+  // ---- Phase 3: auto-tuning raises the background threshold ----
+  //
   // Enabling auto-tuning should fix this, immediately for next manifest writes.
-  // This will allow up to double-ish the size of the compacted manifest,
-  // which last should have been 4000 + some bytes.
+  // This will allow up to roughly double the size of the compacted manifest,
+  // which now includes CF entries plus the piggy-backed background writes.
   ASSERT_EQ(handles.size(), 4U);
-  ASSERT_OK(db_->SetDBOptions({{"max_manifest_space_amp_pct", "105"}}));
+  ASSERT_OK(db_->SetDBOptions({{"max_manifest_space_amp_pct", "95"}}));
 
-  // After 9 CF names should be enough to rotate the manifest
-  for (int i = 1; i <= 5; ++i) {
+  // Auto-tuning lets several more CF names accumulate in the MANIFEST before
+  // the piggy-backed background write crosses the threshold.
+  for (int i = 1; i <= 3; ++i) {
     if ((i % 2) == 1) {
       DropCfFn();
     }
     AddCfFn();
     ASSERT_EQ(prev_manifest_num, cur_manifest_num);
   }
-  TrivialManifestWriteFn();
+  AddCfFn();
   ASSERT_LT(prev_manifest_num, cur_manifest_num);
 
-  // We now have a different last compacted manifest size, should be
-  // able to go beyond 9 CFs named in manifest this time.
+  // We now have a different last compacted manifest size, so the next
+  // threshold should be based on the newly compacted MANIFEST.
   ASSERT_EQ(handles.size(), 6U);
 
   DropCfFn();
@@ -128,11 +182,10 @@ TEST_F(DBEtc3Test, AutoTuneManifestSize) {
     AddCfFn();
     ASSERT_EQ(prev_manifest_num, cur_manifest_num);
   }
-  // We've written 10 named CFs to the manifest. We should be able to
-  // dynamically change the auto-tuning still based on the last "compacted"
-  // manifest size of 7000 + some bytes.
+  // We should be able to dynamically change the auto-tuning still based on
+  // the last "compacted" manifest size.
   ASSERT_OK(db_->SetDBOptions({{"max_manifest_space_amp_pct", "51"}}));
-  TrivialManifestWriteFn();
+  BackgroundManifestWriteFn();
   ASSERT_LT(prev_manifest_num, cur_manifest_num);
   // And the "compacted" manifest size has reset again, so should be changed
   // again sooner.
@@ -141,13 +194,179 @@ TEST_F(DBEtc3Test, AutoTuneManifestSize) {
     AddCfFn();
     ASSERT_EQ(prev_manifest_num, cur_manifest_num);
   }
-  // Enough for manifest change
-  AddCfFn();
+
+  // ---- Phase 4: foreground operations use relaxed threshold ----
+  //
+  // The current MANIFEST is now large enough for the next background manifest
+  // write to rotate it, but still small enough for foreground operations to use
+  // their 25% extra headroom. Assert that each foreground operation stays on
+  // the same MANIFEST, then verify the next background write rotates.
+  const std::string external_files_dir = dbname_ + "/external_files";
+  ASSERT_OK(env_->CreateDirIfMissing(external_files_dir));
+  auto WriteExternalSstFile = [&](const std::string& file_name,
+                                  const std::string& key,
+                                  const std::string& value) {
+    const std::string file_path = external_files_dir + "/" + file_name;
+    Status ignored = env_->DeleteFile(file_path);
+    ignored.PermitUncheckedError();
+
+    SstFileWriter sst_file_writer(EnvOptions(), options);
+    ASSERT_OK(sst_file_writer.Open(file_path));
+    ASSERT_OK(sst_file_writer.Put(key, value));
+    ASSERT_OK(sst_file_writer.Finish());
+  };
+  auto IngestExternalFileForegroundManifestWriteFn =
+      [&](std::string* ingested_key) {
+        uint64_t before_manifest_num = cur_manifest_num;
+        const std::string key = "z" + std::to_string(counter++);
+        const std::string value = "v" + std::to_string(counter++);
+        const std::string file_name =
+            "ingest_" + std::to_string(counter++) + ".sst";
+        const std::string file_path = external_files_dir + "/" + file_name;
+        WriteExternalSstFile(file_name, key, value);
+
+        ASSERT_OK(
+            db_->IngestExternalFile({file_path}, IngestExternalFileOptions()));
+        ASSERT_EQ(value, Get(key));
+        *ingested_key = key;
+        Status ignored = env_->DeleteFile(file_path);
+        ignored.PermitUncheckedError();
+        UpdateManifestNumsFrom(before_manifest_num);
+      };
+  auto CreateColumnFamilyWithImportForegroundManifestWriteFn = [&]() {
+    uint64_t before_manifest_num = cur_manifest_num;
+    const std::string cf_name = "import_cf" + std::to_string(counter++);
+    const std::string key = "import" + std::to_string(counter++);
+    const std::string value = "v" + std::to_string(counter++);
+    const std::string file_name =
+        "import_" + std::to_string(counter++) + ".sst";
+    const std::string file_path = external_files_dir + "/" + file_name;
+    WriteExternalSstFile(file_name, key, value);
+
+    LiveFileMetaData file_metadata;
+    file_metadata.name = file_name;
+    file_metadata.db_path = external_files_dir;
+    file_metadata.smallest_seqno = 0;
+    file_metadata.largest_seqno = 0;
+    file_metadata.level = 0;
+    ExportImportFilesMetaData metadata;
+    metadata.files.push_back(file_metadata);
+    metadata.db_comparator_name = options.comparator->Name();
+
+    ColumnFamilyHandle* import_handle = nullptr;
+    ASSERT_OK(db_->CreateColumnFamilyWithImport(options, cf_name,
+                                                ImportColumnFamilyOptions(),
+                                                metadata, &import_handle));
+    ASSERT_NE(import_handle, nullptr);
+    handles.push_back(import_handle);
+
+    std::string result;
+    ASSERT_OK(db_->Get(ReadOptions(), import_handle, key, &result));
+    ASSERT_EQ(value, result);
+    Status ignored = env_->DeleteFile(file_path);
+    ignored.PermitUncheckedError();
+    UpdateManifestNumsFrom(before_manifest_num);
+  };
+  auto DeleteFilesInRangesForegroundManifestWriteFn =
+      [&](const std::string& key) {
+        uint64_t before_manifest_num = cur_manifest_num;
+        const std::string limit = key + "\xff";
+        std::vector<RangeOpt> ranges;
+        ranges.emplace_back(key, limit);
+        ASSERT_OK(DeleteFilesInRanges(db_.get(), db_->DefaultColumnFamily(),
+                                      ranges.data(), ranges.size(),
+                                      /*include_end=*/false));
+
+        std::string result;
+        ASSERT_TRUE(db_->Get(ReadOptions(), key, &result).IsNotFound());
+        UpdateManifestNumsFrom(before_manifest_num);
+      };
+
+  // Column family manipulation.
+  AddCfFn(/*include_background_manifest_write=*/false);
+  ASSERT_EQ(prev_manifest_num, cur_manifest_num);
+
+  // SetOptions should not write to the MANIFEST. If that regresses, the write
+  // should be treated as a background write and rotate here.
+  {
+    ASSERT_OK(db_->SetOptions({{"level0_slowdown_writes_trigger", "101"}}));
+    UpdateManifestNumsFrom(cur_manifest_num);
+  }
+  ASSERT_EQ(prev_manifest_num, cur_manifest_num);
+
+  // External file ingestion.
+  std::string ingested_key;
+  IngestExternalFileForegroundManifestWriteFn(&ingested_key);
+  ASSERT_EQ(prev_manifest_num, cur_manifest_num);
+
+  // Imported column family creation.
+  CreateColumnFamilyWithImportForegroundManifestWriteFn();
+  ASSERT_EQ(prev_manifest_num, cur_manifest_num);
+
+  // Physical file deletion by range.
+  DeleteFilesInRangesForegroundManifestWriteFn(ingested_key);
+  ASSERT_EQ(prev_manifest_num, cur_manifest_num);
+
+  // Background flush.
+  BackgroundManifestWriteFn();
   ASSERT_LT(prev_manifest_num, cur_manifest_num);
 
+  // ---- Phase 5: persisted compacted manifest size survives close/reopen ----
+  // Close with CFs still live. Reopen with reuse_manifest_on_open so the
+  // manifest is NOT rewritten from scratch. The persisted compacted size
+  // should be loaded and used for auto-tuning.
+  // At this point we have 8 CF handles plus default.
+  ASSERT_EQ(handles.size(), 8U);
+
+  // Collect CF names for reopen, then release handles (Close needs this)
+  std::vector<std::string> cf_names = {"default"};
+  for (auto* h : handles) {
+    cf_names.push_back(h->GetName());
+  }
+  for (auto* h : handles) {
+    ASSERT_OK(db_->DestroyColumnFamilyHandle(h));
+  }
+  handles.clear();
+
+  Close();
+  // Use a max_manifest_file_size below the reused manifest size. Auto-tuning
+  // with the persisted compacted size at 200% amp keeps the tuned threshold
+  // high enough. Without persistence, the threshold would be
+  // max(max_manifest_file_size, 0 * anything) = max_manifest_file_size.
+  //
+  // With max_manifest_file_size set to 3000, missing persisted compacted size
+  // would keep tuned = 3000 and the first AddCf would rotate because the
+  // reused manifest is already larger. With persisted compacted size loaded,
+  // the tuned threshold is based on the compacted size, so no rotation until
+  // the manifest grows well beyond the minimum.
+  options.max_manifest_file_size = 3000;
+  options.max_manifest_space_amp_pct = 200;
+  options.reuse_manifest_on_open = true;
+  uint64_t manifest_num_before_reopen = cur_manifest_num;
+  ReopenWithColumnFamilies(cf_names, options);
+  cur_manifest_num = dbfull()->TEST_Current_Manifest_FileNo();
+
+  // With persistence: the manifest file number should NOT have changed
+  // during reopen, because the persisted compacted size keeps the tuned
+  // threshold high enough. Without persistence, last_compacted = 0, so
+  // tuned = max_manifest_file_size = 3000, and the first LogAndApply
+  // during Open rotates the manifest because it is already larger than 3000.
+  ASSERT_EQ(manifest_num_before_reopen, cur_manifest_num);
+
+  // Adding CFs should still not trigger rotation because the tuned threshold
+  // from the persisted compacted size exceeds the current manifest size.
+  for (int i = 1; i <= 4; ++i) {
+    AddCfFn();
+    ASSERT_EQ(prev_manifest_num, cur_manifest_num);
+  }
+  for (int i = 1; i <= 4; ++i) {
+    AddCfFn();
+    ASSERT_EQ(prev_manifest_num, cur_manifest_num);
+  }
+
   // Wrap up
   while (!handles.empty()) {
-    DropCfFn();
+    DropCfFn(/*include_background_manifest_write=*/false);
   }
 }
 

db/db_impl/db_impl.cc

@@ -102,6 +102,7 @@
 #include "table/get_context.h"
 #include "table/merging_iterator.h"
 #include "table/multiget_context.h"
+#include "table/prepared_file_info.h"
 #include "table/sst_file_dumper.h"
 #include "table/table_builder.h"
 #include "table/two_level_iterator.h"
@@ -175,6 +176,7 @@ DBImpl::DBImpl(const DBOptions& options, const std::string& dbname,
                const bool seq_per_batch, const bool batch_per_txn,
                bool read_only)
     : dbname_(dbname),
+      read_only_(read_only),
       own_info_log_(options.info_log == nullptr),
       initial_db_options_(SanitizeOptions(dbname, options, read_only,
                                           &init_logger_creation_s_)),
@@ -495,6 +497,21 @@ void DBImpl::WaitForAsyncFileOpen() {
   }
 }
 
+void DBImpl::NotifyOnDBShutdownBegin() {
+  mutex_.AssertHeld();
+  if (shutdown_notification_sent_ || immutable_db_options_.listeners.empty()) {
+    return;
+  }
+  shutdown_notification_sent_ = true;
+
+  // release lock while notifying events
+  mutex_.Unlock();
+  for (const auto& listener : immutable_db_options_.listeners) {
+    listener->OnDBShutdownBegin(this);
+  }
+  mutex_.Lock();
+}
+
 // Will lock the mutex_,  will wait for completion if wait is true
 void DBImpl::CancelAllBackgroundWork(bool wait) {
   ROCKS_LOG_INFO(immutable_db_options_.info_log,
@@ -541,6 +558,9 @@ void DBImpl::CancelAllBackgroundWork(bool wait) {
     immutable_db_options_.compaction_service->CancelAwaitingJobs();
   }
 
+  if (!already_shutting_down) {
+    NotifyOnDBShutdownBegin();
+  }
   shutting_down_.store(true, std::memory_order_release);
   bg_cv_.SignalAll();
   if (!wait) {
@@ -653,7 +673,8 @@ void DBImpl::UnregisterBlobDirectWriteColumnFamily() {
 Status DBImpl::MaybeWriteWalMarkersToManifestOnClose() {
   mutex_.AssertHeld();
   if (!mutable_db_options_.optimize_manifest_for_recovery ||
-      !opened_successfully_ || versions_ == nullptr || logs_.empty()) {
+      !opened_successfully_ || read_only_ || versions_ == nullptr ||
+      logs_.empty()) {
     return Status::OK();
   }
 
@@ -789,16 +810,35 @@ Status DBImpl::CloseHelper() {
          bg_flush_scheduled_ || bg_purge_scheduled_ ||
          bg_pressure_callback_in_progress_ ||
          bg_async_file_open_state_ == AsyncFileOpenState::kScheduled ||
+         async_wal_precreate_state_ == AsyncWALPrecreateState::kScheduled ||
          pending_purge_obsolete_files_ ||
          error_handler_.IsRecoveryInProgress()) {
     TEST_SYNC_POINT("DBImpl::~DBImpl:WaitJob");
     bg_cv_.Wait();
   }
 
+  // Release any opened-but-unpublished WAL writer after the in-flight worker
+  // has published its result. Clear the DB-owned async slot while holding
+  // mutex_, but destroy the detached writer after dropping mutex_ because
+  // log::Writer / WritableFileWriter destruction can flush and close the file.
+  // The file itself can be left behind as an empty future WAL; recovery already
+  // tolerates it and marks its file number used if observed.
+  UnpublishedWAL unused_async_wal = std::move(async_wal_precreate_wal_);
+  async_wal_precreate_state_ = AsyncWALPrecreateState::kNotScheduled;
+  if (unused_async_wal.writer) {
+    mutex_.Unlock();
+    unused_async_wal.Reset();
+    mutex_.Lock();
+  }
+
   // Ensure subclasses don't forget to schedule async file opening
   assert(!immutable_db_options_.open_files_async || !opened_successfully_ ||
          bg_async_file_open_state_ != AsyncFileOpenState::kNotScheduled);
 
+  // No FileIngestionHandle from PrepareFileIngestion() may still be
+  // outstanding
+  assert(num_outstanding_prepared_ingestions_.load() == 0);
+
   TEST_SYNC_POINT_CALLBACK("DBImpl::CloseHelper:PendingPurgeFinished",
                            &files_grabbed_for_purge_);
   EraseThreadStatusDbInfo();
@@ -820,6 +860,8 @@ Status DBImpl::CloseHelper() {
   if (default_cf_handle_ != nullptr || persist_stats_cf_handle_ != nullptr) {
     // we need to delete handle outside of lock because it does its own locking
     mutex_.Unlock();
+    TEST_SYNC_POINT("DBImpl::CloseHelper:CFHandleCleanupUnlocked");
+    TEST_SYNC_POINT("DBImpl::CloseHelper:CFHandleCleanupAllowed");
     if (default_cf_handle_) {
       delete default_cf_handle_;
       default_cf_handle_ = nullptr;
@@ -840,7 +882,7 @@ Status DBImpl::CloseHelper() {
   // manifest file), it is not able to identify live files correctly. As a
   // result, all "live" files can get deleted by accident. However, corrupted
   // manifest is recoverable by RepairDB().
-  if (opened_successfully_) {
+  if (opened_successfully_ && !read_only_) {
     JobContext job_context(next_job_id_.fetch_add(1));
     FindObsoleteFiles(&job_context, true);
 
@@ -887,25 +929,6 @@ Status DBImpl::CloseHelper() {
     logs_.clear();
   }
 
-  // Table cache may have table handles holding blocks from the block cache.
-  // We need to release them before the block cache is destroyed. The block
-  // cache may be destroyed inside versions_.reset(), when column family data
-  // list is destroyed, so leaving handles in table cache after
-  // versions_.reset() may cause issues. Here we clean all unreferenced handles
-  // in table cache, and (for certain builds/conditions) assert that no obsolete
-  // files are hanging around unreferenced (leak) in the table/blob file cache.
-  // Now we assume all user queries have finished, so only version set itself
-  // can possibly hold the blocks from block cache. After releasing unreferenced
-  // handles here, only handles held by version set left and inside
-  // versions_.reset(), we will release them. There, we need to make sure every
-  // time a handle is released, we erase it from the cache too. By doing that,
-  // we can guarantee that after versions_.reset(), table cache is empty
-  // so the cache can be safely destroyed.
-#ifndef NDEBUG
-  TEST_VerifyNoObsoleteFilesCached(/*db_mutex_already_held=*/true);
-#endif  // !NDEBUG
-  table_cache_->EraseUnRefEntries();
-
   for (auto& txn_entry : recovered_transactions_) {
     delete txn_entry.second;
   }
@@ -931,6 +954,42 @@ Status DBImpl::CloseHelper() {
     }
   }
 
+  // Drain obsolete-file purge work started AFTER the early CloseHelper wait
+  // near the top of this method. A late SuperVersion cleanup -- a dropped
+  // column family handle, or an in-flight iterator/Get whose ReadOptions or
+  // immutable_db_options_.avoid_unnecessary_blocking_io selected background
+  // purge -- can run in one of the mutex-unlocked windows above and enter the
+  // FindObsoleteFiles() -> PurgeObsoleteFiles(..., true) handoff. During that
+  // handoff pending_purge_obsolete_files_ is already nonzero, but
+  // bg_purge_scheduled_ may still be zero until SchedulePurge() runs at the end
+  // of PurgeObsoleteFiles(). Wait for both states while mutex_/this are still
+  // alive; bg_cv_.Wait() releases mutex_ so pending and scheduled purges can
+  // finish and signal.
+  TEST_SYNC_POINT("DBImpl::CloseHelper:BeforeFinalPurgeDrain");
+  while (pending_purge_obsolete_files_ || bg_purge_scheduled_) {
+    TEST_SYNC_POINT("DBImpl::CloseHelper:FinalPurgeDrainWait");
+    bg_cv_.Wait();
+  }
+
+  // Table cache may have table handles holding blocks from the block cache.
+  // We need to release them before the block cache is destroyed. The block
+  // cache may be destroyed inside versions_.reset(), when column family data
+  // list is destroyed, so leaving handles in table cache after
+  // versions_.reset() may cause issues. Here we clean all unreferenced handles
+  // in table cache, and (for certain builds/conditions) assert that no obsolete
+  // files are hanging around unreferenced (leak) in the table/blob file cache.
+  // Now we assume all user queries have finished, and close-time purge work has
+  // settled, so only version set itself can possibly hold the blocks from block
+  // cache. After releasing unreferenced handles here, only handles held by
+  // version set left and inside versions_.reset(), we will release them. There,
+  // we need to make sure every time a handle is released, we erase it from the
+  // cache too. By doing that, we can guarantee that after versions_.reset(),
+  // table cache is empty so the cache can be safely destroyed.
+#ifndef NDEBUG
+  TEST_VerifyNoObsoleteFilesCached(/*db_mutex_already_held=*/true);
+#endif  // !NDEBUG
+  table_cache_->EraseUnRefEntries();
+
   versions_.reset();
   mutex_.Unlock();
   if (db_lock_ != nullptr) {
@@ -1809,6 +1868,8 @@ Status DBImpl::SetDBOptions(
       // TODO(xiez): clarify why apply optimize for read to write options
       file_options_for_compaction_ = fs_->OptimizeForCompactionTableRead(
           file_options_for_compaction_, immutable_db_options_);
+      TEST_SYNC_POINT_CALLBACK("DBImpl::SetDBOptions:FileOptionsForCompaction",
+                               &file_options_for_compaction_);
       if (wal_other_option_changed || wal_size_option_changed) {
         WriteThread::Writer w;
         write_thread_.EnterUnbatched(&w, &mutex_);
@@ -2440,30 +2501,8 @@ struct SuperVersionHandle {
 static void CleanupSuperVersionHandle(void* arg1, void* /*arg2*/) {
   SuperVersionHandle* sv_handle = static_cast<SuperVersionHandle*>(arg1);
 
-  if (sv_handle->super_version->Unref()) {
-    // Job id == 0 means that this is not our background process, but rather
-    // user thread
-    JobContext job_context(0);
-
-    sv_handle->mu->Lock();
-    sv_handle->super_version->Cleanup();
-    sv_handle->db->FindObsoleteFiles(&job_context, false, true);
-    if (sv_handle->background_purge) {
-      sv_handle->db->ScheduleBgLogWriterClose(&job_context);
-      sv_handle->db->AddSuperVersionsToFreeQueue(sv_handle->super_version);
-      sv_handle->db->SchedulePurge();
-    }
-    sv_handle->mu->Unlock();
-
-    if (!sv_handle->background_purge) {
-      delete sv_handle->super_version;
-    }
-    if (job_context.HaveSomethingToDelete()) {
-      sv_handle->db->PurgeObsoleteFiles(job_context,
-                                        sv_handle->background_purge);
-    }
-    job_context.Clean();
-  }
+  sv_handle->db->CleanupIteratorSuperVersion(sv_handle->super_version,
+                                             sv_handle->background_purge);
 
   delete sv_handle;
 }
@@ -2485,7 +2524,8 @@ static void CleanupGetMergeOperandsState(void* arg1, void* /*arg2*/) {
 InternalIterator* DBImpl::NewInternalIterator(
     const ReadOptions& read_options, ColumnFamilyData* cfd,
     SuperVersion* super_version, Arena* arena, SequenceNumber sequence,
-    bool allow_unprepared_value, ArenaWrappedDBIter* db_iter) {
+    bool allow_unprepared_value, ArenaWrappedDBIter* db_iter,
+    const MultiScanArgs* scan_opts) {
   InternalIterator* internal_iter;
   assert(arena != nullptr);
   auto prefix_extractor =
@@ -2497,47 +2537,62 @@ InternalIterator* DBImpl::NewInternalIterator(
       // here, and no unit test cares about the value provided here.
       !read_options.total_order_seek && prefix_extractor != nullptr,
       read_options.iterate_upper_bound);
-  // Collect iterator for mutable memtable
-  auto mem_iter = super_version->mem->NewIterator(
-      read_options, super_version->GetSeqnoToTimeMapping(), arena,
-      super_version->mutable_cf_options.prefix_extractor.get(),
-      /*for_flush=*/false);
   Status s;
-  if (!read_options.ignore_range_deletions) {
-    std::unique_ptr<TruncatedRangeDelIterator> mem_tombstone_iter;
-    auto range_del_iter = super_version->mem->NewRangeTombstoneIterator(
-        read_options, sequence, false /* immutable_memtable */);
-    if (range_del_iter == nullptr || range_del_iter->empty()) {
-      delete range_del_iter;
+  const Comparator* user_comparator = cfd->user_comparator();
+  const bool mem_intersects =
+      !super_version->mem->IsEmpty() &&
+      MultiScanIntersectsMemTable(super_version->mem, read_options,
+                                  super_version->GetSeqnoToTimeMapping(),
+                                  prefix_extractor, scan_opts, user_comparator);
+  if (scan_opts == nullptr || mem_intersects) {
+    // Collect iterator for mutable memtable
+    auto mem_iter = super_version->mem->NewIterator(
+        read_options, super_version->GetSeqnoToTimeMapping(), arena,
+        super_version->mutable_cf_options.prefix_extractor.get(),
+        /*for_flush=*/false);
+    if (!read_options.ignore_range_deletions) {
+      std::unique_ptr<TruncatedRangeDelIterator> mem_tombstone_iter;
+      std::unique_ptr<FragmentedRangeTombstoneIterator> range_del_iter(
+          super_version->mem->NewRangeTombstoneIterator(
+              read_options, sequence, false /* immutable_memtable */));
+      if (range_del_iter != nullptr && !range_del_iter->empty()) {
+        mem_tombstone_iter = std::make_unique<TruncatedRangeDelIterator>(
+            std::move(range_del_iter), &cfd->ioptions().internal_comparator,
+            nullptr /* smallest */, nullptr /* largest */);
+      }
+      merge_iter_builder.AddPointAndTombstoneIterator(
+          mem_iter, std::move(mem_tombstone_iter));
     } else {
-      mem_tombstone_iter = std::make_unique<TruncatedRangeDelIterator>(
-          std::unique_ptr<FragmentedRangeTombstoneIterator>(range_del_iter),
-          &cfd->ioptions().internal_comparator, nullptr /* smallest */,
-          nullptr /* largest */);
+      merge_iter_builder.AddIterator(mem_iter);
     }
-    merge_iter_builder.AddPointAndTombstoneIterator(
-        mem_iter, std::move(mem_tombstone_iter));
-  } else {
-    merge_iter_builder.AddIterator(mem_iter);
+  } else if (scan_opts != nullptr) {
+    merge_iter_builder.SetMemtablePruned(true);
   }
 
-  // Collect all needed child iterators for immutable memtables
-  if (s.ok()) {
+  if (s.ok() &&
+      (scan_opts == nullptr || super_version->imm->GetTotalNumEntries() > 0)) {
+    // Collect all needed child iterators for immutable memtables. When scan
+    // ranges are provided, AddIterators prunes each immutable memtable
+    // individually.
     super_version->imm->AddIterators(
         read_options, super_version->GetSeqnoToTimeMapping(),
         super_version->mutable_cf_options.prefix_extractor.get(),
-        &merge_iter_builder, !read_options.ignore_range_deletions);
+        &merge_iter_builder, !read_options.ignore_range_deletions, sequence,
+        scan_opts, user_comparator);
   }
   TEST_SYNC_POINT_CALLBACK("DBImpl::NewInternalIterator:StatusCallback", &s);
   if (s.ok()) {
     // Collect iterators for files in L0 - Ln
     if (read_options.read_tier != kMemtableTier) {
-      super_version->current->AddIterators(read_options, file_options_,
-                                           &merge_iter_builder,
-                                           allow_unprepared_value);
+      super_version->current->AddIterators(
+          read_options, file_options_, &merge_iter_builder,
+          allow_unprepared_value, sequence, scan_opts);
     }
     internal_iter = merge_iter_builder.Finish(
         read_options.ignore_range_deletions ? nullptr : db_iter);
+    if (internal_iter == nullptr) {
+      internal_iter = NewEmptyInternalIterator<Slice>(arena);
+    }
     SuperVersionHandle* cleanup = new SuperVersionHandle(
         this, &mutex_, super_version,
         read_options.background_purge_on_iterator_cleanup ||
@@ -2551,6 +2606,35 @@ InternalIterator* DBImpl::NewInternalIterator(
   return NewErrorInternalIterator<Slice>(s, arena);
 }
 
+void DBImpl::CleanupIteratorSuperVersion(SuperVersion* super_version,
+                                         bool background_purge) {
+  background_purge =
+      background_purge || immutable_db_options_.avoid_unnecessary_blocking_io;
+  if (super_version->Unref()) {
+    // Job id == 0 means that this is not our background process, but rather
+    // user thread
+    JobContext job_context(0);
+
+    mutex_.Lock();
+    super_version->Cleanup();
+    FindObsoleteFiles(&job_context, false, true);
+    if (background_purge) {
+      ScheduleBgLogWriterClose(&job_context);
+      AddSuperVersionsToFreeQueue(super_version);
+      SchedulePurge();
+    }
+    mutex_.Unlock();
+
+    if (!background_purge) {
+      delete super_version;
+    }
+    if (job_context.HaveSomethingToDelete()) {
+      PurgeObsoleteFiles(job_context, background_purge);
+    }
+    job_context.Clean();
+  }
+}
+
 ColumnFamilyHandle* DBImpl::DefaultColumnFamily() const {
   return default_cf_handle_;
 }
@@ -5820,6 +5904,7 @@ Status DBImpl::DeleteFilesInRanges(ColumnFamilyHandle* column_family,
   }
 
   VersionEdit edit;
+  edit.MarkForegroundOperation();
   std::set<FileMetaData*> deleted_files;
   JobContext job_context(next_job_id_.fetch_add(1), true);
   {
@@ -5917,6 +6002,79 @@ Status DBImpl::GetLiveFilesChecksumInfo(FileChecksumList* checksum_list) {
   return versions_->GetLiveFilesChecksumInfo(checksum_list);
 }
 
+Status DBImpl::GetPreparedFileInfoForExternalSstIngestion(
+    const std::string& file_path,
+    std::shared_ptr<const PreparedFileInfo>* file_info) {
+  if (file_info == nullptr) {
+    return Status::InvalidArgument("file_info must not be null");
+  }
+  file_info->reset();
+
+  const size_t file_name_pos = file_path.find_last_of("/\\");
+  const std::string file_name = file_name_pos == std::string::npos
+                                    ? file_path
+                                    : file_path.substr(file_name_pos + 1);
+  uint64_t file_number = 0;
+  FileType file_type;
+  if (!ParseFileName(file_name, &file_number, &file_type) ||
+      file_type != kTableFile || file_number == 0) {
+    return Status::InvalidArgument("Invalid table file name: " + file_path);
+  }
+
+  int file_level = -1;
+  FileMetaData* file_meta = nullptr;
+  ColumnFamilyData* file_cfd = nullptr;
+  Version* file_version = nullptr;
+  std::shared_ptr<const TableProperties> table_properties;
+  ReadOptions read_options;
+
+  {
+    InstrumentedMutexLock l(&mutex_);
+    Status s = versions_->GetMetadataForFile(file_number, &file_level,
+                                             &file_meta, &file_cfd);
+    if (!s.ok()) {
+      return s;
+    }
+
+    const std::string expected_file_path = TableFileName(
+        file_cfd->ioptions().cf_paths, file_number, file_meta->fd.GetPathId());
+    if (file_path != expected_file_path) {
+      return Status::InvalidArgument("Path does not match live table file: " +
+                                     file_path);
+    }
+
+    file_cfd->Ref();
+    file_version = file_cfd->current();
+    file_version->Ref();
+  }
+  const Defer cleanup_refs([&]() {
+    InstrumentedMutexLock l(&mutex_);
+    file_version->Unref();
+    file_cfd->UnrefAndTryDelete();
+  });
+
+  Status s = file_cfd->table_cache()->GetTableProperties(
+      file_options_, read_options, file_cfd->internal_comparator(), *file_meta,
+      &table_properties, file_version->GetMutableCFOptions(),
+      false /* no_io */);
+  if (!s.ok()) {
+    return s;
+  }
+  assert(table_properties != nullptr);
+
+  auto prepared_file_info = std::make_shared<PreparedFileInfo>();
+  prepared_file_info->file_size = file_meta->fd.GetFileSize();
+  prepared_file_info->smallest = file_meta->smallest;
+  prepared_file_info->largest = file_meta->largest;
+  prepared_file_info->table_properties = *table_properties;
+  prepared_file_info->table_properties.key_largest_seqno =
+      file_meta->fd.largest_seqno;
+  prepared_file_info->table_properties.key_smallest_seqno =
+      file_meta->fd.smallest_seqno;
+  *file_info = std::move(prepared_file_info);
+  return s;
+}
+
 void DBImpl::GetColumnFamilyMetaData(ColumnFamilyHandle* column_family,
                                      ColumnFamilyMetaData* cf_meta) {
   assert(column_family);
@@ -6639,11 +6797,76 @@ Status DBImpl::IngestExternalFile(
   return IngestExternalFiles({arg});
 }
 
-Status DBImpl::IngestExternalFiles(
-    const std::vector<IngestExternalFileArg>& args) {
-  PERF_TIMER_GUARD(file_ingestion_nanos);
-  // TODO: plumb Env::IOActivity, Env::IOPriority
-  const WriteOptions write_options;
+class FileIngestionHandleImpl : public FileIngestionHandle {
+ public:
+  explicit FileIngestionHandleImpl(DBImpl* db) : db_(db) {}
+  ~FileIngestionHandleImpl() override;
+
+  FileIngestionHandleImpl(const FileIngestionHandleImpl&) = delete;
+  FileIngestionHandleImpl& operator=(const FileIngestionHandleImpl&) = delete;
+  FileIngestionHandleImpl(FileIngestionHandleImpl&&) = delete;
+  FileIngestionHandleImpl& operator=(FileIngestionHandleImpl&&) = delete;
+
+  Status Abort() override;
+
+  DBImpl* const db_;
+  std::vector<ExternalSstFileIngestionJob> jobs_;
+  std::unique_ptr<std::list<uint64_t>::iterator> pending_output_elem_;
+  bool fill_cache_ = true;
+  // Set true once committed or aborted, so the destructor does not roll back.
+  bool consumed_ = false;
+};
+
+FileIngestionHandleImpl::~FileIngestionHandleImpl() {
+  if (consumed_ || db_ == nullptr) {
+    return;
+  }
+  // Dropped without commit or abort: roll back as a safety net.
+  db_->RollbackPreparedFileIngestion(this);
+  ROCKS_LOG_WARN(
+      db_->immutable_db_options_.info_log,
+      "[%zu CF(s)] File ingestion handle destroyed without commit or "
+      "abort; prepared files were rolled back.",
+      jobs_.size());
+}
+
+void DBImpl::RollbackPreparedFileIngestion(FileIngestionHandleImpl* const h) {
+  // Delete the staged internal files and release the reserved file numbers /
+  // pending-output protection, leaving the DB unchanged.
+  const Status rollback_status = Status::Incomplete("file ingestion aborted");
+  for (auto& job : h->jobs_) {
+    job.Cleanup(rollback_status);
+  }
+  {
+    InstrumentedMutexLock l(&mutex_);
+    ReleaseFileNumberFromPendingOutputs(h->pending_output_elem_);
+  }
+  h->consumed_ = true;
+  num_outstanding_prepared_ingestions_.fetch_sub(1);
+}
+
+Status FileIngestionHandleImpl::Abort() {
+  if (consumed_) {
+    return Status::InvalidArgument(
+        "file ingestion handle has already been committed or aborted");
+  }
+  db_->RollbackPreparedFileIngestion(this);
+  return Status::OK();
+}
+
+Status DBImpl::PrepareFileIngestion(
+    const std::vector<IngestExternalFileArg>& args,
+    std::unique_ptr<FileIngestionHandle>* handle) {
+  if (handle == nullptr) {
+    return Status::InvalidArgument("file ingestion handle output is null");
+  }
+  handle->reset();
+  // Recorded as INGEST_EXTERNAL_FILE_PREPARE_TIME on success below.
+  const bool record_ingest_micros =
+      stats_ != nullptr &&
+      stats_->get_stats_level() > StatsLevel::kExceptTimers;
+  const uint64_t prepare_start_micros =
+      record_ingest_micros ? immutable_db_options_.clock->NowMicros() : 0;
 
   if (args.empty()) {
     return Status::InvalidArgument("ingestion arg list is empty");
@@ -6668,6 +6891,17 @@ Status DBImpl::IngestExternalFiles(
           "external_files[" + std::to_string(i) + "] is empty";
       return Status::InvalidArgument(err_msg);
     }
+    if (!args[i].file_infos.empty()) {
+      if (args[i].file_infos.size() != args[i].external_files.size()) {
+        return Status::InvalidArgument("file_infos[" + std::to_string(i) +
+                                       "] size must match external_files[" +
+                                       std::to_string(i) + "] size");
+      }
+      if (args[i].options.write_global_seqno) {
+        return Status::InvalidArgument(
+            "write_global_seqno is not supported when file_infos is set");
+      }
+    }
     if (i && args[i].options.fill_cache != args[i - 1].options.fill_cache) {
       return Status::InvalidArgument(
           "fill_cache should be the same across ingestion options.");
@@ -6744,6 +6978,7 @@ Status DBImpl::IngestExternalFiles(
   }
 
   std::vector<ExternalSstFileIngestionJob> ingestion_jobs;
+  ingestion_jobs.reserve(num_cfs);
   for (const auto& arg : args) {
     auto* cfd = static_cast<ColumnFamilyHandleImpl*>(arg.column_family)->cfd();
     ingestion_jobs.emplace_back(versions_.get(), cfd, immutable_db_options_,
@@ -6761,8 +6996,9 @@ Status DBImpl::IngestExternalFiles(
             this);
     Status es = ingestion_jobs[i].Prepare(
         args[i].external_files, args[i].files_checksums,
-        args[i].files_checksum_func_names, args[i].atomic_replace_range,
-        args[i].file_temperature, start_file_number, super_version);
+        args[i].files_checksum_func_names, args[i].file_infos,
+        args[i].atomic_replace_range, args[i].file_temperature,
+        start_file_number, super_version);
     // capture first error only
     if (!es.ok() && status.ok()) {
       status = es;
@@ -6777,8 +7013,9 @@ Status DBImpl::IngestExternalFiles(
             this);
     Status es = ingestion_jobs[0].Prepare(
         args[0].external_files, args[0].files_checksums,
-        args[0].files_checksum_func_names, args[0].atomic_replace_range,
-        args[0].file_temperature, next_file_number, super_version);
+        args[0].files_checksum_func_names, args[0].file_infos,
+        args[0].atomic_replace_range, args[0].file_temperature,
+        next_file_number, super_version);
     if (!es.ok()) {
       status = es;
     }
@@ -6793,8 +7030,86 @@ Status DBImpl::IngestExternalFiles(
     return status;
   }
 
+  auto handle_impl = std::make_unique<FileIngestionHandleImpl>(this);
+  handle_impl->jobs_ = std::move(ingestion_jobs);
+  handle_impl->pending_output_elem_ = std::move(pending_output_elem);
+  handle_impl->fill_cache_ = args[0].options.fill_cache;
+  if (record_ingest_micros) {
+    RecordTimeToHistogram(
+        stats_, INGEST_EXTERNAL_FILE_PREPARE_TIME,
+        immutable_db_options_.clock->NowMicros() - prepare_start_micros);
+  }
+  num_outstanding_prepared_ingestions_.fetch_add(1);
+  *handle = std::move(handle_impl);
+  return Status::OK();
+}
+
+Status DBImpl::CommitFileIngestionHandles(
+    std::vector<std::unique_ptr<FileIngestionHandle>> handles) {
+  if (handles.empty()) {
+    return Status::InvalidArgument("no file ingestion handles to commit");
+  }
+  // Validate the handles and group their jobs by column family, merging same-CF
+  // jobs into one so each column family commits via a single atomic version
+  // edit and one SuperVersion install.
+  Status status;
+  std::vector<FileIngestionHandleImpl*> hs;
+  hs.reserve(handles.size());
+  std::vector<ExternalSstFileIngestionJob*> ingestion_jobs;
+  const bool fill_cache =
+      static_cast<FileIngestionHandleImpl*>(handles[0].get())->fill_cache_;
+  int max_file_opening_threads = 1;
+
+  {
+    UnorderedMap<ColumnFamilyData*, ExternalSstFileIngestionJob*>
+        primary_job_for_cfd;
+    for (const auto& handle : handles) {
+      assert(handle);
+      auto* h = static_cast<FileIngestionHandleImpl*>(handle.get());
+      assert(h->db_ == this);
+      if (h->consumed_) {
+        status = Status::InvalidArgument(
+            "file ingestion handle has already been committed or aborted");
+      }
+      if (h->fill_cache_ != fill_cache) {
+        status = Status::InvalidArgument(
+            "fill cache arg must be consistent across all handles");
+      }
+      if (!status.ok()) {
+        return status;
+      }
+
+      hs.push_back(h);
+      for (auto& job : h->jobs_) {
+        max_file_opening_threads =
+            std::max(max_file_opening_threads, job.file_opening_threads());
+        auto [it, inserted] =
+            primary_job_for_cfd.try_emplace(job.GetColumnFamilyData(), &job);
+        if (inserted) {
+          ingestion_jobs.push_back(&job);
+        } else {
+          status = it->second->MergeForSameColumnFamily(&job);
+          if (!status.ok()) {
+            return status;
+          }
+        }
+      }
+    }
+  }
+  const size_t num_jobs = ingestion_jobs.size();
+
+  // TODO: plumb Env::IOActivity, Env::IOPriority
+  const WriteOptions write_options;
+  // Decided here (not in Prepare) so the histograms reflect the stats level at
+  // commit time; recorded only on success below.
+  const bool record_ingest_micros =
+      stats_ != nullptr &&
+      stats_->get_stats_level() > StatsLevel::kExceptTimers;
+  const uint64_t run_start_micros =
+      record_ingest_micros ? immutable_db_options_.clock->NowMicros() : 0;
+
   std::vector<SuperVersionContext> sv_ctxs;
-  for (size_t i = 0; i != num_cfs; ++i) {
+  for (size_t i = 0; i != num_jobs; ++i) {
     sv_ctxs.emplace_back(true /* create_superversion */);
   }
   TEST_SYNC_POINT("DBImpl::IngestExternalFiles:BeforeJobsRun:0");
@@ -6805,9 +7120,9 @@ Status DBImpl::IngestExternalFiles(
   // mutex so the lock-acquisition order is ingest_sst_lock -> DB mutex
   // throughout. Use readlock so we still allow concurrent ingestions.
   std::vector<std::unique_ptr<ReadLock>> ingest_read_locks;
-  ingest_read_locks.reserve(num_cfs);
-  for (size_t i = 0; i != num_cfs; ++i) {
-    auto* cfd = ingestion_jobs[i].GetColumnFamilyData();
+  ingest_read_locks.reserve(num_jobs);
+  for (auto* job : ingestion_jobs) {
+    auto* cfd = job->GetColumnFamilyData();
     if (!cfd->IsDropped()) {
       ingest_read_locks.emplace_back(
           std::make_unique<ReadLock>(&cfd->GetIngestSstLock()));
@@ -6833,14 +7148,14 @@ Status DBImpl::IngestExternalFiles(
     // So wait here to ensure there is no pending write to memtable.
     WaitForPendingWrites();
 
-    num_running_ingest_file_ += static_cast<int>(num_cfs);
+    num_running_ingest_file_ += static_cast<int>(num_jobs);
     TEST_SYNC_POINT("DBImpl::IngestExternalFile:AfterIncIngestFileCounter");
     TEST_SYNC_POINT("DBImpl::IngestExternalFile:AfterIncIngestFileCounter:2");
 
     bool at_least_one_cf_need_flush = false;
-    std::vector<bool> need_flush(num_cfs, false);
-    for (size_t i = 0; i != num_cfs; ++i) {
-      auto* cfd = ingestion_jobs[i].GetColumnFamilyData();
+    std::vector<bool> need_flush(num_jobs, false);
+    for (size_t i = 0; i != num_jobs; ++i) {
+      auto* cfd = ingestion_jobs[i]->GetColumnFamilyData();
       if (cfd->IsDropped()) {
         // TODO (yanqin) investigate whether we should abort ingestion or
         // proceed with other non-dropped column families.
@@ -6849,7 +7164,7 @@ Status DBImpl::IngestExternalFiles(
         break;
       }
       bool tmp = false;
-      status = ingestion_jobs[i].NeedsFlush(&tmp, cfd->GetSuperVersion());
+      status = ingestion_jobs[i]->NeedsFlush(&tmp, cfd->GetSuperVersion());
       need_flush[i] = tmp;
       at_least_one_cf_need_flush = (at_least_one_cf_need_flush || tmp);
       if (!status.ok()) {
@@ -6869,11 +7184,11 @@ Status DBImpl::IngestExternalFiles(
             {} /* provided_candidate_cfds */, true /* entered_write_thread */);
         mutex_.Lock();
       } else {
-        for (size_t i = 0; i != num_cfs; ++i) {
+        for (size_t i = 0; i != num_jobs; ++i) {
           if (need_flush[i]) {
             mutex_.Unlock();
             status =
-                FlushMemTable(ingestion_jobs[i].GetColumnFamilyData(),
+                FlushMemTable(ingestion_jobs[i]->GetColumnFamilyData(),
                               flush_opts, FlushReason::kExternalFileIngestion,
                               true /* entered_write_thread */);
             mutex_.Lock();
@@ -6884,22 +7199,22 @@ Status DBImpl::IngestExternalFiles(
         }
       }
       if (status.ok()) {
-        for (size_t i = 0; i != num_cfs; ++i) {
+        for (size_t i = 0; i != num_jobs; ++i) {
           if (immutable_db_options_.atomic_flush || need_flush[i]) {
-            ingestion_jobs[i].SetFlushedBeforeRun();
+            ingestion_jobs[i]->SetFlushedBeforeRun();
           }
         }
       }
     }
     // Run ingestion jobs.
     if (status.ok()) {
-      for (size_t i = 0; i != num_cfs; ++i) {
+      for (size_t i = 0; i != num_jobs; ++i) {
         mutex_.AssertHeld();
-        status = ingestion_jobs[i].Run();
+        status = ingestion_jobs[i]->Run();
         if (!status.ok()) {
           break;
         }
-        ingestion_jobs[i].RegisterRange();
+        ingestion_jobs[i]->RegisterRange();
       }
     }
     // Now that Run() has assigned the actual seqno for each ingested file,
@@ -6911,9 +7226,9 @@ Status DBImpl::IngestExternalFiles(
     // next conversion observes the new barrier and refuses any insert
     // with insert_seq < assigned.
     if (status.ok()) {
-      for (size_t i = 0; i != num_cfs; ++i) {
-        auto* cfd = ingestion_jobs[i].GetColumnFamilyData();
-        SequenceNumber assigned = ingestion_jobs[i].MaxAssignedSequenceNumber();
+      for (auto* job : ingestion_jobs) {
+        auto* cfd = job->GetColumnFamilyData();
+        SequenceNumber assigned = job->MaxAssignedSequenceNumber();
         if (assigned > 0) {
           cfd->mem()->BumpIngestSeqnoBarrier(assigned);
         }
@@ -6921,16 +7236,18 @@ Status DBImpl::IngestExternalFiles(
     }
     if (status.ok()) {
       ReadOptions read_options;
-      read_options.fill_cache = args[0].options.fill_cache;
+      read_options.fill_cache = fill_cache;
       autovector<ColumnFamilyData*> cfds_to_commit;
       autovector<autovector<VersionEdit*>> edit_lists;
       uint32_t num_entries = 0;
-      for (size_t i = 0; i != num_cfs; ++i) {
-        auto* cfd = ingestion_jobs[i].GetColumnFamilyData();
+      for (auto* job : ingestion_jobs) {
+        auto* cfd = job->GetColumnFamilyData();
         assert(!cfd->IsDropped());
         cfds_to_commit.push_back(cfd);
         autovector<VersionEdit*> edit_list;
-        edit_list.push_back(ingestion_jobs[i].edit());
+        auto* edit = job->edit();
+        edit->MarkForegroundOperation();
+        edit_list.push_back(edit);
         edit_lists.push_back(edit_list);
         ++num_entries;
       }
@@ -6943,10 +7260,11 @@ Status DBImpl::IngestExternalFiles(
         }
         assert(0 == num_entries);
       }
-      status =
-          versions_->LogAndApply(cfds_to_commit, read_options, write_options,
-
-                                 edit_lists, &mutex_, directories_.GetDbDir());
+      status = versions_->LogAndApply(
+          cfds_to_commit, read_options, write_options, edit_lists, &mutex_,
+          directories_.GetDbDir(), false /* new_descriptor_log */,
+          nullptr /* new_cf_options */, {} /* manifest_wcbs */, {} /* pre_cb */,
+          max_file_opening_threads);
       // It is safe to update VersionSet last seqno here after LogAndApply since
       // LogAndApply persists last sequence number from VersionEdits,
       // which are from file's largest seqno and not from VersionSet.
@@ -6955,11 +7273,10 @@ Status DBImpl::IngestExternalFiles(
       // mutex when persisting MANIFEST file, and the snapshots taken during
       // that period will not be stable if VersionSet last seqno is updated
       // before LogAndApply.
-      SequenceNumber max_assigned_seqno =
-          ingestion_jobs[0].MaxAssignedSequenceNumber();
-      for (size_t i = 1; i != num_cfs; ++i) {
-        max_assigned_seqno = std::max(
-            max_assigned_seqno, ingestion_jobs[i].MaxAssignedSequenceNumber());
+      SequenceNumber max_assigned_seqno = 0;
+      for (auto* job : ingestion_jobs) {
+        max_assigned_seqno =
+            std::max(max_assigned_seqno, job->MaxAssignedSequenceNumber());
       }
       if (max_assigned_seqno > 0) {
         const SequenceNumber last_seqno = versions_->LastSequence();
@@ -6971,17 +7288,17 @@ Status DBImpl::IngestExternalFiles(
       }
     }
 
-    for (auto& job : ingestion_jobs) {
-      job.UnregisterRange();
+    for (auto* job : ingestion_jobs) {
+      job->UnregisterRange();
     }
 
     if (status.ok()) {
-      for (size_t i = 0; i != num_cfs; ++i) {
-        auto* cfd = ingestion_jobs[i].GetColumnFamilyData();
+      for (size_t i = 0; i != num_jobs; ++i) {
+        auto* cfd = ingestion_jobs[i]->GetColumnFamilyData();
         assert(!cfd->IsDropped());
         InstallSuperVersionAndScheduleWork(cfd, &sv_ctxs[i]);
 #ifndef NDEBUG
-        if (0 == i && num_cfs > 1) {
+        if (0 == i && num_jobs > 1) {
           TEST_SYNC_POINT("DBImpl::IngestExternalFiles:InstallSVForFirstCF:0");
           TEST_SYNC_POINT("DBImpl::IngestExternalFiles:InstallSVForFirstCF:1");
         }
@@ -7006,12 +7323,14 @@ Status DBImpl::IngestExternalFiles(
     PERF_TIMER_STOP(file_ingestion_blocking_live_writes_nanos);
 
     if (status.ok()) {
-      for (auto& job : ingestion_jobs) {
-        job.UpdateStats();
+      for (auto* job : ingestion_jobs) {
+        job->UpdateStats();
+      }
+      for (auto* h : hs) {
+        ReleaseFileNumberFromPendingOutputs(h->pending_output_elem_);
       }
     }
-    ReleaseFileNumberFromPendingOutputs(pending_output_elem);
-    num_running_ingest_file_ -= static_cast<int>(num_cfs);
+    num_running_ingest_file_ -= static_cast<int>(num_jobs);
     if (0 == num_running_ingest_file_) {
       bg_cv_.SignalAll();
     }
@@ -7019,24 +7338,46 @@ Status DBImpl::IngestExternalFiles(
   }
   // mutex_ is unlocked here
 
-  // Cleanup
-  for (size_t i = 0; i != num_cfs; ++i) {
-    sv_ctxs[i].Clean();
-    // This may rollback jobs that have completed successfully. This is
-    // intended for atomicity.
-    ingestion_jobs[i].Cleanup(status);
+  for (auto& sv_ctx : sv_ctxs) {
+    sv_ctx.Clean();
   }
-  if (status.ok()) {
-    for (size_t i = 0; i != num_cfs; ++i) {
-      auto* cfd = ingestion_jobs[i].GetColumnFamilyData();
-      if (!cfd->IsDropped()) {
-        NotifyOnExternalFileIngested(cfd, ingestion_jobs[i]);
-      }
+  if (!status.ok()) {
+    // The atomic commit failed; nothing was made visible. The handles are NOT
+    // consumed, so each one's destructor rolls it back
+    return status;
+  }
+
+  for (auto* job : ingestion_jobs) {
+    job->Cleanup(status);
+    auto* cfd = job->GetColumnFamilyData();
+    if (!cfd->IsDropped()) {
+      NotifyOnExternalFileIngested(cfd, *job);
     }
   }
+  for (auto* h : hs) {
+    h->consumed_ = true;
+    num_outstanding_prepared_ingestions_.fetch_sub(1);
+  }
+  // Record commit latency.
+  if (record_ingest_micros) {
+    RecordTimeToHistogram(
+        stats_, INGEST_EXTERNAL_FILE_RUN_TIME,
+        immutable_db_options_.clock->NowMicros() - run_start_micros);
+  }
   return status;
 }
 
+Status DBImpl::IngestExternalFiles(
+    const std::vector<IngestExternalFileArg>& args) {
+  PERF_TIMER_GUARD(file_ingestion_nanos);
+  std::unique_ptr<FileIngestionHandle> handle;
+  Status status = PrepareFileIngestion(args, &handle);
+  if (!status.ok()) {
+    return status;
+  }
+  return CommitFileIngestionHandle(std::move(handle));
+}
+
 Status DBImpl::CreateColumnFamilyWithImport(
     const ColumnFamilyOptions& options, const std::string& column_family_name,
     const ImportColumnFamilyOptions& import_options,
@@ -7078,6 +7419,7 @@ Status DBImpl::CreateColumnFamilyWithImport(
 
   SuperVersionContext dummy_sv_ctx(/* create_superversion */ true);
   VersionEdit dummy_edit;
+  dummy_edit.MarkForegroundOperation();
   uint64_t next_file_number = 0;
   std::unique_ptr<std::list<uint64_t>::iterator> pending_output_elem;
   {
@@ -7135,9 +7477,10 @@ Status DBImpl::CreateColumnFamilyWithImport(
 
       // Install job edit [Mutex will be unlocked here]
       if (status.ok()) {
-        status = versions_->LogAndApply(cfd, read_options, write_options,
-                                        import_job.edit(), &mutex_,
-                                        directories_.GetDbDir());
+        auto* edit = import_job.edit();
+        edit->MarkForegroundOperation();
+        status = versions_->LogAndApply(cfd, read_options, write_options, edit,
+                                        &mutex_, directories_.GetDbDir());
         if (status.ok()) {
           InstallSuperVersionForConfigChange(cfd, &sv_context);
         }
@@ -7568,6 +7911,7 @@ Status DBImpl::ReserveFileNumbersBeforeIngestion(
   // reuse the file number that has already assigned to the internal file,
   // and this will overwrite the external file. To protect the external
   // file, we have to make sure the file number will never being reused.
+  dummy_edit.MarkForegroundOperation();
   s = versions_->LogAndApply(cfd, read_options, write_options, &dummy_edit,
                              &mutex_, directories_.GetDbDir());
   if (s.ok()) {

db/db_impl/db_impl.h

@@ -15,6 +15,8 @@
 #include <limits>
 #include <list>
 #include <map>
+#include <memory>
+#include <optional>
 #include <set>
 #include <string>
 #include <unordered_map>
@@ -74,6 +76,7 @@ namespace ROCKSDB_NAMESPACE {
 
 class Arena;
 class ArenaWrappedDBIter;
+class FileIngestionHandleImpl;
 class InMemoryStatsHistoryIterator;
 class MemTable;
 class PersistentStatsHistoryIterator;
@@ -573,6 +576,10 @@ class DBImpl : public DB {
       const LiveFilesStorageInfoOptions& opts,
       std::vector<LiveFileStorageInfo>* files) override;
 
+  Status GetPreparedFileInfoForExternalSstIngestion(
+      const std::string& file_path,
+      std::shared_ptr<const PreparedFileInfo>* file_info) override;
+
   // Obtains the meta data of the specified column family of the DB.
   // TODO(yhchiang): output parameter is placed in the end in this codebase.
   void GetColumnFamilyMetaData(ColumnFamilyHandle* column_family,
@@ -602,6 +609,14 @@ class DBImpl : public DB {
   Status IngestExternalFiles(
       const std::vector<IngestExternalFileArg>& args) override;
 
+  using DB::PrepareFileIngestion;
+  Status PrepareFileIngestion(
+      const std::vector<IngestExternalFileArg>& args,
+      std::unique_ptr<FileIngestionHandle>* handle) override;
+
+  Status CommitFileIngestionHandles(
+      std::vector<std::unique_ptr<FileIngestionHandle>> handles) override;
+
   using DB::CreateColumnFamilyWithImport;
   Status CreateColumnFamilyWithImport(
       const ColumnFamilyOptions& options, const std::string& column_family_name,
@@ -855,12 +870,22 @@ class DBImpl : public DB {
   // memtable range tombstone iterator used by the underlying merging iterator.
   // This range tombstone iterator can be refreshed later by db_iter.
   // @param read_options Must outlive the returned iterator.
-  InternalIterator* NewInternalIterator(const ReadOptions& read_options,
-                                        ColumnFamilyData* cfd,
-                                        SuperVersion* super_version,
-                                        Arena* arena, SequenceNumber sequence,
-                                        bool allow_unprepared_value,
-                                        ArenaWrappedDBIter* db_iter = nullptr);
+  // @param sequence The snapshot sequence captured when the DB iterator was
+  // created. Child iterators must use this instead of dereferencing
+  // read_options.snapshot, which may be released before lazy initialization.
+  // @param scan_opts Optional bounded scan ranges used only to prune the
+  // iterator tree during lazy Prepare() initialization.
+  InternalIterator* NewInternalIterator(
+      const ReadOptions& read_options, ColumnFamilyData* cfd,
+      SuperVersion* super_version, Arena* arena, SequenceNumber sequence,
+      bool allow_unprepared_value, ArenaWrappedDBIter* db_iter = nullptr,
+      const MultiScanArgs* scan_opts = nullptr);
+
+  // Release a SuperVersion held by an iterator. This preserves the cleanup
+  // behavior used by materialized internal iterators even when the DB iterator
+  // never needed to lazily build its child iterator tree.
+  void CleanupIteratorSuperVersion(SuperVersion* super_version,
+                                   bool background_purge);
 
   LogsWithPrepTracker* logs_with_prep_tracker() {
     return &logs_with_prep_tracker_;
@@ -1372,6 +1397,7 @@ class DBImpl : public DB {
 
  protected:
   const std::string dbname_;
+  const bool read_only_;
   // TODO(peterd): unify with VersionSet::db_id_
   std::string db_id_;
   // db_session_id_ is an identifier that gets reset
@@ -1453,6 +1479,13 @@ class DBImpl : public DB {
   std::atomic<int> next_job_id_ = 1;
 
   std::atomic<bool> shutting_down_ = false;
+  // Protected by mutex_. This is separate from shutting_down_ because
+  // OnDBShutdownBegin must fire before shutting_down_ is published, and the
+  // notification releases mutex_ while invoking listeners. Marking that the
+  // notification has started prevents a concurrent or reentrant
+  // CancelAllBackgroundWork() call from firing it again during that unlocked
+  // callback window.
+  bool shutdown_notification_sent_ = false;
 
   // No new background jobs can be queued if true. This is used to prevent new
   // background jobs from being queued after WaitForCompact() completes waiting
@@ -1527,6 +1560,16 @@ class DBImpl : public DB {
                                    const Status& st,
                                    const CompactionJobStats& job_stats,
                                    int job_id);
+  // Fires the OnCompactionPreCommit listener callback. Mirrors
+  // NotifyOnCompactionCompleted but is invoked before ReleaseCompactionFiles()
+  // (i.e. while input files still have being_compacted == true). Idempotent:
+  // safe to call from multiple potential release sites; only fires once per
+  // compaction, and only if NotifyOnCompactionBegin previously fired.
+  void NotifyOnCompactionPreCommit(ColumnFamilyData* cfd, Compaction* c,
+                                   const Status& st,
+                                   const CompactionJobStats& job_stats,
+                                   int job_id);
+  void NotifyOnDBShutdownBegin();
   void NotifyOnMemTableSealed(ColumnFamilyData* cfd,
                               const MemTableInfo& mem_table_info);
 
@@ -1887,6 +1930,7 @@ class DBImpl : public DB {
   friend class WriteBatchWithIndex;
   friend class WriteUnpreparedTxnDB;
   friend class WriteUnpreparedTxn;
+  friend class FileIngestionHandleImpl;
 
   friend class ForwardIterator;
   friend struct SuperVersion;
@@ -2224,6 +2268,10 @@ class DBImpl : public DB {
   void ReleaseFileNumberFromPendingOutputs(
       std::unique_ptr<std::list<uint64_t>::iterator>& v);
 
+  // Rolls back one prepared file ingestion (delete its staged files, release
+  // the reserved file numbers)
+  void RollbackPreparedFileIngestion(FileIngestionHandleImpl* const h);
+
   // Similar to pending_outputs, preserve OPTIONS file. Used for remote
   // compaction.
   std::list<uint64_t>::iterator CaptureOptionsFileNumber();
@@ -2574,10 +2622,15 @@ class DBImpl : public DB {
   // Helper function to perform trivial move by updating manifest metadata
   // without rewriting data files. This is called when IsTrivialMove() is true.
   // REQUIRES: mutex held
-  // Returns: Status of the trivial move operation
-  Status PerformTrivialMove(Compaction& c, LogBuffer* log_buffer,
-                            bool& compaction_released, size_t& moved_files,
-                            size_t& moved_bytes);
+  // Populates the Compaction's edit with DeleteFile/AddFile entries for
+  // a trivial move (no data rewriting). Does not commit the edit.
+  void PrepareTrivialMoveEdit(Compaction& c, LogBuffer* log_buffer,
+                              size_t& moved_files, size_t& moved_bytes);
+
+  // REQUIRES: mutex held, PrepareTrivialMoveEdit already called
+  // Commits the trivial move by calling LogAndApply on the prepared edit.
+  // Returns: Status of the manifest write
+  Status CommitTrivialMove(Compaction& c, bool& compaction_released);
 
   // REQUIRES: mutex unlocked
   void TrackOrUntrackFiles(const std::vector<std::string>& existing_data_files,
@@ -2810,7 +2863,68 @@ class DBImpl : public DB {
   size_t GetWalPreallocateBlockSize(uint64_t write_buffer_size) const;
   Env::WriteLifeTimeHint CalculateWALWriteHint() { return Env::WLTH_SHORT; }
 
-  IOStatus CreateWAL(const WriteOptions& write_options, uint64_t log_file_num,
+  // Returns true when async WAL precreation is enabled and compatible with the
+  // active WAL strategy. WAL recycling already avoids file creation latency, so
+  // precreation is disabled when recycle_log_file_num is non-zero.
+  bool AsyncWALPrecreateEnabled() const;
+
+  // A WAL file that has a reserved file number and may have an opened writer,
+  // but has not been added to DBImpl's in-memory logical WAL tracking lists
+  // (logs_ and alive_wal_files_).
+  struct UnpublishedWAL {
+    uint64_t log_number = 0;
+    std::unique_ptr<log::Writer> writer;
+
+    UnpublishedWAL() = default;
+    UnpublishedWAL(const UnpublishedWAL&) = delete;
+    UnpublishedWAL& operator=(const UnpublishedWAL&) = delete;
+
+    UnpublishedWAL(UnpublishedWAL&& other) noexcept {
+      *this = std::move(other);
+    }
+    UnpublishedWAL& operator=(UnpublishedWAL&& other) noexcept {
+      if (this != &other) {
+        log_number = other.log_number;
+        writer = std::move(other.writer);
+        other.Reset();
+      }
+      return *this;
+    }
+
+    void Reset() {
+      log_number = 0;
+      writer.reset();
+    }
+  };
+
+  // Reserves the next WAL file number and schedules a HIGH-priority background
+  // task to precreate that WAL file. A precreated WAL is not a logical WAL
+  // until a foreground WAL rotation consumes it.
+  void MaybeScheduleAsyncWALPrecreate(size_t preallocate_block_size);
+
+  // Background task for opening the reserved future WAL and publishing the
+  // result under mutex_.
+  static void BGWorkAsyncWALPrecreate(void* arg);
+
+  // Waits for an in-flight async WAL precreation and returns a prepared WAL if
+  // one is available. If precreation failed, returns an empty WAL and lets the
+  // foreground rotation create the WAL synchronously. Caller must hold mutex_.
+  UnpublishedWAL WaitForAsyncWALPrecreate();
+
+  // Opens and preallocates a WAL writer without writing logical WAL records.
+  // Used by async WAL precreation and by synchronous WAL creation.
+  IOStatus CreateWALWriter(const DBOptions& db_options, uint64_t log_file_num,
+                           uint64_t recycle_log_number,
+                           size_t preallocate_block_size,
+                           UnpublishedWAL* new_wal);
+
+  // Starts an opened WAL file by writing the initial records required before it
+  // can be installed as the current WAL for foreground writes.
+  IOStatus StartWALFile(const WriteOptions& write_options,
+                        const PredecessorWALInfo& predecessor_wal_info,
+                        log::Writer* new_log);
+  IOStatus CreateWAL(const DBOptions& db_options,
+                     const WriteOptions& write_options, uint64_t log_file_num,
                      uint64_t recycle_log_number, size_t preallocate_block_size,
                      const PredecessorWALInfo& predecessor_wal_info,
                      log::Writer** new_log);
@@ -3306,6 +3420,28 @@ class DBImpl : public DB {
   AsyncFileOpenState bg_async_file_open_state_ =
       AsyncFileOpenState::kNotScheduled;
 
+  // State machine for the single async WAL precreation slot protected by
+  // mutex_. Background precreation failure returns to kNotScheduled; foreground
+  // rotation handles it the same as no prepared WAL and creates one
+  // synchronously. kScheduled owns a reserved file number; kReady owns an
+  // opened writer that has not been started or added to logical WAL tracking.
+  enum class AsyncWALPrecreateState : uint8_t {
+    kNotScheduled = 0,  // No WAL precreate work is in-flight or ready.
+    kScheduled,         // Background task owns creation of the reserved WAL.
+    kReady,             // Reserved WAL writer is open but not logically live.
+  };
+
+  // Protected by mutex_. Tracks at most one background precreated WAL. A
+  // precreated WAL is only reserved empty storage until SwitchMemtable()
+  // consumes it and installs it in DBImpl's in-memory logical WAL tracking
+  // lists (logs_ and alive_wal_files_).
+  AsyncWALPrecreateState async_wal_precreate_state_ =
+      AsyncWALPrecreateState::kNotScheduled;
+
+  // Reserved in-flight/ready precreated WAL. The writer is populated only while
+  // state is kReady.
+  UnpublishedWAL async_wal_precreate_wal_;
+
   std::deque<ManualCompactionState*> manual_compaction_dequeue_;
 
   // shall we disable deletion of obsolete files
@@ -3355,6 +3491,10 @@ class DBImpl : public DB {
   // REQUIRES: mutex held
   int num_running_ingest_file_ = 0;
 
+  // Number of FileIngestionHandle objects produced by PrepareFileIngestion()
+  // that have not been committed or destroyed yet.
+  std::atomic<uint32_t> num_outstanding_prepared_ingestions_{0};
+
   WalManager wal_manager_;
 
   // A value of > 0 temporarily disables scheduling of background work

db/db_impl/db_impl_compaction_flush.cc

@@ -1581,9 +1581,8 @@ Status DBImpl::CompactFiles(const CompactionOptions& compact_options,
   return s;
 }
 
-Status DBImpl::PerformTrivialMove(Compaction& c, LogBuffer* log_buffer,
-                                  bool& compaction_released,
-                                  size_t& moved_files, size_t& moved_bytes) {
+void DBImpl::PrepareTrivialMoveEdit(Compaction& c, LogBuffer* log_buffer,
+                                    size_t& moved_files, size_t& moved_bytes) {
   mutex_.AssertHeld();
 
   ROCKS_LOG_BUFFER(log_buffer, "[%s] Moving %d files to level-%d\n",
@@ -1616,6 +1615,10 @@ Status DBImpl::PerformTrivialMove(Compaction& c, LogBuffer* log_buffer,
     }
     moved_files += c.num_input_files(l);
   }
+}
+
+Status DBImpl::CommitTrivialMove(Compaction& c, bool& compaction_released) {
+  mutex_.AssertHeld();
 
   // Install the new version
   const ReadOptions read_options(Env::IOActivity::kCompaction);
@@ -1740,8 +1743,8 @@ Status DBImpl::CompactFilesImpl(
     bool compaction_released = false;
     size_t moved_files = 0;
     size_t moved_bytes = 0;
-    Status status = PerformTrivialMove(
-        *c.get(), log_buffer, compaction_released, moved_files, moved_bytes);
+    PrepareTrivialMoveEdit(*c.get(), log_buffer, moved_files, moved_bytes);
+    Status status = CommitTrivialMove(*c.get(), compaction_released);
 
     if (status.ok()) {
       InstallSuperVersionAndScheduleWork(
@@ -2009,6 +2012,46 @@ void DBImpl::NotifyOnCompactionCompleted(
   // flush process.
 }
 
+void DBImpl::NotifyOnCompactionPreCommit(
+    ColumnFamilyData* cfd, Compaction* c, const Status& st,
+    const CompactionJobStats& compaction_job_stats, const int job_id) {
+  if (immutable_db_options_.listeners.size() == 0U) {
+    return;
+  }
+  mutex_.AssertHeld();
+  if (shutting_down_.load(std::memory_order_acquire)) {
+    return;
+  }
+
+  // Only fire if OnCompactionBegin has fired for this compaction.
+  if (c->ShouldNotifyOnCompactionCompleted() == false) {
+    return;
+  }
+  // Idempotency: ensure listeners observe at most one
+  // OnCompactionPreCommit per compaction lifecycle, even though the
+  // notify helper is invoked at multiple potential ReleaseCompactionFiles()
+  // sites for safety.
+  if (c->WasNotifyOnCompactionPreCommitCalled()) {
+    return;
+  }
+  c->SetNotifyOnCompactionPreCommitCalled();
+
+  int num_l0_files = cfd->current()->storage_info()->NumLevelFiles(0);
+  // release lock while notifying events
+  mutex_.Unlock();
+  TEST_SYNC_POINT("DBImpl::NotifyOnCompactionPreCommit::UnlockMutex");
+  {
+    CompactionJobInfo info{};
+    info.num_l0_files = num_l0_files;
+    BuildCompactionJobInfo(cfd, c, st, compaction_job_stats, job_id, &info);
+    for (const auto& listener : immutable_db_options_.listeners) {
+      listener->OnCompactionPreCommit(this, info);
+    }
+    info.status.PermitUncheckedError();
+  }
+  mutex_.Lock();
+}
+
 // REQUIREMENT: block all background work by calling PauseBackgroundWork()
 // before calling this function
 // TODO (hx235): Replace Status::NotSupported() with Status::Aborted() for
@@ -3180,7 +3223,7 @@ void DBImpl::ResumeAllCompactions() {
 void DBImpl::MaybeScheduleFlushOrCompaction() {
   mutex_.AssertHeld();
   TEST_SYNC_POINT("DBImpl::MaybeScheduleFlushOrCompaction:Start");
-  if (!opened_successfully_) {
+  if (!opened_successfully_ || read_only_) {
     // Compaction may introduce data race to DB open
     return;
   }
@@ -4024,6 +4067,13 @@ Status DBImpl::BackgroundCompaction(bool* made_progress,
       status = Status::ShutdownInProgress();
     } else if (compaction_aborted_.load(std::memory_order_acquire) > 0) {
       status = Status::Incomplete(Status::SubCode::kCompactionAborted);
+      if (!is_prepicked) {
+        // This automatic compaction was scheduled from compaction_queue_, but
+        // abort happened before PickCompactionFromQueue() could remove the
+        // queued CF. Restore the unscheduled count so ResumeAllCompactions()
+        // can schedule the still-queued work.
+        unscheduled_compactions_++;
+      }
     } else if (is_manual &&
                manual_compaction->canceled.load(std::memory_order_acquire)) {
       status = Status::Incomplete(Status::SubCode::kManualCompactionPaused);
@@ -4031,10 +4081,14 @@ Status DBImpl::BackgroundCompaction(bool* made_progress,
   } else {
     status = error_handler_.GetBGError();
     // If we get here, it means a hard error happened after this compaction
-    // was scheduled by MaybeScheduleFlushOrCompaction(), but before it got
-    // a chance to execute. Since we didn't pop a cfd from the compaction
-    // queue, increment unscheduled_compactions_
-    unscheduled_compactions_++;
+    // was scheduled by MaybeScheduleFlushOrCompaction(), but before it got a
+    // chance to execute. Since non-prepicked work consumed an unscheduled
+    // compaction credit without popping a cfd from the compaction queue,
+    // restore the credit here. Prepicked work was scheduled directly and did
+    // not consume unscheduled_compactions_.
+    if (!is_prepicked) {
+      unscheduled_compactions_++;
+    }
   }
 
   if (!status.ok()) {
@@ -4255,6 +4309,8 @@ Status DBImpl::BackgroundCompaction(bool* made_progress,
     for (const auto& f : *c->inputs(0)) {
       c->edit()->DeleteFile(c->level(), f->fd.GetNumber());
     }
+    NotifyOnCompactionPreCommit(c->column_family_data(), c.get(), status,
+                                compaction_job_stats, job_context->job_id);
     status = versions_->LogAndApply(
         c->column_family_data(), read_options, write_options, c->edit(),
         &mutex_, directories_.GetDbDir(),
@@ -4473,6 +4529,8 @@ Status DBImpl::BackgroundCompaction(bool* made_progress,
         ++out_file_metadata_it;
       }
 
+      NotifyOnCompactionPreCommit(c->column_family_data(), c.get(), status,
+                                  compaction_job_stats, job_context->job_id);
       status = versions_->LogAndApply(
           c->column_family_data(), read_options, write_options, c->edit(),
           &mutex_, directories_.GetDbDir(),
@@ -4549,8 +4607,12 @@ Status DBImpl::BackgroundCompaction(bool* made_progress,
     // Perform the trivial move
     size_t moved_files = 0;
     size_t moved_bytes = 0;
-    status = PerformTrivialMove(*c.get(), log_buffer, compaction_released,
-                                moved_files, moved_bytes);
+    // Populate the edit first so that CompactionJobInfo has output file
+    // info when OnCompactionPreCommit fires.
+    PrepareTrivialMoveEdit(*c.get(), log_buffer, moved_files, moved_bytes);
+    NotifyOnCompactionPreCommit(c->column_family_data(), c.get(), status,
+                                compaction_job_stats, job_context->job_id);
+    status = CommitTrivialMove(*c.get(), compaction_released);
     io_s = versions_->io_status();
     InstallSuperVersionAndScheduleWork(
         c->column_family_data(), job_context->superversion_contexts.data());
@@ -4673,6 +4735,10 @@ Status DBImpl::BackgroundCompaction(bool* made_progress,
       ReleaseOptionsFileNumber(min_options_file_number_elem);
     }
 
+    // Fire OnCompactionPreCommit before compaction_job.Install runs
+    // ReleaseCompactionFiles in its manifest write callback.
+    NotifyOnCompactionPreCommit(c->column_family_data(), c.get(), status,
+                                compaction_job_stats, job_context->job_id);
     status = compaction_job.Install(&compaction_released);
     io_s = compaction_job.io_status();
     if (status.ok()) {
@@ -4692,6 +4758,14 @@ Status DBImpl::BackgroundCompaction(bool* made_progress,
 
   if (c != nullptr) {
     if (!compaction_released) {
+      // Safety net: if any of the paths above did not fire the
+      // OnCompactionPreCommit notification (e.g. status was not OK and
+      // an early-out skipped the LogAndApply / Install), still fire the
+      // notify here while files are about to be released. The notify is
+      // idempotent and is a no-op if it has already fired (or if Begin never
+      // fired).
+      NotifyOnCompactionPreCommit(c->column_family_data(), c.get(), status,
+                                  compaction_job_stats, job_context->job_id);
       c->ReleaseCompactionFiles(status);
     } else {
 #ifndef NDEBUG

db/db_impl/db_impl_files.cc

@@ -833,6 +833,7 @@ void DBImpl::PurgeObsoleteFiles(JobContext& state, bool schedule_only) {
   }
   wal_manager_.PurgeObsoleteWALFiles();
   LogFlush(immutable_db_options_.info_log);
+  TEST_SYNC_POINT("DBImpl::PurgeObsoleteFiles:BeforePendingPurgeFinished");
   InstrumentedMutexLock l(&mutex_);
   --pending_purge_obsolete_files_;
   assert(pending_purge_obsolete_files_ >= 0);

db/db_impl/db_impl_open.cc

@@ -21,6 +21,7 @@
 #include "file/writable_file_writer.h"
 #include "logging/logging.h"
 #include "monitoring/persistent_stats_history.h"
+#include "monitoring/statistics_impl.h"
 #include "monitoring/thread_status_util.h"
 #include "options/options_helper.h"
 #include "rocksdb/options.h"
@@ -122,6 +123,16 @@ DBOptions SanitizeOptions(const std::string& dbname, const DBOptions& src,
     result.recycle_log_file_num = 0;
   }
 
+  if (result.async_wal_precreate && result.recycle_log_file_num != 0) {
+    // Async WAL precreation reserves a future log number, while WAL recycling
+    // chooses from old WAL files. Keep the recycling behavior and disable only
+    // the async optimization.
+    result.async_wal_precreate = false;
+    ROCKS_LOG_WARN(result.info_log,
+                   "async_wal_precreate is disabled since "
+                   "recycle_log_file_num is non-zero");
+  }
+
   if (result.db_paths.size() == 0) {
     result.db_paths.emplace_back(dbname, std::numeric_limits<uint64_t>::max());
   } else if (result.wal_dir.empty()) {
@@ -244,6 +255,17 @@ Status DBImpl::ValidateOptions(const DBOptions& db_options) {
         "then direct I/O reads (use_direct_reads) must be disabled. ");
   }
 
+  if (db_options.allow_mmap_reads &&
+      db_options.use_direct_io_for_compaction_reads) {
+    // mmap reads and direct I/O share the same EnvOptions field, so enabling
+    // both would try to mmap and O_DIRECT the same reads. Reject it here rather
+    // than tripping a lower-level assert.
+    return Status::NotSupported(
+        "If memory mapped reads (allow_mmap_reads) are enabled "
+        "then compaction-only direct I/O reads "
+        "(use_direct_io_for_compaction_reads) must be disabled. ");
+  }
+
   if (db_options.allow_mmap_writes &&
       db_options.use_direct_io_for_flush_and_compaction) {
     return Status::NotSupported(
@@ -495,7 +517,8 @@ Status DBImpl::Recover(
       std::unique_ptr<FSRandomAccessFile> idfile;
       FileOptions customized_fs(file_options_);
       customized_fs.use_direct_reads |=
-          immutable_db_options_.use_direct_io_for_flush_and_compaction;
+          immutable_db_options_.use_direct_io_for_flush_and_compaction ||
+          immutable_db_options_.use_direct_io_for_compaction_reads;
       const std::string& fname =
           manifest_path.empty() ? current_fname : manifest_path;
       s = fs_->NewRandomAccessFile(fname, customized_fs, &idfile, nullptr);
@@ -799,9 +822,19 @@ Status DBImpl::Recover(
 
     if (!wal_files.empty()) {
       if (error_if_wal_file_exists) {
-        return Status::Corruption(
-            "The db was opened in readonly mode with error_if_wal_file_exists"
-            "flag but a WAL file already exists");
+        for (const auto& wal_file : wal_files) {
+          uint64_t bytes;
+          s = env_->GetFileSize(wal_file.second, &bytes);
+          if (!s.ok()) {
+            return s;
+          }
+          if (bytes > 0) {
+            return Status::Corruption(
+                "The db was opened in readonly mode with "
+                "error_if_wal_file_exists flag but a non-empty WAL file "
+                "already exists");
+          }
+        }
       } else if (error_if_data_exists_in_wals) {
         for (auto& wal_file : wal_files) {
           uint64_t bytes;
@@ -2213,6 +2246,16 @@ Status DBImpl::WriteLevel0TableForRecovery(int job_id, ColumnFamilyData* cfd,
 
   constexpr int level = 0;
 
+  if (s.ok() && has_output) {
+    // Before publishing any recovery edit, make the new SST's directory entry
+    // durable. Reused MANIFESTs have no CURRENT update, and fresh MANIFEST
+    // publication should not be relied on to order the recovered SST entry.
+    s = GetDataDir(cfd, meta.fd.GetPathId())
+            ->FsyncWithDirOptions(
+                IOOptions(), nullptr,
+                DirFsyncOptions(DirFsyncOptions::FsyncReason::kNewFileSynced));
+  }
+
   if (s.ok() && has_output) {
     edit->AddFile(level, meta.fd.GetNumber(), meta.fd.GetPathId(),
                   meta.fd.GetFileSize(), meta.smallest, meta.largest,
@@ -2377,16 +2420,17 @@ Status DB::OpenAndTrimHistory(
   return s;
 }
 
-IOStatus DBImpl::CreateWAL(const WriteOptions& write_options,
-                           uint64_t log_file_num, uint64_t recycle_log_number,
-                           size_t preallocate_block_size,
-                           const PredecessorWALInfo& predecessor_wal_info,
-                           log::Writer** new_log) {
+IOStatus DBImpl::CreateWALWriter(const DBOptions& db_options,
+                                 uint64_t log_file_num,
+                                 uint64_t recycle_log_number,
+                                 size_t preallocate_block_size,
+                                 UnpublishedWAL* new_wal) {
+  assert(new_wal);
+  assert(new_wal->log_number == 0);
+  assert(new_wal->writer == nullptr);
   IOStatus io_s;
   std::unique_ptr<FSWritableFile> lfile;
 
-  DBOptions db_options =
-      BuildDBOptions(immutable_db_options_, mutable_db_options_);
   FileOptions opt_file_options =
       fs_->OptimizeForLogWrite(file_options_, db_options);
   opt_file_options.write_hint = CalculateWALWriteHint();
@@ -2424,21 +2468,162 @@ IOStatus DBImpl::CreateWAL(const WriteOptions& write_options,
         Histograms::HISTOGRAM_ENUM_MAX /* hist_type */, listeners, nullptr,
         tmp_set.Contains(FileType::kWalFile),
         tmp_set.Contains(FileType::kWalFile)));
-    *new_log = new log::Writer(std::move(file_writer), log_file_num,
-                               immutable_db_options_.recycle_log_file_num > 0,
-                               immutable_db_options_.manual_wal_flush,
-                               immutable_db_options_.wal_compression,
-                               immutable_db_options_.track_and_verify_wals);
-    io_s = (*new_log)->AddCompressionTypeRecord(write_options);
-    if (io_s.ok()) {
-      io_s = (*new_log)->MaybeAddPredecessorWALInfo(write_options,
-                                                    predecessor_wal_info);
-    }
+    new_wal->log_number = log_file_num;
+    new_wal->writer = std::make_unique<log::Writer>(
+        std::move(file_writer), log_file_num,
+        immutable_db_options_.recycle_log_file_num > 0,
+        immutable_db_options_.manual_wal_flush,
+        immutable_db_options_.wal_compression,
+        immutable_db_options_.track_and_verify_wals);
   }
 
   return io_s;
 }
 
+IOStatus DBImpl::StartWALFile(const WriteOptions& write_options,
+                              const PredecessorWALInfo& predecessor_wal_info,
+                              log::Writer* new_log) {
+  assert(new_log);
+  IOStatus io_s = new_log->AddCompressionTypeRecord(write_options);
+  TEST_SYNC_POINT_CALLBACK("DBImpl::StartWALFile:AfterCompressionTypeRecord",
+                           &io_s);
+  if (io_s.ok()) {
+    io_s = new_log->MaybeAddPredecessorWALInfo(write_options,
+                                               predecessor_wal_info);
+  }
+  return io_s;
+}
+
+IOStatus DBImpl::CreateWAL(const DBOptions& db_options,
+                           const WriteOptions& write_options,
+                           uint64_t log_file_num, uint64_t recycle_log_number,
+                           size_t preallocate_block_size,
+                           const PredecessorWALInfo& predecessor_wal_info,
+                           log::Writer** new_log) {
+  assert(new_log);
+  assert(*new_log == nullptr);
+  UnpublishedWAL wal;
+  IOStatus io_s = CreateWALWriter(db_options, log_file_num, recycle_log_number,
+                                  preallocate_block_size, &wal);
+  if (io_s.ok()) {
+    io_s = StartWALFile(write_options, predecessor_wal_info, wal.writer.get());
+  }
+  if (io_s.ok()) {
+    *new_log = wal.writer.release();
+  }
+  return io_s;
+}
+
+bool DBImpl::AsyncWALPrecreateEnabled() const {
+  return immutable_db_options_.async_wal_precreate &&
+         immutable_db_options_.recycle_log_file_num == 0;
+}
+
+struct AsyncWALPrecreateContext {
+  DBImpl* db = nullptr;
+  DBOptions db_options;
+  uint64_t log_number = 0;
+  size_t preallocate_block_size = 0;
+};
+
+void DBImpl::MaybeScheduleAsyncWALPrecreate(size_t preallocate_block_size) {
+  mutex_.AssertHeld();
+
+  if (!AsyncWALPrecreateEnabled() ||
+      shutting_down_.load(std::memory_order_acquire) ||
+      error_handler_.IsBGWorkStopped() ||
+      async_wal_precreate_state_ != AsyncWALPrecreateState::kNotScheduled) {
+    return;
+  }
+
+  async_wal_precreate_wal_.Reset();
+  async_wal_precreate_wal_.log_number = versions_->NewFileNumber();
+  async_wal_precreate_state_ = AsyncWALPrecreateState::kScheduled;
+
+  auto* ctx = new AsyncWALPrecreateContext();
+  ctx->db = this;
+  ctx->db_options = BuildDBOptions(immutable_db_options_, mutable_db_options_);
+  ctx->log_number = async_wal_precreate_wal_.log_number;
+  ctx->preallocate_block_size = preallocate_block_size;
+  env_->Schedule(&DBImpl::BGWorkAsyncWALPrecreate, ctx, Env::Priority::HIGH,
+                 nullptr);
+}
+
+DBImpl::UnpublishedWAL DBImpl::WaitForAsyncWALPrecreate() {
+  mutex_.AssertHeld();
+  UnpublishedWAL result;
+
+  if (!AsyncWALPrecreateEnabled()) {
+    return result;
+  }
+
+  TEST_SYNC_POINT("DBImpl::WaitForAsyncWALPrecreate:Begin");
+  bool waited = false;
+  uint64_t wait_start_micros = 0;
+  while (async_wal_precreate_state_ == AsyncWALPrecreateState::kScheduled) {
+    TEST_SYNC_POINT("DBImpl::WaitForAsyncWALPrecreate:BeforeWait");
+    if (!waited) {
+      waited = true;
+      wait_start_micros = immutable_db_options_.clock->NowMicros();
+      RecordTick(stats_, WAL_PRECREATE_WAITED);
+    }
+    bg_cv_.Wait();
+  }
+  if (waited) {
+    RecordTick(stats_, WAL_PRECREATE_WAIT_MICROS,
+               immutable_db_options_.clock->NowMicros() - wait_start_micros);
+  }
+
+  if (async_wal_precreate_state_ == AsyncWALPrecreateState::kReady) {
+    result = std::move(async_wal_precreate_wal_);
+    async_wal_precreate_state_ = AsyncWALPrecreateState::kNotScheduled;
+    RecordTick(stats_, WAL_PRECREATE_HIT);
+    return result;
+  }
+
+  assert(async_wal_precreate_state_ == AsyncWALPrecreateState::kNotScheduled);
+  RecordTick(stats_, WAL_PRECREATE_MISS);
+  return result;
+}
+
+void DBImpl::BGWorkAsyncWALPrecreate(void* arg) {
+  TEST_SYNC_POINT("DBImpl::BGWorkAsyncWALPrecreate:Start");
+  std::unique_ptr<AsyncWALPrecreateContext> ctx(
+      static_cast<AsyncWALPrecreateContext*>(arg));
+  DBImpl* db = ctx->db;
+
+  UnpublishedWAL new_wal;
+  IOStatus io_s = db->CreateWALWriter(ctx->db_options, ctx->log_number,
+                                      0 /* recycle_log_number */,
+                                      ctx->preallocate_block_size, &new_wal);
+  if (!io_s.ok()) {
+    RecordTick(db->stats_, WAL_PRECREATE_FAILED);
+    ROCKS_LOG_WARN(db->immutable_db_options_.info_log,
+                   "Async WAL precreate failed for #%" PRIu64 ": %s",
+                   ctx->log_number, io_s.ToString().c_str());
+  }
+
+  TEST_SYNC_POINT("DBImpl::BGWorkAsyncWALPrecreate:BeforePublish");
+  TEST_SYNC_POINT("DBImpl::BGWorkAsyncWALPrecreate:Publish");
+  {
+    InstrumentedMutexLock l(&db->mutex_);
+    if (db->async_wal_precreate_state_ == AsyncWALPrecreateState::kScheduled &&
+        db->async_wal_precreate_wal_.log_number == ctx->log_number) {
+      if (io_s.ok()) {
+        db->async_wal_precreate_wal_ = std::move(new_wal);
+        db->async_wal_precreate_state_ = AsyncWALPrecreateState::kReady;
+      } else {
+        // Async WAL precreation is best-effort. Clear the slot so foreground
+        // rotation falls back to normal synchronous WAL creation.
+        db->async_wal_precreate_wal_.Reset();
+        db->async_wal_precreate_state_ = AsyncWALPrecreateState::kNotScheduled;
+      }
+    }
+    db->bg_cv_.SignalAll();
+  }
+  TEST_SYNC_POINT("DBImpl::BGWorkAsyncWALPrecreate:Done");
+}
+
 void DBImpl::TrackExistingDataFiles(
     const std::vector<std::string>& existing_data_files) {
   TrackOrUntrackFiles(existing_data_files, /*track=*/true);
@@ -2534,8 +2719,10 @@ Status DBImpl::Open(const DBOptions& db_options, const std::string& dbname,
     // created during DB open with predecessor WALs from previous DB session due
     // to `avoid_flush_during_recovery == true`. This can protect the last WAL
     // recovered.
-    s = impl->CreateWAL(write_options, new_log_number, 0 /*recycle_log_number*/,
-                        preallocate_block_size,
+    const DBOptions db_options_snapshot =
+        BuildDBOptions(impl->immutable_db_options_, impl->mutable_db_options_);
+    s = impl->CreateWAL(db_options_snapshot, write_options, new_log_number,
+                        0 /*recycle_log_number*/, preallocate_block_size,
                         PredecessorWALInfo() /* predecessor_wal_info */,
                         &new_log);
     if (s.ok()) {
@@ -2740,6 +2927,8 @@ Status DBImpl::Open(const DBOptions& db_options, const std::string& dbname,
     if (impl->immutable_db_options_.open_files_async) {
       impl->ScheduleAsyncFileOpening();
     }
+    impl->MaybeScheduleAsyncWALPrecreate(
+        impl->GetWalPreallocateBlockSize(max_write_buffer_size));
     impl->mutex_.Unlock();
   }
 

db/db_impl/db_impl_readonly.h

@@ -134,6 +134,19 @@ class DBImplReadOnly : public DBImpl {
     return Status::NotSupported("Not supported operation in read only mode.");
   }
 
+  using DB::PrepareFileIngestion;
+  Status PrepareFileIngestion(
+      const std::vector<IngestExternalFileArg>& /*args*/,
+      std::unique_ptr<FileIngestionHandle>* /*handle*/) override {
+    return Status::NotSupported("Not supported operation in read only mode.");
+  }
+
+  using DB::CommitFileIngestionHandles;
+  Status CommitFileIngestionHandles(
+      std::vector<std::unique_ptr<FileIngestionHandle>> /*handles*/) override {
+    return Status::NotSupported("Not supported operation in read only mode.");
+  }
+
   using DB::CreateColumnFamilyWithImport;
   Status CreateColumnFamilyWithImport(
       const ColumnFamilyOptions& /*options*/,

db/db_impl/db_impl_secondary.cc

@@ -105,10 +105,22 @@ Status DBImplSecondary::FindNewLogNumbers(std::vector<uint64_t>* logs) {
     return s;
   }
 
-  // if log_readers_ is non-empty, it means we have applied all logs with log
-  // numbers smaller than the smallest log in log_readers_, so there is no
-  // need to pass these logs to RecoverLogFiles
-  uint64_t log_number_min = 0;
+  // Drop readers only after MANIFEST replay says the WAL is obsolete. The
+  // primary can create a higher-number empty WAL before it stops appending to a
+  // lower-number current WAL, so seeing a higher-number WAL is not sufficient.
+  const uint64_t min_log_number_to_keep = versions_->min_log_number_to_keep();
+  for (auto iter = log_readers_.begin(); iter != log_readers_.end();) {
+    if (iter->first < min_log_number_to_keep) {
+      iter = log_readers_.erase(iter);
+    } else {
+      break;
+    }
+  }
+
+  // If log_readers_ is non-empty, it means we have applied all logs with log
+  // numbers smaller than the smallest retained log reader. Otherwise, manifest
+  // recovery tells us which older logs can be skipped.
+  uint64_t log_number_min = min_log_number_to_keep;
   if (!log_readers_.empty()) {
     log_number_min = log_readers_.begin()->first;
   }
@@ -321,12 +333,6 @@ Status DBImplSecondary::RecoverLogFiles(
       return status;
     }
   }
-  // remove logreaders from map after successfully recovering the WAL
-  if (log_readers_.size() > 1) {
-    auto erase_iter = log_readers_.begin();
-    std::advance(erase_iter, log_readers_.size() - 1);
-    log_readers_.erase(log_readers_.begin(), erase_iter);
-  }
   return status;
 }
 

db/db_impl/db_impl_secondary.h

@@ -239,6 +239,19 @@ class DBImplSecondary : public DBImpl {
     return Status::NotSupported("Not supported operation in secondary mode.");
   }
 
+  using DB::PrepareFileIngestion;
+  Status PrepareFileIngestion(
+      const std::vector<IngestExternalFileArg>& /*args*/,
+      std::unique_ptr<FileIngestionHandle>* /*handle*/) override {
+    return Status::NotSupported("Not supported operation in secondary mode.");
+  }
+
+  using DB::CommitFileIngestionHandles;
+  Status CommitFileIngestionHandles(
+      std::vector<std::unique_ptr<FileIngestionHandle>> /*handles*/) override {
+    return Status::NotSupported("Not supported operation in secondary mode.");
+  }
+
   // Try to catch up with the primary by reading as much as possible from the
   // log files until there is nothing more to read or encounters an error. If
   // the amount of information in the log files to process is huge, this

db/db_impl/db_impl_write.cc

@@ -16,6 +16,7 @@
 #include "db/error_handler.h"
 #include "db/event_helpers.h"
 #include "db/wide/wide_columns_helper.h"
+#include "file/filename.h"
 #include "logging/logging.h"
 #include "memtable/wbwi_memtable.h"
 #include "monitoring/perf_context_imp.h"
@@ -3107,6 +3108,7 @@ Status DBImpl::SwitchMemtable(ColumnFamilyData* cfd, WriteContext* context,
   const WriteOptions write_options;
 
   log::Writer* new_log = nullptr;
+  UnpublishedWAL prepared_wal;
   MemTable* new_mem = nullptr;
   IOStatus io_s;
 
@@ -3134,8 +3136,16 @@ Status DBImpl::SwitchMemtable(ColumnFamilyData* cfd, WriteContext* context,
       !wal_recycle_files_.empty() && IsFileDeletionsEnabled()) {
     recycle_log_number = wal_recycle_files_.front();
   }
-  uint64_t new_log_number =
-      creating_new_log ? versions_->NewFileNumber() : cur_wal_number_;
+  uint64_t new_log_number = cur_wal_number_;
+  if (creating_new_log) {
+    prepared_wal = WaitForAsyncWALPrecreate();
+    if (prepared_wal.writer) {
+      new_log_number = prepared_wal.log_number;
+    } else {
+      assert(prepared_wal.log_number == 0);
+      new_log_number = versions_->NewFileNumber();
+    }
+  }
   // For use outside of holding DB mutex
   const MutableCFOptions mutable_cf_options_copy =
       cfd->GetLatestMutableCFOptions();
@@ -3156,6 +3166,8 @@ Status DBImpl::SwitchMemtable(ColumnFamilyData* cfd, WriteContext* context,
   // Log this later after lock release. It may be outdated, e.g., if background
   // flush happens before logging, but that should be ok.
   int num_imm_unflushed = cfd->imm()->NumNotFlushed();
+  const DBOptions db_options_snapshot =
+      BuildDBOptions(immutable_db_options_, mutable_db_options_);
   const auto preallocate_block_size =
       GetWalPreallocateBlockSize(mutable_cf_options_copy.write_buffer_size);
   mutex_.Unlock();
@@ -3171,13 +3183,35 @@ Status DBImpl::SwitchMemtable(ColumnFamilyData* cfd, WriteContext* context,
     wal_write_mutex_.Unlock();
     // TODO: Write buffer size passed in should be max of all CF's instead
     // of mutable_cf_options.write_buffer_size.
-    io_s = CreateWAL(write_options, new_log_number, recycle_log_number,
-                     preallocate_block_size, info, &new_log);
+    if (prepared_wal.writer) {
+      io_s = StartWALFile(write_options, info, prepared_wal.writer.get());
+      if (io_s.ok()) {
+        new_log = prepared_wal.writer.release();
+      }
+    } else {
+      io_s =
+          CreateWAL(db_options_snapshot, write_options, new_log_number,
+                    recycle_log_number, preallocate_block_size, info, &new_log);
+    }
     TEST_SYNC_POINT_CALLBACK("DBImpl::SwitchMemtable:AfterCreateWAL", &io_s);
     if (s.ok()) {
       s = io_s;
     }
   }
+  if (!s.ok() && prepared_wal.writer) {
+    const uint64_t prepared_wal_number = prepared_wal.log_number;
+    const std::string prepared_wal_fname =
+        LogFileName(immutable_db_options_.GetWalDir(), prepared_wal_number);
+    prepared_wal.Reset();
+    Status delete_s = env_->DeleteFile(prepared_wal_fname);
+    if (!delete_s.ok() && !delete_s.IsNotFound()) {
+      ROCKS_LOG_WARN(immutable_db_options_.info_log,
+                     "Failed to delete unstarted async precreated WAL #%" PRIu64
+                     ": %s",
+                     prepared_wal_number, delete_s.ToString().c_str());
+    }
+    delete_s.PermitUncheckedError();
+  }
   if (s.ok()) {
     // FIXME: from the comment for GetEarliestSequenceNumber(), any key with
     //  seqno >= earliest_seqno should be in this or later memtable. This means
@@ -3359,6 +3393,7 @@ Status DBImpl::SwitchMemtable(ColumnFamilyData* cfd, WriteContext* context,
   new_mem->Ref();
   cfd->SetMemtable(new_mem);
   InstallSuperVersionAndScheduleWork(cfd, &context->superversion_context);
+  MaybeScheduleAsyncWALPrecreate(preallocate_block_size);
 
   // Notify client that memtable is sealed, now that we have successfully
   // installed a new memtable

db/db_iter.cc

@@ -67,8 +67,9 @@ DBIter::DBIter(Env* _env, const ReadOptions& read_options,
                const MutableCFOptions& mutable_cf_options,
                const Comparator* cmp, InternalIterator* iter,
                const Version* version, SequenceNumber s, bool arena_mode,
-               ReadCallback* read_callback, ColumnFamilyHandleImpl* cfh,
-               bool expose_blob_index, ReadOnlyMemTable* active_mem)
+               ReadCallback* read_callback, DBImpl* db_impl,
+               ColumnFamilyData* cfd, bool expose_blob_index,
+               ReadOnlyMemTable* active_mem)
     : prefix_extractor_(mutable_cf_options.prefix_extractor.get()),
       env_(_env),
       clock_(ioptions.clock),
@@ -76,21 +77,26 @@ DBIter::DBIter(Env* _env, const ReadOptions& read_options,
       user_comparator_(cmp),
       merge_operator_(ioptions.merge_operator.get()),
       iter_(iter),
-      blob_state_(
-          version, read_options.read_tier, read_options.verify_checksums,
-          read_options.fill_cache, read_options.io_activity,
-          cfh ? cfh->cfd()->blob_file_cache() : nullptr,
-          cfh != nullptr && cfh->cfd()->blob_partition_manager() != nullptr),
+      blob_state_(version, read_options.read_tier,
+                  read_options.verify_checksums, read_options.fill_cache,
+                  read_options.io_activity,
+                  cfd ? cfd->blob_file_cache() : nullptr,
+                  cfd != nullptr && cfd->blob_partition_manager() != nullptr),
       read_callback_(read_callback),
       sequence_(s),
-      value_columns_state_(version, read_options, cfh),
+      value_columns_state_(version, read_options, cfd),
       statistics_(ioptions.stats),
       max_skip_(mutable_cf_options.max_sequential_skip_in_iterations),
       max_skippable_internal_keys_(read_options.max_skippable_internal_keys),
       num_internal_keys_skipped_(0),
       iterate_lower_bound_(read_options.iterate_lower_bound),
       iterate_upper_bound_(read_options.iterate_upper_bound),
-      cfh_(cfh),
+      trace_db_(db_impl),
+      trace_cf_id_(cfd != nullptr ? cfd->GetID() : 0),
+      has_trace_state_(db_impl != nullptr && cfd != nullptr),
+      allow_blob_write_path_fallback_(cfd != nullptr &&
+                                      cfd->blob_partition_manager() != nullptr),
+      ingest_sst_lock_(cfd != nullptr ? &cfd->GetIngestSstLock() : nullptr),
       timestamp_ub_(read_options.timestamp),
       timestamp_lb_(read_options.iter_start_ts),
       timestamp_size_(timestamp_ub_ ? timestamp_ub_->size() : 0),
@@ -305,10 +311,8 @@ bool DBIter::SetValueAndColumnsFromBlobImpl(const Slice& user_key,
   // Keep the non-BDW iterator path on the pre-existing Version::GetBlob()
   // fast path. Only enable the direct-write fallback when this CF actually
   // has a write-path partition manager.
-  const bool allow_write_path_fallback =
-      cfh_ != nullptr && cfh_->cfd()->blob_partition_manager() != nullptr;
   const Status s = blob_state_.mut()->reader.RetrieveAndSetBlobValue(
-      user_key, blob_index, allow_write_path_fallback);
+      user_key, blob_index, allow_blob_write_path_fallback_);
   if (!s.ok()) {
     status_ = s;
     valid_ = false;
@@ -1480,7 +1484,9 @@ bool DBIter::FindValueForCurrentKeyUsingSeek() {
     if (timestamp_lb_ != nullptr) {
       saved_key_.SetInternalKey(ikey);
     } else {
-      saved_key_.SetUserKey(ikey.user_key);
+      saved_key_.SetUserKey(
+          ikey.user_key,
+          !pin_thru_lifetime_ || !iter_.iter()->IsKeyPinned() /* copy */);
     }
 
     valid_ = true;
@@ -1615,10 +1621,8 @@ bool DBIter::MergeWithBlobBaseValue(const Slice& blob_index,
     return false;
   }
 
-  const bool allow_write_path_fallback =
-      cfh_ != nullptr && cfh_->cfd()->blob_partition_manager() != nullptr;
   const Status s = blob_state_.mut()->reader.RetrieveAndSetBlobValue(
-      user_key, blob_index, allow_write_path_fallback);
+      user_key, blob_index, allow_blob_write_path_fallback_);
   if (!s.ok()) {
     status_ = s;
     valid_ = false;
@@ -1814,10 +1818,10 @@ void DBIter::MaybeInsertRangeTombstone(const Slice& end_key) {
     }
   }
 
-  assert(cfh_ != nullptr);
+  assert(ingest_sst_lock_ != nullptr);
   if (active_mem_->AddLogicallyRedundantRangeTombstone(
           insert_seq, range_tomb_first_key_.GetUserKey(), end_key,
-          cfh_->cfd()->GetIngestSstLock())) {
+          *ingest_sst_lock_)) {
     RecordTick(statistics_, READ_PATH_RANGE_TOMBSTONES_INSERTED);
     ROCKS_LOG_DEBUG(logger_,
                     "Inserted range tombstone [%s, %s) @ seq %" PRIu64
@@ -1957,10 +1961,10 @@ Status DBIter::ValidateScanOptions(const MultiScanArgs& multiscan_opts) const {
   return Status::OK();
 }
 
-void DBIter::Prepare(const MultiScanArgs& scan_opts) {
+Status DBIter::SetScanOptionsForPrepare(const MultiScanArgs& scan_opts) {
   status_ = ValidateScanOptions(scan_opts);
   if (!status_.ok()) {
-    return;
+    return status_;
   }
   std::optional<MultiScanArgs> new_scan_opts;
   new_scan_opts.emplace(scan_opts);
@@ -1973,14 +1977,27 @@ void DBIter::Prepare(const MultiScanArgs& scan_opts) {
   if (!scan_opts_.value().io_dispatcher) {
     scan_opts_->io_dispatcher.reset(NewIODispatcher());
   }
+  return Status::OK();
+}
 
-  if (!scan_opts.empty()) {
+void DBIter::PrepareInternalChildren() {
+  if (!scan_opts_.has_value() || !status_.ok()) {
+    return;
+  }
+
+  if (scan_opts_.value().HasBoundedScanRanges()) {
     iter_.Prepare(&scan_opts_.value());
   } else {
     iter_.Prepare(nullptr);
   }
 }
 
+void DBIter::Prepare(const MultiScanArgs& scan_opts) {
+  if (SetScanOptionsForPrepare(scan_opts).ok()) {
+    PrepareInternalChildren();
+  }
+}
+
 void DBIter::Seek(const Slice& target) {
   PERF_COUNTER_ADD(iter_seek_count, 1);
   PERF_CPU_TIMER_GUARD(iter_seek_cpu_nanos, clock_);
@@ -2027,7 +2044,7 @@ void DBIter::Seek(const Slice& target) {
     scan_index_++;
   }
 
-  if (cfh_ != nullptr) {
+  if (has_trace_state_) {
     // TODO: What do we do if this returns an error?
     Slice lower_bound, upper_bound;
     if (iterate_lower_bound_ != nullptr) {
@@ -2040,9 +2057,7 @@ void DBIter::Seek(const Slice& target) {
     } else {
       upper_bound = Slice("");
     }
-    cfh_->db()
-        ->TraceIteratorSeek(cfh_->cfd()->GetID(), target, lower_bound,
-                            upper_bound)
+    trace_db_->TraceIteratorSeek(trace_cf_id_, target, lower_bound, upper_bound)
         .PermitUncheckedError();
   }
 
@@ -2095,7 +2110,7 @@ void DBIter::SeekForPrev(const Slice& target) {
   PERF_CPU_TIMER_GUARD(iter_seek_cpu_nanos, clock_);
   StopWatch sw(clock_, statistics_, DB_SEEK);
 
-  if (cfh_ != nullptr) {
+  if (has_trace_state_) {
     // TODO: What do we do if this returns an error?
     Slice lower_bound, upper_bound;
     if (iterate_lower_bound_ != nullptr) {
@@ -2108,8 +2123,8 @@ void DBIter::SeekForPrev(const Slice& target) {
     } else {
       upper_bound = Slice("");
     }
-    cfh_->db()
-        ->TraceIteratorSeekForPrev(cfh_->cfd()->GetID(), target, lower_bound,
+    trace_db_
+        ->TraceIteratorSeekForPrev(trace_cf_id_, target, lower_bound,
                                    upper_bound)
         .PermitUncheckedError();
   }

db/db_iter.h

@@ -32,6 +32,9 @@
 namespace ROCKSDB_NAMESPACE {
 class BlobFileCache;
 class Version;
+namespace port {
+class RWMutex;
+}
 
 // This file declares the factory functions of DBIter, in its original form
 // or a wrapped form with class ArenaWrappedDBIter, which is defined here.
@@ -83,14 +86,19 @@ class DBIter final : public Iterator {
                          ReadCallback* read_callback,
                          ReadOnlyMemTable* active_mem,
                          ColumnFamilyHandleImpl* cfh = nullptr,
-                         bool expose_blob_index = false,
-                         Arena* arena = nullptr) {
+                         bool expose_blob_index = false, Arena* arena = nullptr,
+                         DBImpl* db_impl = nullptr,
+                         ColumnFamilyData* cfd = nullptr) {
+    if (cfh != nullptr) {
+      db_impl = cfh->db();
+      cfd = cfh->cfd();
+    }
     void* mem = arena ? arena->AllocateAligned(sizeof(DBIter))
                       : operator new(sizeof(DBIter));
     DBIter* db_iter = new (mem)
         DBIter(env, read_options, ioptions, mutable_cf_options,
                user_key_comparator, internal_iter, version, sequence, arena,
-               read_callback, cfh, expose_blob_index, active_mem);
+               read_callback, db_impl, cfd, expose_blob_index, active_mem);
     return db_iter;
   }
 
@@ -207,7 +215,7 @@ class DBIter final : public Iterator {
   }
 
   Status status() const override {
-    if (status_.ok()) {
+    if (status_.ok() && iter_.iter() != nullptr) {
       return iter_.status();
     } else {
       assert(!valid_);
@@ -248,19 +256,22 @@ class DBIter final : public Iterator {
     iter_.SetRangeDelReadSeqno(s);
   }
   void set_valid(bool v) { valid_ = v; }
+  void set_status(Status s) { status_ = std::move(s); }
 
   bool PrepareValue() override;
 
   void Prepare(const MultiScanArgs& scan_opts) override;
   Status ValidateScanOptions(const MultiScanArgs& multiscan_opts) const;
+  Status SetScanOptionsForPrepare(const MultiScanArgs& scan_opts);
+  void PrepareInternalChildren();
 
  private:
   DBIter(Env* _env, const ReadOptions& read_options,
          const ImmutableOptions& ioptions,
          const MutableCFOptions& mutable_cf_options, const Comparator* cmp,
          InternalIterator* iter, const Version* version, SequenceNumber s,
-         bool arena_mode, ReadCallback* read_callback,
-         ColumnFamilyHandleImpl* cfh, bool expose_blob_index,
+         bool arena_mode, ReadCallback* read_callback, DBImpl* db_impl,
+         ColumnFamilyData* cfd, bool expose_blob_index,
          ReadOnlyMemTable* active_mem);
 
   class BlobReader {
@@ -318,13 +329,12 @@ class DBIter final : public Iterator {
   class ValueColumnsState {
    public:
     ValueColumnsState(const Version* version, const ReadOptions& read_options,
-                      ColumnFamilyHandleImpl* cfh)
+                      ColumnFamilyData* cfd)
         : entity_blob_resolver_(
               version, read_options.read_tier, read_options.verify_checksums,
               read_options.fill_cache, read_options.io_activity,
-              cfh ? cfh->cfd()->blob_file_cache() : nullptr,
-              cfh != nullptr &&
-                  cfh->cfd()->blob_partition_manager() != nullptr) {}
+              cfd ? cfd->blob_file_cache() : nullptr,
+              cfd != nullptr && cfd->blob_partition_manager() != nullptr) {}
 
     Slice& value() { return value_; }
     const Slice& value() const { return value_; }
@@ -706,7 +716,11 @@ class DBIter final : public Iterator {
   MergeContext merge_context_;
   LocalStatistics local_stats_;
   PinnedIteratorsManager pinned_iters_mgr_;
-  ColumnFamilyHandleImpl* cfh_;
+  DBImpl* trace_db_;
+  uint32_t trace_cf_id_;
+  bool has_trace_state_;
+  bool allow_blob_write_path_fallback_;
+  port::RWMutex* ingest_sst_lock_;
   const Slice* const timestamp_ub_;
   const Slice* const timestamp_lb_;
   const size_t timestamp_size_;

db/db_iter_test.cc

@@ -216,12 +216,19 @@ class TestIterator : public InternalIterator {
   bool IsKeyPinned() const override { return true; }
   bool IsValuePinned() const override { return true; }
 
+  void Prepare(const MultiScanArgs* /*scan_opts*/) override {
+    ++prepare_call_count_;
+  }
+
+  size_t prepare_call_count() const { return prepare_call_count_; }
+
  private:
   bool initialized_;
   bool valid_;
   size_t sequence_number_;
   size_t iter_;
   size_t steps_ = 0;
+  size_t prepare_call_count_ = 0;
 
   InternalKeyComparator cmp;
   std::vector<std::pair<std::string, std::string>> data_;
@@ -243,6 +250,53 @@ class DBIteratorTest : public testing::Test {
   DBIteratorTest() : env_(Env::Default()) {}
 };
 
+TEST_F(DBIteratorTest, PrepareForwardsValidatedScanRanges) {
+  Options options;
+  ImmutableOptions ioptions = ImmutableOptions(options);
+  MutableCFOptions mutable_cf_options = MutableCFOptions(options);
+  TestIterator* internal_iter = new TestIterator(BytewiseComparator());
+  internal_iter->AddPut("a", "val_a");
+  internal_iter->AddPut("b", "val_b");
+  internal_iter->Finish();
+
+  ReadOptions ro;
+  std::unique_ptr<Iterator> db_iter(DBIter::NewIter(
+      env_, ro, ioptions, mutable_cf_options, BytewiseComparator(),
+      internal_iter, nullptr /* version */, 10 /* sequence */,
+      nullptr /* read_callback */, /*active_mem=*/nullptr));
+
+  MultiScanArgs scan_opts(BytewiseComparator());
+  scan_opts.insert("a", "c");
+
+  db_iter->Prepare(scan_opts);
+  ASSERT_OK(db_iter->status());
+  EXPECT_EQ(internal_iter->prepare_call_count(), 1);
+}
+
+TEST_F(DBIteratorTest, PrepareDoesNotForwardInvalidScanRanges) {
+  Options options;
+  ImmutableOptions ioptions = ImmutableOptions(options);
+  MutableCFOptions mutable_cf_options = MutableCFOptions(options);
+  TestIterator* internal_iter = new TestIterator(BytewiseComparator());
+  internal_iter->AddPut("a", "val_a");
+  internal_iter->AddPut("b", "val_b");
+  internal_iter->Finish();
+
+  ReadOptions ro;
+  std::unique_ptr<Iterator> db_iter(DBIter::NewIter(
+      env_, ro, ioptions, mutable_cf_options, BytewiseComparator(),
+      internal_iter, nullptr /* version */, 10 /* sequence */,
+      nullptr /* read_callback */, /*active_mem=*/nullptr));
+
+  MultiScanArgs scan_opts(BytewiseComparator());
+  scan_opts.insert("b", "d");
+  scan_opts.insert("a", "c");
+
+  db_iter->Prepare(scan_opts);
+  EXPECT_TRUE(db_iter->status().IsInvalidArgument());
+  EXPECT_EQ(internal_iter->prepare_call_count(), 0);
+}
+
 TEST_F(DBIteratorTest, DBIteratorPrevNext) {
   Options options;
   ImmutableOptions ioptions = ImmutableOptions(options);

db/db_iterator_test.cc

@@ -10,6 +10,9 @@
 #include <functional>
 #include <iomanip>
 #include <iostream>
+#include <memory>
+#include <utility>
+#include <vector>
 
 #include "db/arena_wrapped_db_iter.h"
 #include "db/db_iter.h"
@@ -87,6 +90,392 @@ TEST_F(DBIteratorBaseTest, APICallsWithPerfContext) {
   delete iter;
 }
 
+TEST_F(DBIteratorBaseTest, PrepareWithMultiScanPrunesNonIntersectingFiles) {
+  Options options = CurrentOptions();
+  options.disable_auto_compactions = true;
+  DestroyAndReopen(options);
+
+  ASSERT_OK(Put("a", "va"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("z", "vz"));
+  ASSERT_OK(Flush());
+  ASSERT_EQ(2, NumTableFilesAtLevel(0));
+
+  int table_iterators_created = 0;
+  int files_added = 0;
+  int block_based_iterators = 0;
+  int level_iterators = 0;
+  SyncPoint::GetInstance()->SetCallBack(
+      "TableCache::NewIterator::BeforeFindTable",
+      [&](void* /*arg*/) { ++table_iterators_created; });
+  SyncPoint::GetInstance()->SetCallBack(
+      "Version::AddIteratorsForLevel:AddedFile",
+      [&](void* /*arg*/) { ++files_added; });
+  SyncPoint::GetInstance()->SetCallBack(
+      "Version::AddIteratorsForLevel:IteratorType", [&](void* arg) {
+        auto* iterator_type = static_cast<std::pair<bool, bool>*>(arg);
+        if (iterator_type->first) {
+          ++block_based_iterators;
+        }
+        if (iterator_type->second) {
+          ++level_iterators;
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  MultiScanArgs scan_opts(BytewiseComparator());
+  scan_opts.insert(Slice("a"), Slice("b"));
+  Slice upper_bound("b");
+  ReadOptions read_options;
+  read_options.iterate_upper_bound = &upper_bound;
+  std::unique_ptr<Iterator> iter(
+      db_->NewIterator(read_options, db_->DefaultColumnFamily()));
+
+  ASSERT_EQ(0, table_iterators_created);
+  ASSERT_EQ(0, files_added);
+  ASSERT_EQ(0, block_based_iterators);
+  ASSERT_EQ(0, level_iterators);
+  iter->Prepare(scan_opts);
+  ASSERT_EQ(1, table_iterators_created);
+  ASSERT_EQ(1, files_added);
+  ASSERT_EQ(1, block_based_iterators);
+  ASSERT_EQ(0, level_iterators);
+  std::vector<std::string> keys;
+  for (iter->Seek("a"); iter->Valid(); iter->Next()) {
+    keys.push_back(iter->key().ToString());
+  }
+  ASSERT_EQ(keys, std::vector<std::string>({"a"}));
+  ASSERT_FALSE(iter->Valid());
+  ASSERT_OK(iter->status());
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+TEST_F(DBIteratorBaseTest, PrepareWithMultiScanPrunesNonIntersectingLevels) {
+  Options options = CurrentOptions();
+  options.disable_auto_compactions = true;
+  DestroyAndReopen(options);
+
+  ASSERT_OK(Put("z", "vz"));
+  ASSERT_OK(Flush());
+  MoveFilesToLevel(2);
+  ASSERT_OK(Put("a", "va"));
+  ASSERT_OK(Flush());
+  MoveFilesToLevel(1);
+  ASSERT_OK(Put("m", "vm"));
+  ASSERT_OK(Flush());
+  MoveFilesToLevel(1);
+
+  int table_iterators_created = 0;
+  int files_added = 0;
+  int block_based_iterators = 0;
+  int level_iterators = 0;
+  SyncPoint::GetInstance()->SetCallBack(
+      "TableCache::NewIterator::BeforeFindTable",
+      [&](void* /*arg*/) { ++table_iterators_created; });
+  SyncPoint::GetInstance()->SetCallBack(
+      "Version::AddIteratorsForLevel:AddedFile",
+      [&](void* /*arg*/) { ++files_added; });
+  SyncPoint::GetInstance()->SetCallBack(
+      "Version::AddIteratorsForLevel:IteratorType", [&](void* arg) {
+        auto* iterator_type = static_cast<std::pair<bool, bool>*>(arg);
+        if (iterator_type->first) {
+          ++block_based_iterators;
+        }
+        if (iterator_type->second) {
+          ++level_iterators;
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  MultiScanArgs scan_opts(BytewiseComparator());
+  scan_opts.insert(Slice("a"), Slice("n"));
+  Slice upper_bound("n");
+  ReadOptions read_options;
+  read_options.iterate_upper_bound = &upper_bound;
+  std::unique_ptr<Iterator> iter(
+      db_->NewIterator(read_options, db_->DefaultColumnFamily()));
+
+  ASSERT_EQ(0, table_iterators_created);
+  ASSERT_EQ(0, files_added);
+  ASSERT_EQ(0, block_based_iterators);
+  ASSERT_EQ(0, level_iterators);
+  iter->Prepare(scan_opts);
+  ASSERT_EQ(0, table_iterators_created);
+  ASSERT_EQ(2, files_added);
+  ASSERT_EQ(0, block_based_iterators);
+  ASSERT_EQ(1, level_iterators);
+  std::vector<std::string> keys;
+  for (iter->Seek("a"); iter->Valid(); iter->Next()) {
+    keys.push_back(iter->key().ToString());
+  }
+  ASSERT_EQ(keys, std::vector<std::string>({"a", "m"}));
+  ASSERT_FALSE(iter->Valid());
+  ASSERT_OK(iter->status());
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+TEST_F(DBIteratorBaseTest, PrepareWithMultiScanAllowsSingleUnboundedRange) {
+  Options options = CurrentOptions();
+  options.disable_auto_compactions = true;
+  DestroyAndReopen(options);
+
+  ASSERT_OK(Put("a", "va"));
+  ASSERT_OK(Put("b", "vb"));
+  ASSERT_OK(Put("c", "vc"));
+  ASSERT_OK(Flush());
+  MoveFilesToLevel(1);
+
+  MultiScanArgs scan_opts(BytewiseComparator());
+  scan_opts.insert(Slice("b"));
+  ReadOptions read_options;
+  std::unique_ptr<Iterator> iter(
+      db_->NewIterator(read_options, db_->DefaultColumnFamily()));
+
+  iter->Prepare(scan_opts);
+  std::vector<std::string> keys;
+  for (iter->Seek("b"); iter->Valid(); iter->Next()) {
+    keys.push_back(iter->key().ToString());
+  }
+  ASSERT_EQ(keys, std::vector<std::string>({"b", "c"}));
+  ASSERT_FALSE(iter->Valid());
+  ASSERT_OK(iter->status());
+}
+
+TEST_F(DBIteratorBaseTest, PrepareWithMultiScanRejectsRepeatedPrepare) {
+  Options options = CurrentOptions();
+  options.disable_auto_compactions = true;
+  DestroyAndReopen(options);
+
+  ASSERT_OK(Put("a", "va"));
+  ASSERT_OK(Flush());
+  MoveFilesToLevel(1);
+
+  MultiScanArgs scan_opts(BytewiseComparator());
+  scan_opts.insert(Slice("a"), Slice("b"));
+  Slice upper_bound("b");
+  ReadOptions read_options;
+  read_options.iterate_upper_bound = &upper_bound;
+  std::unique_ptr<Iterator> iter(
+      db_->NewIterator(read_options, db_->DefaultColumnFamily()));
+
+  iter->Prepare(scan_opts);
+  ASSERT_OK(iter->status());
+  iter->Prepare(scan_opts);
+  ASSERT_NOK(iter->status());
+}
+
+TEST_F(DBIteratorBaseTest, PrepareWithMultiScanDedupsMultipleRangesInSameFile) {
+  Options options = CurrentOptions();
+  options.disable_auto_compactions = true;
+  DestroyAndReopen(options);
+
+  ASSERT_OK(Put("a", "va"));
+  ASSERT_OK(Put("m", "vm"));
+  ASSERT_OK(Put("z", "vz"));
+  ASSERT_OK(Flush());
+  MoveFilesToLevel(1);
+
+  int table_iterators_created = 0;
+  int files_added = 0;
+  int block_based_iterators = 0;
+  int level_iterators = 0;
+  SyncPoint::GetInstance()->SetCallBack(
+      "TableCache::NewIterator::BeforeFindTable",
+      [&](void* /*arg*/) { ++table_iterators_created; });
+  SyncPoint::GetInstance()->SetCallBack(
+      "Version::AddIteratorsForLevel:AddedFile",
+      [&](void* /*arg*/) { ++files_added; });
+  SyncPoint::GetInstance()->SetCallBack(
+      "Version::AddIteratorsForLevel:IteratorType", [&](void* arg) {
+        auto* iterator_type = static_cast<std::pair<bool, bool>*>(arg);
+        if (iterator_type->first) {
+          ++block_based_iterators;
+        }
+        if (iterator_type->second) {
+          ++level_iterators;
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  MultiScanArgs scan_opts(BytewiseComparator());
+  scan_opts.insert(Slice("a"), Slice("b"));
+  scan_opts.insert(Slice("m"), Slice("n"));
+  Slice upper_bound("b");
+  ReadOptions read_options;
+  read_options.iterate_upper_bound = &upper_bound;
+  std::unique_ptr<Iterator> iter(
+      db_->NewIterator(read_options, db_->DefaultColumnFamily()));
+
+  ASSERT_EQ(0, table_iterators_created);
+  ASSERT_EQ(0, files_added);
+  ASSERT_EQ(0, block_based_iterators);
+  ASSERT_EQ(0, level_iterators);
+  iter->Prepare(scan_opts);
+  ASSERT_EQ(1, table_iterators_created);
+  ASSERT_EQ(1, files_added);
+  ASSERT_EQ(1, block_based_iterators);
+  ASSERT_EQ(0, level_iterators);
+
+  std::vector<std::string> keys;
+  for (iter->Seek("a"); iter->Valid(); iter->Next()) {
+    keys.push_back(iter->key().ToString());
+  }
+  ASSERT_OK(iter->status());
+  ASSERT_EQ(keys, std::vector<std::string>({"a"}));
+
+  upper_bound = "n";
+  for (iter->Seek("m"); iter->Valid(); iter->Next()) {
+    keys.push_back(iter->key().ToString());
+  }
+  ASSERT_EQ(keys, std::vector<std::string>({"a", "m"}));
+  ASSERT_FALSE(iter->Valid());
+  ASSERT_OK(iter->status());
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+TEST_F(DBIteratorBaseTest, PrepareWithMultiScanPrunesOverlappingL0Files) {
+  Options options = CurrentOptions();
+  options.disable_auto_compactions = true;
+  DestroyAndReopen(options);
+
+  ASSERT_OK(Put("a", "va"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("m", "vm"));
+  ASSERT_OK(Flush());
+  ASSERT_EQ(2, NumTableFilesAtLevel(0));
+
+  int table_iterators_created = 0;
+  int files_added = 0;
+  int block_based_iterators = 0;
+  int level_iterators = 0;
+  SyncPoint::GetInstance()->SetCallBack(
+      "TableCache::NewIterator::BeforeFindTable",
+      [&](void* /*arg*/) { ++table_iterators_created; });
+  SyncPoint::GetInstance()->SetCallBack(
+      "Version::AddIteratorsForLevel:AddedFile",
+      [&](void* /*arg*/) { ++files_added; });
+  SyncPoint::GetInstance()->SetCallBack(
+      "Version::AddIteratorsForLevel:IteratorType", [&](void* arg) {
+        auto* iterator_type = static_cast<std::pair<bool, bool>*>(arg);
+        if (iterator_type->first) {
+          ++block_based_iterators;
+        }
+        if (iterator_type->second) {
+          ++level_iterators;
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  MultiScanArgs scan_opts(BytewiseComparator());
+  scan_opts.insert(Slice("a"), Slice("n"));
+  Slice upper_bound("n");
+  ReadOptions read_options;
+  read_options.iterate_upper_bound = &upper_bound;
+  std::unique_ptr<Iterator> iter(
+      db_->NewIterator(read_options, db_->DefaultColumnFamily()));
+
+  ASSERT_EQ(0, table_iterators_created);
+  ASSERT_EQ(0, files_added);
+  ASSERT_EQ(0, block_based_iterators);
+  ASSERT_EQ(0, level_iterators);
+  iter->Prepare(scan_opts);
+  ASSERT_EQ(2, table_iterators_created);
+  ASSERT_EQ(2, files_added);
+  ASSERT_EQ(2, block_based_iterators);
+  ASSERT_EQ(0, level_iterators);
+  std::vector<std::string> keys;
+  for (iter->Seek("a"); iter->Valid(); iter->Next()) {
+    keys.push_back(iter->key().ToString());
+  }
+  ASSERT_EQ(keys, std::vector<std::string>({"a", "m"}));
+  ASSERT_FALSE(iter->Valid());
+  ASSERT_OK(iter->status());
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+TEST_F(DBIteratorBaseTest, PrepareWithMultiScanPrunesNonIntersectingMemTables) {
+  Options options = CurrentOptions();
+  options.disable_auto_compactions = true;
+  DestroyAndReopen(options);
+
+  ASSERT_OK(Put("a", "va"));
+  ASSERT_OK(Flush());
+  MoveFilesToLevel(1);
+
+  ASSERT_OK(db_->PauseBackgroundWork());
+  ASSERT_OK(Put("z_imm", "vz"));
+  ASSERT_OK(dbfull()->TEST_SwitchMemtable());
+  ASSERT_OK(Put("z_mem", "vz"));
+
+  int table_iterators_created = 0;
+  int files_added = 0;
+  int block_based_iterators = 0;
+  int level_iterators = 0;
+  int merging_iterators = 0;
+  SyncPoint::GetInstance()->SetCallBack(
+      "TableCache::NewIterator::BeforeFindTable",
+      [&](void* /*arg*/) { ++table_iterators_created; });
+  SyncPoint::GetInstance()->SetCallBack(
+      "Version::AddIteratorsForLevel:AddedFile",
+      [&](void* /*arg*/) { ++files_added; });
+  SyncPoint::GetInstance()->SetCallBack(
+      "Version::AddIteratorsForLevel:IteratorType", [&](void* arg) {
+        auto* iterator_type = static_cast<std::pair<bool, bool>*>(arg);
+        if (iterator_type->first) {
+          ++block_based_iterators;
+        }
+        if (iterator_type->second) {
+          ++level_iterators;
+        }
+      });
+  SyncPoint::GetInstance()->SetCallBack(
+      "MergeIteratorBuilder::Finish:UseMergingIterator", [&](void* arg) {
+        if (*static_cast<bool*>(arg)) {
+          ++merging_iterators;
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  MultiScanArgs scan_opts(BytewiseComparator());
+  scan_opts.insert(Slice("a"), Slice("b"));
+  Slice upper_bound("b");
+  ReadOptions read_options;
+  read_options.iterate_upper_bound = &upper_bound;
+  std::unique_ptr<Iterator> iter(
+      db_->NewIterator(read_options, db_->DefaultColumnFamily()));
+
+  ASSERT_EQ(0, table_iterators_created);
+  ASSERT_EQ(0, files_added);
+  ASSERT_EQ(0, block_based_iterators);
+  ASSERT_EQ(0, level_iterators);
+  ASSERT_EQ(0, merging_iterators);
+  iter->Prepare(scan_opts);
+  ASSERT_EQ(1, table_iterators_created);
+  ASSERT_EQ(1, files_added);
+  ASSERT_EQ(1, block_based_iterators);
+  ASSERT_EQ(0, level_iterators);
+  ASSERT_EQ(0, merging_iterators);
+  std::vector<std::string> keys;
+  for (iter->Seek("a"); iter->Valid(); iter->Next()) {
+    keys.push_back(iter->key().ToString());
+  }
+  ASSERT_EQ(keys, std::vector<std::string>({"a"}));
+  ASSERT_FALSE(iter->Valid());
+  ASSERT_OK(iter->status());
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+  ASSERT_OK(db_->ContinueBackgroundWork());
+}
+
 // Test param:
 //   bool: whether to pass read_callback to NewIterator().
 class DBIteratorTest : public DBIteratorBaseTest,
@@ -945,6 +1334,68 @@ TEST_P(DBIteratorTest, IteratorDeleteAfterCfDelete) {
   delete iter;
 }
 
+TEST_P(DBIteratorTest, IteratorSeekAfterCfDelete) {
+  CreateAndReopenWithCF({"pikachu"}, CurrentOptions());
+
+  ASSERT_OK(Put(1, "foo", "delete-cf-then-seek-iter"));
+  ASSERT_OK(Put(1, "hello", "value2"));
+
+  ColumnFamilyHandle* cf = handles_[1];
+  ReadOptions ro;
+
+  auto* iter = db_->NewIterator(ro, cf);
+
+  // Delete the CF handle before the lazy iterator tree is materialized.
+  EXPECT_OK(db_->DestroyColumnFamilyHandle(cf));
+  handles_.erase(std::begin(handles_) + 1);
+
+  iter->Seek("foo");
+  ASSERT_EQ(IterStatus(iter), "foo->delete-cf-then-seek-iter");
+  iter->SeekForPrev("hello");
+  ASSERT_EQ(IterStatus(iter), "hello->value2");
+  iter->SeekToFirst();
+  ASSERT_EQ(IterStatus(iter), "foo->delete-cf-then-seek-iter");
+  iter->Next();
+  ASSERT_EQ(IterStatus(iter), "hello->value2");
+  delete iter;
+}
+
+TEST_P(DBIteratorTest, IteratorAutoRefreshAfterCfDeleteBeforeLazyInit) {
+  CreateAndReopenWithCF({"pikachu"}, CurrentOptions());
+
+  ASSERT_OK(Put(1, "foo", "delete-cf-then-auto-refresh"));
+  ASSERT_OK(Flush(1));
+
+  ColumnFamilyHandle* cf = handles_[1];
+  const Snapshot* snapshot = db_->GetSnapshot();
+
+  ReadOptions ro;
+  ro.snapshot = snapshot;
+  ro.auto_refresh_iterator_with_snapshot = true;
+
+  auto* iter = db_->NewIterator(ro, cf);
+
+  ASSERT_OK(Put(1, "zzz", "after-snapshot"));
+  ASSERT_OK(Flush(1));
+
+  // Delete the CF handle before the lazy iterator tree is materialized. The
+  // following operations force auto-refresh to acquire a newer SuperVersion.
+  EXPECT_OK(db_->DestroyColumnFamilyHandle(cf));
+  handles_.erase(std::begin(handles_) + 1);
+
+  iter->Seek("foo");
+  ASSERT_EQ(IterStatus(iter), "foo->delete-cf-then-auto-refresh");
+  ASSERT_OK(iter->status());
+  iter->SeekForPrev("foo");
+  ASSERT_EQ(IterStatus(iter), "foo->delete-cf-then-auto-refresh");
+  ASSERT_OK(iter->status());
+  iter->Next();
+  ASSERT_OK(iter->status());
+
+  delete iter;
+  db_->ReleaseSnapshot(snapshot);
+}
+
 TEST_P(DBIteratorTest, IteratorDeleteAfterCfDrop) {
   CreateAndReopenWithCF({"pikachu"}, CurrentOptions());
 
@@ -966,6 +1417,30 @@ TEST_P(DBIteratorTest, IteratorDeleteAfterCfDrop) {
   delete iter;
 }
 
+TEST_P(DBIteratorTest, IteratorSeekAfterCfDrop) {
+  CreateAndReopenWithCF({"pikachu"}, CurrentOptions());
+
+  ASSERT_OK(Put(1, "foo", "drop-cf-then-seek-iter"));
+
+  ReadOptions ro;
+  ColumnFamilyHandle* cf = handles_[1];
+
+  auto* iter = db_->NewIterator(ro, cf);
+
+  // Drop and delete the CF before the lazy iterator tree is materialized.
+  EXPECT_OK(db_->DropColumnFamily(cf));
+  EXPECT_OK(db_->DestroyColumnFamilyHandle(cf));
+  handles_.erase(std::begin(handles_) + 1);
+
+  iter->Seek("foo");
+  ASSERT_EQ(IterStatus(iter), "foo->drop-cf-then-seek-iter");
+  iter->SeekForPrev("foo");
+  ASSERT_EQ(IterStatus(iter), "foo->drop-cf-then-seek-iter");
+  iter->SeekToFirst();
+  ASSERT_EQ(IterStatus(iter), "foo->drop-cf-then-seek-iter");
+  delete iter;
+}
+
 // SetOptions not defined in ROCKSDB LITE
 TEST_P(DBIteratorTest, DBIteratorBoundTest) {
   Options options = CurrentOptions();
@@ -2277,6 +2752,10 @@ TEST_P(DBIteratorTest, ReadAhead) {
   options.env = env_;
   options.disable_auto_compactions = true;
   options.write_buffer_size = 4 << 20;
+  // Pin compression so the readahead byte thresholds below don't depend on the
+  // default compression type. kNoCompression keeps the SST files large enough
+  // to exercise readahead and is always available.
+  options.compression = kNoCompression;
   options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
   BlockBasedTableOptions table_options;
   table_options.block_size = 1024;
@@ -2693,6 +3172,8 @@ TEST_P(DBIteratorTest, CreationFailure) {
 
   Iterator* iter = NewIterator(ReadOptions());
   ASSERT_FALSE(iter->Valid());
+  iter->SeekToFirst();
+  ASSERT_FALSE(iter->Valid());
   ASSERT_TRUE(iter->status().IsCorruption());
   delete iter;
 }
@@ -4395,6 +4876,71 @@ TEST_P(DBMultiScanIteratorTest, RangeAcrossFiles) {
   iter.reset();
 }
 
+TEST_P(DBMultiScanIteratorTest, SortedRangesSkipIODispatcherSort) {
+  auto options = CurrentOptions();
+  options.target_file_size_base = 100 << 10;
+  options.compaction_style = kCompactionStyleUniversal;
+  options.num_levels = 50;
+  options.compression = kNoCompression;
+  DestroyAndReopen(options);
+
+  auto key = [](int i) {
+    std::stringstream ss;
+    ss << "k" << std::setw(5) << std::setfill('0') << i;
+    return ss.str();
+  };
+
+  auto rnd = Random::GetTLSInstance();
+  for (int i = 0; i < 100; ++i) {
+    ASSERT_OK(Put(key(i), rnd->RandomString(2 << 10)));
+  }
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(db_->CompactRange({}, nullptr, nullptr));
+  ASSERT_EQ(2, NumTableFilesAtLevel(49));
+
+  int sort_count = 0;
+  SyncPoint::GetInstance()->SetCallBack(
+      "IODispatcherImpl::SubmitJob:SortBlockHandles",
+      [&](void* /*arg*/) { ++sort_count; });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  auto tracking_dispatcher = std::make_shared<TrackingIODispatcher>();
+  std::vector<std::string> key_ranges({key(10), key(90)});
+  MultiScanArgs scan_options(BytewiseComparator());
+  scan_options.io_dispatcher = tracking_dispatcher;
+  scan_options.use_async_io = false;
+  scan_options.insert(key_ranges[0], key_ranges[1]);
+
+  ReadOptions ro;
+  ro.fill_cache = GetParam();
+  ColumnFamilyHandle* cfh = dbfull()->DefaultColumnFamily();
+  std::unique_ptr<MultiScan> iter =
+      dbfull()->NewMultiScan(ro, cfh, scan_options);
+
+  try {
+    int i = 10;
+    for (auto range : *iter) {
+      for (auto it : range) {
+        ASSERT_EQ(it.first.ToString(), key(i));
+        ++i;
+      }
+    }
+    ASSERT_EQ(i, 90);
+  } catch (MultiScanException& ex) {
+    ASSERT_NOK(ex.status());
+    std::cerr << "Iterator returned status " << ex.what();
+    abort();
+  } catch (std::logic_error& ex) {
+    std::cerr << "Iterator returned logic error " << ex.what();
+    abort();
+  }
+
+  ASSERT_GT(tracking_dispatcher->GetReadSets().size(), 0);
+  EXPECT_EQ(sort_count, 0);
+  iter.reset();
+}
+
 TEST_P(DBMultiScanIteratorTest, FailureTest) {
   auto options = CurrentOptions();
   options.compression = kNoCompression;
@@ -7181,6 +7727,52 @@ TEST_P(ReadPathRangeTombstoneTest,
   db_->ReleaseSnapshot(snap);
 }
 
+// SeekForPrev uses the reverse path: PrevInternal -> FindValueForCurrentKey.
+// When a key has more versions than max_sequential_skip_in_iterations,
+// FindValueForCurrentKey delegates to FindValueForCurrentKeyUsingSeek which
+// re-seeks forward to the newest visible version. That function must preserve
+// key pinning by passing the correct `copy` parameter to
+// saved_key_.SetUserKey().
+TEST_P(DBIteratorTest, SeekForPrevKeyPinnedWithManyVersions) {
+  Options options = CurrentOptions();
+  options.max_sequential_skip_in_iterations = 8;
+  options.statistics = CreateDBStatistics();
+  DestroyAndReopen(options);
+
+  // 20 versions > max_sequential_skip (8), forcing the reseek path.
+  // No flush: keeps all versions alive in the memtable.
+  for (int i = 0; i < 20; i++) {
+    ASSERT_OK(Put("key1", "value" + std::to_string(i)));
+  }
+
+  ReadOptions ro;
+  ro.pin_data = true;
+
+  std::unique_ptr<Iterator> iter(NewIterator(ro));
+  std::string prop_value;
+
+  // Seek (forward) correctly reports key as pinned.
+  iter->Seek("key1");
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_OK(iter->GetProperty("rocksdb.iterator.is-key-pinned", &prop_value));
+  ASSERT_EQ("1", prop_value);
+
+  uint64_t reseeks_before =
+      options.statistics->getTickerCount(NUMBER_OF_RESEEKS_IN_ITERATION);
+
+  iter->SeekForPrev("key1");
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("key1", iter->key().ToString());
+
+  // Confirm FindValueForCurrentKeyUsingSeek was actually taken.
+  uint64_t reseeks_after =
+      options.statistics->getTickerCount(NUMBER_OF_RESEEKS_IN_ITERATION);
+  ASSERT_GT(reseeks_after, reseeks_before);
+
+  ASSERT_OK(iter->GetProperty("rocksdb.iterator.is-key-pinned", &prop_value));
+  ASSERT_EQ("1", prop_value);
+}
+
 }  // namespace ROCKSDB_NAMESPACE
 
 int main(int argc, char** argv) {

db/db_memtable_test.cc

@@ -359,7 +359,7 @@ class DBMemTableTestForSeek : public DBMemTableTest,
 TEST_P(DBMemTableTestForSeek, IntegrityChecks) {
   // Validate key corruption could be detected during seek.
   // We insert many keys into skiplist. Then we corrupt the each key one at a
-  // time. With memtable_veirfy_per_key_checksum_on_seek enabled, when the
+  // time. With memtable_verify_per_key_checksum_on_seek enabled, when the
   // corrupted key is searched, the checksum of every key visited during the
   // seek is validated. It will report data corruption. Otherwise seek returns
   // not found.
@@ -367,7 +367,7 @@ TEST_P(DBMemTableTestForSeek, IntegrityChecks) {
   Options options = CurrentOptions();
   options.allow_data_in_errors = allow_data_in_error;
   options.paranoid_memory_checks = std::get<1>(GetParam());
-  options.memtable_veirfy_per_key_checksum_on_seek = std::get<2>(GetParam());
+  options.memtable_verify_per_key_checksum_on_seek = std::get<2>(GetParam());
   options.memtable_protection_bytes_per_key = 8;
   DestroyAndReopen(options);
 
@@ -410,7 +410,7 @@ TEST_P(DBMemTableTestForSeek, IntegrityChecks) {
   ASSERT_EQ(raw_data_pointer.size(), key_count);
 
   bool enable_key_validation_on_seek =
-      options.memtable_veirfy_per_key_checksum_on_seek;
+      options.memtable_verify_per_key_checksum_on_seek;
 
   // For each key, corrupt it, validate corruption is detected correctly, then
   // revert it.

db/db_open_with_config_test.cc

@@ -151,8 +151,8 @@ TEST_F(DBOpenWithConfigTest, OpenWithComprehensiveConfig) {
   // memtable_op_scan_flush_trigger=0
   options.memtable_op_scan_flush_trigger = 0;
 
-  // memtable_veirfy_per_key_checksum_on_seek=false
-  options.memtable_veirfy_per_key_checksum_on_seek = false;
+  // memtable_verify_per_key_checksum_on_seek=false
+  options.memtable_verify_per_key_checksum_on_seek = false;
 
   // Level and compaction settings
   // num_levels=50

db/db_options_test.cc

@@ -13,6 +13,7 @@
 #include "db/column_family.h"
 #include "db/db_impl/db_impl.h"
 #include "db/db_test_util.h"
+#include "env/mock_env.h"
 #include "options/options_helper.h"
 #include "port/stack_trace.h"
 #include "rocksdb/cache.h"
@@ -102,6 +103,40 @@ TEST_F(DBOptionsTest, ImmutableTrackAndVerifyWalsInManifest) {
   ASSERT_FALSE(s.ok());
 }
 
+TEST_F(DBOptionsTest, ImmutableAsyncWalPrecreate) {
+  Options options;
+  options.env = env_;
+  options.async_wal_precreate = true;
+
+  ImmutableDBOptions db_options(options);
+  ASSERT_TRUE(db_options.async_wal_precreate);
+
+  Reopen(options);
+  ASSERT_TRUE(dbfull()->GetDBOptions().async_wal_precreate);
+
+  Status s = dbfull()->SetDBOptions({{"async_wal_precreate", "false"}});
+  ASSERT_FALSE(s.ok());
+}
+
+TEST_F(DBOptionsTest, SanitizeAsyncWalPrecreateWithWalRecycle) {
+  // This checks the option-level contract for the incompatible WAL paths:
+  // SanitizeOptions() preserves WAL recycling and disables only the
+  // opportunistic async precreation path, so the immutable options reflect
+  // effective behavior.
+  Options options;
+  options.env = env_;
+  options.async_wal_precreate = true;
+  options.recycle_log_file_num = 1;
+
+  Options sanitized_options = SanitizeOptions(dbname_, options);
+  ASSERT_FALSE(sanitized_options.async_wal_precreate);
+  ASSERT_EQ(1U, sanitized_options.recycle_log_file_num);
+
+  options.recycle_log_file_num = 0;
+  sanitized_options = SanitizeOptions(dbname_, options);
+  ASSERT_TRUE(sanitized_options.async_wal_precreate);
+}
+
 TEST_F(DBOptionsTest, ImmutableVerifySstUniqueIdInManifest) {
   Options options;
   options.env = env_;
@@ -1729,6 +1764,171 @@ TEST_F(DBOptionsTest, SetOptionsMultipleColumnFamilies) {
   ASSERT_TRUE(dbfull()->GetOptions(handles_[2]).disable_auto_compactions);
 }
 
+// Confirms the default value and serialization/parse round-trip of the new
+// option. No DB open required; exercises only the options-metadata layer.
+TEST_F(DBOptionsTest, UseDirectIoForCompactionReadsRoundTrip) {
+  // Default value must remain false to preserve existing semantics.
+  ASSERT_FALSE(DBOptions().use_direct_io_for_compaction_reads);
+
+  DBOptions parsed;
+  ConfigOptions config_options;
+  ASSERT_OK(GetDBOptionsFromString(config_options, DBOptions(),
+                                   "use_direct_io_for_compaction_reads=true",
+                                   &parsed));
+  ASSERT_TRUE(parsed.use_direct_io_for_compaction_reads);
+  ASSERT_OK(GetDBOptionsFromString(config_options, DBOptions(),
+                                   "use_direct_io_for_compaction_reads=false",
+                                   &parsed));
+  ASSERT_FALSE(parsed.use_direct_io_for_compaction_reads);
+}
+
+// Validates that Open rejects the documented incompatible combination.
+TEST_F(DBOptionsTest, UseDirectIoForCompactionReadsValidation) {
+  // mmap_reads + use_direct_io_for_compaction_reads is rejected at Open
+  // time, the same way mmap_reads + use_direct_reads has always been
+  // rejected.
+  Options bad_options = CurrentOptions();
+  bad_options.create_if_missing = true;
+  bad_options.allow_mmap_reads = true;
+  bad_options.use_direct_io_for_compaction_reads = true;
+  Status bad_status = TryReopen(bad_options);
+  ASSERT_TRUE(bad_status.IsNotSupported()) << bad_status.ToString();
+}
+
+// Confirms the option is plumbed all the way to the live DB's options API:
+// after opening with the flag set, GetDBOptions() reports it back. Only runs
+// when the environment supports direct I/O.
+TEST_F(DBOptionsTest, UseDirectIoForCompactionReadsLiveReopen) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.use_direct_io_for_compaction_reads = true;
+  // Use a buffered user-read setup so the new flag is the one doing the work.
+  options.use_direct_reads = false;
+  options.use_direct_io_for_flush_and_compaction = false;
+  Status s = TryReopen(options);
+  if (s.IsNotSupported() || s.IsInvalidArgument()) {
+    ROCKSDB_GTEST_BYPASS(
+        "Direct I/O not supported in this test environment; live reopen "
+        "cannot be exercised.");
+    return;
+  }
+  ASSERT_OK(s);
+  ASSERT_TRUE(dbfull()->GetDBOptions().use_direct_io_for_compaction_reads);
+  Close();
+}
+
+TEST_F(DBOptionsTest, UseDirectIoForCompactionReadsUnsupportedFileSystem) {
+  auto fs = std::make_shared<MockFileSystem>(Env::Default()->GetSystemClock(),
+                                             /*supports_direct_io=*/false);
+  std::unique_ptr<Env> mock_env = NewCompositeEnv(fs);
+
+  Options options = CurrentOptions();
+  options.env = mock_env.get();
+  options.create_if_missing = true;
+  options.use_direct_reads = false;
+  options.use_direct_io_for_flush_and_compaction = false;
+  options.use_direct_io_for_compaction_reads = true;
+
+  Status s = TryReopen(options);
+  ASSERT_TRUE(s.IsInvalidArgument()) << s.ToString();
+  ASSERT_NE(s.ToString().find("Direct I/O is not supported"), std::string::npos)
+      << s.ToString();
+}
+
+TEST_F(
+    DBOptionsTest,
+    UseDirectIoForCompactionReadsSetDBOptionsKeepsVerificationReadsBuffered) {
+  auto fs = std::make_shared<MockFileSystem>(Env::Default()->GetSystemClock(),
+                                             /*supports_direct_io=*/true);
+  std::unique_ptr<Env> mock_env = NewCompositeEnv(fs);
+
+  Options options = CurrentOptions();
+  options.env = mock_env.get();
+  options.create_if_missing = true;
+  options.use_direct_reads = false;
+  options.use_direct_io_for_flush_and_compaction = false;
+  options.use_direct_io_for_compaction_reads = true;
+
+  ASSERT_OK(TryReopen(options));
+
+  bool observed = false;
+  FileOptions observed_file_options;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::SetDBOptions:FileOptionsForCompaction", [&](void* arg) {
+        observed = true;
+        observed_file_options = *static_cast<FileOptions*>(arg);
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(dbfull()->SetDBOptions({{"bytes_per_sync", "1024"}}));
+  ASSERT_TRUE(observed);
+  ASSERT_FALSE(observed_file_options.use_direct_reads);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+  Close();
+}
+
+// Exercises the FileSystem::OptimizeForCompactionTableRead and
+// OptimizeForBlobFileRead helpers directly, asserting that the compaction-only
+// flag does not affect shared FileSystem hooks.
+TEST_F(DBOptionsTest, OptimizeForCompactionTableReadUsesGlobalDirectReadsOnly) {
+  FileOptions in_opts;
+  in_opts.use_direct_reads = false;
+
+  {
+    Options check_options;
+    check_options.use_direct_reads = false;
+    check_options.use_direct_io_for_compaction_reads = true;
+    check_options.use_direct_io_for_flush_and_compaction = false;
+    ImmutableDBOptions immutable(check_options);
+    FileOptions sst_read =
+        env_->GetFileSystem()->OptimizeForCompactionTableRead(in_opts,
+                                                              immutable);
+    FileOptions blob_read =
+        env_->GetFileSystem()->OptimizeForBlobFileRead(in_opts, immutable);
+    EXPECT_FALSE(sst_read.use_direct_reads);
+    EXPECT_FALSE(blob_read.use_direct_reads);
+    EXPECT_FALSE(sst_read.use_direct_writes);
+  }
+
+  {
+    Options off_options;
+    off_options.use_direct_reads = false;
+    off_options.use_direct_io_for_compaction_reads = false;
+    off_options.use_direct_io_for_flush_and_compaction = false;
+    ImmutableDBOptions immutable_off(off_options);
+    FileOptions sst_read_off =
+        env_->GetFileSystem()->OptimizeForCompactionTableRead(in_opts,
+                                                              immutable_off);
+    EXPECT_FALSE(sst_read_off.use_direct_reads);
+  }
+
+  {
+    Options global_on_options;
+    global_on_options.use_direct_reads = true;
+    global_on_options.use_direct_io_for_compaction_reads = false;
+    global_on_options.use_direct_io_for_flush_and_compaction = false;
+    ImmutableDBOptions immutable_global(global_on_options);
+    FileOptions sst_read_global =
+        env_->GetFileSystem()->OptimizeForCompactionTableRead(in_opts,
+                                                              immutable_global);
+    EXPECT_TRUE(sst_read_global.use_direct_reads);
+  }
+
+  {
+    Options both_on;
+    both_on.use_direct_reads = true;
+    both_on.use_direct_io_for_compaction_reads = true;
+    both_on.use_direct_io_for_flush_and_compaction = false;
+    ImmutableDBOptions immutable_both(both_on);
+    FileOptions sst_read_both =
+        env_->GetFileSystem()->OptimizeForCompactionTableRead(in_opts,
+                                                              immutable_both);
+    EXPECT_TRUE(sst_read_both.use_direct_reads);
+  }
+}
+
 }  // namespace ROCKSDB_NAMESPACE
 
 int main(int argc, char** argv) {

db/db_secondary_test.cc

@@ -13,6 +13,7 @@
 #include "db/db_with_timestamp_test_util.h"
 #include "db/wide/wide_column_test_util.h"
 #include "db/write_batch_internal.h"
+#include "file/filename.h"
 #include "port/stack_trace.h"
 #include "rocksdb/utilities/transaction_db.h"
 #include "test_util/sync_point.h"
@@ -922,6 +923,48 @@ TEST_F(DBSecondaryTest, OpenAsSecondaryWALTailing) {
   verify_db_func("new_foo_value_1", "new_bar_value");
 }
 
+TEST_F(DBSecondaryTest, CatchUpTailsCurrentWalWhenFutureWalExists) {
+  // Goal: secondary catch-up must keep tailing the current WAL even when a
+  // higher-number empty WAL is present. This reproduces the async WAL
+  // precreation shape by creating a future WAL before opening the secondary,
+  // then appending another write to the current WAL and catching up again.
+  Options options;
+  options.env = env_;
+  Reopen(options);
+
+  ASSERT_OK(Put("foo", "value0"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("foo", "value1"));
+  ASSERT_OK(db_->FlushWAL(/*sync=*/true));
+
+  std::unique_ptr<WalFile> current_wal;
+  ASSERT_OK(db_->GetCurrentWalFile(&current_wal));
+  ASSERT_NE(nullptr, current_wal);
+
+  const uint64_t future_wal_number = current_wal->LogNumber() + 1000;
+  std::unique_ptr<WritableFile> future_wal_file;
+  ASSERT_OK(env_->NewWritableFile(LogFileName(dbname_, future_wal_number),
+                                  &future_wal_file, EnvOptions()));
+  ASSERT_OK(future_wal_file->Close());
+
+  Options secondary_options;
+  secondary_options.env = env_;
+  secondary_options.max_open_files = -1;
+  OpenSecondary(secondary_options);
+
+  ReadOptions read_options;
+  std::string value;
+  ASSERT_OK(db_secondary_->Get(read_options, "foo", &value));
+  ASSERT_EQ("value1", value);
+
+  ASSERT_OK(Put("foo", "value2"));
+  ASSERT_OK(db_->FlushWAL(/*sync=*/true));
+
+  ASSERT_OK(db_secondary_->TryCatchUpWithPrimary());
+  ASSERT_OK(db_secondary_->Get(read_options, "foo", &value));
+  ASSERT_EQ("value2", value);
+}
+
 TEST_F(DBSecondaryTest, SecondaryTailingBug_ISSUE_8467) {
   Options options;
   options.env = env_;

db/db_test.cc

@@ -3516,6 +3516,18 @@ class ModelDB : public DB {
     return Status::NotSupported("Not implemented");
   }
 
+  using DB::PrepareFileIngestion;
+  Status PrepareFileIngestion(
+      const std::vector<IngestExternalFileArg>& /*args*/,
+      std::unique_ptr<FileIngestionHandle>* /*handle*/) override {
+    return Status::NotSupported("Not implemented.");
+  }
+
+  Status CommitFileIngestionHandles(
+      std::vector<std::unique_ptr<FileIngestionHandle>> /*handles*/) override {
+    return Status::NotSupported("Not implemented.");
+  }
+
   using DB::CreateColumnFamilyWithImport;
   Status CreateColumnFamilyWithImport(
       const ColumnFamilyOptions& /*options*/,
@@ -3773,6 +3785,12 @@ class ModelDB : public DB {
     return Status::OK();
   }
 
+  Status GetPreparedFileInfoForExternalSstIngestion(
+      const std::string& /*file_path*/,
+      std::shared_ptr<const PreparedFileInfo>* /*file_info*/) override {
+    return Status::NotSupported();
+  }
+
   Status GetSortedWalFiles(VectorLogPtr& /*files*/) override {
     return Status::OK();
   }
@@ -8071,6 +8089,41 @@ INSTANTIATE_TEST_CASE_P(OpenFilesAsync, OpenFilesAsyncTest,
                                            ::testing::Values(-1, 10),
                                            ::testing::Bool()));
 
+TEST_F(DBTest, ReadOnlyCloseDoesNotDeleteWriterFiles) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.disable_auto_compactions = true;
+  DestroyAndReopen(options);
+
+  ASSERT_OK(Put("before", "value"));
+  ASSERT_OK(Flush());
+
+  std::unique_ptr<DB> read_only_db;
+  ASSERT_OK(DB::OpenForReadOnly(options, dbname_, &read_only_db));
+
+  ASSERT_OK(Put("after", "value"));
+  ASSERT_OK(Flush());
+
+  std::vector<LiveFileMetaData> live_files;
+  db_->GetLiveFilesMetaData(&live_files);
+  ASSERT_GE(live_files.size(), 2);
+
+  std::string newest_sst_path;
+  uint64_t newest_file_number = 0;
+  for (const auto& live_file : live_files) {
+    if (live_file.file_number > newest_file_number) {
+      newest_file_number = live_file.file_number;
+      newest_sst_path = live_file.directory + "/" + live_file.relative_filename;
+    }
+  }
+  ASSERT_FALSE(newest_sst_path.empty());
+  ASSERT_OK(env_->FileExists(newest_sst_path));
+
+  read_only_db.reset();
+
+  ASSERT_OK(env_->FileExists(newest_sst_path));
+}
+
 // Test mix of races with async file open, reads, compactions
 TEST_P(OpenFilesAsyncTest, ConcurrentFileAccess) {
   Options options = CurrentOptions();

db/db_test_util.cc

@@ -13,6 +13,7 @@
 #include "db/forward_iterator.h"
 #include "env/fs_readonly.h"
 #include "env/mock_env.h"
+#include "file/file_util.h"
 #include "port/lang.h"
 #include "rocksdb/cache.h"
 #include "rocksdb/convenience.h"
@@ -90,6 +91,8 @@ DBTestBase::DBTestBase(const std::string path, bool env_do_fsync)
   dbname_ = test::PerThreadDBPath(env_, path);
   alternative_wal_dir_ = dbname_ + "/wal";
   alternative_db_log_dir_ = dbname_ + "/db_log_dir";
+  EXPECT_OK(DestroyDir(env_, alternative_wal_dir_));
+  EXPECT_OK(DestroyDir(env_, alternative_db_log_dir_));
   auto options = CurrentOptions();
   options.env = env_;
   auto delete_options = options;
@@ -117,6 +120,8 @@ DBTestBase::~DBTestBase() {
   if (getenv("KEEP_DB")) {
     printf("DB is still at %s\n", dbname_.c_str());
   } else {
+    EXPECT_OK(DestroyDir(env_, alternative_wal_dir_));
+    EXPECT_OK(DestroyDir(env_, alternative_db_log_dir_));
     EXPECT_OK(DestroyDB(dbname_, options));
   }
   // Ensure SstFileManager (and its DeleteScheduler background thread) is

db/db_wal_test.cc

@@ -7,6 +7,8 @@
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file. See the AUTHORS file for names of contributors.
 
+#include <algorithm>
+
 #include "db/db_test_util.h"
 #include "db/db_with_timestamp_test_util.h"
 #include "options/options_helper.h"
@@ -19,12 +21,62 @@
 #include "utilities/fault_injection_env.h"
 #include "utilities/fault_injection_fs.h"
 
+#if defined(OS_LINUX)
+#include <sys/statfs.h>
+#endif
+
+#if defined(OS_LINUX) && !defined(BTRFS_SUPER_MAGIC)
+#define BTRFS_SUPER_MAGIC 0x9123683E
+#endif
+
+#if defined(OS_LINUX) && !defined(TMPFS_MAGIC)
+#define TMPFS_MAGIC 0x01021994
+#endif
+
+#if defined(OS_LINUX) && !defined(OVERLAYFS_SUPER_MAGIC)
+#define OVERLAYFS_SUPER_MAGIC 0x794c7630
+#endif
+
+#if defined(OS_LINUX) && !defined(ZFS_SUPER_MAGIC)
+#define ZFS_SUPER_MAGIC 0x2fc12fc1
+#endif
+
 namespace ROCKSDB_NAMESPACE {
 class DBWALTestBase : public DBTestBase {
  protected:
   explicit DBWALTestBase(const std::string& dir_name)
       : DBTestBase(dir_name, /*env_do_fsync=*/true) {}
 
+  std::vector<uint64_t> ListWalNumbers() {
+    std::vector<uint64_t> wal_numbers;
+    std::vector<std::string> files;
+    EXPECT_OK(env_->GetChildren(dbname_, &files));
+    for (const auto& file : files) {
+      uint64_t number = 0;
+      FileType type = kWalFile;
+      if (ParseFileName(file, &number, &type) && type == kWalFile) {
+        wal_numbers.push_back(number);
+      }
+    }
+    std::sort(wal_numbers.begin(), wal_numbers.end());
+    return wal_numbers;
+  }
+
+  void CreateEmptyWal(uint64_t wal_number) {
+    std::unique_ptr<WritableFile> file;
+    ASSERT_OK(env_->NewWritableFile(LogFileName(dbname_, wal_number), &file,
+                                    EnvOptions()));
+    ASSERT_OK(file->Close());
+  }
+
+  void CreateWalWithContents(uint64_t wal_number, const Slice& contents) {
+    std::unique_ptr<WritableFile> file;
+    ASSERT_OK(env_->NewWritableFile(LogFileName(dbname_, wal_number), &file,
+                                    EnvOptions()));
+    ASSERT_OK(file->Append(contents));
+    ASSERT_OK(file->Close());
+  }
+
 #if defined(ROCKSDB_PLATFORM_POSIX)
  public:
 #if defined(ROCKSDB_FALLOCATE_PRESENT)
@@ -57,6 +109,31 @@ class DBWALTestBase : public DBTestBase {
     assert(err == 0);
     return sbuf.st_blocks * 512;
   }
+
+#if defined(ROCKSDB_FALLOCATE_PRESENT)
+  bool ShouldSkipAllocationCheck(const std::string& file_name) {
+    (void)file_name;
+#if defined(OS_LINUX)
+    struct statfs fs_stat;
+    if (statfs(file_name.c_str(), &fs_stat) == 0) {
+      if (fs_stat.f_type ==
+          static_cast<decltype(fs_stat.f_type)>(BTRFS_SUPER_MAGIC)) {
+        return true;
+      } else if (fs_stat.f_type ==
+                 static_cast<decltype(fs_stat.f_type)>(ZFS_SUPER_MAGIC)) {
+        return true;
+      } else if (fs_stat.f_type ==
+                 static_cast<decltype(fs_stat.f_type)>(TMPFS_MAGIC)) {
+        return true;
+      } else if (fs_stat.f_type ==
+                 static_cast<decltype(fs_stat.f_type)>(OVERLAYFS_SUPER_MAGIC)) {
+        return true;
+      }
+    }
+#endif
+    return false;
+  }
+#endif  // ROCKSDB_FALLOCATE_PRESENT
 #endif  // ROCKSDB_PLATFORM_POSIX
 };
 
@@ -783,6 +860,8 @@ TEST_F(DBWALTest, WALWithChecksumHandoff) {
   std::shared_ptr<FaultInjectionTestFS> fault_fs(
       new FaultInjectionTestFS(FileSystem::Default()));
   std::unique_ptr<Env> fault_fs_env(NewCompositeEnv(fault_fs));
+  // Close the DB before the local Env is destroyed if an ASSERT exits early.
+  Defer close_db_on_exit([this]() { Close(); });
   do {
     Options options = CurrentOptions();
 
@@ -2752,16 +2831,35 @@ TEST_F(DBWALTest, TruncateLastLogAfterRecoverWithoutFlush) {
   auto& file_before = log_files_before[0];
   ASSERT_LT(file_before->SizeFileBytes(), 1 * kKB);
   // The log file has preallocated space.
-  ASSERT_GE(GetAllocatedFileSize(dbname_ + file_before->PathName()),
-            preallocated_size);
+  {
+    std::string fname = dbname_ + file_before->PathName();
+    uint64_t allocated = GetAllocatedFileSize(fname);
+    if (ShouldSkipAllocationCheck(fname)) {
+      fprintf(stderr,
+              "Skipping preallocation check on this filesystem for %s\n",
+              fname.c_str());
+    } else if (allocated < preallocated_size) {
+      fprintf(stderr,
+              "Warning: allocated size (%lu) less than preallocated size (%lu) "
+              "for %s, skipping check. This may indicate the filesystem does "
+              "not support preallocation or does not report it.\n",
+              allocated, preallocated_size, fname.c_str());
+    } else {
+      ASSERT_GE(allocated, preallocated_size);
+    }
+  }
   Reopen(options);
   VectorLogPtr log_files_after;
   ASSERT_OK(dbfull()->GetSortedWalFiles(log_files_after));
   ASSERT_EQ(1, log_files_after.size());
   ASSERT_LT(log_files_after[0]->SizeFileBytes(), 1 * kKB);
   // The preallocated space should be truncated.
-  ASSERT_LT(GetAllocatedFileSize(dbname_ + file_before->PathName()),
-            preallocated_size);
+  {
+    std::string fname = dbname_ + file_before->PathName();
+    if (!ShouldSkipAllocationCheck(fname)) {
+      ASSERT_LT(GetAllocatedFileSize(fname), preallocated_size);
+    }
+  }
 }
 // Tests that we will truncate the preallocated space of the last log from
 // previous.
@@ -2788,8 +2886,23 @@ TEST_F(DBWALTest, TruncateLastLogAfterRecoverWithFlush) {
   ASSERT_EQ(1, log_files_before.size());
   auto& file_before = log_files_before[0];
   ASSERT_LT(file_before->SizeFileBytes(), 1 * kKB);
-  ASSERT_GE(GetAllocatedFileSize(dbname_ + file_before->PathName()),
-            preallocated_size);
+  {
+    std::string fname = dbname_ + file_before->PathName();
+    uint64_t allocated = GetAllocatedFileSize(fname);
+    if (ShouldSkipAllocationCheck(fname)) {
+      fprintf(stderr,
+              "Skipping preallocation check on this filesystem for %s\n",
+              fname.c_str());
+    } else if (allocated < preallocated_size) {
+      fprintf(stderr,
+              "Warning: allocated size (%lu) less than preallocated size (%lu) "
+              "for %s, skipping check. This may indicate the filesystem does "
+              "not support preallocation or does not report it.\n",
+              allocated, preallocated_size, fname.c_str());
+    } else {
+      ASSERT_GE(allocated, preallocated_size);
+    }
+  }
   // The log file has preallocated space.
   Close();
 
@@ -2807,8 +2920,12 @@ TEST_F(DBWALTest, TruncateLastLogAfterRecoverWithFlush) {
   // if  the process is in a crash loop, the log file may not get
   // deleted and thte preallocated space will keep accumulating. So we need
   // to ensure it gets trtuncated.
-  EXPECT_LT(GetAllocatedFileSize(dbname_ + file_before->PathName()),
-            preallocated_size);
+  {
+    std::string fname = dbname_ + file_before->PathName();
+    if (!ShouldSkipAllocationCheck(fname)) {
+      EXPECT_LT(GetAllocatedFileSize(fname), preallocated_size);
+    }
+  }
   TEST_SYNC_POINT(
       "DBWALTest::TruncateLastLogAfterRecoverWithFlush:AfterTruncate");
   reopen_thread.join();
@@ -2863,14 +2980,31 @@ TEST_F(DBWALTest, TruncateLastLogAfterRecoverWALEmpty) {
   log_file->PrepareWrite(0, 4096);
   log_file.reset();
 
-  ASSERT_GE(GetAllocatedFileSize(last_log), preallocated_size);
+  {
+    uint64_t allocated = GetAllocatedFileSize(last_log);
+    if (ShouldSkipAllocationCheck(last_log)) {
+      fprintf(stderr,
+              "Skipping preallocation check on this filesystem for %s\n",
+              last_log.c_str());
+    } else if (allocated < preallocated_size) {
+      fprintf(stderr,
+              "Warning: allocated size (%lu) less than preallocated size (%lu) "
+              "for %s, skipping check. This may indicate the filesystem does "
+              "not support preallocation or does not report it.\n",
+              allocated, preallocated_size, last_log.c_str());
+    } else {
+      ASSERT_GE(allocated, preallocated_size);
+    }
+  }
 
   port::Thread reopen_thread([&]() { Reopen(options); });
 
   TEST_SYNC_POINT(
       "DBWALTest::TruncateLastLogAfterRecoverWithFlush:AfterRecover");
   // The preallocated space should be truncated.
-  EXPECT_LT(GetAllocatedFileSize(last_log), preallocated_size);
+  if (!ShouldSkipAllocationCheck(last_log)) {
+    EXPECT_LT(GetAllocatedFileSize(last_log), preallocated_size);
+  }
   TEST_SYNC_POINT(
       "DBWALTest::TruncateLastLogAfterRecoverWithFlush:AfterTruncate");
   reopen_thread.join();
@@ -2912,8 +3046,23 @@ TEST_F(DBWALTest, ReadOnlyRecoveryNoTruncate) {
   auto& file_before = log_files_before[0];
   ASSERT_LT(file_before->SizeFileBytes(), 1 * kKB);
   // The log file has preallocated space.
-  auto db_size = GetAllocatedFileSize(dbname_ + file_before->PathName());
-  ASSERT_GE(db_size, preallocated_size);
+  std::string fname = dbname_ + file_before->PathName();
+  auto db_size = GetAllocatedFileSize(fname);
+  {
+    if (ShouldSkipAllocationCheck(fname)) {
+      fprintf(stderr,
+              "Skipping preallocation check on this filesystem for %s\n",
+              fname.c_str());
+    } else if (db_size < preallocated_size) {
+      fprintf(stderr,
+              "Warning: allocated size (%lu) less than preallocated size (%lu) "
+              "for %s, skipping check. This may indicate the filesystem does "
+              "not support preallocation or does not report it.\n",
+              db_size, preallocated_size, fname.c_str());
+    } else {
+      ASSERT_GE(db_size, preallocated_size);
+    }
+  }
   Close();
 
   // enable truncate and open DB as readonly, the file should not be truncated
@@ -2927,8 +3076,12 @@ TEST_F(DBWALTest, ReadOnlyRecoveryNoTruncate) {
   ASSERT_EQ(log_files_after[0]->PathName(), file_before->PathName());
   // The preallocated space should NOT be truncated.
   // the DB size is almost the same.
-  ASSERT_NEAR(GetAllocatedFileSize(dbname_ + file_before->PathName()), db_size,
-              db_size / 100);
+  {
+    std::string fname2 = dbname_ + file_before->PathName();
+    if (!ShouldSkipAllocationCheck(fname2)) {
+      ASSERT_NEAR(GetAllocatedFileSize(fname2), db_size, db_size / 100);
+    }
+  }
   SyncPoint::GetInstance()->DisableProcessing();
   SyncPoint::GetInstance()->ClearAllCallBacks();
 }
@@ -3071,6 +3224,284 @@ TEST_F(DBWALTest, EmptyWalReopenTest) {
   }
 }
 
+TEST_F(DBWALTest, AsyncWalPrecreateWaitsForPendingFile) {
+  Options options = CurrentOptions();
+  options.env = env_;
+  options.statistics = CreateDBStatistics();
+  options.async_wal_precreate = true;
+  options.recycle_log_file_num = 0;
+  options.write_buffer_size = 1024;
+  options.max_write_buffer_number = 4;
+
+  // Goal: cover the race where foreground WAL rotation reaches the switch
+  // before the background precreate publishes its result. The sync points hold
+  // the async worker after it has created the file but before it marks it
+  // ready, then verify the foreground writer waits instead of allocating a
+  // higher WAL number.
+  SyncPoint::GetInstance()->LoadDependency({
+      {"DBImpl::BGWorkAsyncWALPrecreate:BeforePublish",
+       "DBWALTest::AsyncWalPrecreateWaitsForPendingFile:WorkerBlocked"},
+      {"DBWALTest::AsyncWalPrecreateWaitsForPendingFile:AllowPublish",
+       "DBImpl::BGWorkAsyncWALPrecreate:Publish"},
+      {"DBImpl::WaitForAsyncWALPrecreate:BeforeWait",
+       "DBWALTest::AsyncWalPrecreateWaitsForPendingFile:ForegroundWaiting"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+  Defer cleanup_sync_point([] {
+    SyncPoint::GetInstance()->DisableProcessing();
+    SyncPoint::GetInstance()->ClearAllCallBacks();
+  });
+
+  Reopen(options);
+  TEST_SYNC_POINT(
+      "DBWALTest::AsyncWalPrecreateWaitsForPendingFile:WorkerBlocked");
+
+  ASSERT_OK(Put("fill", "value"));
+  Status writer_status;
+  port::Thread writer(
+      [&]() { writer_status = dbfull()->TEST_SwitchMemtable(); });
+
+  TEST_SYNC_POINT(
+      "DBWALTest::AsyncWalPrecreateWaitsForPendingFile:ForegroundWaiting");
+  TEST_SYNC_POINT(
+      "DBWALTest::AsyncWalPrecreateWaitsForPendingFile:AllowPublish");
+  writer.join();
+  ASSERT_OK(writer_status);
+  ASSERT_OK(Put("trigger", "value"));
+  ASSERT_EQ("value", Get("trigger"));
+  ASSERT_EQ(1, options.statistics->getTickerCount(WAL_PRECREATE_HIT));
+  ASSERT_EQ(1, options.statistics->getTickerCount(WAL_PRECREATE_WAITED));
+  ASSERT_EQ(0, options.statistics->getTickerCount(WAL_PRECREATE_MISS));
+}
+
+TEST_F(DBWALTest, AsyncWalPrecreateConsumesReadyFile) {
+  Options options = CurrentOptions();
+  options.env = env_;
+  options.statistics = CreateDBStatistics();
+  options.async_wal_precreate = true;
+  options.recycle_log_file_num = 0;
+  options.write_buffer_size = 1024;
+  options.max_write_buffer_number = 4;
+
+  // Goal: exercise the no-wait consumption path. The background task publishes
+  // a prepared future WAL before the foreground memtable switch, and the switch
+  // must install exactly that reserved WAL number instead of allocating
+  // another.
+  SyncPoint::GetInstance()->LoadDependency({
+      {"DBImpl::BGWorkAsyncWALPrecreate:Done",
+       "DBWALTest::AsyncWalPrecreateConsumesReadyFile:PrecreateDone"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+  Defer cleanup_sync_point([] {
+    SyncPoint::GetInstance()->DisableProcessing();
+    SyncPoint::GetInstance()->ClearAllCallBacks();
+  });
+
+  Reopen(options);
+  TEST_SYNC_POINT(
+      "DBWALTest::AsyncWalPrecreateConsumesReadyFile:PrecreateDone");
+
+  const std::vector<uint64_t> wal_numbers = ListWalNumbers();
+  ASSERT_GE(wal_numbers.size(), 2);
+  const uint64_t prepared_wal_number = wal_numbers.back();
+  ASSERT_NE(prepared_wal_number, dbfull()->TEST_GetCurrentLogNumber());
+
+  ASSERT_OK(Put("fill", "value"));
+  ASSERT_OK(dbfull()->TEST_SwitchMemtable());
+  ASSERT_EQ(prepared_wal_number, dbfull()->TEST_GetCurrentLogNumber());
+  ASSERT_EQ(1, options.statistics->getTickerCount(WAL_PRECREATE_HIT));
+  ASSERT_EQ(0, options.statistics->getTickerCount(WAL_PRECREATE_MISS));
+  ASSERT_EQ(0, options.statistics->getTickerCount(WAL_PRECREATE_WAITED));
+}
+
+TEST_F(DBWALTest, AsyncWalPrecreateRecoveryToleratesActualFutureWal) {
+  Options options = CurrentOptions();
+  options.env = env_;
+  options.async_wal_precreate = true;
+  options.recycle_log_file_num = 0;
+  options.track_and_verify_wals = true;
+  options.track_and_verify_wals_in_manifest = true;
+  options.avoid_flush_during_shutdown = true;
+  options.avoid_flush_during_recovery = true;
+
+  // Goal: cover the actual async precreation path, not a hand-forged empty WAL.
+  // The background task opens the future WAL, clean close releases the
+  // unpublished writer and leaves the empty future WAL behind, and recovery
+  // must tolerate that file while replaying the preceding WAL with real user
+  // data.
+  SyncPoint::GetInstance()->LoadDependency({
+      {"DBImpl::BGWorkAsyncWALPrecreate:Done",
+       "DBWALTest::AsyncWalPrecreateRecoveryToleratesActualFutureWal:"
+       "PrecreateDone"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+  Defer cleanup_sync_point([] {
+    SyncPoint::GetInstance()->DisableProcessing();
+    SyncPoint::GetInstance()->ClearAllCallBacks();
+  });
+
+  DestroyAndReopen(options);
+  TEST_SYNC_POINT(
+      "DBWALTest::AsyncWalPrecreateRecoveryToleratesActualFutureWal:"
+      "PrecreateDone");
+
+  const std::vector<uint64_t> wal_numbers = ListWalNumbers();
+  ASSERT_GE(wal_numbers.size(), 2);
+  const uint64_t future_wal_number = wal_numbers.back();
+  uint64_t future_wal_size = 0;
+  ASSERT_OK(env_->GetFileSize(LogFileName(dbname_, future_wal_number),
+                              &future_wal_size));
+  ASSERT_EQ(0, future_wal_size);
+
+  ASSERT_OK(Put("key", "value"));
+  Close();
+
+  ASSERT_OK(TryReopen(options));
+  ASSERT_EQ("value", Get("key"));
+  ASSERT_OK(Put("after-recovery", "value2"));
+  ASSERT_EQ("value2", Get("after-recovery"));
+}
+
+TEST_F(DBWALTest, AsyncWalPrecreateDeletesFileOnStartFailure) {
+  Options options = CurrentOptions();
+  options.env = env_;
+  options.statistics = CreateDBStatistics();
+  options.async_wal_precreate = true;
+  options.recycle_log_file_num = 0;
+  options.write_buffer_size = 1024;
+  options.max_write_buffer_number = 4;
+
+  // Goal: force failure after consuming a prepared WAL but before it becomes a
+  // logical WAL. The failed, unpublished file must be deleted and its writer
+  // destroyed outside mutex_.
+  SyncPoint::GetInstance()->LoadDependency({
+      {"DBImpl::BGWorkAsyncWALPrecreate:Done",
+       "DBWALTest::AsyncWalPrecreateDeletesFileOnStartFailure:PrecreateDone"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+  Defer cleanup_sync_point([] {
+    SyncPoint::GetInstance()->DisableProcessing();
+    SyncPoint::GetInstance()->ClearAllCallBacks();
+  });
+
+  Reopen(options);
+  TEST_SYNC_POINT(
+      "DBWALTest::AsyncWalPrecreateDeletesFileOnStartFailure:PrecreateDone");
+  SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::StartWALFile:AfterCompressionTypeRecord", [](void* arg) {
+        *static_cast<IOStatus*>(arg) =
+            IOStatus::IOError("injected async WAL start failure");
+      });
+  const std::vector<uint64_t> wal_numbers = ListWalNumbers();
+  ASSERT_GE(wal_numbers.size(), 2);
+  const uint64_t prepared_wal_number = wal_numbers.back();
+
+  ASSERT_OK(Put("fill", "value"));
+  ASSERT_NOK(dbfull()->TEST_SwitchMemtable());
+  ASSERT_TRUE(
+      env_->FileExists(LogFileName(dbname_, prepared_wal_number)).IsNotFound());
+  ASSERT_EQ(1, options.statistics->getTickerCount(WAL_PRECREATE_HIT));
+}
+
+TEST_F(DBWALTest, AsyncWalPrecreateFailureFallsBackToSyncCreation) {
+  Options options = CurrentOptions();
+  options.env = env_;
+  options.statistics = CreateDBStatistics();
+  options.async_wal_precreate = true;
+  options.recycle_log_file_num = 0;
+  options.write_buffer_size = 1024;
+  options.max_write_buffer_number = 4;
+
+  // Goal: fail the background precreate file open after normal DB open has
+  // created the current WAL. The foreground switch should observe no prepared
+  // WAL, create one synchronously, and expose both failure and miss counters.
+  SyncPoint::GetInstance()->LoadDependency({
+      {"DBImpl::BGWorkAsyncWALPrecreate:Done",
+       "DBWALTest::AsyncWalPrecreateFailureFallsBackToSyncCreation:"
+       "PrecreateDone"},
+  });
+  SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::Open:AfterDeleteFiles",
+      [&](void*) { env_->non_writable_count_ = 1; });
+  SyncPoint::GetInstance()->EnableProcessing();
+  Defer cleanup_sync_point([&] {
+    env_->non_writable_count_ = 0;
+    SyncPoint::GetInstance()->DisableProcessing();
+    SyncPoint::GetInstance()->ClearAllCallBacks();
+  });
+
+  Reopen(options);
+  TEST_SYNC_POINT(
+      "DBWALTest::AsyncWalPrecreateFailureFallsBackToSyncCreation:"
+      "PrecreateDone");
+  ASSERT_EQ(1, options.statistics->getTickerCount(WAL_PRECREATE_FAILED));
+
+  ASSERT_OK(Put("fill", "value"));
+  ASSERT_OK(dbfull()->TEST_SwitchMemtable());
+  ASSERT_EQ(1, options.statistics->getTickerCount(WAL_PRECREATE_MISS));
+}
+
+TEST_F(DBWALTest, ReadOnlyErrorIfWalFileExistsIgnoresEmptyWals) {
+  Options options = CurrentOptions();
+  options.env = env_;
+  DestroyAndReopen(options);
+  Close();
+
+  // Goal: preserve error_if_wal_file_exists for WALs containing data while
+  // allowing empty future WALs left by async precreation. This creates empty
+  // WAL files directly so the test isolates the read-only open contract.
+  const std::vector<uint64_t> existing_wal_numbers = ListWalNumbers();
+  const uint64_t empty_wal_number =
+      existing_wal_numbers.empty() ? 1 : existing_wal_numbers.back() + 1;
+  CreateEmptyWal(empty_wal_number);
+
+  DBOptions db_options(options);
+  std::vector<ColumnFamilyDescriptor> column_families;
+  column_families.emplace_back(kDefaultColumnFamilyName,
+                               ColumnFamilyOptions(options));
+  std::vector<ColumnFamilyHandle*> handles;
+  std::unique_ptr<DB> read_only_db;
+  ASSERT_OK(DB::OpenForReadOnly(db_options, dbname_, column_families, &handles,
+                                &read_only_db,
+                                true /* error_if_wal_file_exists */));
+  for (ColumnFamilyHandle* handle : handles) {
+    ASSERT_OK(read_only_db->DestroyColumnFamilyHandle(handle));
+  }
+  read_only_db.reset();
+  handles.clear();
+
+  CreateWalWithContents(empty_wal_number + 1, Slice("not empty"));
+  ASSERT_NOK(DB::OpenForReadOnly(db_options, dbname_, column_families, &handles,
+                                 &read_only_db,
+                                 true /* error_if_wal_file_exists */));
+}
+
+TEST_F(DBWALTest, RecoveryToleratesEmptyFutureWal) {
+  Options options = CurrentOptions();
+  options.env = env_;
+  options.track_and_verify_wals = true;
+  options.track_and_verify_wals_in_manifest = true;
+  options.avoid_flush_during_recovery = true;
+  DestroyAndReopen(options);
+
+  ASSERT_OK(Put("key", "value"));
+  Close();
+
+  const std::vector<uint64_t> wal_numbers = ListWalNumbers();
+  ASSERT_FALSE(wal_numbers.empty());
+  const uint64_t future_wal_number = wal_numbers.back() + 1;
+  CreateEmptyWal(future_wal_number);
+
+  // Goal: model a crash after async WAL precreation but before consumption.
+  // Recovery should process the zero-byte future WAL, mark its file number
+  // used, and keep replaying the real preceding WAL without reporting a WAL
+  // chain or MANIFEST-tracking error.
+  ASSERT_OK(TryReopen(options));
+  ASSERT_EQ("value", Get("key"));
+  ASSERT_OK(Put("after-recovery", "value2"));
+  ASSERT_EQ("value2", Get("after-recovery"));
+}
+
 TEST_F(DBWALTest, RecoveryFlushSwitchWALOnEmptyMemtable) {
   Options options = CurrentOptions();
   auto fault_fs = std::make_shared<FaultInjectionTestFS>(FileSystem::Default());

db/deletefile_test.cc

@@ -7,6 +7,7 @@
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file. See the AUTHORS file for names of contributors.
 
+#include <atomic>
 #include <cstdlib>
 #include <map>
 #include <string>
@@ -258,6 +259,108 @@ TEST_F(DeleteFileTest, BackgroundPurgeIteratorTest) {
   CheckFileTypeCounts(dbname_, 0, 1, 1);
 }
 
+// Regression test for a use-after-free where an obsolete-file purge started
+// DURING DB close -- after CloseHelper's early purge drain -- continued using
+// DBImpl after ~DBImpl destroyed mutex_. Seen as a vhost-user-blk SIGABRT when
+// many RocksDB instances closed concurrently under a Warm Storage fault, with
+// avoid_unnecessary_blocking_io=true on a shared, process-wide Env.
+//
+// Here we make the straggler deterministic: keep a dropped column family handle
+// alive, delete it only after CloseHelper has passed its early purge drain,
+// then park that cleanup after FindObsoleteFiles() has incremented
+// pending_purge_obsolete_files_ but before PurgeObsoleteFiles(..., true)
+// schedules BGWorkPurge. Without the fix, the final drain sees
+// bg_purge_scheduled_ == 0 and can destroy DBImpl while the pending cleanup is
+// still about to lock/use it. With the fix, CloseHelper waits through the
+// pending handoff and then through the scheduled purge.
+TEST_F(DeleteFileTest, CloseWaitsForPurgeScheduledDuringClose) {
+  Options options = CurrentOptions();
+  SetOptions(&options);
+  Destroy(options);
+  options.create_if_missing = true;
+  // Route obsolete-file deletion through the background purge queue, matching
+  // the production config that hit the bug.
+  options.avoid_unnecessary_blocking_io = true;
+  Reopen(options);
+
+  // A dropped CF whose handle we keep alive: deleting the handle schedules a
+  // *background* purge (the dropped-CF teardown path).
+  ColumnFamilyHandle* cfh = nullptr;
+  ASSERT_OK(db_->CreateColumnFamily(ColumnFamilyOptions(), "dropme", &cfh));
+  ASSERT_OK(db_->Put(WriteOptions(), cfh, "key", "value"));
+  ASSERT_OK(db_->Flush(FlushOptions(), cfh));
+  ASSERT_OK(db_->DropColumnFamily(cfh));
+  ASSERT_OK(dbfull()->TEST_WaitForPurge());  // settle setup purges
+
+  std::atomic<bool> block_pending_handoff{false};
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+  SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::PurgeObsoleteFiles:BeforePendingPurgeFinished", [&](void*) {
+        if (!block_pending_handoff.exchange(false)) {
+          return;
+        }
+        TEST_SYNC_POINT(
+            "DeleteFileTest::CloseWaitsForPurge:PendingHandoffReady");
+        TEST_SYNC_POINT(
+            "DeleteFileTest::CloseWaitsForPurge:PendingHandoffBlocked");
+      });
+  SyncPoint::GetInstance()->LoadDependency({
+      // Delete the dropped CF handle only after CloseHelper has passed its
+      // early purge drain and entered a mutex-unlocked close window.
+      {"DBImpl::CloseHelper:CFHandleCleanupUnlocked",
+       "DeleteFileTest::CloseWaitsForPurge:DropDuringCloseUnlock"},
+      // Let the test thread observe that the deleter is parked in the pending
+      // handoff window.
+      {"DeleteFileTest::CloseWaitsForPurge:PendingHandoffReady",
+       "DeleteFileTest::CloseWaitsForPurge:WaitForPendingHandoff"},
+      // Keep CloseHelper in the mutex-unlocked close window until the
+      // dropped-CF cleanup reaches the pending-to-scheduled handoff.
+      {"DeleteFileTest::CloseWaitsForPurge:PendingHandoffReady",
+       "DBImpl::CloseHelper:CFHandleCleanupAllowed"},
+      // Observe that CloseHelper is actually waiting in the final purge drain
+      // while the dropped-CF cleanup is still parked in the pending handoff.
+      {"DBImpl::CloseHelper:FinalPurgeDrainWait",
+       "DeleteFileTest::CloseWaitsForPurge:WaitForFinalPurgeDrain"},
+      // Keep the dropped-CF cleanup parked in the pending handoff window until
+      // the test releases it.
+      {"DeleteFileTest::CloseWaitsForPurge:ReleasePendingHandoff",
+       "DeleteFileTest::CloseWaitsForPurge:PendingHandoffBlocked"},
+      // Park the scheduled BGWorkPurge before it locks mutex_ until we release
+      // it, so CloseHelper also has to wait for bg_purge_scheduled_.
+      {"DeleteFileTest::CloseWaitsForPurge:ReleaseBackgroundPurge",
+       "DBImpl::BackgroundCallPurge:beforeMutexLock"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  port::Thread closer([&]() { db_.reset(); });
+
+  // Unblocks once CloseHelper is in a mutex-unlocked close window; then
+  // deleting the dropped CF handle starts purge work the early drain missed.
+  TEST_SYNC_POINT("DeleteFileTest::CloseWaitsForPurge:DropDuringCloseUnlock");
+  block_pending_handoff.store(true);
+  ColumnFamilyHandle* dropped_cfh = cfh;
+  cfh = nullptr;
+  port::Thread dropped_cf_deleter([dropped_cfh]() { delete dropped_cfh; });
+
+  TEST_SYNC_POINT("DeleteFileTest::CloseWaitsForPurge:WaitForPendingHandoff");
+
+  // Wait until CloseHelper is blocked in the final drain with no scheduled
+  // purge yet. Without the fix, it would proceed while the dropped-CF cleanup
+  // is still pending. With the fix, it waits for pending_purge_obsolete_files_.
+  TEST_SYNC_POINT("DeleteFileTest::CloseWaitsForPurge:WaitForFinalPurgeDrain");
+
+  // Release the pending cleanup, then the BGWorkPurge it schedules. With the
+  // fix, the closer waits for both transitions and completes cleanly.
+  TEST_SYNC_POINT("DeleteFileTest::CloseWaitsForPurge:ReleasePendingHandoff");
+  dropped_cf_deleter.join();
+  TEST_SYNC_POINT("DeleteFileTest::CloseWaitsForPurge:ReleaseBackgroundPurge");
+  closer.join();
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
 TEST_F(DeleteFileTest, PurgeDuringOpen) {
   Options options = CurrentOptions();
   CheckFileTypeCounts(dbname_, -1, 0, -1);

db/external_sst_file_basic_test.cc

@@ -1311,7 +1311,7 @@ TEST_F(ExternalSSTFileBasicTest, SyncFailure) {
     });
     if (i == 0) {
       SyncPoint::GetInstance()->SetCallBack(
-          "ExternalSstFileIngestionJob::Prepare:Reopen", [&](void* s) {
+          "ExternalSstFileIngestionJob::CheckSyncReturnCode", [&](void* s) {
             Status* status = static_cast<Status*>(s);
             if (status->IsNotSupported()) {
               no_sync = true;
@@ -1367,13 +1367,13 @@ TEST_F(ExternalSSTFileBasicTest, SyncFailure) {
   }
 }
 
-TEST_F(ExternalSSTFileBasicTest, ReopenNotSupported) {
+TEST_F(ExternalSSTFileBasicTest, SyncFileNotSupported) {
   Options options;
   options.create_if_missing = true;
   options.env = env_;
 
   SyncPoint::GetInstance()->SetCallBack(
-      "ExternalSstFileIngestionJob::Prepare:Reopen", [&](void* arg) {
+      "ExternalSstFileIngestionJob::CheckSyncReturnCode", [&](void* arg) {
         Status* s = static_cast<Status*>(arg);
         *s = Status::NotSupported();
       });
@@ -1386,7 +1386,7 @@ TEST_F(ExternalSSTFileBasicTest, ReopenNotSupported) {
   std::unique_ptr<SstFileWriter> sst_file_writer(
       new SstFileWriter(EnvOptions(), sst_file_writer_options));
   std::string file_name =
-      sst_files_dir_ + "reopen_not_supported_test_" + ".sst";
+      sst_files_dir_ + "sync_file_not_supported_test_" + ".sst";
   ASSERT_OK(sst_file_writer->Open(file_name));
   ASSERT_OK(sst_file_writer->Put("bar", "v2"));
   ASSERT_OK(sst_file_writer->Finish());

db/external_sst_file_ingestion_job.cc

@@ -19,6 +19,7 @@
 #include "logging/logging.h"
 #include "monitoring/statistics_impl.h"
 #include "table/merging_iterator.h"
+#include "table/prepared_file_info.h"
 #include "table/sst_file_writer_collectors.h"
 #include "table/table_builder.h"
 #include "table/unique_id_impl.h"
@@ -31,18 +32,24 @@ Status ExternalSstFileIngestionJob::Prepare(
     const std::vector<std::string>& external_files_paths,
     const std::vector<std::string>& files_checksums,
     const std::vector<std::string>& files_checksum_func_names,
+    const std::vector<const PreparedFileInfo*>& file_infos,
     const std::optional<RangeOpt>& atomic_replace_range,
     const Temperature& file_temperature, uint64_t next_file_number,
     SuperVersion* sv) {
   Status status;
 
-  // Read the information of files we are ingesting
-  for (const std::string& file_path : external_files_paths) {
+  // Read the information of files we are ingesting.
+  for (size_t i = 0; i < external_files_paths.size(); i++) {
+    const std::string& file_path = external_files_paths[i];
     IngestedFileInfo file_to_ingest;
     // For temperature, first assume it matches provided hint
     file_to_ingest.file_temperature = file_temperature;
-    status =
-        GetIngestedFileInfo(file_path, next_file_number++, &file_to_ingest, sv);
+    file_to_ingest.prefetch_lmax_index_and_filter_blocks =
+        ingestion_options_.prefetch_lmax_index_and_filter_blocks;
+    const PreparedFileInfo* prepared_file_info =
+        file_infos.empty() ? nullptr : file_infos[i];
+    status = GetIngestedFileInfo(file_path, next_file_number++,
+                                 prepared_file_info, &file_to_ingest, sv);
     if (!status.ok()) {
       ROCKS_LOG_WARN(db_options_.info_log,
                      "Failed to get ingested file info: %s: %s",
@@ -76,23 +83,10 @@ Status ExternalSstFileIngestionJob::Prepare(
   auto num_files = files_to_ingest_.size();
   if (num_files == 0) {
     return Status::InvalidArgument("The list of files is empty");
-  } else if (num_files > 1) {
-    // Verify that passed files don't have overlapping ranges
-    autovector<const IngestedFileInfo*> sorted_files;
-    for (size_t i = 0; i < num_files; i++) {
-      sorted_files.push_back(&files_to_ingest_[i]);
-    }
-
-    std::sort(sorted_files.begin(), sorted_files.end(), file_range_checker_);
-
-    for (size_t i = 0; i + 1 < num_files; i++) {
-      if (file_range_checker_.Overlaps(*sorted_files[i], *sorted_files[i + 1],
-                                       /* known_sorted= */ true)) {
-        files_overlap_ = true;
-        break;
-      }
-    }
   }
+  // Detect whether the input files overlap one another; this drives how they
+  // are divided into batches below.
+  files_overlap_ = ComputeFilesOverlap(files_to_ingest_);
 
   if (atomic_replace_range.has_value()) {
     atomic_replace_range_.emplace();
@@ -165,35 +159,21 @@ Status ExternalSstFileIngestionJob::Prepare(
         // It is unsafe to assume application had sync the file and file
         // directory before ingest the file. For integrity of RocksDB we need
         // to sync the file.
-
-        // TODO(xingbo), We should in general be moving away from production
-        // uses of ReuseWritableFile (except explicitly for WAL recycling),
-        // ReopenWritableFile, and NewRandomRWFile. We should create a
-        // FileSystem::SyncFile/FsyncFile API that by default does the
-        // re-open+sync+close combo but can (a) be reused easily, and (b) be
-        // overridden to do that more cleanly, e.g. in EncryptedEnv.
-        // https://github.com/facebook/rocksdb/issues/13741
-        std::unique_ptr<FSWritableFile> file_to_sync;
-        Status s = fs_->ReopenWritableFile(path_inside_db, env_options_,
-                                           &file_to_sync, nullptr);
-        TEST_SYNC_POINT_CALLBACK("ExternalSstFileIngestionJob::Prepare:Reopen",
-                                 &s);
+        TEST_SYNC_POINT("ExternalSstFileIngestionJob::BeforeSyncIngestedFile");
+        Status s = fs_->SyncFile(path_inside_db, env_options_, IOOptions(),
+                                 db_options_.use_fsync, nullptr);
+        TEST_SYNC_POINT("ExternalSstFileIngestionJob::AfterSyncIngestedFile");
+        TEST_SYNC_POINT_CALLBACK(
+            "ExternalSstFileIngestionJob::CheckSyncReturnCode", &s);
         // Some file systems (especially remote/distributed) don't support
-        // reopening a file for writing and don't require reopening and
-        // syncing the file. Ignore the NotSupported error in that case.
+        // explicitly syncing the file and don't require it. Ignore the
+        // NotSupported error in that case.
         if (!s.IsNotSupported()) {
           status = s;
-          if (status.ok()) {
-            TEST_SYNC_POINT(
-                "ExternalSstFileIngestionJob::BeforeSyncIngestedFile");
-            status = SyncIngestedFile(file_to_sync.get());
-            TEST_SYNC_POINT(
-                "ExternalSstFileIngestionJob::AfterSyncIngestedFile");
-            if (!status.ok()) {
-              ROCKS_LOG_WARN(db_options_.info_log,
-                             "Failed to sync ingested file %s: %s",
-                             path_inside_db.c_str(), status.ToString().c_str());
-            }
+          if (!status.ok()) {
+            ROCKS_LOG_WARN(db_options_.info_log,
+                           "Failed to sync ingested file %s: %s",
+                           path_inside_db.c_str(), status.ToString().c_str());
           }
         }
       } else if (status.IsNotSupported() &&
@@ -447,6 +427,60 @@ void ExternalSstFileIngestionJob::DivideInputFilesIntoBatches() {
   }
 }
 
+bool ExternalSstFileIngestionJob::ComputeFilesOverlap(
+    const autovector<IngestedFileInfo>& files) const {
+  const size_t num_files = files.size();
+  if (num_files <= 1) {
+    return false;
+  }
+  // Verify whether the files have overlapping ranges by sorting copies of the
+  // file ranges and checking adjacent pairs.
+  autovector<const IngestedFileInfo*> sorted_files;
+  for (size_t i = 0; i < num_files; i++) {
+    sorted_files.push_back(&files[i]);
+  }
+  std::sort(sorted_files.begin(), sorted_files.end(), file_range_checker_);
+  for (size_t i = 0; i + 1 < num_files; i++) {
+    if (file_range_checker_.Overlaps(*sorted_files[i], *sorted_files[i + 1],
+                                     /* known_sorted= */ true)) {
+      return true;
+    }
+  }
+  return false;
+}
+
+Status ExternalSstFileIngestionJob::MergeForSameColumnFamily(
+    ExternalSstFileIngestionJob* other) {
+  assert(other != nullptr);
+  assert(other != this);
+  assert(cfd_ == other->cfd_);
+  if (atomic_replace_range_.has_value() ||
+      other->atomic_replace_range_.has_value()) {
+    return Status::NotSupported(
+        "cannot merge file ingestion handles for the same column family when "
+        "atomic_replace_range is used");
+  }
+  if (!(ingestion_options_ == other->ingestion_options_)) {
+    return Status::InvalidArgument(
+        "file ingestion handles for the same column family must be prepared "
+        "with the same IngestExternalFileOptions");
+  }
+  // Append the other job's prepared files after this job's so that, for any
+  // overlapping keys, the other job's data wins via a higher assigned sequence
+  // number -- the same semantics as passing all the files to a single ingestion
+  // call in this order. Recompute overlap and rebuild the batches over the
+  // union.
+  for (IngestedFileInfo& file : other->files_to_ingest_) {
+    files_to_ingest_.push_back(std::move(file));
+  }
+  other->files_to_ingest_.clear();
+  other->file_batches_to_ingest_.clear();
+  files_overlap_ = ComputeFilesOverlap(files_to_ingest_);
+  file_batches_to_ingest_.clear();
+  DivideInputFilesIntoBatches();
+  return Status::OK();
+}
+
 Status ExternalSstFileIngestionJob::NeedsFlush(bool* flush_needed,
                                                SuperVersion* super_version) {
   Status status;
@@ -716,6 +750,9 @@ Status ExternalSstFileIngestionJob::AssignLevelsForOneBatch(
         tail_size, file->user_defined_timestamps_persisted, "", "");
     f_metadata.temperature = file->file_temperature;
     f_metadata.marked_for_compaction = marked_for_compaction;
+    f_metadata.skip_index_and_filter_blocks_prefetch =
+        !file->prefetch_lmax_index_and_filter_blocks &&
+        file->picked_level == cfd_->NumberLevels() - 1;
     // Extract min/max timestamps from table properties for UDT support.
     // This ensures ingested files have proper timestamp ranges in FileMetaData,
     // similar to files created by flush and compaction.
@@ -970,18 +1007,14 @@ Status ExternalSstFileIngestionJob::ResetTableReader(
 }
 
 Status ExternalSstFileIngestionJob::SanityCheckTableProperties(
-    const std::string& external_file, uint64_t new_file_number,
-    SuperVersion* sv, IngestedFileInfo* file_to_ingest,
-    std::unique_ptr<TableReader>* table_reader) {
-  // Get the external file properties
-  auto props = table_reader->get()->GetTableProperties();
-  assert(props.get());
-  const auto& uprops = props->user_collected_properties;
+    const std::string& external_file, const TableProperties& props,
+    IngestedFileInfo* file_to_ingest) {
+  const auto& uprops = props.user_collected_properties;
 
   // Get table version
   auto version_iter = uprops.find(ExternalSstFilePropertyNames::kVersion);
   if (version_iter == uprops.end()) {
-    assert(!SstFileWriter::CreatedBySstFileWriter(*props));
+    assert(!SstFileWriter::CreatedBySstFileWriter(props));
     if (!ingestion_options_.allow_db_generated_files) {
       return Status::Corruption("External file version not found");
     } else {
@@ -990,7 +1023,7 @@ Status ExternalSstFileIngestionJob::SanityCheckTableProperties(
       file_to_ingest->version = 0;
     }
   } else {
-    assert(SstFileWriter::CreatedBySstFileWriter(*props));
+    assert(SstFileWriter::CreatedBySstFileWriter(props));
     file_to_ingest->version = DecodeFixed32(version_iter->second.c_str());
   }
 
@@ -1004,12 +1037,17 @@ Status ExternalSstFileIngestionJob::SanityCheckTableProperties(
 
     // Set the global sequence number
     file_to_ingest->original_seqno = DecodeFixed64(seqno_iter->second.c_str());
-    if (props->external_sst_file_global_seqno_offset == 0) {
-      file_to_ingest->global_seqno_offset = 0;
+    file_to_ingest->global_seqno_offset =
+        static_cast<size_t>(props.external_sst_file_global_seqno_offset);
+    // The on-disk offset is only needed if we will write the global seqno back
+    // into the file (write_global_seqno). The metadata fast-path does not open
+    // the file, and its in-memory table properties do not carry the offset
+    // (it is only computed while reading the file back); that is fine as long
+    // as write_global_seqno is not requested.
+    if (ingestion_options_.write_global_seqno &&
+        file_to_ingest->global_seqno_offset == 0) {
       return Status::Corruption("Was not able to find file global seqno field");
     }
-    file_to_ingest->global_seqno_offset =
-        static_cast<size_t>(props->external_sst_file_global_seqno_offset);
   } else if (file_to_ingest->version == 1) {
     // SST file V1 should not have global seqno field
     assert(seqno_iter == uprops.end());
@@ -1030,23 +1068,20 @@ Status ExternalSstFileIngestionJob::SanityCheckTableProperties(
                                    " is not supported");
   }
 
-  file_to_ingest->cf_id = static_cast<uint32_t>(props->column_family_id);
-  // This assignment works fine even though `table_reader` may later be reset,
-  // since that will not affect how table properties are parsed, and this
-  // assignment is making a copy.
-  file_to_ingest->table_properties = *props;
+  file_to_ingest->cf_id = static_cast<uint32_t>(props.column_family_id);
+  file_to_ingest->table_properties = props;
 
   // Get number of entries in table
-  file_to_ingest->num_entries = props->num_entries;
-  file_to_ingest->num_range_deletions = props->num_range_deletions;
+  file_to_ingest->num_entries = props.num_entries;
+  file_to_ingest->num_range_deletions = props.num_range_deletions;
 
   // Validate table properties related to comparator name and user defined
   // timestamps persisted flag.
   file_to_ingest->user_defined_timestamps_persisted =
-      static_cast<bool>(props->user_defined_timestamps_persisted);
+      static_cast<bool>(props.user_defined_timestamps_persisted);
   bool mark_sst_file_has_no_udt = false;
   Status s = ValidateUserDefinedTimestampsOptions(
-      cfd_->user_comparator(), props->comparator_name,
+      cfd_->user_comparator(), props.comparator_name,
       cfd_->ioptions().persist_user_defined_timestamps,
       file_to_ingest->user_defined_timestamps_persisted,
       &mark_sst_file_has_no_udt);
@@ -1054,7 +1089,9 @@ Status ExternalSstFileIngestionJob::SanityCheckTableProperties(
     // A column family that enables user-defined timestamps in Memtable only
     // feature can also ingest external files created by a setting that disables
     // user-defined timestamps. In that case, we need to re-mark the
-    // user_defined_timestamps_persisted flag for the file.
+    // user_defined_timestamps_persisted flag for the file. The open-and-scan
+    // caller is then responsible for reopening its `TableReader` with the
+    // updated flag.
     file_to_ingest->user_defined_timestamps_persisted = false;
   } else if (!s.ok()) {
     ROCKS_LOG_WARN(
@@ -1064,22 +1101,84 @@ Status ExternalSstFileIngestionJob::SanityCheckTableProperties(
     return s;
   }
 
-  // `TableReader` is initialized with `user_defined_timestamps_persisted` flag
-  // to be true. If its value changed to false after this sanity check, we
-  // need to reset the `TableReader`.
-  if (ucmp_->timestamp_size() > 0 &&
-      !file_to_ingest->user_defined_timestamps_persisted) {
-    s = ResetTableReader(external_file, new_file_number,
-                         file_to_ingest->user_defined_timestamps_persisted, sv,
-                         file_to_ingest, table_reader);
-  }
   return s;
 }
 
-Status ExternalSstFileIngestionJob::GetIngestedFileInfo(
+Status ExternalSstFileIngestionJob::GetIngestedFileInfoFromFileInfo(
+    const std::string& external_file,
+    const PreparedFileInfo& prepared_file_info,
+    IngestedFileInfo* file_to_ingest) {
+  // Boundaries, size, and table properties were obtained without opening the
+  // file (e.g. produced by SstFileWriter::Finish). We reuse them directly.
+  file_to_ingest->file_size = prepared_file_info.file_size;
+  Status status = SanityCheckTableProperties(
+      external_file, prepared_file_info.table_properties, file_to_ingest);
+  if (!status.ok()) {
+    ROCKS_LOG_WARN(
+        db_options_.info_log,
+        "Failed to sanity check table properties for external file %s: %s",
+        external_file.c_str(), status.ToString().c_str());
+    return status;
+  }
+
+  const size_t ts_sz = ucmp_->timestamp_size();
+  if (ts_sz > 0 && !file_to_ingest->user_defined_timestamps_persisted) {
+    auto pad_timestamp = [ts_sz](std::string* result, const Slice& key) {
+      assert(result->empty());
+      if (ExtractValueType(key) == kTypeRangeDeletion) {
+        PadInternalKeyWithMaxTimestamp(result, key, ts_sz);
+      } else {
+        PadInternalKeyWithMinTimestamp(result, key, ts_sz);
+      }
+    };
+    pad_timestamp(file_to_ingest->smallest_internal_key.rep(),
+                  prepared_file_info.smallest.Encode());
+    pad_timestamp(file_to_ingest->largest_internal_key.rep(),
+                  prepared_file_info.largest.Encode());
+  } else {
+    file_to_ingest->smallest_internal_key = prepared_file_info.smallest;
+    file_to_ingest->largest_internal_key = prepared_file_info.largest;
+  }
+
+  if (ingestion_options_.allow_db_generated_files) {
+    // Sequence numbers are preserved (not reassigned), so the bounds must be
+    // known before we skip the GetSeqnoBoundaryForFile scan.
+    if (!file_to_ingest->table_properties.HasKeyLargestSeqno()) {
+      return Status::Corruption("Unknown largest seqno for db generated file.");
+    }
+    file_to_ingest->largest_seqno =
+        file_to_ingest->table_properties.key_largest_seqno;
+    if (file_to_ingest->largest_seqno == 0) {
+      file_to_ingest->smallest_seqno = 0;
+    } else {
+      if (!file_to_ingest->table_properties.HasKeySmallestSeqno()) {
+        return Status::Corruption(
+            "Unknown smallest seqno for db generated file.");
+      }
+      file_to_ingest->smallest_seqno =
+          file_to_ingest->table_properties.key_smallest_seqno;
+    }
+  } else {
+    // Normal ingestion reassigns a global sequence number later, so the file's
+    // keys must currently be at seqno 0 (mirror of the open-and-scan check).
+    SequenceNumber largest_seqno =
+        file_to_ingest->table_properties.key_largest_seqno;
+    // UINT64_MAX means unknown and the file is generated before table property
+    // `key_largest_seqno` is introduced.
+    if (largest_seqno != UINT64_MAX && largest_seqno > 0) {
+      return Status::Corruption(
+          "External file has non zero largest sequence number " +
+          std::to_string(largest_seqno));
+    }
+  }
+  return Status::OK();
+}
+
+Status ExternalSstFileIngestionJob::GetIngestedFileInfoFromFile(
     const std::string& external_file, uint64_t new_file_number,
-    IngestedFileInfo* file_to_ingest, SuperVersion* sv) {
-  file_to_ingest->external_file_path = external_file;
+    IngestedFileInfo* file_to_ingest, SuperVersion* sv,
+    std::unique_ptr<TableReader>* out_table_reader) {
+  TEST_SYNC_POINT("ExternalSstFileIngestionJob::GetIngestedFileInfo:ReadPath");
 
   // Get external file size
   Status status = fs_->GetFileSize(external_file, IOOptions(),
@@ -1091,15 +1190,11 @@ Status ExternalSstFileIngestionJob::GetIngestedFileInfo(
     return status;
   }
 
-  // Assign FD with number
-  file_to_ingest->fd =
-      FileDescriptor(new_file_number, 0, file_to_ingest->file_size);
-
-  // Create TableReader for external file
+  // Create TableReader for external file.
   std::unique_ptr<TableReader> table_reader;
   // Initially create the `TableReader` with flag
-  // `user_defined_timestamps_persisted` to be true since that's the most common
-  // case
+  // `user_defined_timestamps_persisted` to be true since that's the most
+  // common case
   status = ResetTableReader(external_file, new_file_number,
                             /*user_defined_timestamps_persisted=*/true, sv,
                             file_to_ingest, &table_reader);
@@ -1110,8 +1205,8 @@ Status ExternalSstFileIngestionJob::GetIngestedFileInfo(
     return status;
   }
 
-  status = SanityCheckTableProperties(external_file, new_file_number, sv,
-                                      file_to_ingest, &table_reader);
+  status = SanityCheckTableProperties(
+      external_file, *table_reader->GetTableProperties(), file_to_ingest);
   if (!status.ok()) {
     ROCKS_LOG_WARN(
         db_options_.info_log,
@@ -1120,6 +1215,23 @@ Status ExternalSstFileIngestionJob::GetIngestedFileInfo(
     return status;
   }
 
+  // The `TableReader` above was opened with `user_defined_timestamps_persisted`
+  // assumed true. If the sanity check determined the file has no persisted
+  // timestamps (UDT-in-Memtable-only feature), reopen it with the corrected
+  // flag so keys are parsed properly by the scan below.
+  if (ucmp_->timestamp_size() > 0 &&
+      !file_to_ingest->user_defined_timestamps_persisted) {
+    status = ResetTableReader(external_file, new_file_number,
+                              file_to_ingest->user_defined_timestamps_persisted,
+                              sv, file_to_ingest, &table_reader);
+    if (!status.ok()) {
+      ROCKS_LOG_WARN(db_options_.info_log,
+                     "Failed to reset table reader for external file %s: %s",
+                     external_file.c_str(), status.ToString().c_str());
+      return status;
+    }
+  }
+
   const bool allow_data_in_errors = db_options_.allow_data_in_errors;
   ParsedInternalKey key;
   if (ingestion_options_.allow_db_generated_files) {
@@ -1141,8 +1253,8 @@ Status ExternalSstFileIngestionJob::GetIngestedFileInfo(
   } else {
     SequenceNumber largest_seqno =
         table_reader.get()->GetTableProperties()->key_largest_seqno;
-    // UINT64_MAX means unknown and the file is generated before table property
-    // `key_largest_seqno` is introduced.
+    // UINT64_MAX means unknown and the file is generated before table
+    // property `key_largest_seqno` is introduced.
     if (largest_seqno != UINT64_MAX && largest_seqno > 0) {
       return Status::Corruption(
           "External file has non zero largest sequence number " +
@@ -1150,24 +1262,6 @@ Status ExternalSstFileIngestionJob::GetIngestedFileInfo(
     }
   }
 
-  if (ingestion_options_.verify_checksums_before_ingest) {
-    // If customized readahead size is needed, we can pass a user option
-    // all the way to here. Right now we just rely on the default readahead
-    // to keep things simple.
-    // TODO: plumb Env::IOActivity, Env::IOPriority
-    ReadOptions ro;
-    ro.readahead_size = ingestion_options_.verify_checksums_readahead_size;
-    ro.fill_cache = ingestion_options_.fill_cache;
-    status = table_reader->VerifyChecksum(
-        ro, TableReaderCaller::kExternalSSTIngestion);
-    if (!status.ok()) {
-      ROCKS_LOG_WARN(db_options_.info_log,
-                     "Failed to verify checksum for table reader: %s",
-                     status.ToString().c_str());
-      return status;
-    }
-  }
-
   // TODO: plumb Env::IOActivity, Env::IOPriority
   ReadOptions ro;
   ro.fill_cache = ingestion_options_.fill_cache;
@@ -1247,6 +1341,22 @@ Status ExternalSstFileIngestionJob::GetIngestedFileInfo(
             "External file has a range deletion with non zero sequence "
             "number.");
       }
+#ifndef NDEBUG
+      // To keep aligned with the fast path that expects range deletion keys to
+      // always use max ts.
+      const size_t ts_sz = ucmp_->timestamp_size();
+      if (ts_sz > 0) {
+        const std::string max_ts(ts_sz, '\xff');
+        assert(key.user_key.size() >= ts_sz);
+        assert(ucmp_->CompareTimestamp(
+                   ExtractTimestampFromUserKey(key.user_key, ts_sz), max_ts) ==
+               0);
+        assert(range_del_iter->value().size() >= ts_sz);
+        assert(ucmp_->CompareTimestamp(
+                   ExtractTimestampFromUserKey(range_del_iter->value(), ts_sz),
+                   max_ts) == 0);
+      }
+#endif
       RangeTombstone tombstone(key, range_del_iter->value());
       file_range_checker_.MaybeUpdateRange(tombstone.SerializeKey(),
                                            tombstone.SerializeEndKey(),
@@ -1254,6 +1364,75 @@ Status ExternalSstFileIngestionJob::GetIngestedFileInfo(
     }
   }
 
+  *out_table_reader = std::move(table_reader);
+  return Status::OK();
+}
+
+Status ExternalSstFileIngestionJob::GetIngestedFileInfo(
+    const std::string& external_file, uint64_t new_file_number,
+    const PreparedFileInfo* prepared_file_info,
+    IngestedFileInfo* file_to_ingest, SuperVersion* sv) {
+  file_to_ingest->external_file_path = external_file;
+
+  std::unique_ptr<TableReader> table_reader;
+  Status status =
+      prepared_file_info != nullptr
+          ? GetIngestedFileInfoFromFileInfo(external_file, *prepared_file_info,
+                                            file_to_ingest)
+          : GetIngestedFileInfoFromFile(external_file, new_file_number,
+                                        file_to_ingest, sv, &table_reader);
+  if (!status.ok()) {
+    return status;
+  }
+
+  assert(file_to_ingest->file_size > 0);
+  assert(!file_to_ingest->unset());
+  assert(file_to_ingest->table_properties.num_entries > 0 ||
+         file_to_ingest->table_properties.num_range_deletions > 0);
+  if (ingestion_options_.allow_db_generated_files) {
+    // These files keep their original sequence numbers (derived, not
+    // reassigned), so the bounds must be valid here.
+    assert(file_to_ingest->smallest_seqno <= file_to_ingest->largest_seqno);
+    assert(file_to_ingest->largest_seqno < kMaxSequenceNumber);
+  }
+
+  // Verify the file checksum if requested. The open-and-scan path already has a
+  // `TableReader` open; the fast-path opens one here only when verification is
+  // requested -- otherwise it performs no file I/O.
+  if (ingestion_options_.verify_checksums_before_ingest) {
+    if (table_reader == nullptr) {
+      status =
+          ResetTableReader(external_file, new_file_number,
+                           file_to_ingest->user_defined_timestamps_persisted,
+                           sv, file_to_ingest, &table_reader);
+      if (!status.ok()) {
+        ROCKS_LOG_WARN(db_options_.info_log,
+                       "Failed to reset table reader for external file %s: %s",
+                       external_file.c_str(), status.ToString().c_str());
+        return status;
+      }
+    }
+    // If customized readahead size is needed, we can pass a user option all the
+    // way to here. Right now we just rely on the default readahead to keep
+    // things simple.
+    // TODO: plumb Env::IOActivity, Env::IOPriority
+    ReadOptions ro;
+    ro.readahead_size = ingestion_options_.verify_checksums_readahead_size;
+    ro.fill_cache = ingestion_options_.fill_cache;
+    status = table_reader->VerifyChecksum(
+        ro, TableReaderCaller::kExternalSSTIngestion);
+    if (!status.ok()) {
+      ROCKS_LOG_WARN(db_options_.info_log,
+                     "Failed to verify checksum for external file %s: %s",
+                     external_file.c_str(), status.ToString().c_str());
+      return status;
+    }
+  }
+
+  // Assign FD with number.
+  file_to_ingest->fd =
+      FileDescriptor(new_file_number, 0, file_to_ingest->file_size);
+
   const size_t ts_sz = ucmp_->timestamp_size();
   Slice smallest = file_to_ingest->smallest_internal_key.user_key();
   Slice largest = file_to_ingest->largest_internal_key.user_key();

db/external_sst_file_ingestion_job.h

@@ -181,6 +181,9 @@ struct IngestedFileInfo : public KeyRangeInfo {
   // setting.
   bool user_defined_timestamps_persisted = true;
 
+  // Whether Lmax commit-time table opening should prefetch index/filter blocks.
+  bool prefetch_lmax_index_and_filter_blocks = true;
+
   SequenceNumber largest_seqno = kMaxSequenceNumber;
   SequenceNumber smallest_seqno = kMaxSequenceNumber;
 };
@@ -248,6 +251,7 @@ class ExternalSstFileIngestionJob {
   Status Prepare(const std::vector<std::string>& external_files_paths,
                  const std::vector<std::string>& files_checksums,
                  const std::vector<std::string>& files_checksum_func_names,
+                 const std::vector<const PreparedFileInfo*>& file_infos,
                  const std::optional<RangeOpt>& atomic_replace_range,
                  const Temperature& file_temperature, uint64_t next_file_number,
                  SuperVersion* sv);
@@ -301,6 +305,20 @@ class ExternalSstFileIngestionJob {
     return max_assigned_seqno_;
   }
 
+  // Max threads requested for opening this job's files during commit, from the
+  // per-CF IngestExternalFileOptions.
+  int file_opening_threads() const {
+    return ingestion_options_.file_opening_threads;
+  }
+
+  // Merge another already-Prepare()d job for the SAME column family into this
+  // one so both sets of files are committed by a single Run(). The other job's
+  // files are appended after this job's, so for any overlapping keys the other
+  // job's data wins via a higher assigned sequence number -- the same semantics
+  // as passing all the files to a single ingestion call in this order. The
+  // other job is left empty.
+  Status MergeForSameColumnFamily(ExternalSstFileIngestionJob* other);
+
  private:
   Status ResetTableReader(const std::string& external_file,
                           uint64_t new_file_number,
@@ -315,17 +333,37 @@ class ExternalSstFileIngestionJob {
   // different options. For example: when external file does not contain
   // timestamps while column family enables UDT in Memtables only feature.
   Status SanityCheckTableProperties(const std::string& external_file,
-                                    uint64_t new_file_number, SuperVersion* sv,
-                                    IngestedFileInfo* file_to_ingest,
-                                    std::unique_ptr<TableReader>* table_reader);
+                                    const TableProperties& props,
+                                    IngestedFileInfo* file_to_ingest);
 
   // Open the external file and populate `file_to_ingest` with all the
-  // external information we need to ingest this file.
+  // external information we need to ingest this file. When
+  // `prepared_file_info` is non-null, its caller-supplied metadata is reused
+  // instead of opening and scanning the file.
   Status GetIngestedFileInfo(const std::string& external_file,
                              uint64_t new_file_number,
+                             const PreparedFileInfo* prepared_file_info,
                              IngestedFileInfo* file_to_ingest,
                              SuperVersion* sv);
 
+  // Acquire the per-file metadata from the caller-supplied opaque
+  // `PreparedFileInfo` (produced by SstFileWriter::Finish) instead of opening
+  // the file.
+  Status GetIngestedFileInfoFromFileInfo(
+      const std::string& external_file,
+      const PreparedFileInfo& prepared_file_info,
+      IngestedFileInfo* file_to_ingest);
+
+  // Acquire the per-file metadata by opening the external file and scanning it
+  // (table properties, sequence number bounds, and boundary keys including any
+  // range-tombstone extensions). Used when no file_info is available. The
+  // opened `TableReader` is returned via `*table_reader` so the caller can
+  // reuse it (e.g. to verify the file checksum) without re-opening the file.
+  Status GetIngestedFileInfoFromFile(
+      const std::string& external_file, uint64_t new_file_number,
+      IngestedFileInfo* file_to_ingest, SuperVersion* sv,
+      std::unique_ptr<TableReader>* out_table_reader);
+
   // If the input files' key range overlaps themselves, this function divides
   // them in the user specified order into multiple batches. Where the files
   // within a batch do not overlap with each other, but key range could overlap
@@ -334,6 +372,10 @@ class ExternalSstFileIngestionJob {
   // make one batch.
   void DivideInputFilesIntoBatches();
 
+  // Returns whether any two files in `files` have overlapping key ranges, by
+  // sorting the file ranges and checking adjacent pairs.
+  bool ComputeFilesOverlap(const autovector<IngestedFileInfo>& files) const;
+
   // Assign level for the files in one batch. The files within one batch are not
   // overlapping, and we assign level to each file one after another.
   // If `prev_batch_uppermost_level` is specified, all files in this batch will
@@ -407,7 +449,7 @@ class ExternalSstFileIngestionJob {
   SnapshotList* db_snapshots_;
   autovector<IngestedFileInfo> files_to_ingest_;
   std::vector<FileBatchInfo> file_batches_to_ingest_;
-  const IngestExternalFileOptions& ingestion_options_;
+  const IngestExternalFileOptions ingestion_options_;
   std::optional<KeyRangeInfo> atomic_replace_range_;
   Directories* directories_;
   EventLogger* event_logger_;

db/external_sst_file_test.cc

@@ -5,6 +5,7 @@
 
 #include <table/block_based/block_based_table_factory.h>
 
+#include <atomic>
 #include <functional>
 #include <memory>
 #include <sstream>
@@ -17,6 +18,7 @@
 #include "port/stack_trace.h"
 #include "rocksdb/sst_file_reader.h"
 #include "rocksdb/sst_file_writer.h"
+#include "table/prepared_file_info.h"
 #include "test_util/testutil.h"
 #include "util/random.h"
 #include "util/thread_guard.h"
@@ -92,10 +94,17 @@ class ExternSSTFileLinkFailFallbackTest
 
 class ExternalSSTFileTest
     : public ExternalSSTFileTestBase,
-      public ::testing::WithParamInterface<std::tuple<bool, bool>> {
+      public ::testing::WithParamInterface<std::tuple<bool, bool, bool>> {
  public:
   ExternalSSTFileTest() = default;
 
+  void SetUp() override {
+    const auto* info = ::testing::UnitTest::GetInstance()->current_test_info();
+    if (info != nullptr && info->value_param() != nullptr) {
+      two_phase_ingest_ = std::get<2>(GetParam());
+    }
+  }
+
   Status GenerateOneExternalFile(
       const Options& options, ColumnFamilyHandle* cfh,
       std::vector<std::pair<std::string, std::string>>& data, int file_id,
@@ -152,7 +161,8 @@ class ExternalSSTFileTest
       bool verify_checksums_before_ingest = true, bool ingest_behind = false,
       bool sort_data = false,
       std::map<std::string, std::string>* true_data = nullptr,
-      ColumnFamilyHandle* cfh = nullptr, bool fill_cache = false) {
+      ColumnFamilyHandle* cfh = nullptr, bool fill_cache = false,
+      bool ingest_with_file_info = false) {
     // Generate a file id if not provided
     if (file_id == -1) {
       file_id = last_file_id_ + 1;
@@ -188,7 +198,12 @@ class ExternalSSTFileTest
         return s;
       }
     }
-    s = sst_file_writer.Finish();
+    ExternalSstFileInfo file_info;
+    if (ingest_with_file_info) {
+      s = sst_file_writer.Finish(&file_info);
+    } else {
+      s = sst_file_writer.Finish();
+    }
 
     if (s.ok()) {
       IngestExternalFileOptions ifo;
@@ -197,7 +212,21 @@ class ExternalSSTFileTest
       ifo.verify_checksums_before_ingest = verify_checksums_before_ingest;
       ifo.ingest_behind = ingest_behind;
       ifo.fill_cache = fill_cache;
-      if (cfh) {
+      if (two_phase_ingest_) {
+        std::unique_ptr<FileIngestionHandle> handle;
+        s = db_->PrepareFileIngestion(cfh ? cfh : db_->DefaultColumnFamily(),
+                                      {file_path}, ifo, &handle);
+        if (s.ok()) {
+          s = db_->CommitFileIngestionHandle(std::move(handle));
+        }
+      } else if (ingest_with_file_info) {
+        IngestExternalFileArg arg;
+        arg.column_family = cfh ? cfh : db_->DefaultColumnFamily();
+        arg.external_files = {file_path};
+        arg.file_infos = {file_info.prepared_file_info.get()};
+        arg.options = ifo;
+        s = db_->IngestExternalFiles({arg});
+      } else if (cfh) {
         s = db_->IngestExternalFile(cfh, {file_path}, ifo);
       } else {
         s = db_->IngestExternalFile({file_path}, ifo);
@@ -243,6 +272,14 @@ class ExternalSSTFileTest
       args[i].external_files.push_back(external_file_path);
       args[i].options = ifos[i];
     }
+    if (two_phase_ingest_) {
+      std::unique_ptr<FileIngestionHandle> handle;
+      Status s = db_->PrepareFileIngestion(args, &handle);
+      if (!s.ok()) {
+        return s;
+      }
+      return db_->CommitFileIngestionHandle(std::move(handle));
+    }
     return db_->IngestExternalFiles(args);
   }
 
@@ -294,10 +331,619 @@ class ExternalSSTFileTest
     return db_->IngestExternalFile(files, opts);
   }
 
+  // Writes `data` (sorted by user key) into a new external SST and returns its
+  // path in *file_path WITHOUT ingesting it, so a test can drive the two-phase
+  // PrepareFileIngestion()/CommitFileIngestion() API directly.
+  Status GenerateExternalFileOnly(
+      const Options& options,
+      std::vector<std::pair<std::string, std::string>> data,
+      std::string* file_path) {
+    std::sort(data.begin(), data.end(),
+              [&](const std::pair<std::string, std::string>& e1,
+                  const std::pair<std::string, std::string>& e2) {
+                return options.comparator->Compare(e1.first, e2.first) < 0;
+              });
+    *file_path = sst_files_dir_ + "only_" + std::to_string(++last_file_id_);
+    SstFileWriter sst_file_writer(EnvOptions(), options);
+    Status s = sst_file_writer.Open(*file_path);
+    if (!s.ok()) {
+      return s;
+    }
+    for (const auto& entry : data) {
+      s = sst_file_writer.Put(entry.first, entry.second);
+      if (!s.ok()) {
+        sst_file_writer.Finish().PermitUncheckedError();
+        return s;
+      }
+    }
+    return sst_file_writer.Finish();
+  }
+
  protected:
   int last_file_id_ = 0;
+  bool two_phase_ingest_ = false;
 };
 
+TEST_F(ExternalSSTFileTest, IngestionTimingHistogram) {
+  Options options = CurrentOptions();
+  options.statistics = CreateDBStatistics();
+  options.statistics->set_stats_level(StatsLevel::kAll);  // enable timers
+  DestroyAndReopen(options);
+
+  ASSERT_OK(GenerateAndAddExternalFile(
+      options, {{"k1", "v1"}, {"k2", "v2"}, {"k3", "v3"}}, /*file_id=*/-1,
+      /*allow_global_seqno=*/true, /*write_global_seqno=*/false,
+      /*verify_checksums_before_ingest=*/true, /*ingest_behind=*/false,
+      /*sort_data=*/true));
+
+  auto hist = [&](Histograms h) {
+    HistogramData hd;
+    options.statistics->histogramData(h, &hd);
+    return hd;
+  };
+
+  // One sample per call (not per CF) in each phase histogram.
+  HistogramData prepare_hd = hist(INGEST_EXTERNAL_FILE_PREPARE_TIME);
+  HistogramData run_hd = hist(INGEST_EXTERNAL_FILE_RUN_TIME);
+  ASSERT_EQ(1, prepare_hd.count);
+  ASSERT_EQ(1, run_hd.count);
+  ASSERT_GT(prepare_hd.max, 0.0);
+  ASSERT_GT(run_hd.max, 0.0);
+
+  // A second call adds exactly one more sample to each histogram.
+  ASSERT_OK(GenerateAndAddExternalFile(
+      options, {{"k4", "v4"}, {"k5", "v5"}}, /*file_id=*/-1,
+      /*allow_global_seqno=*/true, /*write_global_seqno=*/false,
+      /*verify_checksums_before_ingest=*/true, /*ingest_behind=*/false,
+      /*sort_data=*/true));
+  ASSERT_EQ(2, hist(INGEST_EXTERNAL_FILE_PREPARE_TIME).count);
+  ASSERT_EQ(2, hist(INGEST_EXTERNAL_FILE_RUN_TIME).count);
+
+  // A failed ingestion records nothing (latency is logged only on success).
+  ASSERT_NOK(db_->IngestExternalFile({"/path/does/not/exist.sst"},
+                                     IngestExternalFileOptions()));
+  ASSERT_EQ(2, hist(INGEST_EXTERNAL_FILE_PREPARE_TIME).count);
+  ASSERT_EQ(2, hist(INGEST_EXTERNAL_FILE_RUN_TIME).count);
+}
+
+TEST_F(ExternalSSTFileTest, PrepareThenCommit) {
+  Options options = CurrentOptions();
+  DestroyAndReopen(options);
+
+  std::string file_path;
+  ASSERT_OK(GenerateExternalFileOnly(
+      options, {{"k1", "v1"}, {"k2", "v2"}, {"k3", "v3"}}, &file_path));
+
+  IngestExternalFileArg arg;
+  arg.column_family = db_->DefaultColumnFamily();
+  arg.external_files = {file_path};
+
+  std::unique_ptr<FileIngestionHandle> handle;
+  ASSERT_OK(db_->PrepareFileIngestion({arg}, &handle));
+  ASSERT_TRUE(handle != nullptr);
+
+  // The prepared file is staged in the DB but not yet visible.
+  ASSERT_EQ("NOT_FOUND", Get("k1"));
+
+  ASSERT_OK(db_->CommitFileIngestionHandle(std::move(handle)));
+  ASSERT_EQ("v1", Get("k1"));
+  ASSERT_EQ("v2", Get("k2"));
+  ASSERT_EQ("v3", Get("k3"));
+}
+
+TEST_F(ExternalSSTFileTest, ParallelFileOpenWithFileOpeningThreads) {
+  // Ingest many non-overlapping files with file_opening_threads > 1 so commit
+  // opens table readers with multiple threads. max_open_files == -1 makes
+  // commit open every new file.
+  Options options = CurrentOptions();
+  options.max_open_files = -1;
+  DestroyAndReopen(options);
+
+  constexpr int kNumFiles = 8;
+  constexpr int kKeysPerFile = 50;
+  std::vector<std::string> files;
+  std::map<std::string, std::string> true_data;
+  for (int f = 0; f < kNumFiles; f++) {
+    std::vector<std::pair<std::string, std::string>> data;
+    for (int k = 0; k < kKeysPerFile; k++) {
+      const int key = f * kKeysPerFile + k;
+      const std::string value = "val" + std::to_string(key);
+      data.emplace_back(Key(key), value);
+      true_data[Key(key)] = value;
+    }
+    std::string file_path;
+    ASSERT_OK(GenerateExternalFileOnly(options, data, &file_path));
+    files.push_back(file_path);
+  }
+
+  IngestExternalFileOptions ifo;
+  ifo.file_opening_threads = 4;
+  ASSERT_OK(db_->IngestExternalFile(files, ifo));
+
+  for (const auto& [key, value] : true_data) {
+    ASSERT_EQ(value, Get(key));
+  }
+}
+
+TEST_F(ExternalSSTFileTest, LmaxPrefetchSkipDoesNotDisableL0Prefetch) {
+  LRUCacheOptions co;
+  co.capacity = 32 << 20;
+  std::shared_ptr<Cache> cache = NewLRUCache(co);
+  BlockBasedTableOptions table_options;
+  table_options.block_cache = cache;
+  table_options.cache_index_and_filter_blocks = true;
+  table_options.filter_policy.reset(NewBloomFilterPolicy(10));
+
+  Options options = CurrentOptions();
+  options.disable_auto_compactions = true;
+  options.max_open_files = -1;
+  options.num_levels = 2;
+  options.optimize_filters_for_hits = false;
+  options.statistics = CreateDBStatistics();
+  options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+  DestroyAndReopen(options);
+
+  const auto write_file = [&](const std::string& file_path,
+                              const std::string& value,
+                              ExternalSstFileInfo* file_info) {
+    SstFileWriter writer(EnvOptions(), options);
+    ASSERT_OK(writer.Open(file_path));
+    ASSERT_OK(writer.Put(Key(10), value));
+    ASSERT_OK(writer.Finish(file_info));
+  };
+
+  const auto ingest_file = [&](const std::string& file_path,
+                               const ExternalSstFileInfo& file_info,
+                               bool fail_if_not_bottommost_level) {
+    IngestExternalFileArg arg;
+    arg.column_family = db_->DefaultColumnFamily();
+    arg.external_files = {file_path};
+    arg.file_infos = {file_info.prepared_file_info.get()};
+    arg.options.fail_if_not_bottommost_level = fail_if_not_bottommost_level;
+    arg.options.verify_checksums_before_ingest = false;
+    arg.options.prefetch_lmax_index_and_filter_blocks = false;
+    ASSERT_OK(db_->IngestExternalFiles({arg}));
+  };
+
+  const std::string lmax_file_path = sst_files_dir_ + "lazy_lmax.sst";
+  ExternalSstFileInfo lmax_file_info;
+  write_file(lmax_file_path, "v10", &lmax_file_info);
+
+  ASSERT_EQ(0, options.statistics->getAndResetTickerCount(
+                   Tickers::BLOCK_CACHE_INDEX_ADD));
+  ASSERT_EQ(0, options.statistics->getAndResetTickerCount(
+                   Tickers::BLOCK_CACHE_FILTER_ADD));
+
+  ingest_file(lmax_file_path, lmax_file_info,
+              true /* fail_if_not_bottommost_level */);
+  ASSERT_EQ("0,1", FilesPerLevel());
+
+  ASSERT_EQ(0, options.statistics->getAndResetTickerCount(
+                   Tickers::BLOCK_CACHE_INDEX_ADD));
+  ASSERT_EQ(0, options.statistics->getAndResetTickerCount(
+                   Tickers::BLOCK_CACHE_FILTER_ADD));
+
+  const std::string l0_file_path = sst_files_dir_ + "l0.sst";
+  ExternalSstFileInfo l0_file_info;
+  write_file(l0_file_path, "v11", &l0_file_info);
+
+  ASSERT_EQ(0, options.statistics->getAndResetTickerCount(
+                   Tickers::BLOCK_CACHE_INDEX_ADD));
+  ASSERT_EQ(0, options.statistics->getAndResetTickerCount(
+                   Tickers::BLOCK_CACHE_FILTER_ADD));
+
+  ingest_file(l0_file_path, l0_file_info,
+              false /* fail_if_not_bottommost_level */);
+  ASSERT_EQ("1,1", FilesPerLevel());
+
+  EXPECT_GT(options.statistics->getAndResetTickerCount(
+                Tickers::BLOCK_CACHE_INDEX_ADD),
+            0);
+  EXPECT_GT(options.statistics->getAndResetTickerCount(
+                Tickers::BLOCK_CACHE_FILTER_ADD),
+            0);
+}
+
+TEST_F(ExternalSSTFileTest, AbortPreparedIngestion) {
+  Options options = CurrentOptions();
+  DestroyAndReopen(options);
+
+  std::string file_path;
+  ASSERT_OK(GenerateExternalFileOnly(options, {{"x", "1"}}, &file_path));
+  IngestExternalFileArg arg;
+  arg.column_family = db_->DefaultColumnFamily();
+  arg.external_files = {file_path};
+
+  std::unique_ptr<FileIngestionHandle> handle;
+  ASSERT_OK(db_->PrepareFileIngestion({arg}, &handle));
+  ASSERT_OK(handle->Abort());
+  ASSERT_EQ("NOT_FOUND", Get("x"));
+
+  // Committing an already-aborted handle returns an error.
+  ASSERT_TRUE(
+      db_->CommitFileIngestionHandle(std::move(handle)).IsInvalidArgument());
+
+  // The DB remains healthy after an abort: a fresh prepare + commit works.
+  std::unique_ptr<FileIngestionHandle> handle2;
+  ASSERT_OK(db_->PrepareFileIngestion({arg}, &handle2));
+  ASSERT_OK(db_->CommitFileIngestionHandle(std::move(handle2)));
+  ASSERT_EQ("1", Get("x"));
+}
+
+TEST_F(ExternalSSTFileTest, MultiHandleAtomicCommit) {
+  Options options = CurrentOptions();
+  CreateAndReopenWithCF({"cf1", "cf2", "cf3"}, options);
+
+  // Two handles whose column families overlap on cf2:
+  //   handle 1 -> cf1, cf2     handle 2 -> cf2, cf3
+  // Committing them together installs cf1 and cf3 each from one handle, while
+  // cf2 receives files from BOTH handles, merged into a single per-CF commit.
+  std::string f_cf1;
+  std::string f_cf2_h1;
+  std::string f_cf2_h2;
+  std::string f_cf3;
+  ASSERT_OK(GenerateExternalFileOnly(options, {{"a", "1"}}, &f_cf1));
+  // The cf2 files share key "b": handle 2 is committed later, so its value must
+  // win via a higher assigned sequence number.
+  ASSERT_OK(GenerateExternalFileOnly(options, {{"b", "h1"}, {"b1", "10"}},
+                                     &f_cf2_h1));
+  ASSERT_OK(GenerateExternalFileOnly(options, {{"b", "h2"}, {"b2", "20"}},
+                                     &f_cf2_h2));
+  ASSERT_OK(GenerateExternalFileOnly(options, {{"c", "3"}}, &f_cf3));
+
+  std::vector<IngestExternalFileArg> args1(2);
+  args1[0].column_family = handles_[1];  // cf1
+  args1[0].external_files = {f_cf1};
+  args1[1].column_family = handles_[2];  // cf2
+  args1[1].external_files = {f_cf2_h1};
+
+  std::vector<IngestExternalFileArg> args2(2);
+  args2[0].column_family =
+      handles_[2];  // cf2 -> merges with handle 1's cf2 job
+  args2[0].external_files = {f_cf2_h2};
+  args2[1].column_family = handles_[3];  // cf3
+  args2[1].external_files = {f_cf3};
+
+  std::unique_ptr<FileIngestionHandle> h1;
+  std::unique_ptr<FileIngestionHandle> h2;
+  ASSERT_OK(db_->PrepareFileIngestion(args1, &h1));
+  ASSERT_OK(db_->PrepareFileIngestion(args2, &h2));
+  ASSERT_EQ("NOT_FOUND", Get(1, "a"));
+  ASSERT_EQ("NOT_FOUND", Get(2, "b1"));
+  ASSERT_EQ("NOT_FOUND", Get(3, "c"));
+
+  // Commit both handles atomically.
+  std::vector<std::unique_ptr<FileIngestionHandle>> batch;
+  batch.push_back(std::move(h1));
+  batch.push_back(std::move(h2));
+  ASSERT_OK(db_->CommitFileIngestionHandles(std::move(batch)));
+
+  ASSERT_EQ("1", Get(1, "a"));
+  ASSERT_EQ("h2", Get(2, "b"));   // overlapping key: later handle (h2) wins
+  ASSERT_EQ("10", Get(2, "b1"));  // cf2 from handle 1
+  ASSERT_EQ("20", Get(2, "b2"));  // cf2 from handle 2 (merged)
+  ASSERT_EQ("3", Get(3, "c"));
+}
+
+TEST_F(ExternalSSTFileTest, MultiHandleSameColumnFamilyIncompatibleOptions) {
+  Options options = CurrentOptions();
+  DestroyAndReopen(options);
+
+  std::string f1;
+  std::string f2;
+  ASSERT_OK(GenerateExternalFileOnly(options, {{"a", "1"}}, &f1));
+  ASSERT_OK(GenerateExternalFileOnly(options, {{"b", "2"}}, &f2));
+
+  // Same column family but with options the merged Run() would consult set
+  // differently across the two handles.
+  IngestExternalFileOptions opts_fill;
+  IngestExternalFileOptions opts_no_fill;
+  opts_no_fill.fill_cache = false;
+  std::unique_ptr<FileIngestionHandle> h1;
+  std::unique_ptr<FileIngestionHandle> h2;
+  ASSERT_OK(db_->PrepareFileIngestion(db_->DefaultColumnFamily(), {f1},
+                                      opts_fill, &h1));
+  ASSERT_OK(db_->PrepareFileIngestion(db_->DefaultColumnFamily(), {f2},
+                                      opts_no_fill, &h2));
+
+  std::vector<std::unique_ptr<FileIngestionHandle>> batch;
+  batch.push_back(std::move(h1));
+  batch.push_back(std::move(h2));
+  ASSERT_TRUE(
+      db_->CommitFileIngestionHandles(std::move(batch)).IsInvalidArgument());
+
+  // The rejected commit rolled both handles back; the CF is still usable.
+  ASSERT_EQ("NOT_FOUND", Get("a"));
+  ASSERT_EQ("NOT_FOUND", Get("b"));
+  std::unique_ptr<FileIngestionHandle> h3;
+  ASSERT_OK(db_->PrepareFileIngestion(db_->DefaultColumnFamily(), {f1},
+                                      IngestExternalFileOptions(), &h3));
+  ASSERT_OK(db_->CommitFileIngestionHandle(std::move(h3)));
+  ASSERT_EQ("1", Get("a"));
+}
+
+TEST_F(ExternalSSTFileTest, PrepareFileIngestionValidationErrors) {
+  Options options = CurrentOptions();
+  DestroyAndReopen(options);
+
+  std::unique_ptr<FileIngestionHandle> handle;
+
+  ASSERT_TRUE(
+      db_->PrepareFileIngestion(std::vector<IngestExternalFileArg>{}, nullptr)
+          .IsInvalidArgument());
+
+  // Empty arg list.
+  ASSERT_TRUE(
+      db_->PrepareFileIngestion(std::vector<IngestExternalFileArg>{}, &handle)
+          .IsInvalidArgument());
+  ASSERT_TRUE(handle == nullptr);
+
+  // move_files and link_files are mutually exclusive.
+  std::string file_path;
+  ASSERT_OK(GenerateExternalFileOnly(options, {{"k", "v"}}, &file_path));
+  IngestExternalFileArg arg;
+  arg.column_family = db_->DefaultColumnFamily();
+  arg.external_files = {file_path};
+  arg.options.move_files = true;
+  arg.options.link_files = true;
+  ASSERT_TRUE(db_->PrepareFileIngestion({arg}, &handle).IsInvalidArgument());
+  ASSERT_TRUE(handle == nullptr);
+}
+
+// Ingestion that reuses the SstFileWriter's metadata via
+// IngestExternalFileArg::file_infos skips the open-and-scan path and still
+// produces correct data, including overlapping data that is reassigned a new
+// global sequence number.
+TEST_F(ExternalSSTFileTest, IngestWithFileInfo) {
+  Options options = CurrentOptions();
+  DestroyAndReopen(options);
+  std::map<std::string, std::string> true_data;
+
+  // Count entries into the open-and-scan path so we can assert it is skipped.
+  std::atomic<int> read_path_count{0};
+  SyncPoint::GetInstance()->SetCallBack(
+      "ExternalSstFileIngestionJob::GetIngestedFileInfo:ReadPath",
+      [&](void*) { read_path_count.fetch_add(1); });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  // Reuse the writer's metadata, with the normal ingestion options: a global
+  // sequence number is assigned (allow_db_generated_files is NOT set) and the
+  // checksum is verified (the fast-path opens the file only to verify it, which
+  // is not a metadata scan).
+  auto ingest_fi = [&](std::vector<std::pair<std::string, std::string>> data) {
+    return GenerateAndAddExternalFile(
+        options, std::move(data), /*file_id=*/-1, /*allow_global_seqno=*/true,
+        /*write_global_seqno=*/false, /*verify_checksums_before_ingest=*/true,
+        /*ingest_behind=*/false, /*sort_data=*/true, &true_data,
+        /*cfh=*/nullptr, /*fill_cache=*/false, /*ingest_with_file_info=*/true);
+  };
+
+  // Non-overlapping ingest: data reads back and the file is not scanned.
+  ASSERT_OK(ingest_fi({{Key(1), "a1"}, {Key(2), "a2"}, {Key(3), "a3"}}));
+  ASSERT_EQ(read_path_count.load(), 0);
+
+  // Overlap existing (flushed) data so a non-zero global seqno is assigned and
+  // the newer values win -- the normal seqno-reassignment path through
+  // file_infos.
+  ASSERT_OK(db_->Put(WriteOptions(), Key(2), "old2"));
+  true_data[Key(2)] = "old2";
+  ASSERT_OK(Flush());
+  ASSERT_OK(ingest_fi({{Key(2), "new2"}, {Key(4), "new4"}}));
+  ASSERT_EQ(read_path_count.load(), 0);
+
+  // Control: ingesting the same way but WITHOUT file_infos does enter the scan.
+  ASSERT_OK(GenerateAndAddExternalFile(
+      options, {{Key(8), "c8"}}, /*file_id=*/-1, /*allow_global_seqno=*/true,
+      /*write_global_seqno=*/false, /*verify_checksums_before_ingest=*/true,
+      /*ingest_behind=*/false, /*sort_data=*/true, &true_data));
+  ASSERT_GT(read_path_count.load(), 0);
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+  size_t kcnt = 0;
+  VerifyDBFromMap(true_data, &kcnt, false);
+}
+
+// Range deletions are honored on the file_infos fast-path: the writer's
+// range-del bounds are carried in the metadata, so a DeleteRange covering
+// existing keys still shadows them.
+TEST_F(ExternalSSTFileTest, IngestWithFileInfoRangeDeletion) {
+  Options options = CurrentOptions();
+  DestroyAndReopen(options);
+  ASSERT_OK(db_->Put(WriteOptions(), Key(20), "live20"));
+  ASSERT_OK(db_->Put(WriteOptions(), Key(25), "live25"));
+  ASSERT_OK(Flush());
+
+  const std::string f = sst_files_dir_ + "rangedel.sst";
+  SstFileWriter w(EnvOptions(), options);
+  ASSERT_OK(w.Open(f));
+  ASSERT_OK(w.Put(Key(30), "v30"));
+  ASSERT_OK(w.DeleteRange(Key(20), Key(26)));  // covers keys 20..25
+  ExternalSstFileInfo file_info;
+  ASSERT_OK(w.Finish(&file_info));
+
+  IngestExternalFileArg arg;
+  arg.column_family = db_->DefaultColumnFamily();
+  arg.external_files = {f};
+  arg.file_infos = {file_info.prepared_file_info.get()};
+  arg.options.allow_global_seqno = true;
+  ASSERT_OK(db_->IngestExternalFiles({arg}));
+
+  ReadOptions ro;
+  std::string val;
+  ASSERT_TRUE(db_->Get(ro, Key(20), &val).IsNotFound());
+  ASSERT_TRUE(db_->Get(ro, Key(25), &val).IsNotFound());
+  ASSERT_OK(db_->Get(ro, Key(30), &val));
+  ASSERT_EQ(val, "v30");
+}
+
+// verify_checksums_before_ingest is honored on the file_infos fast-path: it
+// opens the file just to verify, so corruption is still detected.
+TEST_F(ExternalSSTFileTest, IngestWithFileInfoVerifiesChecksum) {
+  Options options = CurrentOptions();
+  DestroyAndReopen(options);
+
+  const std::string f = sst_files_dir_ + "corrupt.sst";
+  SstFileWriter w(EnvOptions(), options);
+  ASSERT_OK(w.Open(f));
+  for (int k = 0; k < 100; k++) {
+    ASSERT_OK(w.Put(Key(k), "v" + Key(k)));
+  }
+  ExternalSstFileInfo file_info;
+  ASSERT_OK(w.Finish(&file_info));
+  ASSERT_OK(test::CorruptFile(options.env, f, /*offset=*/64,
+                              /*bytes_to_corrupt=*/8));
+
+  IngestExternalFileArg arg;
+  arg.column_family = db_->DefaultColumnFamily();
+  arg.external_files = {f};
+  arg.file_infos = {file_info.prepared_file_info.get()};
+  arg.options.verify_checksums_before_ingest = true;
+  ASSERT_TRUE(db_->IngestExternalFiles({arg}).IsCorruption());
+}
+
+// write_global_seqno is incompatible with file_infos: the file is not opened,
+// so the seqno cannot be written back into it.
+TEST_F(ExternalSSTFileTest, IngestWithFileInfoRejectsWriteGlobalSeqno) {
+  Options options = CurrentOptions();
+  DestroyAndReopen(options);
+
+  const std::string f = sst_files_dir_ + "wgs.sst";
+  SstFileWriter w(EnvOptions(), options);
+  ASSERT_OK(w.Open(f));
+  ASSERT_OK(w.Put(Key(1), "v1"));
+  ExternalSstFileInfo file_info;
+  ASSERT_OK(w.Finish(&file_info));
+
+  IngestExternalFileArg arg;
+  arg.column_family = db_->DefaultColumnFamily();
+  arg.external_files = {f};
+  arg.file_infos = {file_info.prepared_file_info.get()};
+  arg.options.write_global_seqno = true;
+  ASSERT_TRUE(db_->IngestExternalFiles({arg}).IsInvalidArgument());
+}
+
+TEST_F(ExternalSSTFileTest, GetPreparedFileInfoForExternalSstIngestion) {
+  Options options = CurrentOptions();
+  DestroyAndReopen(options);
+
+  const std::string source_dbname = dbname_ + "_source";
+  ASSERT_OK(DestroyDB(source_dbname, options));
+  std::unique_ptr<DB> source_db;
+  ASSERT_OK(DB::Open(options, source_dbname, &source_db));
+
+  ASSERT_OK(source_db->Put(WriteOptions(), Key(10), "v10"));
+  ASSERT_OK(source_db->Flush(FlushOptions()));
+
+  std::vector<LiveFileMetaData> live_meta;
+  source_db->GetLiveFilesMetaData(&live_meta);
+  ASSERT_EQ(live_meta.size(), 1);
+  ASSERT_GT(live_meta[0].largest_seqno, 0);
+  const std::string file_path =
+      live_meta[0].directory + "/" + live_meta[0].relative_filename;
+
+  std::shared_ptr<const PreparedFileInfo> prepared_file_info;
+  ASSERT_TRUE(source_db
+                  ->GetPreparedFileInfoForExternalSstIngestion(
+                      live_meta[0].relative_filename, &prepared_file_info)
+                  .IsInvalidArgument());
+  ASSERT_EQ(prepared_file_info, nullptr);
+  prepared_file_info = std::make_shared<PreparedFileInfo>();
+  ASSERT_TRUE(source_db
+                  ->GetPreparedFileInfoForExternalSstIngestion(
+                      dbname_ + "_wrong/" + live_meta[0].relative_filename,
+                      &prepared_file_info)
+                  .IsInvalidArgument());
+  ASSERT_EQ(prepared_file_info, nullptr);
+
+  ASSERT_OK(source_db->GetPreparedFileInfoForExternalSstIngestion(
+      file_path, &prepared_file_info));
+  ASSERT_TRUE(prepared_file_info != nullptr);
+  ASSERT_EQ(prepared_file_info->file_size, live_meta[0].size);
+  ASSERT_EQ(prepared_file_info->table_properties.key_largest_seqno,
+            live_meta[0].largest_seqno);
+
+  IngestExternalFileArg arg;
+  arg.column_family = db_->DefaultColumnFamily();
+  arg.external_files = {file_path};
+  arg.file_infos = {prepared_file_info.get()};
+  arg.options.allow_db_generated_files = true;
+  arg.options.snapshot_consistency = false;
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "ExternalSstFileIngestionJob::GetIngestedFileInfo:ReadPath",
+      [](void*) { FAIL() << "Prepared file info should skip the read path"; });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(db_->IngestExternalFiles({arg}));
+
+  ASSERT_EQ(Get(Key(10)), "v10");
+
+  ASSERT_OK(source_db->Close());
+  source_db.reset();
+  ASSERT_OK(DestroyDB(source_dbname, options));
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+TEST_F(ExternalSSTFileTest,
+       IngestWithFileInfoRejectsMissingDbGeneratedSeqnoBounds) {
+  Options options = CurrentOptions();
+  DestroyAndReopen(options);
+
+  const std::string f = sst_files_dir_ + "missing_seqno_bounds.sst";
+  SstFileWriter w(EnvOptions(), options);
+  ASSERT_OK(w.Open(f));
+  ASSERT_OK(w.Put(Key(1), "v1"));
+  ExternalSstFileInfo file_info;
+  ASSERT_OK(w.Finish(&file_info));
+  auto prepared_file_info = file_info.prepared_file_info;
+
+  auto missing_seqno_bounds =
+      std::make_shared<PreparedFileInfo>(*prepared_file_info);
+  missing_seqno_bounds->table_properties.key_smallest_seqno = UINT64_MAX;
+  missing_seqno_bounds->table_properties.key_largest_seqno = UINT64_MAX;
+
+  IngestExternalFileArg arg;
+  arg.column_family = db_->DefaultColumnFamily();
+  arg.external_files = {f};
+  arg.file_infos = {missing_seqno_bounds.get()};
+  arg.options.allow_db_generated_files = true;
+  arg.options.snapshot_consistency = false;
+  Status s = db_->IngestExternalFiles({arg});
+  ASSERT_TRUE(s.IsCorruption());
+  ASSERT_NE(s.ToString().find("Unknown largest seqno"), std::string::npos);
+}
+
+TEST_F(ExternalSSTFileTest, IngestWithFileInfoPreparedKeyBoundTypes) {
+  Options options = CurrentOptions();
+  DestroyAndReopen(options);
+
+  const std::string f = sst_files_dir_ + "prepared_key_bound_types.sst";
+  SstFileWriter w(EnvOptions(), options);
+  ASSERT_OK(w.Open(f));
+  ASSERT_OK(w.Put(Key(10), "v10"));
+  ASSERT_OK(w.DeleteRange(Key(20), Key(30)));
+  ExternalSstFileInfo file_info;
+  ASSERT_OK(w.Finish(&file_info));
+  auto prepared_file_info = file_info.prepared_file_info;
+
+  ParsedInternalKey smallest;
+  ASSERT_OK(ParseInternalKey(prepared_file_info->smallest.Encode(), &smallest,
+                             false /* log_err_key */));
+  ASSERT_EQ(0, smallest.sequence);
+  ASSERT_EQ(kTypeValue, smallest.type);
+  ASSERT_EQ(Key(10), smallest.user_key.ToString());
+
+  ParsedInternalKey largest;
+  ASSERT_OK(ParseInternalKey(prepared_file_info->largest.Encode(), &largest,
+                             false /* log_err_key */));
+  ASSERT_EQ(kMaxSequenceNumber, largest.sequence);
+  ASSERT_EQ(kTypeRangeDeletion, largest.type);
+  ASSERT_EQ(Key(30), largest.user_key.ToString());
+}
+
 TEST_F(ExternalSSTFileTest, ComparatorMismatch) {
   Options options = CurrentOptions();
   Options options_diff_ucmp = options;
@@ -1816,7 +2462,8 @@ TEST_P(ExternalSSTFileTest, IngestFileWithGlobalSeqnoRandomized) {
     for (int i = 0; i < 500; i++) {
       std::vector<std::pair<std::string, std::string>> random_data;
       for (int j = 0; j < 100; j++) {
-        std::string k = rnd.RandomString(rnd.Next() % 20);
+        std::string k =
+            rnd.RandomString(rnd.Next() % 20 + 1);  // requires non-empty keys
         std::string v = rnd.RandomString(rnd.Next() % 50);
         random_data.emplace_back(k, v);
       }
@@ -3247,6 +3894,20 @@ class ExternalSSTFileWithTimestampTest : public ExternalSSTFileTest {
     return dst;
   }
 
+  std::string MaxTimestamp() {
+    Slice ts = MaxU64Ts();
+    return ts.ToString();
+  }
+
+  Options UDTMemtableOnlyOptions() {
+    Options options = CurrentOptions();
+    options.comparator = test::BytewiseComparatorWithU64TsWrapper();
+    options.persist_user_defined_timestamps = false;
+    options.allow_concurrent_memtable_write = false;
+    options.disable_auto_compactions = true;
+    return options;
+  }
+
   Status IngestExternalUDTFile(const std::vector<std::string>& files,
                                bool allow_global_seqno = true) {
     IngestExternalFileOptions opts;
@@ -3815,8 +4476,74 @@ TEST_F(ExternalSSTFileWithTimestampTest, TimestampsNotPersistedBasic) {
                          kRangeDelSkipConfigs));
 }
 
+TEST_F(ExternalSSTFileWithTimestampTest,
+       TimestampsPersistedIngestWithFileInfoRangeDeletionBounds) {
+  Options options = CurrentOptions();
+  options.comparator = test::BytewiseComparatorWithU64TsWrapper();
+  options.persist_user_defined_timestamps = true;
+  options.disable_auto_compactions = true;
+
+  DestroyAndReopen(options);
+
+  const std::string ts = EncodeAsUint64(10);
+  const std::string f = sst_files_dir_ + "udt_persisted_rangedel_bounds.sst";
+  SstFileWriter w(EnvOptions(), options);
+  ASSERT_OK(w.Open(f));
+  ASSERT_OK(w.Put(Key(30), ts, "v30"));
+  ASSERT_OK(w.DeleteRange(Key(20), Key(26), ts));
+  ExternalSstFileInfo file_info;
+  ASSERT_OK(w.Finish(&file_info));
+  auto prepared_file_info = file_info.prepared_file_info;
+  ASSERT_EQ(Key(20) + MaxTimestamp(),
+            prepared_file_info->smallest.user_key().ToString());
+  ASSERT_EQ(Key(30) + ts, prepared_file_info->largest.user_key().ToString());
+
+  IngestExternalFileArg arg;
+  arg.column_family = db_->DefaultColumnFamily();
+  arg.external_files = {f};
+  arg.file_infos = {file_info.prepared_file_info.get()};
+  ASSERT_OK(db_->IngestExternalFiles({arg}));
+
+  std::vector<LiveFileMetaData> metadata;
+  db_->GetLiveFilesMetaData(&metadata);
+  ASSERT_EQ(1, metadata.size());
+  ASSERT_EQ(Key(20) + MaxTimestamp(), metadata[0].smallestkey);
+  ASSERT_EQ(Key(30) + ts, metadata[0].largestkey);
+}
+
+TEST_F(ExternalSSTFileWithTimestampTest,
+       TimestampsNotPersistedIngestWithFileInfoRangeDeletionBounds) {
+  Options options = UDTMemtableOnlyOptions();
+
+  DestroyAndReopen(options);
+
+  const std::string f = sst_files_dir_ + "udt_rangedel_bounds.sst";
+  SstFileWriter w(EnvOptions(), options);
+  ASSERT_OK(w.Open(f));
+  ASSERT_OK(w.Put(Key(30), EncodeAsUint64(0), "v30"));
+  ASSERT_OK(w.DeleteRange(Key(20), Key(26), EncodeAsUint64(0)));
+  ExternalSstFileInfo file_info;
+  ASSERT_OK(w.Finish(&file_info));
+  auto prepared_file_info = file_info.prepared_file_info;
+  ASSERT_EQ(Key(20), prepared_file_info->smallest.user_key().ToString());
+  ASSERT_EQ(Key(30), prepared_file_info->largest.user_key().ToString());
+
+  IngestExternalFileArg arg;
+  arg.column_family = db_->DefaultColumnFamily();
+  arg.external_files = {f};
+  arg.file_infos = {file_info.prepared_file_info.get()};
+  ASSERT_OK(db_->IngestExternalFiles({arg}));
+
+  std::vector<LiveFileMetaData> metadata;
+  db_->GetLiveFilesMetaData(&metadata);
+  ASSERT_EQ(1, metadata.size());
+  ASSERT_EQ(Key(20) + MaxTimestamp(), metadata[0].smallestkey);
+  ASSERT_EQ(Key(30) + EncodeAsUint64(0), metadata[0].largestkey);
+}
+
 INSTANTIATE_TEST_CASE_P(ExternalSSTFileTest, ExternalSSTFileTest,
-                        testing::Combine(testing::Bool(), testing::Bool()));
+                        testing::Combine(testing::Bool(), testing::Bool(),
+                                         testing::Bool()));
 
 class IngestDBGeneratedFileTest
     : public ExternalSSTFileTestBase,

db/fault_injection_test.cc

@@ -646,6 +646,78 @@ TEST(FaultInjectionFSTest, ReadUnsyncedData) {
     ASSERT_EQ(0, sl.compare(Slice(data.data() + first_read_len, sl.size())));
   }
 }
+
+TEST(FaultInjectionFSTest, FileOpenContractRejectsReadersWhileOpenForWrite) {
+  std::shared_ptr<FaultInjectionTestFS> fault_fs =
+      std::make_shared<FaultInjectionTestFS>(FileSystem::Default());
+  const std::string fname = test::PerThreadDBPath(
+      "file_open_contract_no_readers_while_open_for_write");
+
+  FileOptions writer_options;
+  writer_options.open_contract = FileOpenContract::kNoReadersWhileOpenForWrite;
+
+  std::unique_ptr<FSWritableFile> writer;
+  ASSERT_OK(fault_fs->NewWritableFile(fname, writer_options, &writer, nullptr));
+
+  std::unique_ptr<FSRandomAccessFile> reader;
+  IOStatus s =
+      fault_fs->NewRandomAccessFile(fname, FileOptions(), &reader, nullptr);
+  ASSERT_TRUE(s.IsNotSupported());
+
+  ASSERT_OK(writer->Close(IOOptions(), nullptr));
+  ASSERT_OK(
+      fault_fs->NewRandomAccessFile(fname, FileOptions(), &reader, nullptr));
+}
+
+TEST(FaultInjectionFSTest,
+     FileOpenContractAllowsSyncFileButRejectsReopenedWrite) {
+  std::shared_ptr<FaultInjectionTestFS> fault_fs =
+      std::make_shared<FaultInjectionTestFS>(FileSystem::Default());
+  const std::string fname =
+      test::PerThreadDBPath("file_open_contract_no_reopen_for_write");
+
+  FileOptions writer_options;
+  writer_options.open_contract = FileOpenContract::kNoReopenForWrite;
+
+  std::unique_ptr<FSWritableFile> writer;
+  ASSERT_OK(fault_fs->NewWritableFile(fname, writer_options, &writer, nullptr));
+  ASSERT_OK(writer->Append("abc", IOOptions(), nullptr));
+  ASSERT_OK(writer->Close(IOOptions(), nullptr));
+
+  ASSERT_OK(fault_fs->SyncFile(fname, FileOptions(), IOOptions(),
+                               /*use_fsync=*/false, nullptr));
+
+  std::unique_ptr<FSWritableFile> reopened_writer;
+  ASSERT_OK(fault_fs->ReopenWritableFile(fname, FileOptions(), &reopened_writer,
+                                         nullptr));
+  IOStatus s = reopened_writer->Append("d", IOOptions(), nullptr);
+  ASSERT_TRUE(s.IsNotSupported());
+  ASSERT_OK(reopened_writer->Close(IOOptions(), nullptr));
+}
+
+TEST(FaultInjectionFSTest, FileOpenContractAllowsCreateAfterDelete) {
+  std::shared_ptr<FaultInjectionTestFS> fault_fs =
+      std::make_shared<FaultInjectionTestFS>(FileSystem::Default());
+  const std::string fname =
+      test::PerThreadDBPath("file_open_contract_recreate_after_delete");
+
+  FileOptions writer_options;
+  writer_options.open_contract = FileOpenContract::kNoReopenForWrite;
+
+  std::unique_ptr<FSWritableFile> writer;
+  ASSERT_OK(fault_fs->NewWritableFile(fname, writer_options, &writer, nullptr));
+  ASSERT_OK(writer->Append("abc", IOOptions(), nullptr));
+  ASSERT_OK(writer->Close(IOOptions(), nullptr));
+
+  ASSERT_OK(FileSystem::Default()->DeleteFile(fname, IOOptions(), nullptr));
+
+  std::unique_ptr<FSWritableFile> recreated_writer;
+  ASSERT_OK(fault_fs->NewWritableFile(fname, writer_options, &recreated_writer,
+                                      nullptr));
+  ASSERT_OK(recreated_writer->Append("def", IOOptions(), nullptr));
+  ASSERT_OK(recreated_writer->Close(IOOptions(), nullptr));
+  ASSERT_OK(fault_fs->DeleteFile(fname, IOOptions(), nullptr));
+}
 }  // namespace ROCKSDB_NAMESPACE
 
 int main(int argc, char** argv) {

db/listener_test.cc

@@ -31,6 +31,7 @@
 #include "util/mutexlock.h"
 #include "util/rate_limiter_impl.h"
 #include "util/string_util.h"
+#include "utilities/fault_injection_fs.h"
 #include "utilities/merge_operators.h"
 
 namespace ROCKSDB_NAMESPACE {
@@ -201,6 +202,194 @@ TEST_F(EventListenerTest, OnSingleDBCompactionTest) {
   }
 }
 
+// Listener that asserts OnCompactionPreCommit fires strictly between
+// OnCompactionBegin and OnCompactionCompleted, and that input files'
+// being_compacted flag is still true at that point.
+class TestCompactionPreCommitListener : public EventListener {
+ public:
+  explicit TestCompactionPreCommitListener(EventListenerTest* test)
+      : test_(test) {}
+
+  void OnCompactionBegin(DB* /*db*/, const CompactionJobInfo& ci) override {
+    std::lock_guard<std::mutex> lock(mutex_);
+    ++begin_count_;
+    last_begin_job_id_ = ci.job_id;
+    EXPECT_EQ(begin_count_, pre_commit_count_ + 1);
+    EXPECT_EQ(begin_count_, completed_count_ + 1);
+  }
+
+  void OnCompactionPreCommit(DB* db, const CompactionJobInfo& ci) override {
+    std::lock_guard<std::mutex> lock(mutex_);
+    ++pre_commit_count_;
+    // Must fire after Begin and before Completed for this compaction.
+    EXPECT_EQ(pre_commit_count_, begin_count_);
+    EXPECT_EQ(pre_commit_count_, completed_count_ + 1);
+    EXPECT_EQ(ci.job_id, last_begin_job_id_);
+    EXPECT_GT(ci.input_files.size(), 0U);
+
+    // Verify input files are still marked being_compacted.
+    std::vector<std::vector<FileMetaData>> files_by_level;
+    test_->dbfull()->TEST_GetFilesMetaData(test_->handles_[ci.cf_id],
+                                           &files_by_level);
+    EXPECT_EQ(test_->db_.get(), db);
+    for (const auto& info : ci.input_file_infos) {
+      bool found = false;
+      for (const auto& level_files : files_by_level) {
+        for (const auto& meta : level_files) {
+          if (meta.fd.GetNumber() == info.file_number) {
+            found = true;
+            EXPECT_TRUE(meta.being_compacted)
+                << "input file " << info.file_number
+                << " should still be being_compacted in "
+                   "OnCompactionPreCommit";
+          }
+        }
+      }
+      EXPECT_TRUE(found) << "input file " << info.file_number
+                         << " not found in DB";
+    }
+  }
+
+  void OnCompactionCompleted(DB* /*db*/, const CompactionJobInfo& ci) override {
+    std::lock_guard<std::mutex> lock(mutex_);
+    ++completed_count_;
+    // Must fire after the matching PreCommit.
+    EXPECT_EQ(completed_count_, pre_commit_count_);
+    EXPECT_EQ(completed_count_, begin_count_);
+    EXPECT_EQ(ci.job_id, last_begin_job_id_);
+  }
+
+  size_t BeginCount() {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return begin_count_;
+  }
+  size_t PreCommitCount() {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return pre_commit_count_;
+  }
+  size_t CompletedCount() {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return completed_count_;
+  }
+
+ private:
+  EventListenerTest* test_;
+  std::mutex mutex_;
+  size_t begin_count_ = 0;
+  size_t pre_commit_count_ = 0;
+  size_t completed_count_ = 0;
+  int last_begin_job_id_ = -1;
+};
+
+TEST_F(EventListenerTest, OnCompactionPreCommitOrdering) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.compaction_style = kCompactionStyleLevel;
+  options.compression = kNoCompression;
+  options.level0_file_num_compaction_trigger = 4;
+  auto* listener = new TestCompactionPreCommitListener(this);
+  options.listeners.emplace_back(listener);
+
+  // Verify sync-point ordering: Begin -> PreCommit -> Completed.
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+      {{"DBImpl::NotifyOnCompactionBegin::UnlockMutex",
+        "DBImpl::NotifyOnCompactionPreCommit::UnlockMutex"},
+       {"DBImpl::NotifyOnCompactionPreCommit::UnlockMutex",
+        "DBImpl::NotifyOnCompactionCompleted::UnlockMutex"}});
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  CreateAndReopenWithCF({"pikachu"}, options);
+
+  Random rnd(301);
+  for (int i = 0; i < 4; i++) {
+    for (int j = 0; j < 10; j++) {
+      ASSERT_OK(Put(1, rnd.RandomString(10), rnd.RandomString(10)));
+    }
+    ASSERT_OK(Flush(1));
+  }
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  EXPECT_GT(listener->BeginCount(), 0U);
+  EXPECT_EQ(listener->PreCommitCount(), listener->BeginCount());
+  EXPECT_EQ(listener->CompletedCount(), listener->BeginCount());
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+class TestDBShutdownBeginListener : public EventListener {
+ public:
+  void OnDBShutdownBegin(DB* db) override {
+    std::lock_guard<std::mutex> lock(mutex_);
+    ++shutdown_count_;
+    last_db_ = db;
+  }
+
+  int ShutdownCount() const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return shutdown_count_;
+  }
+
+  DB* LastDB() const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return last_db_;
+  }
+
+ private:
+  mutable std::mutex mutex_;
+  int shutdown_count_ = 0;
+  DB* last_db_ = nullptr;
+};
+
+TEST_F(EventListenerTest, OnDBShutdownBeginOnceForCancelAndClose) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  auto listener = std::make_shared<TestDBShutdownBeginListener>();
+  options.listeners.emplace_back(listener);
+  DestroyAndReopen(options);
+
+  DB* db = db_.get();
+  dbfull()->CancelAllBackgroundWork(false);
+  EXPECT_EQ(1, listener->ShutdownCount());
+  EXPECT_EQ(db, listener->LastDB());
+
+  Close();
+  EXPECT_EQ(1, listener->ShutdownCount());
+}
+
+TEST_F(EventListenerTest, OnDBShutdownBeginOnFailedOpen) {
+  Close();
+
+  std::shared_ptr<FaultInjectionTestFS> fs(
+      new FaultInjectionTestFS(env_->GetFileSystem()));
+  std::unique_ptr<Env> env(NewCompositeEnv(fs));
+
+  Options options = CurrentOptions();
+  options.env = env.get();
+  options.create_if_missing = true;
+  auto listener = std::make_shared<TestDBShutdownBeginListener>();
+  options.listeners.emplace_back(listener);
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "PersistRocksDBOptions:create",
+      [&](void* /*arg*/) { fs->SetFilesystemActive(false); });
+  SyncPoint::GetInstance()->SetCallBack(
+      "PersistRocksDBOptions:written",
+      [&](void* /*arg*/) { fs->SetFilesystemActive(true); });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  std::unique_ptr<DB> db;
+  Status s = DB::Open(options, dbname_, &db);
+  ASSERT_TRUE(s.IsIOError()) << s.ToString();
+  EXPECT_EQ(1, listener->ShutdownCount());
+  EXPECT_NE(nullptr, listener->LastDB());
+  EXPECT_EQ(nullptr, db.get());
+
+  fs->SetFilesystemActive(true);
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
 // This simple Listener can only handle one flush at a time.
 class TestFlushListener : public EventListener {
  public:
@@ -1754,7 +1943,7 @@ TEST_F(EventListenerTest, BackgroundJobPressure) {
   listener->Reset();
   sleeping_task.WakeUp();
   sleeping_task.WaitUntilDone();
-  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+  ASSERT_OK(dbfull()->TEST_WaitForBackgroundWork());
 
   snapshots = listener->GetSnapshots();
   CheckInvariants(snapshots);

db/memtable.cc

@@ -159,8 +159,8 @@ ImmutableMemTableOptions::ImmutableMemTableOptions(
           mutable_cf_options.memtable_protection_bytes_per_key),
       allow_data_in_errors(ioptions.allow_data_in_errors),
       paranoid_memory_checks(mutable_cf_options.paranoid_memory_checks),
-      memtable_veirfy_per_key_checksum_on_seek(
-          mutable_cf_options.memtable_veirfy_per_key_checksum_on_seek),
+      memtable_verify_per_key_checksum_on_seek(
+          mutable_cf_options.memtable_verify_per_key_checksum_on_seek),
       memtable_batch_lookup_optimization(
           ioptions.memtable_batch_lookup_optimization) {}
 
@@ -256,7 +256,7 @@ MemTable::MemTable(const InternalKeyComparator& cmp,
           mutable_cf_options.memtable_max_range_deletions),
       key_validation_callback_(
           (moptions_.protection_bytes_per_key != 0 &&
-           moptions_.memtable_veirfy_per_key_checksum_on_seek)
+           moptions_.memtable_verify_per_key_checksum_on_seek)
               ? std::bind(&MemTable::ValidateKey, this, std::placeholders::_1,
                           std::placeholders::_2)
               : std::function<Status(const char*, bool)>(nullptr)) {
@@ -528,7 +528,7 @@ class MemTableIterator : public InternalIterator {
         paranoid_memory_checks_(mem.moptions_.paranoid_memory_checks),
         validate_on_seek_(
             mem.moptions_.paranoid_memory_checks ||
-            mem.moptions_.memtable_veirfy_per_key_checksum_on_seek),
+            mem.moptions_.memtable_verify_per_key_checksum_on_seek),
         allow_data_in_error_(mem.moptions_.allow_data_in_errors),
         key_validation_callback_(mem.key_validation_callback_) {
     if (kind == kRangeDelEntries) {
@@ -1682,7 +1682,7 @@ void MemTable::GetFromTable(const LookupKey& key,
   saver.protection_bytes_per_key = moptions_.protection_bytes_per_key;
 
   if (!moptions_.paranoid_memory_checks &&
-      !moptions_.memtable_veirfy_per_key_checksum_on_seek) {
+      !moptions_.memtable_verify_per_key_checksum_on_seek) {
     table_->Get(key, &saver, SaveValue);
   } else {
     Status check_s = table_->GetAndValidate(
@@ -1750,7 +1750,7 @@ void MemTable::MultiGet(const ReadOptions& read_options, MultiGetRange* range,
   // Use batch lookup optimization when enabled
   bool use_batch_optimization = moptions_.memtable_batch_lookup_optimization;
   bool validate = moptions_.paranoid_memory_checks ||
-                  moptions_.memtable_veirfy_per_key_checksum_on_seek;
+                  moptions_.memtable_verify_per_key_checksum_on_seek;
 
   if (use_batch_optimization) {
     // Phase 1: Handle range tombstones and set up Savers for batched lookup

db/memtable.h

@@ -68,7 +68,7 @@ struct ImmutableMemTableOptions {
   uint32_t protection_bytes_per_key;
   bool allow_data_in_errors;
   bool paranoid_memory_checks;
-  bool memtable_veirfy_per_key_checksum_on_seek;
+  bool memtable_verify_per_key_checksum_on_seek;
   bool memtable_batch_lookup_optimization;
 };
 

db/memtable_list.cc

@@ -17,6 +17,7 @@
 #include "db/version_set.h"
 #include "logging/log_buffer.h"
 #include "logging/logging.h"
+#include "memory/arena.h"
 #include "monitoring/thread_status_util.h"
 #include "rocksdb/db.h"
 #include "rocksdb/env.h"
@@ -31,6 +32,78 @@ class InternalKeyComparator;
 class Mutex;
 class VersionSet;
 
+bool MultiScanOverlapsUserKeyRange(const MultiScanArgs* scan_opts,
+                                   const Comparator* user_comparator,
+                                   const Slice& smallest_user_key,
+                                   const Slice& largest_user_key) {
+  if (scan_opts == nullptr || !scan_opts->HasBoundedScanRanges()) {
+    return true;
+  }
+
+  for (const ScanOptions& scan_range : scan_opts->GetScanRanges()) {
+    if (user_comparator->CompareWithoutTimestamp(
+            scan_range.range.limit.value(), /*a_has_ts=*/false,
+            smallest_user_key, /*b_has_ts=*/true) <= 0) {
+      continue;
+    }
+    if (user_comparator->CompareWithoutTimestamp(
+            scan_range.range.start.value(), /*a_has_ts=*/false,
+            largest_user_key, /*b_has_ts=*/true) > 0) {
+      continue;
+    }
+    return true;
+  }
+  return false;
+}
+
+bool MultiScanIteratorOverlapsUserKeyRange(InternalIterator* iter,
+                                           const MultiScanArgs* scan_opts,
+                                           const Comparator* user_comparator) {
+  iter->SeekToFirst();
+  if (!iter->status().ok()) {
+    return true;
+  }
+  if (!iter->Valid()) {
+    return true;
+  }
+  const std::string smallest_user_key = ExtractUserKey(iter->key()).ToString();
+  iter->SeekToLast();
+  if (!iter->status().ok()) {
+    return true;
+  }
+  if (!iter->Valid()) {
+    return true;
+  }
+  const std::string largest_user_key = ExtractUserKey(iter->key()).ToString();
+  return MultiScanOverlapsUserKeyRange(scan_opts, user_comparator,
+                                       smallest_user_key, largest_user_key);
+}
+
+bool MultiScanIntersectsMemTable(
+    ReadOnlyMemTable* memtable, const ReadOptions& read_options,
+    UnownedPtr<const SeqnoToTimeMapping> seqno_to_time_mapping,
+    const SliceTransform* prefix_extractor, const MultiScanArgs* scan_opts,
+    const Comparator* user_comparator) {
+  if (scan_opts == nullptr || !scan_opts->HasBoundedScanRanges()) {
+    return true;
+  }
+  if (memtable->NumRangeDeletion() > 0) {
+    return true;
+  }
+
+  Arena arena;
+  ReadOptions intersect_read_options(read_options);
+  intersect_read_options.total_order_seek = true;
+  intersect_read_options.iterate_lower_bound = nullptr;
+  intersect_read_options.iterate_upper_bound = nullptr;
+  InternalIterator* iter =
+      memtable->NewIterator(intersect_read_options, seqno_to_time_mapping,
+                            &arena, prefix_extractor, /*for_flush=*/false);
+  ScopedArenaPtr<InternalIterator> iter_guard(iter);
+  return MultiScanIteratorOverlapsUserKeyRange(iter, scan_opts,
+                                               user_comparator);
+}
+
 void MemTableListVersion::AddMemTable(ReadOnlyMemTable* m) {
   if (!memlist_.empty()) {
     // ID can be equal for MemPurge
@@ -229,33 +302,42 @@ void MemTableListVersion::AddIterators(
     const ReadOptions& options,
     UnownedPtr<const SeqnoToTimeMapping> seqno_to_time_mapping,
     const SliceTransform* prefix_extractor,
-    MergeIteratorBuilder* merge_iter_builder, bool add_range_tombstone_iter) {
+    MergeIteratorBuilder* merge_iter_builder, bool add_range_tombstone_iter,
+    SequenceNumber read_seq, const MultiScanArgs* scan_opts,
+    const Comparator* user_comparator) {
   for (auto& m : memlist_) {
+    const bool should_probe_scan_intersection =
+        scan_opts != nullptr && scan_opts->HasBoundedScanRanges() &&
+        m->NumRangeDeletion() == 0;
+    if (should_probe_scan_intersection &&
+        !MultiScanIntersectsMemTable(m, options, seqno_to_time_mapping,
+                                     prefix_extractor, scan_opts,
+                                     user_comparator)) {
+      continue;
+    }
     auto mem_iter =
         m->NewIterator(options, seqno_to_time_mapping,
                        merge_iter_builder->GetArena(), prefix_extractor,
                        /*for_flush=*/false);
+    ScopedArenaPtr<InternalIterator> mem_iter_guard(mem_iter);
     if (!add_range_tombstone_iter || options.ignore_range_deletions) {
-      merge_iter_builder->AddIterator(mem_iter);
+      merge_iter_builder->AddIterator(mem_iter_guard.release());
     } else {
-      // Except for snapshot read, using kMaxSequenceNumber is OK because these
-      // are immutable memtables.
-      SequenceNumber read_seq = options.snapshot != nullptr
-                                    ? options.snapshot->GetSequenceNumber()
-                                    : kMaxSequenceNumber;
+      const SequenceNumber range_del_read_seq =
+          read_seq != kMaxSequenceNumber || options.snapshot == nullptr
+              ? read_seq
+              : options.snapshot->GetSequenceNumber();
       std::unique_ptr<TruncatedRangeDelIterator> mem_tombstone_iter;
-      auto range_del_iter = m->NewRangeTombstoneIterator(
-          options, read_seq, true /* immutale_memtable */);
-      if (range_del_iter == nullptr || range_del_iter->empty()) {
-        delete range_del_iter;
-      } else {
+      std::unique_ptr<FragmentedRangeTombstoneIterator> range_del_iter(
+          m->NewRangeTombstoneIterator(options, range_del_read_seq,
+                                       true /* immutale_memtable */));
+      if (range_del_iter != nullptr && !range_del_iter->empty()) {
         mem_tombstone_iter = std::make_unique<TruncatedRangeDelIterator>(
-            std::unique_ptr<FragmentedRangeTombstoneIterator>(range_del_iter),
-            &m->GetInternalKeyComparator(), nullptr /* smallest */,
-            nullptr /* largest */);
+            std::move(range_del_iter), &m->GetInternalKeyComparator(),
+            nullptr /* smallest */, nullptr /* largest */);
       }
       merge_iter_builder->AddPointAndTombstoneIterator(
-          mem_iter, std::move(mem_tombstone_iter));
+          mem_iter_guard.release(), std::move(mem_tombstone_iter));
     }
   }
 }

db/memtable_list.h

@@ -34,6 +34,21 @@ class MemTableList;
 
 struct FlushJobInfo;
 
+// Returns true when at least one bounded scan range overlaps the user-key
+// range. Unbounded scan options conservatively overlap.
+bool MultiScanOverlapsUserKeyRange(const MultiScanArgs* scan_opts,
+                                   const Comparator* user_comparator,
+                                   const Slice& smallest_user_key,
+                                   const Slice& largest_user_key);
+
+// Returns true when the memtable may contain keys or range deletions relevant
+// to the scan ranges. Iterator/status errors conservatively overlap.
+bool MultiScanIntersectsMemTable(
+    ReadOnlyMemTable* memtable, const ReadOptions& read_options,
+    UnownedPtr<const SeqnoToTimeMapping> seqno_to_time_mapping,
+    const SliceTransform* prefix_extractor, const MultiScanArgs* scan_opts,
+    const Comparator* user_comparator);
+
 // keeps a list of immutable memtables (ReadOnlyMemtable*) in a vector.
 // The list is immutable if refcount is bigger than one. It is used as
 // a state for Get() and iterator code paths.
@@ -122,11 +137,17 @@ class MemTableListVersion {
                     std::vector<InternalIterator*>* iterator_list,
                     Arena* arena);
 
+  // read_seq controls range tombstone visibility. It may be captured before
+  // lazy iterator initialization. scan_opts can prune immutable memtables that
+  // cannot intersect the requested scan ranges.
   void AddIterators(const ReadOptions& options,
                     UnownedPtr<const SeqnoToTimeMapping> seqno_to_time_mapping,
                     const SliceTransform* prefix_extractor,
                     MergeIteratorBuilder* merge_iter_builder,
-                    bool add_range_tombstone_iter);
+                    bool add_range_tombstone_iter,
+                    SequenceNumber read_seq = kMaxSequenceNumber,
+                    const MultiScanArgs* scan_opts = nullptr,
+                    const Comparator* user_comparator = nullptr);
 
   uint64_t GetTotalNumEntries() const;
 

db/range_tombstone_fragmenter.cc

@@ -497,6 +497,9 @@ FragmentedRangeTombstoneIterator::SplitBySnapshot(
       splits;
   SequenceNumber lower = 0;
   SequenceNumber upper;
+  // Safe to dereference icmp_ below: SplitBySnapshot is construction-time work
+  // (called synchronously from the range-del aggregator while the reader that
+  // owns *icmp_ is alive). See the icmp_ note in the header.
   for (size_t i = 0; i <= snapshots.size(); i++) {
     if (i >= snapshots.size()) {
       upper = kMaxSequenceNumber;

db/range_tombstone_fragmenter.h

@@ -199,7 +199,9 @@ class FragmentedRangeTombstoneIterator : public InternalIterator {
 
   RangeTombstone Tombstone() const {
     assert(Valid());
-    if (icmp_->user_comparator()->timestamp_size()) {
+    // Use ucmp_ (long-lived; see the icmp_ note below) rather than
+    // icmp_->user_comparator(). They are equal for all constructors.
+    if (ucmp_->timestamp_size()) {
       return RangeTombstone(start_key(), end_key(), seq(), timestamp());
     }
     return RangeTombstone(start_key(), end_key(), seq());
@@ -316,6 +318,12 @@ class FragmentedRangeTombstoneIterator : public InternalIterator {
 
   const RangeTombstoneStackStartComparator tombstone_start_cmp_;
   const RangeTombstoneStackEndComparator tombstone_end_cmp_;
+  // icmp_ is a borrowed pointer and is not guaranteed to outlive this iterator:
+  // it may be built from a comparator with a shorter lifetime than an iterator
+  // some owner keeps alive. Only dereference it during construction-time work
+  // (e.g. SplitBySnapshot), when the referent is known alive. Post-construction
+  // navigation must use ucmp_ (the long-lived user comparator) and tombstones_;
+  // do not add new icmp_ dereferences in Next()/Seek()/Tombstone().
   const InternalKeyComparator* icmp_;
   const Comparator* ucmp_;
   std::shared_ptr<FragmentedRangeTombstoneList> tombstones_ref_;

db/table_cache.cc

@@ -98,7 +98,7 @@ Status TableCache::GetTableReader(
     const MutableCFOptions& mutable_cf_options, bool skip_filters, int level,
     bool prefetch_index_and_filter_in_cache,
     size_t max_file_size_for_l0_meta_pin, Temperature file_temperature,
-    std::string* file_open_metadata) {
+    std::string* file_open_metadata, bool avoid_shared_metadata_cache) {
   std::string fname = TableFileName(
       ioptions_.cf_paths, file_meta.fd.GetNumber(), file_meta.fd.GetPathId());
   std::unique_ptr<FSRandomAccessFile> file;
@@ -190,7 +190,8 @@ Status TableCache::GetTableReader(
             block_cache_tracer_, max_file_size_for_l0_meta_pin, db_session_id_,
             file_meta.fd.GetNumber(), expected_unique_id,
             file_meta.fd.largest_seqno, file_meta.tail_size,
-            file_meta.user_defined_timestamps_persisted),
+            file_meta.user_defined_timestamps_persisted,
+            avoid_shared_metadata_cache),
         std::move(file_reader), file_meta.fd.GetFileSize(), table_reader,
         prefetch_index_and_filter_in_cache);
     TEST_SYNC_POINT("TableCache::GetTableReader:0");
@@ -213,11 +214,53 @@ Status TableCache::FindTable(
     const bool no_io, HistogramImpl* file_read_hist, bool skip_filters,
     int level, bool prefetch_index_and_filter_in_cache,
     size_t max_file_size_for_l0_meta_pin, Temperature file_temperature,
-    bool pin_table_handle, std::string* file_open_metadata) {
+    bool pin_table_handle, std::string* file_open_metadata,
+    std::unique_ptr<TableReader>* fresh_table_reader_owner,
+    const TableCacheOpenOptions& open_options) {
   assert(out_table_reader != nullptr && *out_table_reader == nullptr);
   assert(handle != nullptr && *handle == nullptr);
+  // open_ephemeral_table_reader requests a fresh reader;
+  // fresh_table_reader_owner is where we return it. The two must agree.
+  assert(open_options.open_ephemeral_table_reader ==
+         (fresh_table_reader_owner != nullptr));
   PERF_TIMER_GUARD_WITH_CLOCK(find_table_nanos, ioptions_.clock);
 
+  // Bypass path: open a fresh TableReader, skipping the pinned-reader fast path
+  // and the shared cache. The caller takes ownership via
+  // fresh_table_reader_owner. no_io is not allowed here since opening a new
+  // reader always needs I/O.
+  if (fresh_table_reader_owner != nullptr) {
+    assert(!no_io);
+    if (no_io) {
+      // Defensive in release builds: the caller violated the contract.
+      return Status::Incomplete(
+          "fresh TableReader requested but no_io is set; cannot open file");
+    }
+    std::unique_ptr<TableReader> table_reader;
+    const bool effective_skip_filters =
+        skip_filters || open_options.skip_filters;
+    TEST_SYNC_POINT_CALLBACK("TableCache::FindTable:FreshTableReader",
+                             const_cast<TableCacheOpenOptions*>(&open_options));
+    Status s = GetTableReader(
+        ro, file_options, internal_comparator, file_meta,
+        false /* sequential mode */, file_read_hist, &table_reader,
+        mutable_cf_options, effective_skip_filters, level,
+        prefetch_index_and_filter_in_cache &&
+            !open_options.avoid_shared_metadata_cache,
+        max_file_size_for_l0_meta_pin, file_temperature, file_open_metadata,
+        open_options.avoid_shared_metadata_cache);
+    if (!s.ok()) {
+      assert(table_reader == nullptr);
+      RecordTick(ioptions_.stats, NO_FILE_ERRORS);
+      IGNORE_STATUS_IF_ERROR(s);
+      return s;
+    }
+    *out_table_reader = table_reader.get();
+    *fresh_table_reader_owner = std::move(table_reader);
+    *handle = nullptr;
+    return s;
+  }
+
   // Fast path: if table reader is already pinned, return it directly without a
   // cache lookup.
   auto pinned_reader = file_meta.fd.pinned_reader.Get();
@@ -313,15 +356,23 @@ InternalIterator* TableCache::NewIterator(
     const InternalKey* largest_compaction_key, bool allow_unprepared_value,
     const SequenceNumber* read_seqno,
     std::unique_ptr<TruncatedRangeDelIterator>* range_del_iter,
-    bool maybe_pin_table_handle, std::string* file_open_metadata) {
+    bool maybe_pin_table_handle, std::string* file_open_metadata,
+    const TableCacheOpenOptions& open_options) {
   PERF_TIMER_GUARD(new_table_iterator_nanos);
 
   Status s;
   TableReader* table_reader = nullptr;
   TypedHandle* handle = nullptr;
+  assert(!open_options.open_ephemeral_table_reader ||
+         table_reader_ptr == nullptr);
+  // Holds ownership of a freshly-opened TableReader when the caller asked us
+  // to bypass the shared cache. When non-empty, the iterator we hand back must
+  // arrange to free it on destruction.
+  std::unique_ptr<TableReader> ephemeral_reader;
   if (table_reader_ptr != nullptr) {
     *table_reader_ptr = nullptr;
   }
+  const bool effective_skip_filters = skip_filters || open_options.skip_filters;
   bool for_compaction = caller == TableReaderCaller::kCompaction;
   TEST_SYNC_POINT_CALLBACK("TableCache::NewIterator::BeforeFindTable",
                            const_cast<FileDescriptor*>(&file_meta.fd));
@@ -329,9 +380,13 @@ InternalIterator* TableCache::NewIterator(
       options, file_options, icomparator, file_meta, &handle,
       mutable_cf_options, &table_reader,
       options.read_tier == kBlockCacheTier /* no_io */, file_read_hist,
-      skip_filters, level, true /* prefetch_index_and_filter_in_cache */,
-      max_file_size_for_l0_meta_pin, file_meta.temperature,
-      maybe_pin_table_handle && should_pin_table_handles_, file_open_metadata);
+      effective_skip_filters, level,
+      /*prefetch_index_and_filter_in_cache=*/
+      !open_options.avoid_shared_metadata_cache, max_file_size_for_l0_meta_pin,
+      file_meta.temperature,
+      maybe_pin_table_handle && should_pin_table_handles_, file_open_metadata,
+      open_options.open_ephemeral_table_reader ? &ephemeral_reader : nullptr,
+      open_options);
   InternalIterator* result = nullptr;
   if (s.ok()) {
     if (options.table_filter &&
@@ -340,13 +395,17 @@ InternalIterator* TableCache::NewIterator(
     } else {
       result = table_reader->NewIterator(
           options, mutable_cf_options.prefix_extractor.get(), arena,
-          skip_filters, caller, file_options.compaction_readahead_size,
-          allow_unprepared_value);
+          effective_skip_filters, caller,
+          file_options.compaction_readahead_size, allow_unprepared_value);
     }
     if (handle != nullptr) {
       cache_.RegisterReleaseAsCleanup(handle, *result);
       handle = nullptr;  // prevent from releasing below
     }
+    // Don't hand ephemeral_reader to result's cleanup yet: range-del
+    // processing below can set s to non-OK, in which case result is replaced
+    // by an error iterator at function end. Transfer ownership only once we
+    // know s stays OK; otherwise ephemeral_reader's destructor frees it.
 
     if (for_compaction) {
       table_reader->SetupForCompaction();
@@ -398,8 +457,37 @@ InternalIterator* TableCache::NewIterator(
   if (handle != nullptr) {
     cache_.Release(handle);
   }
+  // Range-del processing is done and s is final. Hand the ephemeral reader's
+  // lifetime to the returned iterator; if s is non-OK, leave it for
+  // ephemeral_reader's destructor. RegisterCleanup gets the raw pointer before
+  // release(), so if its allocation throws the reader is still owned by
+  // ephemeral_reader and freed on unwind.
+  if (s.ok() && ephemeral_reader && result != nullptr) {
+    TableReader* raw = ephemeral_reader.get();
+    result->RegisterCleanup(
+        [](void* arg1, void* /*arg2*/) {
+          delete static_cast<TableReader*>(arg1);
+        },
+        raw, nullptr);
+    // release() returns raw, which the cleanup now owns; drop the unique_ptr's
+    // ownership so the reader isn't freed twice.
+    [[maybe_unused]] TableReader* released = ephemeral_reader.release();
+    assert(released == raw);
+  }
   if (!s.ok()) {
+    // Today result is always null here: it is only set when s was OK, and the
+    // only later change to s, new_range_del_iter->status(), is always OK for a
+    // FragmentedRangeTombstoneIterator. The assert documents that; the cleanup
+    // still disposes of any stray result before the error iterator replaces it.
     assert(result == nullptr);
+    if (result != nullptr) {
+      if (arena != nullptr) {
+        result->~InternalIterator();
+      } else {
+        delete result;
+      }
+      result = nullptr;
+    }
     result = NewErrorInternalIterator<Slice>(s, arena);
   }
   return result;

db/table_cache.h

@@ -35,6 +35,18 @@ struct FileDescriptor;
 class GetContext;
 class HistogramImpl;
 
+struct TableCacheOpenOptions {
+  // Open a new TableReader owned by the returned iterator instead of reusing
+  // the pinned reader or shared TableCache entry.
+  bool open_ephemeral_table_reader = false;
+
+  // Avoid inserting open-time metadata blocks into the shared block cache.
+  bool avoid_shared_metadata_cache = false;
+
+  // Disable loading/accessing filter blocks for this open/iterator.
+  bool skip_filters = false;
+};
+
 // Manages caching for TableReader objects for a column family. The actual
 // cache is allocated separately and passed to the constructor. TableCache
 // wraps around the underlying SST file readers by providing Get(),
@@ -73,7 +85,9 @@ class TableCache {
   // underlying the returned iterator, or nullptr if no Table object underlies
   // the returned iterator.  The returned "*table_reader_ptr" object is owned
   // by the cache and should not be deleted, and is valid for as long as the
-  // returned iterator is live.
+  // returned iterator is live. `table_reader_ptr` must be nullptr when
+  // `open_options.open_ephemeral_table_reader` is set because that reader is
+  // owned by iterator cleanup rather than by the cache.
   // If !options.ignore_range_deletions, and range_del_iter is non-nullptr,
   // then range_del_iter is set to a TruncatedRangeDelIterator for range
   // tombstones in the SST file corresponding to the specified file number. The
@@ -82,12 +96,16 @@ class TableCache {
   // @param options Must outlive the returned iterator.
   // @param range_del_agg If non-nullptr, adds range deletions to the
   //    aggregator. If an error occurs, returns it in a NewErrorInternalIterator
-  // @param for_compaction If true, a new TableReader may be allocated (but
-  //                       not cached), depending on the CF options
+  // @param caller Identifies the high-level caller for table-reader behavior
+  //               such as compaction preparation.
   // @param skip_filters Disables loading/accessing the filter block
   // @param level The level this table is at, -1 for "not set / don't know"
   // @param range_del_read_seqno If non-nullptr, will be used to create
   // *range_del_iter.
+  // @param open_options Optional table-open policy overrides. The caller must
+  //                       not combine `open_ephemeral_table_reader` with
+  //                       `ReadOptions::read_tier = kBlockCacheTier` (no_io);
+  //                       opening an ephemeral reader inherently requires I/O.
   InternalIterator* NewIterator(
       const ReadOptions& options, const FileOptions& toptions,
       const InternalKeyComparator& internal_comparator,
@@ -103,7 +121,8 @@ class TableCache {
       bool maybe_pin_table_handle = false,
       // If non-null, and the table reader is newly opened (not cached),
       // retrieves file open metadata via GetFileOpenMetadata().
-      std::string* file_open_metadata = nullptr);
+      std::string* file_open_metadata = nullptr,
+      const TableCacheOpenOptions& open_options = TableCacheOpenOptions());
 
   // If a seek to internal key "k" in specified file finds an entry,
   // call get_context->SaveValue() repeatedly until
@@ -190,18 +209,30 @@ class TableCache {
   // @param pin_table_handle If true, pins the table reader on file_meta so
   //              future lookups bypass the cache. *handle is set to nullptr
   //              on return in this case.
-  Status FindTable(const ReadOptions& ro, const FileOptions& toptions,
-                   const InternalKeyComparator& internal_comparator,
-                   const FileMetaData& file_meta, TypedHandle**,
-                   const MutableCFOptions& mutable_cf_options,
-                   TableReader** table_reader, const bool no_io = false,
-                   HistogramImpl* file_read_hist = nullptr,
-                   bool skip_filters = false, int level = -1,
-                   bool prefetch_index_and_filter_in_cache = true,
-                   size_t max_file_size_for_l0_meta_pin = 0,
-                   Temperature file_temperature = Temperature::kUnknown,
-                   bool pin_table_handle = false,
-                   std::string* file_open_metadata = nullptr);
+  // @param fresh_table_reader_owner If non-null, FindTable always opens a new
+  //              TableReader (skipping the pinned-reader fast path and the
+  //              shared cache) and writes ownership into this unique_ptr.
+  //              `*handle` will be nullptr on return and `*table_reader` will
+  //              point to the freshly allocated reader. Callers that pass this
+  //              are responsible for keeping the unique_ptr alive for the
+  //              lifetime of any iterators built on top.
+  // @param open_options Optional table-open policy overrides (e.g.
+  //              avoid_shared_metadata_cache, skip_filters).
+  //              `open_options.open_ephemeral_table_reader` must be set if and
+  //              only if `fresh_table_reader_owner` is non-null.
+  Status FindTable(
+      const ReadOptions& ro, const FileOptions& toptions,
+      const InternalKeyComparator& internal_comparator,
+      const FileMetaData& file_meta, TypedHandle**,
+      const MutableCFOptions& mutable_cf_options, TableReader** table_reader,
+      const bool no_io = false, HistogramImpl* file_read_hist = nullptr,
+      bool skip_filters = false, int level = -1,
+      bool prefetch_index_and_filter_in_cache = true,
+      size_t max_file_size_for_l0_meta_pin = 0,
+      Temperature file_temperature = Temperature::kUnknown,
+      bool pin_table_handle = false, std::string* file_open_metadata = nullptr,
+      std::unique_ptr<TableReader>* fresh_table_reader_owner = nullptr,
+      const TableCacheOpenOptions& open_options = TableCacheOpenOptions());
 
   // Get the table properties of a given table.
   // @no_io: indicates if we should load table to the cache if it is not present
@@ -286,7 +317,8 @@ class TableCache {
                         bool prefetch_index_and_filter_in_cache = true,
                         size_t max_file_size_for_l0_meta_pin = 0,
                         Temperature file_temperature = Temperature::kUnknown,
-                        std::string* file_open_metadata = nullptr);
+                        std::string* file_open_metadata = nullptr,
+                        bool avoid_shared_metadata_cache = false);
 
   // Update the max_covering_tombstone_seq in the GetContext for each key based
   // on the range deletions in the table

db/version_edit.cc

@@ -240,6 +240,13 @@ bool VersionEdit::EncodeTo(std::string* dst,
     PutLengthPrefixedSlice(dst, progress_data);
   }
 
+  if (has_last_compacted_manifest_file_size_) {
+    PutVarint32(dst, kLastCompactedManifestFileSize);
+    std::string varint_size;
+    PutVarint64(&varint_size, last_compacted_manifest_file_size_);
+    PutLengthPrefixedSlice(dst, Slice(varint_size));
+  }
+
   return true;
 }
 
@@ -890,6 +897,17 @@ Status VersionEdit::DecodeFrom(const Slice& src) {
         break;
       }
 
+      case kLastCompactedManifestFileSize: {
+        Slice encoded;
+        if (GetLengthPrefixedSlice(&input, &encoded) &&
+            GetVarint64(&encoded, &last_compacted_manifest_file_size_)) {
+          has_last_compacted_manifest_file_size_ = true;
+        } else {
+          msg = "last compacted manifest file size";
+        }
+        break;
+      }
+
       default:
         if (tag & kTagSafeIgnoreMask) {
           // Tag from future which can be safely ignored.
@@ -1064,6 +1082,10 @@ std::string VersionEdit::DebugString(bool hex_key) const {
     r.append("\n SubcompactionProgress: ");
     r.append(subcompaction_progress_.ToString());
   }
+  if (has_last_compacted_manifest_file_size_) {
+    r.append("\n  LastCompactedManifestFileSize: ");
+    AppendNumberTo(&r, last_compacted_manifest_file_size_);
+  }
   r.append("\n}\n");
   return r;
 }
@@ -1220,6 +1242,10 @@ std::string VersionEdit::DebugJSON(int edit_num, bool hex_key) const {
     jw << "SubcompactionProgress" << subcompaction_progress_.ToString();
   }
 
+  if (has_last_compacted_manifest_file_size_) {
+    jw << "LastCompactedManifestFileSize" << last_compacted_manifest_file_size_;
+  }
+
   jw.EndObject();
 
   return jw.Get();

db/version_edit.h

@@ -74,6 +74,7 @@ enum Tag : uint32_t {
   kWalDeletion2,
   kPersistUserDefinedTimestamps,
   kSubcompactionProgress,
+  kLastCompactedManifestFileSize,
 };
 
 enum SubcompactionProgressPerLevelCustomTag : uint32_t {
@@ -336,6 +337,11 @@ struct FileMetaData {
   // via FileOptions::file_metadata on subsequent opens. Empty if not available.
   std::string file_open_metadata;
 
+  // Skips prefetching index and filter blocks into block cache on file open,
+  // not persisted in MANIFEST. NOTE: false does not guarantee prefetching
+  // either (e.g. when index and filter blocks are not cached in block cache).
+  bool skip_index_and_filter_blocks_prefetch = false;
+
   FileMetaData() = default;
 
   FileMetaData(uint64_t file, uint32_t file_path_id, uint64_t file_size,
@@ -977,6 +983,11 @@ class VersionEdit {
 
   bool IsColumnFamilyDrop() const { return is_column_family_drop_; }
 
+  void MarkForegroundOperation() { is_foreground_operation_ = true; }
+  bool IsForegroundOperation() const {
+    return is_foreground_operation_ || IsColumnFamilyManipulation();
+  }
+
   void MarkNoManifestWriteDummy() { is_no_manifest_write_dummy_ = true; }
   bool IsNoManifestWriteDummy() const { return is_no_manifest_write_dummy_; }
 
@@ -1016,6 +1027,17 @@ class VersionEdit {
     subcompaction_progress_.Clear();
   }
 
+  void SetLastCompactedManifestFileSize(uint64_t size) {
+    has_last_compacted_manifest_file_size_ = true;
+    last_compacted_manifest_file_size_ = size;
+  }
+  bool HasLastCompactedManifestFileSize() const {
+    return has_last_compacted_manifest_file_size_;
+  }
+  uint64_t GetLastCompactedManifestFileSize() const {
+    return last_compacted_manifest_file_size_;
+  }
+
   // Recovery-time per-column-family edits only need to be written when they
   // advance the CF's log number or carry some other manifest state.
   bool ShouldEmitPerColumnFamilyRecoveryEdit(uint64_t current_log_number) const;
@@ -1099,6 +1121,7 @@ class VersionEdit {
   // it also includes column family name.
   bool is_column_family_drop_ = false;
   bool is_column_family_add_ = false;
+  bool is_foreground_operation_ = false;
   std::string column_family_name_;
 
   uint32_t remaining_entries_ = 0;
@@ -1111,6 +1134,9 @@ class VersionEdit {
   bool has_subcompaction_progress_ = false;
   SubcompactionProgress subcompaction_progress_;
 
+  bool has_last_compacted_manifest_file_size_ = false;
+  uint64_t last_compacted_manifest_file_size_ = 0;
+
   // Newly created table files and blob files are eligible for deletion if they
   // are not registered as live files after the background jobs creating them
   // have finished. In case committing the VersionEdit containing such changes

db/version_edit_handler.cc

@@ -642,6 +642,11 @@ Status VersionEditHandler::ExtractInfoFromVersionEdit(ColumnFamilyData* cfd,
              version_edit_params_.GetLastSequence() <= edit.GetLastSequence());
       version_edit_params_.SetLastSequence(edit.GetLastSequence());
     }
+    if (edit.HasLastCompactedManifestFileSize()) {
+      version_set_->last_compacted_manifest_file_size_ =
+          edit.GetLastCompactedManifestFileSize();
+      version_set_->TuneMaxManifestFileSize();
+    }
     if (!version_edit_params_.HasPrevLogNumber()) {
       version_edit_params_.SetPrevLogNumber(0);
     }

db/version_set.cc

@@ -18,7 +18,9 @@
 #include <set>
 #include <stdexcept>
 #include <string>
+#include <type_traits>
 #include <unordered_map>
+#include <utility>
 #include <vector>
 
 #include "db/blob/blob_fetcher.h"
@@ -69,7 +71,6 @@
 #include "table/multiget_context.h"
 #include "table/plain/plain_table_factory.h"
 #include "table/table_reader.h"
-#include "table/two_level_iterator.h"
 #include "table/unique_id_impl.h"
 #include "test_util/sync_point.h"
 #include "util/cast_util.h"
@@ -110,6 +111,181 @@ int FindFileInRange(const InternalKeyComparator& icmp,
   return static_cast<int>(std::lower_bound(b + left, b + right, key, cmp) - b);
 }
 
+InternalKey MultiScanInternalKey(const Slice& user_key,
+                                 const Comparator* user_comparator) {
+  const size_t timestamp_size = user_comparator->timestamp_size();
+  if (timestamp_size == 0) {
+    return InternalKey(user_key, kMaxSequenceNumber, kValueTypeForSeek);
+  }
+
+  std::string key_with_ts;
+  AppendKeyWithMaxTimestamp(&key_with_ts, user_key, timestamp_size);
+  return InternalKey(key_with_ts, kMaxSequenceNumber, kValueTypeForSeek);
+}
+
+bool MultiScanRangeOverlapsFile(const UserComparatorWrapper& user_comparator,
+                                const ScanOptions& scan_opts,
+                                const FdWithKeyRange& file) {
+  assert(scan_opts.range.start.has_value());
+  assert(scan_opts.range.limit.has_value());
+
+  if (user_comparator.CompareWithoutTimestamp(
+          scan_opts.range.limit.value(), /*a_has_ts=*/false,
+          ExtractUserKey(file.smallest_key), /*b_has_ts=*/true) <= 0) {
+    return false;
+  }
+
+  if (user_comparator.CompareWithoutTimestamp(
+          scan_opts.range.start.value(), /*a_has_ts=*/false,
+          ExtractUserKey(file.largest_key), /*b_has_ts=*/true) > 0) {
+    return false;
+  }
+
+  return true;
+}
+
+void AddMultiScanFileIndex(std::vector<size_t>* file_indexes,
+                           std::vector<char>* selected, size_t file_index) {
+  if ((*selected)[file_index]) {
+    return;
+  }
+  (*selected)[file_index] = true;
+  file_indexes->push_back(file_index);
+}
+
+template <typename OnOverlappingFile>
+bool NotifyMultiScanOverlappingFile(OnOverlappingFile& on_overlapping_file,
+                                    size_t file_index,
+                                    const ScanOptions& scan_range) {
+  if constexpr (std::is_same_v<std::invoke_result_t<OnOverlappingFile, size_t,
+                                                    const ScanOptions&>,
+                               bool>) {
+    return on_overlapping_file(file_index, scan_range);
+  } else {
+    on_overlapping_file(file_index, scan_range);
+    return true;
+  }
+}
+
+template <typename OnOverlappingFile>
+void ForEachMultiScanOverlappingFile(const InternalKeyComparator& icmp,
+                                     const UserComparatorWrapper& ucmp,
+                                     const LevelFilesBrief& file_level,
+                                     int level, const MultiScanArgs& scan_opts,
+                                     OnOverlappingFile on_overlapping_file) {
+  assert(scan_opts.HasBoundedScanRanges());
+  if (file_level.num_files == 0) {
+    return;
+  }
+
+  if (level == 0) {
+    for (size_t i = 0; i < file_level.num_files; ++i) {
+      for (const ScanOptions& scan_range : scan_opts.GetScanRanges()) {
+        if (MultiScanRangeOverlapsFile(ucmp, scan_range, file_level.files[i])) {
+          if (!NotifyMultiScanOverlappingFile(on_overlapping_file, i,
+                                              scan_range)) {
+            return;
+          }
+        }
+
+        if (ucmp.CompareWithoutTimestamp(
+                scan_range.range.start.value(), /*a_has_ts=*/false,
+                ExtractUserKey(file_level.files[i].largest_key),
+                /*b_has_ts=*/true) > 0) {
+          break;
+        }
+      }
+    }
+    return;
+  }
+
+  for (const ScanOptions& scan_range : scan_opts.GetScanRanges()) {
+    const InternalKey start_key = MultiScanInternalKey(
+        scan_range.range.start.value(), ucmp.user_comparator());
+    size_t first_file = FindFile(icmp, file_level, start_key.Encode());
+    if (first_file >= file_level.num_files) {
+      continue;
+    }
+
+    const InternalKey limit_key = MultiScanInternalKey(
+        scan_range.range.limit.value(), ucmp.user_comparator());
+    size_t last_file = FindFile(icmp, file_level, limit_key.Encode());
+    if (last_file >= file_level.num_files) {
+      last_file = file_level.num_files - 1;
+    }
+
+    for (size_t i = first_file; i <= last_file; ++i) {
+      if (MultiScanRangeOverlapsFile(ucmp, scan_range, file_level.files[i])) {
+        if (!NotifyMultiScanOverlappingFile(on_overlapping_file, i,
+                                            scan_range)) {
+          return;
+        }
+      }
+    }
+  }
+}
+
+std::vector<size_t> GetMultiScanOverlappingFiles(
+    const InternalKeyComparator& icmp, const UserComparatorWrapper& ucmp,
+    const LevelFilesBrief& file_level, int level,
+    const MultiScanArgs& scan_opts,
+    size_t max_file_indexes = std::numeric_limits<size_t>::max()) {
+  std::vector<size_t> file_indexes;
+  if (!scan_opts.HasBoundedScanRanges() || file_level.num_files == 0) {
+    return file_indexes;
+  }
+  file_indexes.reserve(
+      std::min(static_cast<size_t>(file_level.num_files), max_file_indexes));
+
+  std::vector<char> selected(file_level.num_files, false);
+  ForEachMultiScanOverlappingFile(
+      icmp, ucmp, file_level, level, scan_opts,
+      [&](size_t file_index, const ScanOptions& /*scan_range*/) {
+        if (file_indexes.size() >= max_file_indexes) {
+          return false;
+        }
+        AddMultiScanFileIndex(&file_indexes, &selected, file_index);
+        return file_indexes.size() < max_file_indexes;
+      });
+
+  return file_indexes;
+}
+
+void AddTableIteratorForLevel(
+    ColumnFamilyData* cfd, const ReadOptions& read_options,
+    const FileOptions& soptions, const MutableCFOptions& mutable_cf_options,
+    MergeIteratorBuilder* merge_iter_builder, Arena* arena,
+    const LevelFilesBrief& file_level, int level, size_t file_index,
+    bool skip_filters, size_t max_file_size_for_l0_meta_pin,
+    bool allow_unprepared_value, SequenceNumber read_seq) {
+  std::unique_ptr<TruncatedRangeDelIterator> tombstone_iter = nullptr;
+  const auto& file = file_level.files[file_index];
+  auto table_iter = cfd->table_cache()->NewIterator(
+      read_options, soptions, cfd->internal_comparator(), *file.file_metadata,
+      /*range_del_agg=*/nullptr, mutable_cf_options, nullptr,
+      cfd->internal_stats()->GetFileReadHist(level),
+      TableReaderCaller::kUserIterator, arena, skip_filters, level,
+      max_file_size_for_l0_meta_pin,
+      /*smallest_compaction_key=*/nullptr,
+      /*largest_compaction_key=*/nullptr, allow_unprepared_value, &read_seq,
+      &tombstone_iter,
+      /*maybe_pin_table_handle=*/true);
+#ifndef NDEBUG
+  TEST_SYNC_POINT_CALLBACK("Version::AddIteratorsForLevel:AddedFile",
+                           file.file_metadata);
+  std::pair<bool, bool> iterator_type(true /* is_block_based_table_iterator */,
+                                      false /* is_level_iterator */);
+  TEST_SYNC_POINT_CALLBACK("Version::AddIteratorsForLevel:IteratorType",
+                           &iterator_type);
+#endif
+  if (read_options.ignore_range_deletions) {
+    merge_iter_builder->AddIterator(table_iter);
+  } else {
+    merge_iter_builder->AddPointAndTombstoneIterator(table_iter,
+                                                     std::move(tombstone_iter));
+  }
+}
+
 Status OverlapWithIterator(const Comparator* ucmp,
                            const Slice& smallest_user_key,
                            const Slice& largest_user_key,
@@ -966,6 +1142,15 @@ bool SomeFileOverlapsRange(const InternalKeyComparator& icmp,
 
 namespace {
 
+TableCacheOpenOptions GetCompactionTableCacheOpenOptions(
+    bool open_ephemeral_table_reader) {
+  TableCacheOpenOptions options;
+  options.open_ephemeral_table_reader = open_ephemeral_table_reader;
+  options.avoid_shared_metadata_cache = open_ephemeral_table_reader;
+  options.skip_filters = open_ephemeral_table_reader;
+  return options;
+}
+
 class LevelIterator final : public InternalIterator {
  public:
   // NOTE: many of the const& parameters are saved in this object (so
@@ -981,7 +1166,9 @@ class LevelIterator final : public InternalIterator {
       bool allow_unprepared_value = false,
       std::unique_ptr<TruncatedRangeDelIterator>*** range_tombstone_iter_ptr_ =
           nullptr,
-      Statistics* db_statistics = nullptr, SystemClock* clock = nullptr)
+      SequenceNumber read_seq = kMaxSequenceNumber,
+      Statistics* db_statistics = nullptr, SystemClock* clock = nullptr,
+      bool open_ephemeral_table_reader = false)
       : table_cache_(table_cache),
         read_options_(read_options),
         file_options_(file_options),
@@ -997,9 +1184,11 @@ class LevelIterator final : public InternalIterator {
         pinned_iters_mgr_(nullptr),
         compaction_boundaries_(compaction_boundaries),
         range_tombstone_iter_(nullptr),
-        read_seq_(read_options.snapshot
-                      ? read_options.snapshot->GetSequenceNumber()
-                      : kMaxSequenceNumber),
+        range_tombstone_iter_required_(range_tombstone_iter_ptr_ != nullptr),
+        read_seq_(read_seq != kMaxSequenceNumber ||
+                          read_options.snapshot == nullptr
+                      ? read_seq
+                      : read_options.snapshot->GetSequenceNumber()),
         level_(level),
         skip_filters_(skip_filters),
         allow_unprepared_value_(allow_unprepared_value),
@@ -1007,7 +1196,9 @@ class LevelIterator final : public InternalIterator {
         to_return_sentinel_(false),
         scan_opts_(nullptr),
         db_statistics_(db_statistics),
-        clock_(clock) {
+        clock_(clock),
+        table_cache_open_options_(
+            GetCompactionTableCacheOpenOptions(open_ephemeral_table_reader)) {
     // Empty level is not supported.
     assert(flevel_ != nullptr && flevel_->num_files > 0);
     if (range_tombstone_iter_ptr_) {
@@ -1146,69 +1337,18 @@ class LevelIterator final : public InternalIterator {
     assert(so->GetComparator() == user_comparator_.user_comparator());
 
     file_to_scan_opts_ = std::make_unique<ScanOptionsMap>();
-    for (size_t k = 0; k < scan_opts_->size(); k++) {
-      const ScanOptions& opt = scan_opts_->GetScanRanges().at(k);
-      auto start = opt.range.start;
-      auto end = opt.range.limit;
-
-      if (!start.has_value()) {
-        continue;
-      }
-
-      // We can capture this case in the future, but for now lets skip this.
-      if (!end.has_value()) {
-        continue;
-      }
-
-      const size_t timestamp_size =
-          user_comparator_.user_comparator()->timestamp_size();
-      InternalKey istart, iend;
-      if (timestamp_size == 0) {
-        istart =
-            InternalKey(start.value(), kMaxSequenceNumber, kValueTypeForSeek);
-        // end key is exclusive for multiscan
-        iend = InternalKey(end.value(), kMaxSequenceNumber, kValueTypeForSeek);
-      } else {
-        std::string start_key_with_ts, end_key_with_ts;
-        AppendKeyWithMaxTimestamp(&start_key_with_ts, start.value(),
-                                  timestamp_size);
-        AppendKeyWithMaxTimestamp(&end_key_with_ts, end.value(),
-                                  timestamp_size);
-        istart = InternalKey(start_key_with_ts, kMaxSequenceNumber,
-                             kValueTypeForSeek);
-        // end key is exclusive for multiscan
-        iend =
-            InternalKey(end_key_with_ts, kMaxSequenceNumber, kValueTypeForSeek);
-      }
-
-      // TODO: This needs to be optimized, right now we iterate twice, which
-      // we dont need to. We can do this in N rather than 2N.
-      size_t fstart = FindFile(icomparator_, *flevel_, istart.Encode());
-      size_t fend = FindFile(icomparator_, *flevel_, iend.Encode());
-
-      // We need to check the relevant cases
-      // Cases:
-      // 1. [  S        E  ]
-      // 2. [  S  ]  [  E  ]
-      // 3. [  S  ] ...... [  E  ]
-      for (auto i = fstart; i <= fend; i++) {
-        if (i < flevel_->num_files) {
-          // FindFile only compares against the largest_key, so we need this
-          // additional check to ensure the scan range overlaps the file
-          if (icomparator_.InternalKeyComparator::Compare(
-                  iend.Encode(), flevel_->files[i].smallest_key) < 0) {
-            continue;
-          }
-          auto const metadata = flevel_->files[i].file_metadata;
+    ForEachMultiScanOverlappingFile(
+        icomparator_, user_comparator_, *flevel_, level_, *scan_opts_,
+        [&](size_t file_index, const ScanOptions& scan_range) {
+          auto const metadata = flevel_->files[file_index].file_metadata;
           if (metadata->FileIsStandAloneRangeTombstone()) {
             // Skip stand alone range deletion files.
-            continue;
+            return;
           }
-          auto& args = GetMultiScanArgForFile(i);
-          args.insert(start.value(), end.value(), opt.property_bag);
-        }
-      }
-    }
+          auto& args = GetMultiScanArgForFile(file_index);
+          args.insert(scan_range.range.start.value(),
+                      scan_range.range.limit.value(), scan_range.property_bag);
+        });
 
     StopWatch timer(clock_, db_statistics_, MULTISCAN_PREPARE_ITERATORS);
 
@@ -1221,7 +1361,8 @@ class LevelIterator final : public InternalIterator {
                           file_to_arg.first));
     }
 
-    if (so->use_async_io) {
+    if (so->use_async_io &&
+        (!range_tombstone_iter_required_ || range_tombstone_iter_ != nullptr)) {
       auto before = file_index_;
       // Pre-create and prepare only relevant file iterators
       for (auto& file_to_arg : *file_to_scan_opts_) {
@@ -1314,7 +1455,9 @@ class LevelIterator final : public InternalIterator {
         /*max_file_size_for_l0_meta_pin=*/0, smallest_compaction_key,
         largest_compaction_key, allow_unprepared_value_, &read_seq_,
         range_tombstone_iter_,
-        /*maybe_pin_table_handle=*/true);
+        /*maybe_pin_table_handle=*/
+        !table_cache_open_options_.open_ephemeral_table_reader,
+        /*file_open_metadata=*/nullptr, table_cache_open_options_);
   }
 
   // Check if current file being fully within iterate_lower_bound.
@@ -1371,6 +1514,7 @@ class LevelIterator final : public InternalIterator {
   //
   // *range_tombstone_iter_ points to range tombstones of the current SST file
   std::unique_ptr<TruncatedRangeDelIterator>* range_tombstone_iter_;
+  bool range_tombstone_iter_required_;
 
   // The sentinel key to be returned
   Slice sentinel_;
@@ -1391,6 +1535,7 @@ class LevelIterator final : public InternalIterator {
 
   Statistics* db_statistics_ = nullptr;
   SystemClock* clock_ = nullptr;
+  TableCacheOpenOptions table_cache_open_options_;
 
   // Our stored scan_opts for each prefix
   std::unique_ptr<ScanOptionsMap> file_to_scan_opts_ = nullptr;
@@ -1680,22 +1825,20 @@ bool LevelIterator::SkipEmptyFileForward() {
       if (scan_opts_ && FileHasMultiScanArg(file_index_)) {
         const ScanOptions& opts =
             GetMultiScanArgForFile(file_index_).GetScanRanges().front();
-        if (opts.range.start.has_value()) {
-          InternalKey target;
-          const size_t ts_size =
-              user_comparator_.user_comparator()->timestamp_size();
-          if (ts_size == 0) {
-            target = InternalKey(opts.range.start.value(), kMaxSequenceNumber,
-                                 kValueTypeForSeek);
-          } else {
-            std::string seek_key;
-            AppendKeyWithMaxTimestamp(&seek_key, opts.range.start.value(),
-                                      ts_size);
-            target =
-                InternalKey(seek_key, kMaxSequenceNumber, kValueTypeForSeek);
-          }
-          file_iter_.Seek(target.Encode());
+        assert(opts.range.start.has_value());
+        InternalKey target;
+        const size_t ts_size =
+            user_comparator_.user_comparator()->timestamp_size();
+        if (ts_size == 0) {
+          target = InternalKey(opts.range.start.value(), kMaxSequenceNumber,
+                               kValueTypeForSeek);
+        } else {
+          std::string seek_key;
+          AppendKeyWithMaxTimestamp(&seek_key, opts.range.start.value(),
+                                    ts_size);
+          target = InternalKey(seek_key, kMaxSequenceNumber, kValueTypeForSeek);
         }
+        file_iter_.Seek(target.Encode());
       } else {
         file_iter_.SeekToFirst();
       }
@@ -1742,14 +1885,8 @@ void LevelIterator::SkipEmptyFileBackward() {
 
 #ifndef NDEBUG
 bool LevelIterator::OverlapRange(const ScanOptions& opts, size_t file_index) {
-  return (user_comparator_.CompareWithoutTimestamp(
-              opts.range.start.value(), /*a_has_ts=*/false,
-              ExtractUserKey(flevel_->files[file_index].largest_key),
-              /*b_has_ts=*/true) <= 0 &&
-          user_comparator_.CompareWithoutTimestamp(
-              opts.range.limit.value(), /*a_has_ts=*/false,
-              ExtractUserKey(flevel_->files[file_index].smallest_key),
-              /*b_has_ts=*/true) > 0);
+  return MultiScanRangeOverlapsFile(user_comparator_, opts,
+                                    flevel_->files[file_index]);
 }
 #endif
 
@@ -2271,7 +2408,9 @@ InternalIterator* Version::TEST_GetLevelIterator(
       mutable_cf_options_, cfd_->internal_stats()->GetFileReadHist(level),
       TableReaderCaller::kUserIterator, IsFilterSkipped(level), level,
       nullptr /* range_del_agg */, nullptr /* compaction_boundaries */,
-      allow_unprepared_value, &tombstone_iter_ptr, db_statistics_, clock_);
+      allow_unprepared_value,
+      read_options.ignore_range_deletions ? nullptr : &tombstone_iter_ptr,
+      kMaxSequenceNumber, db_statistics_, clock_);
   if (read_options.ignore_range_deletions) {
     merge_iter_builder->AddIterator(level_iter);
   } else {
@@ -2342,19 +2481,22 @@ double VersionStorageInfo::GetEstimatedCompressionRatioAtLevel(
 void Version::AddIterators(const ReadOptions& read_options,
                            const FileOptions& soptions,
                            MergeIteratorBuilder* merge_iter_builder,
-                           bool allow_unprepared_value) {
+                           bool allow_unprepared_value, SequenceNumber read_seq,
+                           const MultiScanArgs* scan_opts) {
   assert(storage_info_.finalized_);
 
   for (int level = 0; level < storage_info_.num_non_empty_levels(); level++) {
     AddIteratorsForLevel(read_options, soptions, merge_iter_builder, level,
-                         allow_unprepared_value);
+                         allow_unprepared_value, read_seq, scan_opts);
   }
 }
 
 void Version::AddIteratorsForLevel(const ReadOptions& read_options,
                                    const FileOptions& soptions,
                                    MergeIteratorBuilder* merge_iter_builder,
-                                   int level, bool allow_unprepared_value) {
+                                   int level, bool allow_unprepared_value,
+                                   SequenceNumber read_seq,
+                                   const MultiScanArgs* scan_opts) {
   assert(storage_info_.finalized_);
   if (level >= storage_info_.num_non_empty_levels()) {
     // This is an empty level
@@ -2365,37 +2507,95 @@ void Version::AddIteratorsForLevel(const ReadOptions& read_options,
   }
 
   auto* arena = merge_iter_builder->GetArena();
+  const LevelFilesBrief& file_level = storage_info_.LevelFilesBrief(level);
+  const UserComparatorWrapper user_comparator(
+      cfd_->internal_comparator().user_comparator());
+
+  if (scan_opts != nullptr) {
+    const std::vector<size_t> file_indexes = GetMultiScanOverlappingFiles(
+        cfd_->internal_comparator(), user_comparator, file_level, level,
+        *scan_opts, level == 0 ? std::numeric_limits<size_t>::max() : 2);
+    if (file_indexes.empty()) {
+      return;
+    }
+
+    if (level == 0) {
+      const bool should_sample = should_sample_file_read();
+      for (size_t file_index : file_indexes) {
+        AddTableIteratorForLevel(
+            cfd_, read_options, soptions, mutable_cf_options_,
+            merge_iter_builder, arena, file_level, level, file_index,
+            /*skip_filters=*/false, max_file_size_for_l0_meta_pin_,
+            allow_unprepared_value, read_seq);
+        if (should_sample) {
+          sample_file_read_inc(file_level.files[file_index].file_metadata);
+        }
+      }
+      return;
+    }
+
+    if (file_indexes.size() == 1) {
+      AddTableIteratorForLevel(cfd_, read_options, soptions,
+                               mutable_cf_options_, merge_iter_builder, arena,
+                               file_level, level, file_indexes.front(),
+                               IsFilterSkipped(level),
+                               /*max_file_size_for_l0_meta_pin=*/0,
+                               allow_unprepared_value, read_seq);
+      return;
+    }
+
+#ifndef NDEBUG
+    for (size_t file_index = 0; file_index < file_level.num_files;
+         ++file_index) {
+      TEST_SYNC_POINT_CALLBACK("Version::AddIteratorsForLevel:AddedFile",
+                               file_level.files[file_index].file_metadata);
+    }
+#endif
+    auto* mem = arena->AllocateAligned(sizeof(LevelIterator));
+    std::unique_ptr<TruncatedRangeDelIterator>** tombstone_iter_ptr = nullptr;
+    auto level_iter = new (mem) LevelIterator(
+        cfd_->table_cache(), read_options, soptions,
+        cfd_->internal_comparator(), &file_level, mutable_cf_options_,
+        cfd_->internal_stats()->GetFileReadHist(level),
+        TableReaderCaller::kUserIterator, IsFilterSkipped(level), level,
+        /*range_del_agg=*/nullptr,
+        /*compaction_boundaries=*/nullptr, allow_unprepared_value,
+        read_options.ignore_range_deletions ? nullptr : &tombstone_iter_ptr,
+        read_seq, db_statistics_, clock_);
+#ifndef NDEBUG
+    std::pair<bool, bool> iterator_type(
+        false /* is_block_based_table_iterator */,
+        true /* is_level_iterator */);
+    TEST_SYNC_POINT_CALLBACK("Version::AddIteratorsForLevel:IteratorType",
+                             &iterator_type);
+#endif
+    if (read_options.ignore_range_deletions) {
+      merge_iter_builder->AddIterator(level_iter);
+    } else {
+      assert(tombstone_iter_ptr);
+      merge_iter_builder->AddPointAndTombstoneIterator(
+          level_iter, nullptr /* tombstone_iter */, tombstone_iter_ptr);
+    }
+    return;
+  }
+
   if (level == 0) {
     // Merge all level zero files together since they may overlap
-    std::unique_ptr<TruncatedRangeDelIterator> tombstone_iter = nullptr;
-    for (size_t i = 0; i < storage_info_.LevelFilesBrief(0).num_files; i++) {
-      const auto& file = storage_info_.LevelFilesBrief(0).files[i];
-      auto table_iter = cfd_->table_cache()->NewIterator(
-          read_options, soptions, cfd_->internal_comparator(),
-          *file.file_metadata, /*range_del_agg=*/nullptr, mutable_cf_options_,
-          nullptr, cfd_->internal_stats()->GetFileReadHist(0),
-          TableReaderCaller::kUserIterator, arena,
-          /*skip_filters=*/false, /*level=*/0, max_file_size_for_l0_meta_pin_,
-          /*smallest_compaction_key=*/nullptr,
-          /*largest_compaction_key=*/nullptr, allow_unprepared_value,
-          /*range_del_read_seqno=*/nullptr, &tombstone_iter,
-          /*maybe_pin_table_handle=*/true);
-      if (read_options.ignore_range_deletions) {
-        merge_iter_builder->AddIterator(table_iter);
-      } else {
-        merge_iter_builder->AddPointAndTombstoneIterator(
-            table_iter, std::move(tombstone_iter));
+    const bool should_sample = should_sample_file_read();
+    for (size_t i = 0; i < file_level.num_files; i++) {
+      AddTableIteratorForLevel(
+          cfd_, read_options, soptions, mutable_cf_options_, merge_iter_builder,
+          arena, file_level, level, i,
+          /*skip_filters=*/false, max_file_size_for_l0_meta_pin_,
+          allow_unprepared_value, read_seq);
+      if (should_sample) {
+        // Count once for every L0 file. This is done per iterator creation
+        // rather than Seek(), while files in other levels are sampled on
+        // seek/next/prev.
+        sample_file_read_inc(file_level.files[i].file_metadata);
       }
     }
-    if (should_sample_file_read()) {
-      // Count ones for every L0 files. This is done per iterator creation
-      // rather than Seek(), while files in other levels are sampled on
-      // seek/next/prev.
-      for (FileMetaData* meta : storage_info_.LevelFiles(0)) {
-        sample_file_read_inc(meta);
-      }
-    }
-  } else if (storage_info_.LevelFilesBrief(level).num_files > 0) {
+  } else if (file_level.num_files > 0) {
     // For levels > 0, we can use a concatenating iterator that sequentially
     // walks through the non-overlapping files in the level, opening them
     // lazily.
@@ -2403,12 +2603,13 @@ void Version::AddIteratorsForLevel(const ReadOptions& read_options,
     std::unique_ptr<TruncatedRangeDelIterator>** tombstone_iter_ptr = nullptr;
     auto level_iter = new (mem) LevelIterator(
         cfd_->table_cache(), read_options, soptions,
-        cfd_->internal_comparator(), &storage_info_.LevelFilesBrief(level),
-        mutable_cf_options_, cfd_->internal_stats()->GetFileReadHist(level),
+        cfd_->internal_comparator(), &file_level, mutable_cf_options_,
+        cfd_->internal_stats()->GetFileReadHist(level),
         TableReaderCaller::kUserIterator, IsFilterSkipped(level), level,
         /*range_del_agg=*/nullptr,
         /*compaction_boundaries=*/nullptr, allow_unprepared_value,
-        &tombstone_iter_ptr, db_statistics_, clock_);
+        read_options.ignore_range_deletions ? nullptr : &tombstone_iter_ptr,
+        read_seq, db_statistics_, clock_);
     if (read_options.ignore_range_deletions) {
       merge_iter_builder->AddIterator(level_iter);
     } else {
@@ -2464,7 +2665,8 @@ Status Version::OverlapWithLevelIterator(const ReadOptions& read_options,
         cfd_->internal_comparator(), &storage_info_.LevelFilesBrief(level),
         mutable_cf_options_, cfd_->internal_stats()->GetFileReadHist(level),
         TableReaderCaller::kUserIterator, IsFilterSkipped(level), level,
-        &range_del_agg, nullptr, false, nullptr, db_statistics_, clock_));
+        &range_del_agg, nullptr, false, nullptr, kMaxSequenceNumber,
+        db_statistics_, clock_));
     status = OverlapWithIterator(ucmp, smallest_user_key, largest_user_key,
                                  iter.get(), overlap);
   }
@@ -5520,16 +5722,19 @@ struct VersionSet::ManifestWriter {
   ColumnFamilyData* cfd;
   const autovector<VersionEdit*>& edit_list;
   const std::function<void(const Status&)> manifest_write_callback;
+  int max_file_opening_threads;
 
   explicit ManifestWriter(
       InstrumentedMutex* mu, ColumnFamilyData* _cfd,
       const autovector<VersionEdit*>& e,
-      const std::function<void(const Status&)>& manifest_wcb)
+      const std::function<void(const Status&)>& manifest_wcb,
+      int _max_file_opening_threads = 1)
       : done(false),
         cv(mu),
         cfd(_cfd),
         edit_list(e),
-        manifest_write_callback(manifest_wcb) {}
+        manifest_write_callback(manifest_wcb),
+        max_file_opening_threads(_max_file_opening_threads) {}
   ~ManifestWriter() { status.PermitUncheckedError(); }
 
   bool IsAllWalEdits() const {
@@ -5904,6 +6109,10 @@ Status VersionSet::ProcessManifestWrites(
   ManifestWriter& first_writer = writers.front();
   ManifestWriter* last_writer = &first_writer;
 
+  // Supports opening table readers with multiple threads, mostly useful for
+  // batched external sst file ingestion.
+  int batch_max_file_opening_threads = 1;
+
   assert(!manifest_writers_.empty());
   assert(manifest_writers_.front() == &first_writer);
 
@@ -5940,6 +6149,9 @@ Status VersionSet::ProcessManifestWrites(
     for (;;) {
       assert(!(*it)->edit_list.front()->IsColumnFamilyManipulation());
       last_writer = *it;
+      batch_max_file_opening_threads =
+          std::max(batch_max_file_opening_threads,
+                   last_writer->max_file_opening_threads);
       assert(last_writer != nullptr);
       assert(last_writer->cfd != nullptr);
       if (last_writer->cfd->IsDropped()) {
@@ -6111,9 +6323,32 @@ Status VersionSet::ProcessManifestWrites(
 
   uint64_t prev_manifest_file_size = manifest_file_size_;
   assert(pending_manifest_file_number_ == 0);
+  bool has_foreground_operation = false;
+  for (const VersionEdit* e : batch_edits) {
+    if (e->IsForegroundOperation()) {
+      has_foreground_operation = true;
+      break;
+    }
+  }
+  // For MANIFEST write batches with any foreground operation (external file
+  // ingestion/import, DeleteFilesInRange(s), and column family manipulations
+  // like CreateColumnFamily and DropColumnFamily), relax the size limit by 25%
+  // to reduce the likelihood of a user operation blocking on MANIFEST rotation.
+  // Background-only batches (flush/compaction) still rotate at the normal
+  // threshold.
+  // TODO/future: for workloads like atomic-replace ingestion-only, with zero
+  // or few flushes and compactions, it might be nice to trigger background
+  // manifest rotation if we are beyond the soft limit. But the vast majority
+  // of workloads should have plenty of background manifest ops to avoid
+  // foreground rotation.
+  uint64_t enforced_limit = tuned_max_manifest_file_size_;
+  if (has_foreground_operation) {
+    uint64_t new_limit = enforced_limit + enforced_limit / 4;
+    // don't keep in case of overflow
+    enforced_limit = std::max(enforced_limit, new_limit);
+  }
   if (!skip_manifest_write &&
-      (!descriptor_log_ ||
-       prev_manifest_file_size >= tuned_max_manifest_file_size_)) {
+      (!descriptor_log_ || prev_manifest_file_size >= enforced_limit)) {
     TEST_SYNC_POINT("VersionSet::ProcessManifestWrites:BeforeNewManifest");
     new_descriptor_log = true;
   } else {
@@ -6175,7 +6410,7 @@ Status VersionSet::ProcessManifestWrites(
                builder_guards.size() == versions.size());
         ColumnFamilyData* cfd = versions[i]->cfd_;
         s = builder_guards[i]->version_builder()->LoadTableHandlers(
-            cfd->internal_stats(), 1 /* max_threads */,
+            cfd->internal_stats(), batch_max_file_opening_threads,
             true /* prefetch_index_and_filter_in_cache */,
             false /* is_initial_load */, versions[i]->GetMutableCFOptions(),
             MaxFileSizeForL0MetaPin(versions[i]->GetMutableCFOptions()),
@@ -6284,6 +6519,8 @@ Status VersionSet::ProcessManifestWrites(
       io_s = SetCurrentFile(
           write_options, fs_.get(), dbname_, pending_manifest_file_number_,
           file_options_.temperature, dir_contains_current_file);
+      TEST_SYNC_POINT_CALLBACK(
+          "VersionSet::ProcessManifestWrites:AfterSetCurrentFile", &io_s);
       if (!io_s.ok()) {
         s = io_s;
         // Quarantine old manifest file in case new manifest file's CURRENT
@@ -6536,7 +6773,7 @@ Status VersionSet::LogAndApply(
     InstrumentedMutex* mu, FSDirectory* dir_contains_current_file,
     bool new_descriptor_log, const ColumnFamilyOptions* new_cf_options,
     const std::vector<std::function<void(const Status&)>>& manifest_wcbs,
-    const std::function<Status()>& pre_cb) {
+    const std::function<Status()>& pre_cb, int max_file_opening_threads) {
   mu->AssertHeld();
   int num_edits = 0;
   for (const auto& elist : edit_lists) {
@@ -6568,7 +6805,8 @@ Status VersionSet::LogAndApply(
   for (int i = 0; i < num_cfds; ++i) {
     const auto wcb =
         manifest_wcbs.empty() ? [](const Status&) {} : manifest_wcbs[i];
-    writers.emplace_back(mu, column_family_datas[i], edit_lists[i], wcb);
+    writers.emplace_back(mu, column_family_datas[i], edit_lists[i], wcb,
+                         max_file_opening_threads);
     manifest_writers_.push_back(&writers[i]);
   }
   assert(!writers.empty());
@@ -7496,6 +7734,28 @@ Status VersionSet::WriteCurrentStateToManifest(
       }
     }
   }
+
+  // Record the approximate compacted manifest size so it's available at
+  // recovery time for TuneMaxManifestFileSize(). This record must come last
+  // in WriteCurrentStateToManifest for an accurate size estimate.
+  {
+    // Include a rough estimate of this record's own size (~15 bytes for the
+    // VersionEdit payload + log record header).
+    constexpr uint64_t kEstimatedRecordOverhead = 15;
+    VersionEdit edit;
+    edit.SetLastCompactedManifestFileSize(log->file()->GetFileSize() +
+                                          kEstimatedRecordOverhead);
+    std::string record;
+    if (!edit.EncodeTo(&record)) {
+      return Status::Corruption("Unable to Encode VersionEdit:" +
+                                edit.DebugString(true));
+    }
+    io_s = log->AddRecord(write_options, record);
+    if (!io_s.ok()) {
+      return io_s;
+    }
+  }
+
   return Status::OK();
 }
 
@@ -7812,7 +8072,7 @@ InternalIterator* VersionSet::MakeInputIterator(
     RangeDelAggregator* range_del_agg,
     const FileOptions& file_options_compactions,
     const std::optional<const Slice>& start,
-    const std::optional<const Slice>& end) {
+    const std::optional<const Slice>& end, bool open_ephemeral_table_reader) {
   auto cfd = c->column_family_data();
   // Level-0 files have to be merged together.  For other levels,
   // we will make a concatenating iterator per level.
@@ -7831,6 +8091,8 @@ InternalIterator* VersionSet::MakeInputIterator(
       range_tombstones;
   size_t num = 0;
   [[maybe_unused]] size_t num_input_files = 0;
+  const TableCacheOpenOptions table_cache_open_options =
+      GetCompactionTableCacheOpenOptions(open_ephemeral_table_reader);
   for (size_t which = 0; which < c->num_input_levels(); which++) {
     const LevelFilesBrief* flevel = c->input_levels(which);
     num_input_files += flevel->num_files;
@@ -7867,7 +8129,9 @@ InternalIterator* VersionSet::MakeInputIterator(
               /*largest_compaction_key=*/nullptr,
               /*allow_unprepared_value=*/false,
               /*range_del_read_seqno=*/nullptr,
-              /*range_del_iter=*/&range_tombstone_iter);
+              /*range_del_iter=*/&range_tombstone_iter,
+              /*maybe_pin_table_handle=*/false,
+              /*file_open_metadata=*/nullptr, table_cache_open_options);
           range_tombstones.emplace_back(std::move(range_tombstone_iter),
                                         nullptr);
         }
@@ -7882,7 +8146,8 @@ InternalIterator* VersionSet::MakeInputIterator(
             TableReaderCaller::kCompaction, /*skip_filters=*/false,
             /*level=*/static_cast<int>(c->level(which)), range_del_agg,
             c->boundaries(which), false, &tombstone_iter_ptr,
-            db_options_->statistics.get(), clock_);
+            kMaxSequenceNumber, db_options_->statistics.get(), clock_,
+            open_ephemeral_table_reader);
         range_tombstones.emplace_back(nullptr, tombstone_iter_ptr);
       }
     }

db/version_set.h

@@ -85,6 +85,7 @@ class MergeIteratorBuilder;
 class SystemClock;
 class ManifestTailer;
 class FilePickerMultiGet;
+class MultiScanArgs;
 
 // VersionEdit is always supposed to be valid and it is used to point at
 // entries in Manifest. Ideally it should not be used as a container to
@@ -916,17 +917,27 @@ class Version {
   // yield the contents of this Version when merged together.
   // @param read_options Must outlive any iterator built by
   // `merger_iter_builder`.
+  // @param read_seq Snapshot sequence to use for range tombstone visibility.
+  // This is passed separately because lazy iterator initialization may happen
+  // after read_options.snapshot has been released by the caller.
+  // @param scan_opts Optional bounded scan ranges used to prune levels/files
+  // while building the iterator tree.
   void AddIterators(const ReadOptions& read_options,
                     const FileOptions& soptions,
                     MergeIteratorBuilder* merger_iter_builder,
-                    bool allow_unprepared_value);
+                    bool allow_unprepared_value, SequenceNumber read_seq,
+                    const MultiScanArgs* scan_opts = nullptr);
 
   // @param read_options Must outlive any iterator built by
   // `merger_iter_builder`.
+  // @param read_seq Snapshot sequence to use for range tombstone visibility.
+  // @param scan_opts Optional bounded scan ranges used to prune this level.
   void AddIteratorsForLevel(const ReadOptions& read_options,
                             const FileOptions& soptions,
                             MergeIteratorBuilder* merger_iter_builder,
-                            int level, bool allow_unprepared_value);
+                            int level, bool allow_unprepared_value,
+                            SequenceNumber read_seq,
+                            const MultiScanArgs* scan_opts = nullptr);
 
   Status OverlapWithLevelIterator(const ReadOptions&, const FileOptions&,
                                   const Slice& smallest_user_key,
@@ -1319,7 +1330,8 @@ class VersionSet {
       bool new_descriptor_log = false,
       const ColumnFamilyOptions* new_cf_options = nullptr,
       const std::vector<std::function<void(const Status&)>>& manifest_wcbs = {},
-      const std::function<Status()>& pre_cb = {});
+      const std::function<Status()>& pre_cb = {},
+      int max_file_opening_threads = 1);
 
   void WakeUpWaitingManifestWriters();
 
@@ -1558,12 +1570,16 @@ class VersionSet {
   // The caller should delete the iterator when no longer needed.
   // @param read_options Must outlive the returned iterator.
   // @param start, end indicates compaction range
+  // @param open_ephemeral_table_reader When true, the per-file iterators
+  //              bypass the shared TableCache and open fresh TableReaders
+  //              using `file_options_compactions`.
   InternalIterator* MakeInputIterator(
       const ReadOptions& read_options, const Compaction* c,
       RangeDelAggregator* range_del_agg,
       const FileOptions& file_options_compactions,
       const std::optional<const Slice>& start,
-      const std::optional<const Slice>& end);
+      const std::optional<const Slice>& end,
+      bool open_ephemeral_table_reader = false);
 
   // Add all files listed in any live version to *live_table_files and
   // *live_blob_files. Note that these lists may contain duplicates.
@@ -1950,7 +1966,8 @@ class ReactiveVersionSet : public VersionSet {
       InstrumentedMutex* /*mu*/, FSDirectory* /*dir_contains_current_file*/,
       bool /*new_descriptor_log*/, const ColumnFamilyOptions* /*new_cf_option*/,
       const std::vector<std::function<void(const Status&)>>& /*manifest_wcbs*/,
-      const std::function<Status()>& /*pre_cb*/) override {
+      const std::function<Status()>& /*pre_cb*/,
+      int /*max_file_opening_threads*/) override {
     return Status::NotSupported("not supported in reactive mode");
   }
 

db/version_util.cc

@@ -62,11 +62,14 @@ Status LoadTableHandlersHelper(
 
       TableCache::TypedHandle* handle = nullptr;
       TableReader* table_reader = nullptr;
+      const bool prefetch_index_and_filter_for_file =
+          prefetch_index_and_filter_in_cache &&
+          !file_meta->skip_index_and_filter_blocks_prefetch;
       auto status = table_cache->FindTable(
           read_options, file_options, internal_comparator, *file_meta, &handle,
           mutable_cf_options, &table_reader, false /* no_io */,
           internal_stats->GetFileReadHist(level), false /* skip_filters */,
-          level, prefetch_index_and_filter_in_cache,
+          level, prefetch_index_and_filter_for_file,
           max_file_size_for_l0_meta_pin, file_meta->temperature,
           true /* pin_table_handle */);
 

db_stress_tool/batched_ops_stress.cc

@@ -13,7 +13,9 @@
 namespace ROCKSDB_NAMESPACE {
 class BatchedOpsStressTest : public StressTest {
  public:
-  BatchedOpsStressTest() = default;
+  BatchedOpsStressTest(int db_index, const std::string& db_path,
+                       const std::string& ev_path, const std::string& sec_path)
+      : StressTest(db_index, db_path, ev_path, sec_path) {}
   virtual ~BatchedOpsStressTest() = default;
 
   bool IsStateTracked() const override { return false; }
@@ -722,7 +724,11 @@ class BatchedOpsStressTest : public StressTest {
   }
 };
 
-StressTest* CreateBatchedOpsStressTest() { return new BatchedOpsStressTest(); }
+StressTest* CreateBatchedOpsStressTest(int db_index, const std::string& db_path,
+                                       const std::string& ev_path,
+                                       const std::string& sec_path) {
+  return new BatchedOpsStressTest(db_index, db_path, ev_path, sec_path);
+}
 
 }  // namespace ROCKSDB_NAMESPACE
 #endif  // GFLAGS

db_stress_tool/cf_consistency_stress.cc

@@ -21,7 +21,10 @@
 namespace ROCKSDB_NAMESPACE {
 class CfConsistencyStressTest : public StressTest {
  public:
-  CfConsistencyStressTest() : batch_id_(0) {}
+  CfConsistencyStressTest(int db_index, const std::string& db_path,
+                          const std::string& ev_path,
+                          const std::string& sec_path)
+      : StressTest(db_index, db_path, ev_path, sec_path), batch_id_(0) {}
 
   ~CfConsistencyStressTest() override = default;
 
@@ -227,10 +230,10 @@ class CfConsistencyStressTest : public StressTest {
                    key, &value0);
 
       // Temporarily disable error injection for verification
-      if (fault_fs_guard) {
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kMetadataRead);
       }
 
@@ -279,10 +282,10 @@ class CfConsistencyStressTest : public StressTest {
       }
 
       //  Enable back error injection disabled for verification
-      if (fault_fs_guard) {
-        fault_fs_guard->EnableThreadLocalErrorInjection(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->EnableThreadLocalErrorInjection(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->EnableThreadLocalErrorInjection(
+        db_fault_injection_fs_->EnableThreadLocalErrorInjection(
             FaultInjectionIOType::kMetadataRead);
       }
       db_->ReleaseSnapshot(snapshot);
@@ -397,10 +400,10 @@ class CfConsistencyStressTest : public StressTest {
                          &cmp_result);
 
       //  Temporarily disable error injection for verification
-      if (fault_fs_guard) {
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kMetadataRead);
       }
 
@@ -559,10 +562,10 @@ class CfConsistencyStressTest : public StressTest {
       }
 
       //  Enable back error injection disabled for verification
-      if (fault_fs_guard) {
-        fault_fs_guard->EnableThreadLocalErrorInjection(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->EnableThreadLocalErrorInjection(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->EnableThreadLocalErrorInjection(
+        db_fault_injection_fs_->EnableThreadLocalErrorInjection(
             FaultInjectionIOType::kMetadataRead);
       }
     }
@@ -1543,8 +1546,11 @@ class CfConsistencyStressTest : public StressTest {
   std::deque<DebugEvent> recent_debug_events_;
 };
 
-StressTest* CreateCfConsistencyStressTest() {
-  return new CfConsistencyStressTest();
+StressTest* CreateCfConsistencyStressTest(int db_index,
+                                          const std::string& db_path,
+                                          const std::string& ev_path,
+                                          const std::string& sec_path) {
+  return new CfConsistencyStressTest(db_index, db_path, ev_path, sec_path);
 }
 
 }  // namespace ROCKSDB_NAMESPACE

db_stress_tool/db_stress_common.cc

@@ -13,16 +13,17 @@
 
 #include <cmath>
 
+#include "db_stress_tool/db_stress_test_base.h"
 #include "file/file_util.h"
 #include "rocksdb/secondary_cache.h"
 #include "util/file_checksum_helper.h"
 #include "util/xxhash.h"
 
-ROCKSDB_NAMESPACE::Env* db_stress_listener_env = nullptr;
-ROCKSDB_NAMESPACE::Env* db_stress_env = nullptr;
-std::shared_ptr<ROCKSDB_NAMESPACE::FaultInjectionTestFS> fault_fs_guard;
+ROCKSDB_NAMESPACE::Env* raw_env = nullptr;
 std::shared_ptr<ROCKSDB_NAMESPACE::SecondaryCache> compressed_secondary_cache;
 std::shared_ptr<ROCKSDB_NAMESPACE::Cache> block_cache;
+std::shared_ptr<ROCKSDB_NAMESPACE::WriteBufferManager> wbm;
+std::shared_ptr<ROCKSDB_NAMESPACE::RateLimiter> rate_limiter;
 enum ROCKSDB_NAMESPACE::CompressionType compression_type_e =
     ROCKSDB_NAMESPACE::kSnappyCompression;
 enum ROCKSDB_NAMESPACE::CompressionType bottommost_compression_type_e =
@@ -114,10 +115,10 @@ void PoolSizeChangeThread(void* v) {
     if (new_thread_pool_size < 1) {
       new_thread_pool_size = 1;
     }
-    db_stress_env->SetBackgroundThreads(new_thread_pool_size,
-                                        ROCKSDB_NAMESPACE::Env::Priority::LOW);
+    raw_env->SetBackgroundThreads(new_thread_pool_size,
+                                  ROCKSDB_NAMESPACE::Env::Priority::LOW);
     // Sleep up to 3 seconds
-    db_stress_env->SleepForMicroseconds(
+    raw_env->SleepForMicroseconds(
         thread->rand.Next() % FLAGS_compaction_thread_pool_adjust_interval *
             1000 +
         1);
@@ -144,7 +145,7 @@ void DbVerificationThread(void* v) {
     if (!shared->HasVerificationFailedYet()) {
       stress_test->ContinuouslyVerifyDb(thread);
     }
-    db_stress_env->SleepForMicroseconds(
+    raw_env->SleepForMicroseconds(
         thread->rand.Next() % FLAGS_continuous_verification_interval * 1000 +
         1);
   }
@@ -166,7 +167,7 @@ void CompressedCacheSetCapacityThread(void* v) {
         return;
       }
     }
-    db_stress_env->SleepForMicroseconds(FLAGS_secondary_cache_update_interval);
+    raw_env->SleepForMicroseconds(FLAGS_secondary_cache_update_interval);
     if (FLAGS_compressed_secondary_cache_size > 0) {
       Status s = compressed_secondary_cache->SetCapacity(0);
       size_t capacity;
@@ -174,7 +175,7 @@ void CompressedCacheSetCapacityThread(void* v) {
         s = compressed_secondary_cache->GetCapacity(capacity);
         assert(capacity == 0);
       }
-      db_stress_env->SleepForMicroseconds(10 * 1000 * 1000);
+      raw_env->SleepForMicroseconds(10 * 1000 * 1000);
       if (s.ok()) {
         s = compressed_secondary_cache->SetCapacity(
             FLAGS_compressed_secondary_cache_size);
@@ -201,7 +202,7 @@ void CompressedCacheSetCapacityThread(void* v) {
         block_cache->SetCapacity(capacity - adjustment);
         fprintf(stdout, "New cache capacity = %zu\n",
                 block_cache->GetCapacity());
-        db_stress_env->SleepForMicroseconds(10 * 1000 * 1000);
+        raw_env->SleepForMicroseconds(10 * 1000 * 1000);
         block_cache->SetCapacity(capacity);
       } else {
         Status s;
@@ -214,7 +215,7 @@ void CompressedCacheSetCapacityThread(void* v) {
         s = UpdateTieredCache(block_cache, /*capacity*/ -1,
                               new_comp_cache_ratio);
         if (s.ok()) {
-          db_stress_env->SleepForMicroseconds(10 * 1000 * 1000);
+          raw_env->SleepForMicroseconds(10 * 1000 * 1000);
         }
         if (s.ok()) {
           s = UpdateTieredCache(block_cache, /*capacity*/ -1,
@@ -231,36 +232,37 @@ void CompressedCacheSetCapacityThread(void* v) {
 
 #ifndef NDEBUG
 static void SetupFaultInjectionForRemoteCompaction(SharedState* shared) {
-  if (!fault_fs_guard) {
+  auto fault_fs = shared->GetStressTest()->GetDbFaultInjectionFs();
+  if (!fault_fs) {
     return;
   }
 
-  fault_fs_guard->SetThreadLocalErrorContext(
+  fault_fs->SetThreadLocalErrorContext(
       FaultInjectionIOType::kRead, shared->GetSeed(), FLAGS_read_fault_one_in,
       FLAGS_inject_error_severity == 1 /* retryable */,
       FLAGS_inject_error_severity == 2 /* has_data_loss*/);
-  fault_fs_guard->EnableThreadLocalErrorInjection(FaultInjectionIOType::kRead);
+  fault_fs->EnableThreadLocalErrorInjection(FaultInjectionIOType::kRead);
 
-  fault_fs_guard->SetThreadLocalErrorContext(
+  fault_fs->SetThreadLocalErrorContext(
       FaultInjectionIOType::kWrite, shared->GetSeed(), FLAGS_write_fault_one_in,
       FLAGS_inject_error_severity == 1 /* retryable */,
       FLAGS_inject_error_severity == 2 /* has_data_loss*/);
-  fault_fs_guard->EnableThreadLocalErrorInjection(FaultInjectionIOType::kWrite);
+  fault_fs->EnableThreadLocalErrorInjection(FaultInjectionIOType::kWrite);
 
-  fault_fs_guard->SetThreadLocalErrorContext(
+  fault_fs->SetThreadLocalErrorContext(
       FaultInjectionIOType::kMetadataRead, shared->GetSeed(),
       FLAGS_metadata_read_fault_one_in,
       FLAGS_inject_error_severity == 1 /* retryable */,
       FLAGS_inject_error_severity == 2 /* has_data_loss*/);
-  fault_fs_guard->EnableThreadLocalErrorInjection(
+  fault_fs->EnableThreadLocalErrorInjection(
       FaultInjectionIOType::kMetadataRead);
 
-  fault_fs_guard->SetThreadLocalErrorContext(
+  fault_fs->SetThreadLocalErrorContext(
       FaultInjectionIOType::kMetadataWrite, shared->GetSeed(),
       FLAGS_metadata_write_fault_one_in,
       FLAGS_inject_error_severity == 1 /* retryable */,
       FLAGS_inject_error_severity == 2 /* has_data_loss*/);
-  fault_fs_guard->EnableThreadLocalErrorInjection(
+  fault_fs->EnableThreadLocalErrorInjection(
       FaultInjectionIOType::kMetadataWrite);
 }
 #endif  // NDEBUG
@@ -268,7 +270,7 @@ static void SetupFaultInjectionForRemoteCompaction(SharedState* shared) {
 static CompactionServiceOptionsOverride CreateOverrideOptions(
     const Options& options, const CompactionServiceJobInfo& job_info) {
   CompactionServiceOptionsOverride override_options{
-      .env = db_stress_env,
+      .env = options.env,
       .file_checksum_gen_factory = options.file_checksum_gen_factory,
       .merge_operator = options.merge_operator,
       .compaction_filter = options.compaction_filter,
@@ -311,14 +313,8 @@ static CompactionServiceOptionsOverride CreateOverrideOptions(
 }
 
 static Status CleanupOutputDirectory(const std::string& output_directory) {
-#ifndef NDEBUG
-  // Temporarily disable fault injection to ensure deletion always succeeds
-  if (fault_fs_guard) {
-    fault_fs_guard->DisableAllThreadLocalErrorInjection();
-  }
-#endif  // NDEBUG
-
-  Status s = DestroyDir(db_stress_env, output_directory);
+  // Uses raw_env (no fault injection); cleanup must always succeed.
+  Status s = DestroyDir(raw_env, output_directory);
   if (!s.ok()) {
     fprintf(stderr,
             "Failed to destroy output directory %s when allow_resumption is "
@@ -327,7 +323,7 @@ static Status CleanupOutputDirectory(const std::string& output_directory) {
   }
 
   if (s.ok()) {
-    s = db_stress_env->CreateDir(output_directory);
+    s = raw_env->CreateDir(output_directory);
     if (!s.ok()) {
       fprintf(stderr,
               "Failed to recreate output directory %s when allow_resumption is "
@@ -336,13 +332,6 @@ static Status CleanupOutputDirectory(const std::string& output_directory) {
     }
   }
 
-#ifndef NDEBUG
-  // Re-enable fault injection after deletion
-  if (fault_fs_guard) {
-    fault_fs_guard->EnableAllThreadLocalErrorInjection();
-  }
-#endif  // NDEBUG
-
   return s;
 }
 
@@ -497,7 +486,7 @@ void RemoteCompactionWorkerThread(void* v) {
           shared, stress_test, rand, successful_compaction_end_to_end_micros);
     }
 
-    db_stress_env->SleepForMicroseconds(
+    raw_env->SleepForMicroseconds(
         thread->rand.Next() % FLAGS_remote_compaction_worker_interval * 1000 +
         1);
   }
@@ -679,7 +668,7 @@ bool VerifyIteratorAttributeGroups(
 }
 
 std::string GetNowNanos() {
-  uint64_t t = db_stress_env->NowNanos();
+  uint64_t t = raw_env->NowNanos();
   std::string ret;
   PutFixed64(&ret, t);
   return ret;
@@ -691,7 +680,7 @@ uint64_t GetWriteUnixTime(ThreadState* thread) {
                FLAGS_preclude_last_level_data_seconds);
   static uint64_t kFallbackTime = std::numeric_limits<uint64_t>::max();
   int64_t write_time = 0;
-  Status s = db_stress_env->GetCurrentTime(&write_time);
+  Status s = raw_env->GetCurrentTime(&write_time);
   uint32_t write_time_mode = thread->rand.Uniform(3);
   if (write_time_mode == 0 || !s.ok()) {
     return kFallbackTime;
@@ -806,6 +795,8 @@ std::shared_ptr<FileChecksumGenFactory> GetFileChecksumImpl(
   return std::make_shared<DbStressChecksumGenFactory>(internal_name);
 }
 
+// Expected values state files are always on local filesystem (Python owns
+// this dir), so use Env::Default() (PosixEnv) even when raw_env is remote.
 Status DeleteFilesInDirectory(const std::string& dirname) {
   std::vector<std::string> filenames;
   Status s = Env::Default()->GetChildren(dirname, &filenames);
@@ -880,8 +871,12 @@ Status DestroyUnverifiedSubdir(const std::string& dirname) {
 Status DbStressDestroyDb(const std::string& db_path) {
   Status s;
   Options options;
-  // NOTE: using db_stress_listener_env in order to see obsolete MANIFEST files
-  options.env = db_stress_listener_env;
+  // Use raw_env (no DbStressFSWrapper) because DbStressFSWrapper renames
+  // MANIFEST files to MANIFEST-xxx_renamed_ instead of deleting them during
+  // normal DB operation (to preserve history for debugging). Using raw_env
+  // here ensures DestroyDB actually deletes these files. DestroyDir below
+  // catches any remaining files not known to DestroyDB.
+  options.env = raw_env;
   // Remove DB files in a principled way to avoid issues
   if (FLAGS_use_blob_db) {
     s = blob_db::DestroyBlobDB(db_path, options, blob_db::BlobDBOptions());
@@ -891,9 +886,8 @@ Status DbStressDestroyDb(const std::string& db_path) {
   if (!s.ok()) {
     return s;
   }
-  // Remove everything else recursively, only reporting success if able to
-  // delete everything
-  return DestroyDir(db_stress_listener_env, db_path);
+  // Remove everything else recursively (catches MANIFEST_renamed_ files)
+  return DestroyDir(raw_env, db_path);
 }
 
 }  // namespace ROCKSDB_NAMESPACE

db_stress_tool/db_stress_common.h

@@ -49,6 +49,7 @@
 #include "rocksdb/advanced_options.h"
 #include "rocksdb/cache.h"
 #include "rocksdb/env.h"
+#include "rocksdb/rate_limiter.h"
 #include "rocksdb/slice.h"
 #include "rocksdb/slice_transform.h"
 #include "rocksdb/statistics.h"
@@ -61,6 +62,7 @@
 #include "rocksdb/utilities/transaction.h"
 #include "rocksdb/utilities/transaction_db.h"
 #include "rocksdb/write_batch.h"
+#include "rocksdb/write_buffer_manager.h"
 #include "test_util/testutil.h"
 #include "util/coding.h"
 #include "util/compression.h"
@@ -114,6 +116,7 @@ DECLARE_double(memtable_prefix_bloom_size_ratio);
 DECLARE_bool(memtable_whole_key_filtering);
 DECLARE_int32(open_files);
 DECLARE_bool(open_files_async);
+DECLARE_bool(async_wal_precreate);
 DECLARE_uint64(compressed_secondary_cache_size);
 DECLARE_int32(compressed_secondary_cache_numshardbits);
 DECLARE_int32(secondary_cache_update_interval);
@@ -190,6 +193,7 @@ DECLARE_string(db);
 DECLARE_string(secondaries_base);
 DECLARE_bool(test_secondary);
 DECLARE_string(expected_values_dir);
+DECLARE_int32(num_dbs);
 DECLARE_bool(expected_state_trace_debug);
 DECLARE_int64(expected_state_trace_debug_key);
 DECLARE_int32(expected_state_trace_debug_max_logs);
@@ -197,6 +201,7 @@ DECLARE_bool(verify_checksum);
 DECLARE_bool(mmap_read);
 DECLARE_bool(mmap_write);
 DECLARE_bool(use_direct_reads);
+DECLARE_bool(use_direct_io_for_compaction_reads);
 DECLARE_bool(use_direct_io_for_flush_and_compaction);
 DECLARE_bool(mock_direct_io);
 DECLARE_bool(statistics);
@@ -224,6 +229,8 @@ DECLARE_uint64(backup_max_size);
 DECLARE_int32(checkpoint_one_in);
 DECLARE_int32(ingest_external_file_one_in);
 DECLARE_int32(ingest_external_file_width);
+DECLARE_int32(ingest_external_file_prepare_commit_one_in);
+DECLARE_int32(ingest_external_file_use_file_info_one_in);
 DECLARE_int32(compact_files_one_in);
 DECLARE_int32(compact_range_one_in);
 DECLARE_int32(promote_l0_one_in);
@@ -299,7 +306,7 @@ DECLARE_string(default_write_temperature);
 DECLARE_string(default_temperature);
 DECLARE_uint32(verify_output_flags);
 DECLARE_bool(paranoid_memory_checks);
-DECLARE_bool(memtable_veirfy_per_key_checksum_on_seek);
+DECLARE_bool(memtable_verify_per_key_checksum_on_seek);
 DECLARE_bool(memtable_batch_lookup_optimization);
 
 // Options for transaction dbs.
@@ -462,6 +469,7 @@ DECLARE_uint32(ingest_wbwi_one_in);
 DECLARE_bool(universal_reduce_file_locking);
 DECLARE_bool(use_multiscan);
 DECLARE_bool(multiscan_use_async_io);
+DECLARE_bool(read_scoped_block_buffer_provider);
 DECLARE_uint64(multiscan_max_prefetch_memory_bytes);
 
 // Compaction deletion trigger declarations for stress testing
@@ -472,18 +480,23 @@ DECLARE_int32(compaction_on_deletion_window_size);
 DECLARE_double(compaction_on_deletion_ratio);
 DECLARE_double(read_triggered_compaction_threshold);
 
+DECLARE_string(listener_uri);
+
 constexpr long KB = 1024;
 constexpr int kRandomValueMaxFactor = 3;
 constexpr int kValueMaxLen = 100;
 constexpr uint32_t kLargePrimeForCommonFactorSkew = 1872439133;
 
-// wrapped posix environment
-extern ROCKSDB_NAMESPACE::Env* db_stress_env;
-extern ROCKSDB_NAMESPACE::Env* db_stress_listener_env;
-extern std::shared_ptr<ROCKSDB_NAMESPACE::FaultInjectionTestFS> fault_fs_guard;
+// Base env from --env_uri/--fs_uri. No wrappers (no DbStressFSWrapper, no
+// fault injection). Could be PosixEnv or remote. Used for infrastructure ops
+// (threads, time, dirs, cleanup). Per-StressTest env adds fault injection +
+// DbStressFSWrapper on top for DB I/O.
+extern ROCKSDB_NAMESPACE::Env* raw_env;
 extern std::shared_ptr<ROCKSDB_NAMESPACE::SecondaryCache>
     compressed_secondary_cache;
 extern std::shared_ptr<ROCKSDB_NAMESPACE::Cache> block_cache;
+extern std::shared_ptr<ROCKSDB_NAMESPACE::WriteBufferManager> wbm;
+extern std::shared_ptr<ROCKSDB_NAMESPACE::RateLimiter> rate_limiter;
 
 extern enum ROCKSDB_NAMESPACE::CompressionType compression_type_e;
 extern enum ROCKSDB_NAMESPACE::CompressionType bottommost_compression_type_e;
@@ -828,10 +841,21 @@ AttributeGroups GenerateAttributeGroups(
     const std::vector<ColumnFamilyHandle*>& cfhs, uint32_t value_base,
     const Slice& slice);
 
-StressTest* CreateCfConsistencyStressTest();
-StressTest* CreateBatchedOpsStressTest();
-StressTest* CreateNonBatchedOpsStressTest();
-StressTest* CreateMultiOpsTxnsStressTest();
+StressTest* CreateCfConsistencyStressTest(int db_index,
+                                          const std::string& db_path,
+                                          const std::string& ev_path,
+                                          const std::string& sec_path);
+StressTest* CreateBatchedOpsStressTest(int db_index, const std::string& db_path,
+                                       const std::string& ev_path,
+                                       const std::string& sec_path);
+StressTest* CreateNonBatchedOpsStressTest(int db_index,
+                                          const std::string& db_path,
+                                          const std::string& ev_path,
+                                          const std::string& sec_path);
+StressTest* CreateMultiOpsTxnsStressTest(int db_index,
+                                         const std::string& db_path,
+                                         const std::string& ev_path,
+                                         const std::string& sec_path);
 void CheckAndSetOptionsForMultiOpsTxnStressTest();
 void InitializeHotKeyGenerator(double alpha);
 int64_t GetOneHotKeyID(double rand_seed, int64_t max_key);

db_stress_tool/db_stress_compaction_service.cc

@@ -1,4 +1,4 @@
-//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  Copyright (c) Meta Platforms, Inc. and affiliates.
 //  This source code is licensed under both the GPLv2 (found in the
 //  COPYING file in the root directory) and Apache 2.0 License
 //  (found in the LICENSE.Apache file in the root directory).

db_stress_tool/db_stress_driver.cc

@@ -10,12 +10,14 @@
 
 #include "db_stress_tool/db_stress_shared_state.h"
 #ifdef GFLAGS
+#include <mutex>
+
 #include "db_stress_tool/db_stress_common.h"
 #include "utilities/fault_injection_fs.h"
 
 namespace ROCKSDB_NAMESPACE {
 void ThreadBody(void* v) {
-  ThreadStatusUtil::RegisterThread(db_stress_env, ThreadStatus::USER);
+  ThreadStatusUtil::RegisterThread(raw_env, ThreadStatus::USER);
   ThreadState* thread = static_cast<ThreadState*>(v);
   SharedState* shared = thread->shared;
 
@@ -64,17 +66,21 @@ void ThreadBody(void* v) {
   ThreadStatusUtil::UnregisterThread();
 }
 bool RunStressTestImpl(SharedState* shared) {
-  SystemClock* clock = db_stress_env->GetSystemClock().get();
+  SystemClock* clock = raw_env->GetSystemClock().get();
   StressTest* stress = shared->GetStressTest();
+  const std::string db_label =
+      (FLAGS_num_dbs > 1) ? "[" + stress->GetDbLabel() + "] " : "";
 
   if (shared->ShouldVerifyAtBeginning() && FLAGS_preserve_unverified_changes) {
-    Status s = InitUnverifiedSubdir(FLAGS_db);
-    if (s.ok() && !FLAGS_expected_values_dir.empty()) {
-      s = InitUnverifiedSubdir(FLAGS_expected_values_dir);
+    const auto& db_path = stress->GetDbPath();
+    const auto& ev_dir = stress->GetExpectedValuesDir();
+    Status s = InitUnverifiedSubdir(db_path);
+    if (s.ok() && !ev_dir.empty()) {
+      s = InitUnverifiedSubdir(ev_dir);
     }
     if (!s.ok()) {
-      fprintf(stderr, "Failed to setup unverified state dir: %s\n",
-              s.ToString().c_str());
+      fprintf(stderr, "%sFailed to setup unverified state dir: %s\n",
+              db_label.c_str(), s.ToString().c_str());
       exit(1);
     }
   }
@@ -84,24 +90,15 @@ bool RunStressTestImpl(SharedState* shared) {
 
   uint32_t n = FLAGS_threads;
   uint64_t now = clock->NowMicros();
-  fprintf(stdout, "%s Initializing worker threads\n",
-          clock->TimeToString(now / 1000000).c_str());
+  fprintf(stdout, "%s %sInitializing worker threads\n",
+          clock->TimeToString(now / 1000000).c_str(), db_label.c_str());
 
   shared->SetThreads(n);
 
-  if (FLAGS_compaction_thread_pool_adjust_interval > 0) {
-    shared->IncBgThreads();
-  }
-
   if (FLAGS_continuous_verification_interval > 0) {
     shared->IncBgThreads();
   }
 
-  if (FLAGS_compressed_secondary_cache_size > 0 ||
-      FLAGS_compressed_secondary_cache_ratio > 0.0) {
-    shared->IncBgThreads();
-  }
-
   uint32_t remote_compaction_worker_thread_count =
       FLAGS_remote_compaction_worker_threads;
   if (remote_compaction_worker_thread_count > 0) {
@@ -113,25 +110,39 @@ bool RunStressTestImpl(SharedState* shared) {
   std::vector<ThreadState*> threads(n);
   for (uint32_t i = 0; i < n; i++) {
     threads[i] = new ThreadState(i, shared);
-    db_stress_env->StartThread(ThreadBody, threads[i]);
+    raw_env->StartThread(ThreadBody, threads[i]);
   }
 
+  // Spawn at most one PoolSizeChangeThread globally. All DBs share the
+  // same raw_env, so one thread resizing the pool is sufficient.
+  // IncBgThreads inside call_once so only the launching DB's SharedState
+  // tracks this thread (avoids deadlock for non-first DBs).
   ThreadState bg_thread(0, shared);
+  static std::once_flag pool_size_thread_flag;
   if (FLAGS_compaction_thread_pool_adjust_interval > 0) {
-    db_stress_env->StartThread(PoolSizeChangeThread, &bg_thread);
+    std::call_once(pool_size_thread_flag, [&]() {
+      shared->IncBgThreads();
+      raw_env->StartThread(PoolSizeChangeThread, &bg_thread);
+    });
   }
 
   ThreadState continuous_verification_thread(0, shared);
   if (FLAGS_continuous_verification_interval > 0) {
-    db_stress_env->StartThread(DbVerificationThread,
-                               &continuous_verification_thread);
+    raw_env->StartThread(DbVerificationThread, &continuous_verification_thread);
   }
 
+  // Spawn at most one CompressedCacheSetCapacityThread globally. The cache
+  // is shared across all DBs, and the thread's SetCapacity(0)/assert(==0)
+  // sequence would race if multiple DBs each spawned their own copy.
   ThreadState compressed_cache_set_capacity_thread(0, shared);
+  static std::once_flag compressed_cache_thread_flag;
   if (FLAGS_compressed_secondary_cache_size > 0 ||
       FLAGS_compressed_secondary_cache_ratio > 0.0) {
-    db_stress_env->StartThread(CompressedCacheSetCapacityThread,
-                               &compressed_cache_set_capacity_thread);
+    std::call_once(compressed_cache_thread_flag, [&]() {
+      shared->IncBgThreads();
+      raw_env->StartThread(CompressedCacheSetCapacityThread,
+                           &compressed_cache_set_capacity_thread);
+    });
   }
 
   std::vector<ThreadState*> remote_compaction_worker_threads;
@@ -141,7 +152,7 @@ bool RunStressTestImpl(SharedState* shared) {
     for (uint32_t i = 0; i < remote_compaction_worker_thread_count; i++) {
       ThreadState* ts = new ThreadState(i, shared);
       remote_compaction_worker_threads.push_back(ts);
-      db_stress_env->StartThread(RemoteCompactionWorkerThread, ts);
+      raw_env->StartThread(RemoteCompactionWorkerThread, ts);
     }
   }
 
@@ -156,16 +167,19 @@ bool RunStressTestImpl(SharedState* shared) {
     }
     if (shared->ShouldVerifyAtBeginning()) {
       if (shared->HasVerificationFailedYet()) {
-        fprintf(stderr, "Crash-recovery verification failed :(\n");
+        fprintf(stderr, "%sCrash-recovery verification failed :(\n",
+                db_label.c_str());
       } else {
-        fprintf(stdout, "Crash-recovery verification passed :)\n");
-        Status s = DestroyUnverifiedSubdir(FLAGS_db);
-        if (s.ok() && !FLAGS_expected_values_dir.empty()) {
-          s = DestroyUnverifiedSubdir(FLAGS_expected_values_dir);
+        fprintf(stdout, "%sCrash-recovery verification passed :)\n",
+                db_label.c_str());
+        const auto& ev_dir = stress->GetExpectedValuesDir();
+        Status s = DestroyUnverifiedSubdir(stress->GetDbPath());
+        if (s.ok() && !ev_dir.empty()) {
+          s = DestroyUnverifiedSubdir(ev_dir);
         }
         if (!s.ok()) {
-          fprintf(stderr, "Failed to cleanup unverified state dir: %s\n",
-                  s.ToString().c_str());
+          fprintf(stderr, "%sFailed to cleanup unverified state dir: %s\n",
+                  db_label.c_str(), s.ToString().c_str());
           exit(1);
         }
       }
@@ -174,15 +188,16 @@ bool RunStressTestImpl(SharedState* shared) {
     if (!FLAGS_verification_only) {
       // This is after the verification step to avoid making all those `Get()`s
       // and `MultiGet()`s contend on the DB-wide trace mutex.
-      if (!FLAGS_expected_values_dir.empty()) {
+      if (!stress->GetExpectedValuesDir().empty()) {
         stress->TrackExpectedState(shared);
       }
 
       if (FLAGS_sync_fault_injection || FLAGS_write_fault_one_in > 0) {
-        fault_fs_guard->SetFilesystemDirectWritable(false);
-        fault_fs_guard->SetInjectUnsyncedDataLoss(FLAGS_sync_fault_injection);
+        auto fault_fs = stress->GetDbFaultInjectionFs();
+        fault_fs->SetFilesystemDirectWritable(false);
+        fault_fs->SetInjectUnsyncedDataLoss(FLAGS_sync_fault_injection);
         if (FLAGS_exclude_wal_from_write_fault_injection) {
-          fault_fs_guard->SetFileTypesExcludedFromWriteFaultInjection(
+          fault_fs->SetFileTypesExcludedFromWriteFaultInjection(
               {FileType::kWalFile});
         }
       }
@@ -190,8 +205,8 @@ bool RunStressTestImpl(SharedState* shared) {
         Status s = stress->EnableAutoCompaction();
         assert(s.ok());
       }
-      fprintf(stdout, "%s Starting database operations\n",
-              clock->TimeToString(now / 1000000).c_str());
+      fprintf(stdout, "%s %sStarting database operations\n",
+              clock->TimeToString(now / 1000000).c_str(), db_label.c_str());
 
       shared->SetStart();
       shared->GetCondVar()->SignalAll();
@@ -201,14 +216,17 @@ bool RunStressTestImpl(SharedState* shared) {
 
       now = clock->NowMicros();
       if (FLAGS_test_batches_snapshots) {
-        fprintf(stdout, "%s Limited verification already done during gets\n",
-                clock->TimeToString((uint64_t)now / 1000000).c_str());
+        fprintf(stdout, "%s %sLimited verification already done during gets\n",
+                clock->TimeToString((uint64_t)now / 1000000).c_str(),
+                db_label.c_str());
       } else if (FLAGS_skip_verifydb) {
-        fprintf(stdout, "%s Verification skipped\n",
-                clock->TimeToString((uint64_t)now / 1000000).c_str());
+        fprintf(stdout, "%s %sVerification skipped\n",
+                clock->TimeToString((uint64_t)now / 1000000).c_str(),
+                db_label.c_str());
       } else {
-        fprintf(stdout, "%s Starting verification\n",
-                clock->TimeToString((uint64_t)now / 1000000).c_str());
+        fprintf(stdout, "%s %sStarting verification\n",
+                clock->TimeToString((uint64_t)now / 1000000).c_str(),
+                db_label.c_str());
       }
 
       shared->SetStartVerify();
@@ -230,7 +248,7 @@ bool RunStressTestImpl(SharedState* shared) {
     for (unsigned int i = 1; i < n; i++) {
       threads[0]->stats.Merge(threads[i]->stats);
     }
-    threads[0]->stats.Report("Stress Test");
+    threads[0]->stats.Report((db_label + "Stress Test").c_str());
   }
 
   for (unsigned int i = 0; i < n; i++) {
@@ -241,8 +259,8 @@ bool RunStressTestImpl(SharedState* shared) {
   now = clock->NowMicros();
   if (!FLAGS_skip_verifydb && !FLAGS_test_batches_snapshots &&
       !shared->HasVerificationFailedYet()) {
-    fprintf(stdout, "%s Verification successful\n",
-            clock->TimeToString(now / 1000000).c_str());
+    fprintf(stdout, "%s %sVerification successful\n",
+            clock->TimeToString(now / 1000000).c_str(), db_label.c_str());
   }
 
   if (!FLAGS_verification_only) {
@@ -260,6 +278,9 @@ bool RunStressTestImpl(SharedState* shared) {
       shared->GetCondVar()->Wait();
     }
   }
+  now = clock->NowMicros();
+  fprintf(stdout, "%s %sStress test bg threads finished\n",
+          clock->TimeToString(now / 1000000).c_str(), db_label.c_str());
 
   assert(remote_compaction_worker_threads.size() ==
          remote_compaction_worker_thread_count);
@@ -271,13 +292,13 @@ bool RunStressTestImpl(SharedState* shared) {
   }
 
   if (shared->HasVerificationFailedYet()) {
-    fprintf(stderr, "Verification failed :(\n");
+    fprintf(stderr, "%sVerification failed :(\n", db_label.c_str());
     return false;
   }
   return true;
 }
 bool RunStressTest(SharedState* shared) {
-  ThreadStatusUtil::RegisterThread(db_stress_env, ThreadStatus::USER);
+  ThreadStatusUtil::RegisterThread(raw_env, ThreadStatus::USER);
   bool result = RunStressTestImpl(shared);
   ThreadStatusUtil::UnregisterThread();
   return result;

db_stress_tool/db_stress_gflags.cc

@@ -206,6 +206,10 @@ DEFINE_int32(open_files, ROCKSDB_NAMESPACE::Options().max_open_files,
 DEFINE_bool(open_files_async, ROCKSDB_NAMESPACE::Options().open_files_async,
             "Options.open_files_async");
 
+DEFINE_bool(async_wal_precreate,
+            ROCKSDB_NAMESPACE::Options().async_wal_precreate,
+            "Options.async_wal_precreate");
+
 DEFINE_uint64(compressed_secondary_cache_size, 0,
               "Number of bytes to use as a cache of compressed data."
               " 0 means use default settings.");
@@ -728,6 +732,11 @@ DEFINE_bool(mmap_write, ROCKSDB_NAMESPACE::Options().allow_mmap_writes,
 DEFINE_bool(use_direct_reads, ROCKSDB_NAMESPACE::Options().use_direct_reads,
             "Use O_DIRECT for reading data");
 
+DEFINE_bool(use_direct_io_for_compaction_reads,
+            ROCKSDB_NAMESPACE::Options().use_direct_io_for_compaction_reads,
+            "Use O_DIRECT for compaction-input SST reads only, while keeping "
+            "user reads buffered");
+
 DEFINE_bool(use_direct_io_for_flush_and_compaction,
             ROCKSDB_NAMESPACE::Options().use_direct_io_for_flush_and_compaction,
             "Use O_DIRECT for writing data");
@@ -858,6 +867,19 @@ DEFINE_int32(ingest_external_file_one_in, 0,
 DEFINE_int32(ingest_external_file_width, 100,
              "The width of the ingested external files.");
 
+DEFINE_int32(ingest_external_file_prepare_commit_one_in, 0,
+             "If non-zero, an ingestion that would call IngestExternalFile() "
+             "instead uses the two-phase PrepareFileIngestion()/"
+             "CommitFileIngestion() API once for every N such ingestions on "
+             "average, occasionally dropping the prepared handle without "
+             "committing to exercise the rollback path. 0 disables it.");
+
+DEFINE_int32(ingest_external_file_use_file_info_one_in, 0,
+             "If non-zero, the ingestexternalfile flow reuses each file's "
+             "metadata via IngestExternalFileArg::file_infos (from "
+             "SstFileWriter::Finish) once every N ingestions on average, so "
+             "ingestion skips re-opening and scanning the files.");
+
 DEFINE_int32(compact_files_one_in, 0,
              "If non-zero, then CompactFiles() will be called once for every N "
              "operations on average.  0 indicates CompactFiles() is disabled.");
@@ -1017,7 +1039,8 @@ DEFINE_int32(iterpercent, 10,
 static const bool FLAGS_iterpercent_dummy __attribute__((__unused__)) =
     RegisterFlagValidator(&FLAGS_iterpercent, &ValidateInt32Percent);
 
-DEFINE_uint64(num_iterations, 10, "Number of iterations per MultiIterate run");
+DEFINE_uint64(num_iterations, 10,
+              "Number of iterations per iterator or MultiScan run");
 static const bool FLAGS_num_iterations_dummy __attribute__((__unused__)) =
     RegisterFlagValidator(&FLAGS_num_iterations, &ValidateUint32Range);
 
@@ -1626,9 +1649,9 @@ DEFINE_bool(paranoid_memory_checks,
             "Sets CF option paranoid_memory_checks.");
 
 DEFINE_bool(
-    memtable_veirfy_per_key_checksum_on_seek,
-    ROCKSDB_NAMESPACE::Options().memtable_veirfy_per_key_checksum_on_seek,
-    "Sets CF option memtable_veirfy_per_key_checksum_on_seek.");
+    memtable_verify_per_key_checksum_on_seek,
+    ROCKSDB_NAMESPACE::Options().memtable_verify_per_key_checksum_on_seek,
+    "Sets CF option memtable_verify_per_key_checksum_on_seek.");
 
 DEFINE_bool(memtable_batch_lookup_optimization,
             ROCKSDB_NAMESPACE::Options().memtable_batch_lookup_optimization,
@@ -1664,6 +1687,10 @@ DEFINE_double(read_triggered_compaction_threshold,
               ROCKSDB_NAMESPACE::Options().read_triggered_compaction_threshold,
               "Sets CF option read_triggered_compaction_threshold.");
 
+DEFINE_string(listener_uri, "",
+              "URI for an additional EventListener to attach (e.g. "
+              "auto_tuner://). Empty means none.");
+
 DEFINE_bool(
     auto_refresh_iterator_with_snapshot,
     ROCKSDB_NAMESPACE::ReadOptions().auto_refresh_iterator_with_snapshot,
@@ -1697,10 +1724,27 @@ DEFINE_bool(use_multiscan, false,
 DEFINE_bool(multiscan_use_async_io, false,
             "If set, enable async_io for MultiScan operations.");
 
+DEFINE_bool(read_scoped_block_buffer_provider, false,
+            "If set, configure ReadOptions::read_scoped_block_buffer_provider "
+            "with a stress-test provider for supported scan reads.");
+
 DEFINE_uint64(multiscan_max_prefetch_memory_bytes, 0,
               "If non-zero, sets the max_prefetch_memory_bytes on the "
               "IODispatcher used for MultiScan. This limits the total memory "
               "used for prefetching data blocks across all concurrent "
               "MultiScan ReadSets.");
 
+DEFINE_int32(num_dbs, 1,
+             "Number of DB instances to run in parallel. "
+             "For num_dbs=1: --db is the DB path directly (default). "
+             "For num_dbs>1: --db is a parent directory; db_stress "
+             "auto-creates db_0/, db_1/, ... subdirs underneath "
+             "(same for --expected_values_dir and secondaries). "
+             "Per-DB flags: threads, max_key, ops_per_thread, reopen, "
+             "column_families, and all DB options. "
+             "Shared across all instances: background env threads "
+             "(compaction, flush pool), cache, write_buffer_manager, "
+             "compressed_secondary_cache, rate_limiter, and "
+             "compaction_thread_pool_adjust_interval.");
+
 #endif  // GFLAGS

db_stress_tool/db_stress_listener.cc

@@ -7,6 +7,7 @@
 
 #include <cstdint>
 
+#include "db_stress_tool/db_stress_test_base.h"
 #include "file/file_util.h"
 #include "rocksdb/file_system.h"
 #include "util/coding_lean.h"
@@ -15,6 +16,21 @@ namespace ROCKSDB_NAMESPACE {
 
 #ifdef GFLAGS
 
+DbStressListener::DbStressListener(
+    const std::string& db_name, const std::vector<DbPath>& db_paths,
+    const std::vector<ColumnFamilyDescriptor>& column_families,
+    SharedState* shared)
+    : db_name_(db_name),
+      db_paths_(db_paths),
+      column_families_(column_families),
+      num_pending_file_creations_(0),
+      unique_ids_(shared->GetStressTest()->GetExpectedValuesDir().empty()
+                      ? db_name
+                      : shared->GetStressTest()->GetExpectedValuesDir()),
+      shared_(shared),
+      db_fault_injection_fs_(shared->GetStressTest()->GetDbFaultInjectionFs()) {
+}
+
 UniqueIdVerifier::UniqueIdVerifier(const std::string& dir)
     : path_(dir + "/.unique_ids") {
   // We expect such a small number of files generated during this test

db_stress_tool/db_stress_listener.h

@@ -6,7 +6,9 @@
 #ifdef GFLAGS
 #pragma once
 
+#include <cinttypes>
 #include <mutex>
+#include <unordered_map>
 #include <unordered_set>
 
 #include "db_stress_tool/db_stress_compaction_service.h"
@@ -19,13 +21,12 @@
 #include "rocksdb/listener.h"
 #include "rocksdb/table_properties.h"
 #include "rocksdb/unique_id.h"
+#include "util/atomic.h"
 #include "util/gflags_compat.h"
 #include "util/random.h"
 #include "utilities/fault_injection_fs.h"
 DECLARE_int32(compact_files_one_in);
 
-extern std::shared_ptr<ROCKSDB_NAMESPACE::FaultInjectionTestFS> fault_fs_guard;
-
 namespace ROCKSDB_NAMESPACE {
 
 // Verify across process executions that all seen IDs are unique
@@ -55,27 +56,22 @@ class DbStressListener : public EventListener {
   DbStressListener(const std::string& db_name,
                    const std::vector<DbPath>& db_paths,
                    const std::vector<ColumnFamilyDescriptor>& column_families,
-                   SharedState* shared)
-      : db_name_(db_name),
-        db_paths_(db_paths),
-        column_families_(column_families),
-        num_pending_file_creations_(0),
-        unique_ids_(FLAGS_expected_values_dir.empty()
-                        ? db_name
-                        : FLAGS_expected_values_dir),
-        shared_(shared) {}
+                   SharedState* shared);
 
   const char* Name() const override { return kClassName(); }
   static const char* kClassName() { return "DBStressListener"; }
 
   ~DbStressListener() override { assert(num_pending_file_creations_ == 0); }
+
+  void OnDBShutdownBegin(DB* /*db*/) override { shutting_down_.Store(true); }
+
   void OnFlushCompleted(DB* /*db*/, const FlushJobInfo& info) override {
     assert(IsValidColumnFamilyName(info.cf_name));
     VerifyFilePath(info.file_path);
     // pretending doing some work here
     RandomSleep();
-    if (fault_fs_guard) {
-      fault_fs_guard->DisableAllThreadLocalErrorInjection();
+    if (db_fault_injection_fs_) {
+      db_fault_injection_fs_->DisableAllThreadLocalErrorInjection();
     }
     shared_->SetPersistedSeqno(info.largest_seqno);
   }
@@ -83,46 +79,118 @@ class DbStressListener : public EventListener {
   void OnFlushBegin(DB* /*db*/,
                     const FlushJobInfo& /*flush_job_info*/) override {
     RandomSleep();
-    if (fault_fs_guard) {
-      fault_fs_guard->SetThreadLocalErrorContext(
+    if (db_fault_injection_fs_) {
+      db_fault_injection_fs_->SetThreadLocalErrorContext(
           FaultInjectionIOType::kRead, static_cast<uint32_t>(FLAGS_seed),
           FLAGS_read_fault_one_in,
           FLAGS_inject_error_severity == 1 /* retryable */,
           FLAGS_inject_error_severity == 2 /* has_data_loss*/);
-      fault_fs_guard->EnableThreadLocalErrorInjection(
+      db_fault_injection_fs_->EnableThreadLocalErrorInjection(
           FaultInjectionIOType::kRead);
 
-      fault_fs_guard->SetThreadLocalErrorContext(
+      db_fault_injection_fs_->SetThreadLocalErrorContext(
           FaultInjectionIOType::kWrite, static_cast<uint32_t>(FLAGS_seed),
           FLAGS_write_fault_one_in,
           FLAGS_inject_error_severity == 1 /* retryable */,
           FLAGS_inject_error_severity == 2 /* has_data_loss*/);
-      fault_fs_guard->EnableThreadLocalErrorInjection(
+      db_fault_injection_fs_->EnableThreadLocalErrorInjection(
           FaultInjectionIOType::kWrite);
 
-      fault_fs_guard->SetThreadLocalErrorContext(
+      db_fault_injection_fs_->SetThreadLocalErrorContext(
           FaultInjectionIOType::kMetadataRead,
           static_cast<uint32_t>(FLAGS_seed), FLAGS_metadata_read_fault_one_in,
           FLAGS_inject_error_severity == 1 /* retryable */,
           FLAGS_inject_error_severity == 2 /* has_data_loss*/);
-      fault_fs_guard->EnableThreadLocalErrorInjection(
+      db_fault_injection_fs_->EnableThreadLocalErrorInjection(
           FaultInjectionIOType::kMetadataRead);
 
-      fault_fs_guard->SetThreadLocalErrorContext(
+      db_fault_injection_fs_->SetThreadLocalErrorContext(
           FaultInjectionIOType::kMetadataWrite,
           static_cast<uint32_t>(FLAGS_seed), FLAGS_metadata_write_fault_one_in,
           FLAGS_inject_error_severity == 1 /* retryable */,
           FLAGS_inject_error_severity == 2 /* has_data_loss*/);
-      fault_fs_guard->EnableThreadLocalErrorInjection(
+      db_fault_injection_fs_->EnableThreadLocalErrorInjection(
           FaultInjectionIOType::kMetadataWrite);
     }
   }
 
-  void OnTableFileDeleted(const TableFileDeletionInfo& /*info*/) override {
+  void OnTableFileDeleted(const TableFileDeletionInfo& info) override {
+    // A file must not be deleted while it is still between
+    // OnCompactionBegin and OnCompactionPreCommit (i.e. actively being
+    // compacted with being_compacted == true). (Perhaps more realistically,
+    // this is checking for failure to call OnCompactionPreCommit.)
+    //
+    // During shutdown, compaction notifications are skipped so tracking
+    // state may be stale -- skip the check.
+    if (!shutting_down_.Load()) {
+      std::lock_guard<std::mutex> lock(compacting_files_mu_);
+      uint64_t file_number = FileNumberFromPath(info.file_path);
+      if (file_number != 0 &&
+          compacting_files_.find(file_number) != compacting_files_.end()) {
+        fprintf(stderr,
+                "OnTableFileDeleted for file tracked as being compacted "
+                "(between Begin and PreCommit): file_number=%" PRIu64 "\n",
+                file_number);
+        fflush(stderr);
+        std::abort();
+      }
+    }
     RandomSleep();
   }
 
-  void OnCompactionBegin(DB* /*db*/, const CompactionJobInfo& /*ci*/) override {
+  void OnCompactionBegin(DB* /*db*/, const CompactionJobInfo& ci) override {
+    // Sanity check inspired by a Meta-internal check: the same input file must
+    // not be in two concurrent compactions. Inserting input file numbers into a
+    // shared set on Begin and removing them on PreCommit (rather than on
+    // Completed) is what exercises the new OnCompactionPreCommit
+    // callback. Removing on Completed would race with another picker that
+    // grabs the same file as soon as ReleaseCompactionFiles flips
+    // FileMetaData::being_compacted back to false; PreCommit runs
+    // while being_compacted is still true and so closes that race.
+    {
+      std::lock_guard<std::mutex> lock(compacting_files_mu_);
+      for (const auto& info : ci.input_file_infos) {
+        auto [_, inserted] = compacting_files_.insert(info.file_number);
+        if (!inserted) {
+          fprintf(stderr,
+                  "Concurrent compaction of SST file detected: cf=%s "
+                  "file_number=%" PRIu64 "\n",
+                  ci.cf_name.c_str(), info.file_number);
+          fflush(stderr);
+          std::abort();
+        }
+      }
+    }
+    RandomSleep();
+  }
+
+  void OnCompactionPreCommit(DB* /*db*/, const CompactionJobInfo& ci) override {
+    // Pair with OnCompactionBegin's bookkeeping: move files from
+    // compacting_files_ and record the job for Completed verification.
+    {
+      std::lock_guard<std::mutex> lock(compacting_files_mu_);
+      std::unordered_set<uint64_t> job_files;
+      for (const auto& info : ci.input_file_infos) {
+        size_t erased = compacting_files_.erase(info.file_number);
+        if (erased != 1) {
+          fprintf(stderr,
+                  "OnCompactionPreCommit for file not in tracking set: "
+                  "cf=%s file_number=%" PRIu64 "\n",
+                  ci.cf_name.c_str(), info.file_number);
+          fflush(stderr);
+          std::abort();
+        }
+        job_files.insert(info.file_number);
+      }
+      auto [_, inserted] =
+          precommitted_jobs_.emplace(ci.job_id, std::move(job_files));
+      if (!inserted) {
+        fprintf(stderr, "OnCompactionPreCommit: duplicate job_id %d\n",
+                ci.job_id);
+        fflush(stderr);
+        std::abort();
+      }
+    }
     RandomSleep();
   }
 
@@ -135,49 +203,77 @@ class DbStressListener : public EventListener {
     for (const auto& file_path : ci.output_files) {
       VerifyFilePath(file_path);
     }
+
+    // Verify that OnCompactionPreCommit fired before OnCompactionCompleted
+    // for the same job, with matching input files.
+    {
+      std::lock_guard<std::mutex> lock(compacting_files_mu_);
+      auto it = precommitted_jobs_.find(ci.job_id);
+      if (it == precommitted_jobs_.end()) {
+        fprintf(stderr,
+                "OnCompactionCompleted without prior OnCompactionPreCommit: "
+                "cf=%s job_id=%d\n",
+                ci.cf_name.c_str(), ci.job_id);
+        fflush(stderr);
+        std::abort();
+      }
+      // Verify input file sets match between PreCommit and Completed.
+      for (const auto& info : ci.input_file_infos) {
+        if (it->second.find(info.file_number) == it->second.end()) {
+          fprintf(stderr,
+                  "OnCompactionCompleted: input file %" PRIu64
+                  " not in PreCommit set for job_id=%d\n",
+                  info.file_number, ci.job_id);
+          fflush(stderr);
+          std::abort();
+        }
+      }
+      precommitted_jobs_.erase(it);
+    }
+
     // pretending doing some work here
     RandomSleep();
   }
 
   void OnSubcompactionBegin(const SubcompactionJobInfo& /* si */) override {
-    if (fault_fs_guard) {
-      fault_fs_guard->SetThreadLocalErrorContext(
+    if (db_fault_injection_fs_) {
+      db_fault_injection_fs_->SetThreadLocalErrorContext(
           FaultInjectionIOType::kRead, static_cast<uint32_t>(FLAGS_seed),
           FLAGS_read_fault_one_in,
           FLAGS_inject_error_severity == 1 /* retryable */,
           FLAGS_inject_error_severity == 2 /* has_data_loss*/);
-      fault_fs_guard->EnableThreadLocalErrorInjection(
+      db_fault_injection_fs_->EnableThreadLocalErrorInjection(
           FaultInjectionIOType::kRead);
 
-      fault_fs_guard->SetThreadLocalErrorContext(
+      db_fault_injection_fs_->SetThreadLocalErrorContext(
           FaultInjectionIOType::kWrite, static_cast<uint32_t>(FLAGS_seed),
           FLAGS_write_fault_one_in,
           FLAGS_inject_error_severity == 1 /* retryable */,
           FLAGS_inject_error_severity == 2 /* has_data_loss*/);
-      fault_fs_guard->EnableThreadLocalErrorInjection(
+      db_fault_injection_fs_->EnableThreadLocalErrorInjection(
           FaultInjectionIOType::kWrite);
 
-      fault_fs_guard->SetThreadLocalErrorContext(
+      db_fault_injection_fs_->SetThreadLocalErrorContext(
           FaultInjectionIOType::kMetadataRead,
           static_cast<uint32_t>(FLAGS_seed), FLAGS_metadata_read_fault_one_in,
           FLAGS_inject_error_severity == 1 /* retryable */,
           FLAGS_inject_error_severity == 2 /* has_data_loss*/);
-      fault_fs_guard->EnableThreadLocalErrorInjection(
+      db_fault_injection_fs_->EnableThreadLocalErrorInjection(
           FaultInjectionIOType::kMetadataRead);
 
-      fault_fs_guard->SetThreadLocalErrorContext(
+      db_fault_injection_fs_->SetThreadLocalErrorContext(
           FaultInjectionIOType::kMetadataWrite,
           static_cast<uint32_t>(FLAGS_seed), FLAGS_metadata_write_fault_one_in,
           FLAGS_inject_error_severity == 1 /* retryable */,
           FLAGS_inject_error_severity == 2 /* has_data_loss*/);
-      fault_fs_guard->EnableThreadLocalErrorInjection(
+      db_fault_injection_fs_->EnableThreadLocalErrorInjection(
           FaultInjectionIOType::kMetadataWrite);
     }
   }
 
   void OnSubcompactionCompleted(const SubcompactionJobInfo& /* si */) override {
-    if (fault_fs_guard) {
-      fault_fs_guard->DisableAllThreadLocalErrorInjection();
+    if (db_fault_injection_fs_) {
+      db_fault_injection_fs_->DisableAllThreadLocalErrorInjection();
     }
   }
 
@@ -270,11 +366,12 @@ class DbStressListener : public EventListener {
                             Status /* bg_error */,
                             bool* /* auto_recovery */) override {
     RandomSleep();
-    if (FLAGS_error_recovery_with_no_fault_injection && fault_fs_guard) {
-      fault_fs_guard->DisableAllThreadLocalErrorInjection();
+    if (FLAGS_error_recovery_with_no_fault_injection &&
+        db_fault_injection_fs_) {
+      db_fault_injection_fs_->DisableAllThreadLocalErrorInjection();
       // TODO(hx235): only exempt the flush thread during error recovery instead
       // of all the flush threads from error injection
-      fault_fs_guard->SetIOActivitiesExcludedFromFaultInjection(
+      db_fault_injection_fs_->SetIOActivitiesExcludedFromFaultInjection(
           {Env::IOActivity::kFlush});
     }
   }
@@ -282,9 +379,10 @@ class DbStressListener : public EventListener {
   void OnErrorRecoveryEnd(
       const BackgroundErrorRecoveryInfo& /*info*/) override {
     RandomSleep();
-    if (FLAGS_error_recovery_with_no_fault_injection && fault_fs_guard) {
-      fault_fs_guard->EnableAllThreadLocalErrorInjection();
-      fault_fs_guard->SetIOActivitiesExcludedFromFaultInjection({});
+    if (FLAGS_error_recovery_with_no_fault_injection &&
+        db_fault_injection_fs_) {
+      db_fault_injection_fs_->EnableAllThreadLocalErrorInjection();
+      db_fault_injection_fs_->SetIOActivitiesExcludedFromFaultInjection({});
     }
   }
 
@@ -375,8 +473,43 @@ class DbStressListener : public EventListener {
   std::atomic<int> num_pending_file_creations_;
   UniqueIdVerifier unique_ids_;
   SharedState* shared_;
+  std::shared_ptr<FaultInjectionTestFS> db_fault_injection_fs_;
   mutable std::mutex bg_pressure_mu_;
   BackgroundJobPressure last_bg_pressure_;
+  // Files (by file_number) currently in flight from OnCompactionBegin to
+  // OnCompactionPreCommit. Used to detect concurrent compaction of
+  // the same SST file -- the bug fixed by the OnCompactionPreCommit
+  // callback. Protected by compacting_files_mu_.
+  std::mutex compacting_files_mu_;
+  std::unordered_set<uint64_t> compacting_files_;
+  // Set when DBImpl begins shutdown to suppress false positives from stale
+  // tracking when compaction callbacks are skipped during shutdown.
+  Atomic<bool> shutting_down_{false};
+  // Jobs that have passed OnCompactionPreCommit but not yet
+  // OnCompactionCompleted. Maps job_id -> input file numbers.
+  // Used to verify Begin -> PreCommit -> Completed ordering per job.
+  // Protected by compacting_files_mu_.
+  std::unordered_map<int, std::unordered_set<uint64_t>> precommitted_jobs_;
+
+  // Extract file number from a file path like "/path/to/000123.sst".
+  // Returns 0 if the path cannot be parsed.
+  static uint64_t FileNumberFromPath(const std::string& file_path) {
+    size_t pos = file_path.find_last_of('/');
+    // Avoid copying file_path when no '/' separator is found
+    std::string file_name_buf;
+    if (pos != std::string::npos) {
+      file_name_buf = file_path.substr(pos + 1);
+    }
+    const std::string& file_name =
+        (pos == std::string::npos) ? file_path : file_name_buf;
+    uint64_t file_number = 0;
+    FileType file_type;
+    if (ParseFileName(file_name, &file_number, &file_type) &&
+        file_type == kTableFile) {
+      return file_number;
+    }
+    return 0;
+  }
 };
 }  // namespace ROCKSDB_NAMESPACE
 #endif  // GFLAGS

db_stress_tool/db_stress_shared_state.cc

@@ -11,7 +11,109 @@
 #ifdef GFLAGS
 #include "db_stress_tool/db_stress_shared_state.h"
 
+#include "db_stress_tool/db_stress_test_base.h"
+
 namespace ROCKSDB_NAMESPACE {
 thread_local bool SharedState::ignore_read_error;
+
+SharedState::SharedState(Env* /*env*/, StressTest* stress_test)
+    : cv_(&mu_),
+      seed_(static_cast<uint32_t>(FLAGS_seed)),
+      max_key_(FLAGS_max_key),
+      log2_keys_per_lock_(static_cast<uint32_t>(FLAGS_log2_keys_per_lock)),
+      num_threads_(0),
+      num_initialized_(0),
+      num_populated_(0),
+      vote_reopen_(0),
+      num_done_(0),
+      start_(false),
+      start_verify_(false),
+      num_bg_threads_(0),
+      should_stop_bg_thread_(false),
+      bg_thread_finished_(0),
+      stress_test_(stress_test),
+      verification_failure_(false),
+      should_stop_test_(false),
+      no_overwrite_ids_(GenerateNoOverwriteIds()),
+      expected_state_manager_(nullptr),
+      printing_verification_results_(false),
+      start_timestamp_(Env::Default()->NowNanos()) {
+  Status status;
+  // TODO: We should introduce a way to explicitly disable verification
+  // during shutdown. When that is disabled and FLAGS_expected_values_dir
+  // is empty (disabling verification at startup), we can skip tracking
+  // expected state. Only then should we permit bypassing the below feature
+  // compatibility checks.
+  const auto& expected_values_dir = stress_test_->GetExpectedValuesDir();
+  if (!expected_values_dir.empty()) {
+    if (!std::atomic<uint32_t>{}.is_lock_free() ||
+        !std::atomic<uint64_t>{}.is_lock_free()) {
+      std::ostringstream status_s;
+      status_s << "Cannot use --expected_values_dir on platforms without "
+                  "lock-free "
+               << (!std::atomic<uint32_t>{}.is_lock_free()
+                       ? "std::atomic<uint32_t>"
+                       : "std::atomic<uint64_t>");
+      status = Status::InvalidArgument(status_s.str());
+    }
+
+    if (status.ok() && FLAGS_clear_column_family_one_in > 0) {
+      status = Status::InvalidArgument(
+          "Cannot use --expected_values_dir on when "
+          "--clear_column_family_one_in is greater than zero.");
+    }
+  }
+  if (status.ok()) {
+    if (expected_values_dir.empty()) {
+      expected_state_manager_.reset(
+          new AnonExpectedStateManager(FLAGS_max_key, FLAGS_column_families));
+    } else {
+      expected_state_manager_.reset(new FileExpectedStateManager(
+          FLAGS_max_key, FLAGS_column_families, expected_values_dir));
+    }
+    status = expected_state_manager_->Open();
+  }
+  if (!status.ok()) {
+    fprintf(stderr, "Failed setting up expected state with error: %s\n",
+            status.ToString().c_str());
+    exit(1);  // NOLINT(concurrency-mt-unsafe)
+  }
+
+  if (FLAGS_test_batches_snapshots) {
+    fprintf(stdout, "No lock creation because test_batches_snapshots set\n");
+    return;
+  }
+
+  long num_locks = static_cast<long>(max_key_ >> log2_keys_per_lock_);
+  if (max_key_ & ((1 << log2_keys_per_lock_) - 1)) {
+    num_locks++;
+  }
+  fprintf(stdout, "Creating %ld locks\n", num_locks * FLAGS_column_families);
+  key_locks_.resize(FLAGS_column_families);
+
+  for (int i = 0; i < FLAGS_column_families; ++i) {
+    key_locks_[i].reset(new port::Mutex[num_locks]);
+  }
+  if (FLAGS_read_fault_one_in || FLAGS_metadata_read_fault_one_in) {
+#ifdef NDEBUG
+    // Unsupported in release mode because it relies on
+    // `IGNORE_STATUS_IF_ERROR` to distinguish faults not expected to lead to
+    // failure.
+    fprintf(stderr,
+            "Cannot set nonzero value for --read_fault_one_in in "
+            "release mode.");
+    exit(1);  // NOLINT(concurrency-mt-unsafe)
+#else         // NDEBUG
+    SyncPoint::GetInstance()->SetCallBack("FaultInjectionIgnoreError",
+                                          IgnoreReadErrorCallback);
+    SyncPoint::GetInstance()->EnableProcessing();
+#endif        // NDEBUG
+  }
+}
+
+bool SharedState::ShouldVerifyAtBeginning() const {
+  return !stress_test_->GetExpectedValuesDir().empty();
+}
+
 }  // namespace ROCKSDB_NAMESPACE
 #endif  // GFLAGS

db_stress_tool/db_stress_shared_state.h

@@ -24,7 +24,6 @@ DECLARE_uint64(log2_keys_per_lock);
 DECLARE_int32(threads);
 DECLARE_int32(column_families);
 DECLARE_int32(nooverwritepercent);
-DECLARE_string(expected_values_dir);
 DECLARE_int32(clear_column_family_one_in);
 DECLARE_bool(test_batches_snapshots);
 DECLARE_int32(compaction_thread_pool_adjust_interval);
@@ -78,99 +77,7 @@ class SharedState {
   // for those calls
   static thread_local bool ignore_read_error;
 
-  SharedState(Env* /*env*/, StressTest* stress_test)
-      : cv_(&mu_),
-        seed_(static_cast<uint32_t>(FLAGS_seed)),
-        max_key_(FLAGS_max_key),
-        log2_keys_per_lock_(static_cast<uint32_t>(FLAGS_log2_keys_per_lock)),
-        num_threads_(0),
-        num_initialized_(0),
-        num_populated_(0),
-        vote_reopen_(0),
-        num_done_(0),
-        start_(false),
-        start_verify_(false),
-        num_bg_threads_(0),
-        should_stop_bg_thread_(false),
-        bg_thread_finished_(0),
-        stress_test_(stress_test),
-        verification_failure_(false),
-        should_stop_test_(false),
-        no_overwrite_ids_(GenerateNoOverwriteIds()),
-        expected_state_manager_(nullptr),
-        printing_verification_results_(false),
-        start_timestamp_(Env::Default()->NowNanos()) {
-    Status status;
-    // TODO: We should introduce a way to explicitly disable verification
-    // during shutdown. When that is disabled and FLAGS_expected_values_dir
-    // is empty (disabling verification at startup), we can skip tracking
-    // expected state. Only then should we permit bypassing the below feature
-    // compatibility checks.
-    if (!FLAGS_expected_values_dir.empty()) {
-      if (!std::atomic<uint32_t>{}.is_lock_free() ||
-          !std::atomic<uint64_t>{}.is_lock_free()) {
-        std::ostringstream status_s;
-        status_s << "Cannot use --expected_values_dir on platforms without "
-                    "lock-free "
-                 << (!std::atomic<uint32_t>{}.is_lock_free()
-                         ? "std::atomic<uint32_t>"
-                         : "std::atomic<uint64_t>");
-        status = Status::InvalidArgument(status_s.str());
-      }
-
-      if (status.ok() && FLAGS_clear_column_family_one_in > 0) {
-        status = Status::InvalidArgument(
-            "Cannot use --expected_values_dir on when "
-            "--clear_column_family_one_in is greater than zero.");
-      }
-    }
-    if (status.ok()) {
-      if (FLAGS_expected_values_dir.empty()) {
-        expected_state_manager_.reset(
-            new AnonExpectedStateManager(FLAGS_max_key, FLAGS_column_families));
-      } else {
-        expected_state_manager_.reset(new FileExpectedStateManager(
-            FLAGS_max_key, FLAGS_column_families, FLAGS_expected_values_dir));
-      }
-      status = expected_state_manager_->Open();
-    }
-    if (!status.ok()) {
-      fprintf(stderr, "Failed setting up expected state with error: %s\n",
-              status.ToString().c_str());
-      exit(1);
-    }
-
-    if (FLAGS_test_batches_snapshots) {
-      fprintf(stdout, "No lock creation because test_batches_snapshots set\n");
-      return;
-    }
-
-    long num_locks = static_cast<long>(max_key_ >> log2_keys_per_lock_);
-    if (max_key_ & ((1 << log2_keys_per_lock_) - 1)) {
-      num_locks++;
-    }
-    fprintf(stdout, "Creating %ld locks\n", num_locks * FLAGS_column_families);
-    key_locks_.resize(FLAGS_column_families);
-
-    for (int i = 0; i < FLAGS_column_families; ++i) {
-      key_locks_[i].reset(new port::Mutex[num_locks]);
-    }
-    if (FLAGS_read_fault_one_in || FLAGS_metadata_read_fault_one_in) {
-#ifdef NDEBUG
-      // Unsupported in release mode because it relies on
-      // `IGNORE_STATUS_IF_ERROR` to distinguish faults not expected to lead to
-      // failure.
-      fprintf(stderr,
-              "Cannot set nonzero value for --read_fault_one_in in "
-              "release mode.");
-      exit(1);
-#else   // NDEBUG
-      SyncPoint::GetInstance()->SetCallBack("FaultInjectionIgnoreError",
-                                            IgnoreReadErrorCallback);
-      SyncPoint::GetInstance()->EnableProcessing();
-#endif  // NDEBUG
-    }
-  }
+  SharedState(Env* env, StressTest* stress_test);
 
   ~SharedState() {
 #ifndef NDEBUG
@@ -430,9 +337,7 @@ class SharedState {
     return bg_thread_finished_ == num_bg_threads_;
   }
 
-  bool ShouldVerifyAtBeginning() const {
-    return !FLAGS_expected_values_dir.empty();
-  }
+  bool ShouldVerifyAtBeginning() const;
 
   bool PrintingVerificationResults() {
     bool tmp = false;

db_stress_tool/db_stress_stat.h

@@ -15,6 +15,7 @@
 #include "rocksdb/statistics.h"
 #include "rocksdb/system_clock.h"
 #include "util/gflags_compat.h"
+#include "util/mutexlock.h"
 #include "util/random.h"
 
 DECLARE_bool(histogram);
@@ -181,6 +182,9 @@ class Stats {
     double rate = bytes_mb / elapsed;
     double throughput = (double)done_ / elapsed;
 
+    // Lock stdout so multi-DB reports don't interleave.
+    static port::Mutex report_mutex;
+    MutexLock l(&report_mutex);
     fprintf(stdout, "%-12s: ", name);
     fprintf(stdout, "%.3f micros/op %ld ops/sec\n", seconds_ * 1e6 / done_,
             (long)throughput);

db_stress_tool/db_stress_test_base.h

@@ -13,6 +13,7 @@
 
 #include "db_stress_tool/db_stress_common.h"
 #include "db_stress_tool/db_stress_shared_state.h"
+#include "env/composite_env_wrapper.h"
 #include "rocksdb/experimental.h"
 #include "rocksdb/user_defined_index.h"
 #include "utilities/fault_injection_fs.h"
@@ -22,6 +23,7 @@ class SystemClock;
 class Transaction;
 class TransactionDB;
 class OptimisticTransactionDB;
+class DbStressFSWrapper;
 struct TransactionDBOptions;
 using experimental::SstQueryFilterConfigsManager;
 
@@ -40,11 +42,26 @@ class StressTest {
   // from optimistic transactions when conflict detection retries are exhausted.
   static bool IsExpectedTxnError(const Status& s);
 
-  StressTest();
+  StressTest(int db_index, const std::string& db_path,
+             const std::string& ev_path, const std::string& sec_path);
 
   virtual ~StressTest() {}
 
-  std::shared_ptr<Cache> NewCache(size_t capacity, int32_t num_shard_bits);
+  const std::string& GetDbLabel() const;
+  const std::string& GetDbPath() const;
+  const std::string& GetExpectedValuesDir() const;
+  const std::string& GetSecondariesBase() const;
+
+  // See db_fault_injection_fs_ member.
+  std::shared_ptr<FaultInjectionTestFS> GetDbFaultInjectionFs() const {
+    return db_fault_injection_fs_;
+  }
+
+  // See db_env_ member.
+  Env* GetDbEnv() const;
+
+  static std::shared_ptr<Cache> NewCache(size_t capacity,
+                                         int32_t num_shard_bits);
 
   static std::vector<std::string> GetBlobCompressionTags();
 
@@ -71,6 +88,11 @@ class StressTest {
   Options GetOptions(int cf_id);
   void CleanUp();
 
+ private:
+  void InitializeListenersForOpen(
+      SharedState* shared,
+      const std::vector<ColumnFamilyDescriptor>& cf_descriptors);
+
  protected:
   static int GetMinInjectedErrorCount(int error_count_1, int error_count_2) {
     if (error_count_1 > 0 && error_count_2 > 0) {
@@ -84,39 +106,34 @@ class StressTest {
     }
   }
 
-  void UpdateIfInitialWriteFails(Env* db_stress_env, const Status& write_s,
+  void UpdateIfInitialWriteFails(Env* env, const Status& write_s,
                                  Status* initial_write_s,
                                  bool* initial_wal_write_may_succeed,
                                  uint64_t* wait_for_recover_start_time,
                                  bool commit_bypass_memtable = false) {
-    assert(db_stress_env && initial_write_s && initial_wal_write_may_succeed &&
+    assert(env && initial_write_s && initial_wal_write_may_succeed &&
            wait_for_recover_start_time);
-    // Only update `initial_write_s`, `initial_wal_write_may_succeed` when the
-    // first write fails
     if (!write_s.ok() && (*initial_write_s).ok()) {
       *initial_write_s = write_s;
-      // With commit_bypass_memtable, we create a new WAL after WAL write
-      // succeeds, that wal creation may fail due to injected error. So the
-      // initial wal write may succeed even if status is failed to write to wal
       *initial_wal_write_may_succeed =
           commit_bypass_memtable ||
           !FaultInjectionTestFS::IsFailedToWriteToWALError(*initial_write_s);
-      *wait_for_recover_start_time = db_stress_env->NowMicros();
+      *wait_for_recover_start_time = env->NowMicros();
     }
   }
 
-  void PrintWriteRecoveryWaitTimeIfNeeded(Env* db_stress_env,
+  void PrintWriteRecoveryWaitTimeIfNeeded(Env* env,
                                           const Status& initial_write_s,
                                           bool initial_wal_write_may_succeed,
                                           uint64_t wait_for_recover_start_time,
                                           const std::string& thread_name) {
-    assert(db_stress_env);
+    assert(env);
     bool waited_for_recovery = !initial_write_s.ok() &&
                                IsErrorInjectedAndRetryable(initial_write_s) &&
                                initial_wal_write_may_succeed;
     if (waited_for_recovery) {
       uint64_t elapsed_sec =
-          (db_stress_env->NowMicros() - wait_for_recover_start_time) / 1000000;
+          (env->NowMicros() - wait_for_recover_start_time) / 1000000;
       if (elapsed_sec > 10) {
         fprintf(stdout,
                 "%s thread slept to wait for write recovery for "
@@ -303,6 +320,35 @@ class StressTest {
     kLastOpSeekToLast
   };
 
+  // Enum used to track MANIFEST verification mode during DB reopen
+  enum ManifestVerifyMode {
+    MANIFEST_VERIFY_NONE,
+    // MANIFEST file should be reused (same file number), CURRENT should not
+    // change. Used when reuse_manifest_on_open=1. Warnings are logged on
+    // failure but test continues.
+    MANIFEST_VERIFY_REUSE,
+    // MANIFEST file should be reused AND no recovery writes should occur.
+    // Used when both reuse_manifest_on_open=1 and
+    // optimize_manifest_for_recovery=1. Warnings are logged on failure but test
+    // continues.
+    MANIFEST_VERIFY_NO_WRITE,
+    // Strict verification: MANIFEST file must be reused with ZERO writes
+    // (complete avoidance). Used when ALL conditions for complete avoidance
+    // are met. Failures are FATAL and will terminate the test.
+    // Conditions for STRICT mode:
+    // - Both reuse_manifest_on_open=1 and optimize_manifest_for_recovery=1
+    // - Not in best_efforts_recovery mode
+    // - avoid_flush_during_recovery=true (no flush during recovery)
+    // - write_dbid_to_manifest=0 (no DB_ID write on open)
+    // - metadata_write_fault_one_in=0 (no fault injection)
+    // - open_metadata_write_fault_one_in=0 (no fault injection)
+    // - MANIFEST not corrupted, not at size limit
+    // Note: avoid_flush_during_shutdown is NOT required. If it leaves data
+    // in WAL but avoid_flush_during_recovery=true prevents flushing it,
+    // MANIFEST still won't be written.
+    MANIFEST_VERIFY_STRICT
+  };
+
   // Compare the two iterator, iter and cmp_iter are in the same position,
   // unless iter might be made invalidate or undefined because of
   // upper or lower bounds, or prefix extractor.
@@ -413,7 +459,20 @@ class StressTest {
 
   void CleanUpColumnFamilies();
 
-  std::shared_ptr<Cache> cache_;
+  void RecordManifestStateBeforeReopen();
+  void VerifyManifestNotRewritten();
+
+  int db_index_;
+  std::string db_label_;
+  std::string db_path_;
+  std::string expected_values_path_;
+  std::string secondaries_path_;
+  // Wraps raw_env->GetFileSystem(). See DbStressFSWrapper.
+  std::shared_ptr<DbStressFSWrapper> db_stress_fs_;
+  // Wraps db_stress_fs_ when NeedsFaultInjection(). Null otherwise.
+  std::shared_ptr<FaultInjectionTestFS> db_fault_injection_fs_;
+  // Wraps the outermost FS above. Set as options_.env for all DB I/O.
+  std::unique_ptr<CompositeEnvWrapper> db_env_;
   std::shared_ptr<Cache> compressed_cache_;
   std::shared_ptr<const FilterPolicy> filter_policy_;
   std::unique_ptr<DB> db_owner_;
@@ -439,6 +498,12 @@ class StressTest {
   std::unique_ptr<DB> secondary_db_;
   std::vector<ColumnFamilyHandle*> secondary_cfhs_;
   bool is_db_stopped_;
+
+  // MANIFEST verification state for reopen
+  ManifestVerifyMode manifest_verify_mode_;
+  uint64_t manifest_file_number_before_reopen_;
+  uint64_t manifest_file_size_before_reopen_;
+  std::string current_file_content_before_reopen_;
 };
 
 // Load options from OPTIONS file and populate `options`.

db_stress_tool/db_stress_tool.cc

@@ -22,6 +22,7 @@
 
 #ifdef GFLAGS
 #include <iostream>
+#include <thread>
 
 #include "db_stress_tool/db_stress_common.h"
 #include "db_stress_tool/db_stress_driver.h"
@@ -33,11 +34,79 @@
 namespace ROCKSDB_NAMESPACE {
 namespace {
 static std::shared_ptr<ROCKSDB_NAMESPACE::Env> env_guard;
-static std::shared_ptr<ROCKSDB_NAMESPACE::Env> env_wrapper_guard;
 static std::shared_ptr<ROCKSDB_NAMESPACE::Env> legacy_env_wrapper_guard;
-static std::shared_ptr<ROCKSDB_NAMESPACE::CompositeEnvWrapper>
-    dbsl_env_wrapper_guard;
-static std::shared_ptr<CompositeEnvWrapper> fault_env_guard;
+// Raw pointers for signal-safe crash callback. Signal handlers can only
+// access file-static/global variables; can't capture StressTest instances.
+static std::vector<ROCKSDB_NAMESPACE::FaultInjectionTestFS*>
+    fault_fs_for_crash_report;
+
+int ValidateNumDbsFlags() {
+  if (FLAGS_num_dbs < 1) {
+    fprintf(stderr, "Error: --num_dbs must be >= 1\n");
+    return 1;
+  }
+  if (FLAGS_num_dbs > 1) {
+    if (FLAGS_clear_column_family_one_in > 0) {
+      fprintf(stderr,
+              "Error: --num_dbs > 1 incompatible with "
+              "--clear_column_family_one_in\n");
+      return 1;
+    }
+    if (FLAGS_test_multi_ops_txns) {
+      fprintf(stderr,
+              "Error: --num_dbs > 1 incompatible with "
+              "--test_multi_ops_txns\n");
+      return 1;
+    }
+  }
+  return 0;
+}
+
+int DestroyAllDbs() {
+  bool all_ok = true;
+  const int num_dbs = FLAGS_num_dbs;
+  auto destroy_one = [&](const std::string& db_path) {
+    Status s = DbStressDestroyDb(db_path);
+    if (s.ok()) {
+      fprintf(stdout, "Successfully destroyed db at %s\n", db_path.c_str());
+    } else {
+      fprintf(stderr, "Failed to destroy db at %s: %s\n", db_path.c_str(),
+              s.ToString().c_str());
+      all_ok = false;
+    }
+  };
+  if (num_dbs == 1) {
+    destroy_one(FLAGS_db);
+  } else {
+    for (int i = 0; i < num_dbs; i++) {
+      destroy_one(FLAGS_db + "/db_" + std::to_string(i));
+    }
+    DestroyDir(raw_env, FLAGS_db);
+  }
+  return all_ok ? 0 : 1;
+}
+
+void RegisterCrashCallbacks(
+    const std::vector<std::unique_ptr<StressTest>>& stress_tests, int num_dbs) {
+  fault_fs_for_crash_report.resize(num_dbs, nullptr);
+  bool any_fault_fs = false;
+  for (int i = 0; i < num_dbs; i++) {
+    fault_fs_for_crash_report[i] =
+        stress_tests[i]->GetDbFaultInjectionFs().get();
+    if (fault_fs_for_crash_report[i]) {
+      any_fault_fs = true;
+    }
+  }
+  if (any_fault_fs) {
+    port::RegisterCrashCallback([]() {
+      for (auto* fs : fault_fs_for_crash_report) {
+        if (fs) {
+          fs->PrintRecentInjectedErrors();
+        }
+      }
+    });
+  }
+}
 
 int ReturnFlagValidationError(const char* message) {
   std::cerr << "Error: " << message << '\n';
@@ -67,8 +136,6 @@ int db_stress_tool(int argc, char** argv) {
       StringToCompressionType(FLAGS_bottommost_compression_type.c_str());
   checksum_type_e = StringToChecksumType(FLAGS_checksum_type.c_str());
 
-  Env* raw_env;
-
   int env_opts = !FLAGS_env_uri.empty() + !FLAGS_fs_uri.empty();
   if (env_opts > 1) {
     fprintf(stderr, "Error: --env_uri and --fs_uri are mutually exclusive\n");
@@ -82,57 +149,10 @@ int db_stress_tool(int argc, char** argv) {
             s.ToString().c_str());
     exit(1);
   }
-  dbsl_env_wrapper_guard = std::make_shared<CompositeEnvWrapper>(raw_env);
-  db_stress_listener_env = dbsl_env_wrapper_guard.get();
 
-  if (FLAGS_open_metadata_read_fault_one_in ||
-      FLAGS_open_metadata_write_fault_one_in || FLAGS_open_read_fault_one_in ||
-      FLAGS_open_write_fault_one_in || FLAGS_metadata_read_fault_one_in ||
-      FLAGS_metadata_write_fault_one_in || FLAGS_read_fault_one_in ||
-      FLAGS_write_fault_one_in || FLAGS_sync_fault_injection) {
-    FaultInjectionTestFS* fs =
-        new FaultInjectionTestFS(raw_env->GetFileSystem());
-    fault_fs_guard.reset(fs);
-    // Info logs are debugging artifacts, so exclude them from fault injection
-    // and keep error accounting focused on DB data and metadata.
-    fault_fs_guard->SetFileTypesExcludedFromFaultInjection(
-        {FileType::kInfoLogFile});
-    // Set it to direct writable here to initially bypass any fault injection
-    // during DB open This will correspondingly be overwritten in
-    // StressTest::Open() for open fault injection and in RunStressTestImpl()
-    // for proper fault injection setup.
-    fault_fs_guard->SetFilesystemDirectWritable(true);
-    fault_env_guard =
-        std::make_shared<CompositeEnvWrapper>(raw_env, fault_fs_guard);
-    raw_env = fault_env_guard.get();
-
-    // Register a crash callback so that recently injected errors are
-    // printed to stderr when the process crashes (SIGABRT, SIGSEGV, etc.).
-    // This helps diagnose stress test failures caused by fault injection.
-    port::RegisterCrashCallback([]() {
-      if (fault_fs_guard) {
-        fault_fs_guard->PrintRecentInjectedErrors();
-      }
-    });
-  }
-
-  auto db_stress_fs =
-      std::make_shared<DbStressFSWrapper>(raw_env->GetFileSystem());
-  env_wrapper_guard =
-      std::make_shared<CompositeEnvWrapper>(raw_env, db_stress_fs);
-  db_stress_env = env_wrapper_guard.get();
-
-  // Handle --destroy_db_and_exit early, before other option validation
+  // Handle --destroy_db_and_exit early
   if (FLAGS_destroy_db_and_exit) {
-    s = DbStressDestroyDb(FLAGS_db);
-    if (s.ok()) {
-      fprintf(stdout, "Successfully destroyed db at %s\n", FLAGS_db.c_str());
-      return 0;
-    } else {
-      fprintf(stderr, "Failed to destroy db at %s: %s\n", FLAGS_db.c_str(),
-              s.ToString().c_str());
-      return 1;
-    }
+    return DestroyAllDbs();
   }
 
   // Handle --delete_dir_and_exit early, before other option validation
@@ -149,14 +169,21 @@ int db_stress_tool(int argc, char** argv) {
     }
   }
 
+  {
+    int rc = ValidateNumDbsFlags();
+    if (rc != 0) {
+      return rc;
+    }
+  }
+
   FLAGS_rep_factory = StringToRepFactory(FLAGS_memtablerep.c_str());
 
   // The number of background threads should be at least as much the
   // max number of concurrent compactions.
-  db_stress_env->SetBackgroundThreads(FLAGS_max_background_compactions,
-                                      ROCKSDB_NAMESPACE::Env::Priority::LOW);
-  db_stress_env->SetBackgroundThreads(FLAGS_num_bottom_pri_threads,
-                                      ROCKSDB_NAMESPACE::Env::Priority::BOTTOM);
+  raw_env->SetBackgroundThreads(FLAGS_max_background_compactions,
+                                ROCKSDB_NAMESPACE::Env::Priority::LOW);
+  raw_env->SetBackgroundThreads(FLAGS_num_bottom_pri_threads,
+                                ROCKSDB_NAMESPACE::Env::Priority::BOTTOM);
   if (FLAGS_prefixpercent > 0 && FLAGS_prefix_size < 0) {
     fprintf(stderr,
             "Error: prefixpercent is non-zero while prefix_size is "
@@ -356,42 +383,75 @@ int db_stress_tool(int argc, char** argv) {
   // Choose a location for the test database if none given with --db=<path>
   if (FLAGS_db.empty()) {
     std::string default_db_path;
-    db_stress_env->GetTestDirectory(&default_db_path);
+    raw_env->GetTestDirectory(&default_db_path);
     default_db_path += "/dbstress";
     FLAGS_db = default_db_path;
   }
 
-  // Now that FLAGS_db is resolved, set the fault injection log file path
-  // so that PrintAll() writes to a file instead of stderr (signal-safe).
-  // Store the log in TEST_TMPDIR (outside the DB directory) so it survives
-  // DB reopen (which cleans untracked files) and gets included in the
-  // sandcastle db.tar.gz artifact for post-failure analysis.
-  if (fault_fs_guard) {
-    std::string log_dir;
-    const char* test_tmpdir = getenv("TEST_TMPDIR");
-    if (test_tmpdir && test_tmpdir[0] != '\0') {
-      log_dir = test_tmpdir;
-    } else {
-      log_dir = "/tmp";
-    }
-    std::string log_path = log_dir + "/fault_injection_" +
-                           std::to_string(getpid()) + "_" +
-                           std::to_string(time(nullptr)) + ".log";
-    fault_fs_guard->SetInjectedErrorLogPath(log_path);
-  }
-
-  if ((FLAGS_test_secondary || FLAGS_continuous_verification_interval > 0) &&
-      FLAGS_secondaries_base.empty()) {
-    std::string default_secondaries_path;
-    db_stress_env->GetTestDirectory(&default_secondaries_path);
-    default_secondaries_path += "/dbstress_secondaries";
-    s = db_stress_env->CreateDirIfMissing(default_secondaries_path);
+  // For num_dbs=1: --db and --expected_values_dir are paths used as-is.
+  // For num_dbs>1: they are parent directories; C++ creates db_0/, db_1/, ...
+  // subdirs underneath. DB::Open(create_if_missing) creates each DB dir.
+  // Python (db_crashtest.py) also creates EV dirs; C++ creates them here as a
+  // fallback for direct CLI usage.
+  const int num_dbs = FLAGS_num_dbs;
+  if (num_dbs > 1) {
+    s = raw_env->CreateDirIfMissing(FLAGS_db);
     if (!s.ok()) {
-      fprintf(stderr, "Failed to create directory %s: %s\n",
-              default_secondaries_path.c_str(), s.ToString().c_str());
+      fprintf(stderr, "Failed to create directory %s: %s\n", FLAGS_db.c_str(),
+              s.ToString().c_str());
       exit(1);
     }
-    FLAGS_secondaries_base = default_secondaries_path;
+  }
+  std::vector<std::string> db_paths;
+  std::vector<std::string> ev_paths;
+  for (int i = 0; i < num_dbs; i++) {
+    std::string suffix = (num_dbs == 1) ? "" : "/db_" + std::to_string(i);
+    db_paths.push_back(FLAGS_db + suffix);
+    if (!FLAGS_expected_values_dir.empty()) {
+      std::string ep = FLAGS_expected_values_dir + suffix;
+      s = Env::Default()->CreateDirIfMissing(ep);
+      if (!s.ok()) {
+        fprintf(stderr, "Failed to create directory %s: %s\n", ep.c_str(),
+                s.ToString().c_str());
+        exit(1);
+      }
+      ev_paths.push_back(std::move(ep));
+    }
+  }
+  if (ev_paths.empty()) {
+    ev_paths.resize(num_dbs);
+  }
+
+  // Secondary paths: C++ owns creation entirely.
+  std::vector<std::string> sec_paths;
+  if (FLAGS_test_secondary || FLAGS_continuous_verification_interval > 0) {
+    std::string sec_parent;
+    if (!FLAGS_secondaries_base.empty()) {
+      sec_parent = FLAGS_secondaries_base;
+    } else {
+      raw_env->GetTestDirectory(&sec_parent);
+      sec_parent += "/dbstress_secondaries";
+    }
+    s = raw_env->CreateDirIfMissing(sec_parent);
+    if (!s.ok()) {
+      fprintf(stderr, "Failed to create directory %s: %s\n", sec_parent.c_str(),
+              s.ToString().c_str());
+      exit(1);
+    }
+    for (int i = 0; i < num_dbs; i++) {
+      std::string suffix = (num_dbs == 1) ? "" : "/db_" + std::to_string(i);
+      std::string sec_path = sec_parent + suffix;
+      s = raw_env->CreateDirIfMissing(sec_path);
+      if (!s.ok()) {
+        fprintf(stderr, "Failed to create directory %s: %s\n", sec_path.c_str(),
+                s.ToString().c_str());
+        exit(1);
+      }
+      sec_paths.push_back(std::move(sec_path));
+    }
+  }
+  if (sec_paths.empty()) {
+    sec_paths.resize(num_dbs);
   }
 
   if (FLAGS_best_efforts_recovery &&
@@ -511,25 +571,72 @@ int db_stress_tool(int argc, char** argv) {
     key_gen_ctx.weights.emplace_back(key_gen_ctx.window -
                                      keys_per_level * (levels - 1));
   }
-  std::unique_ptr<ROCKSDB_NAMESPACE::SharedState> shared;
-  std::unique_ptr<ROCKSDB_NAMESPACE::StressTest> stress;
-  if (FLAGS_test_cf_consistency) {
-    stress.reset(CreateCfConsistencyStressTest());
-  } else if (FLAGS_test_batches_snapshots) {
-    stress.reset(CreateBatchedOpsStressTest());
-  } else if (FLAGS_test_multi_ops_txns) {
-    stress.reset(CreateMultiOpsTxnsStressTest());
-  } else {
-    stress.reset(CreateNonBatchedOpsStressTest());
-  }
-  // Initialize the Zipfian pre-calculated array
+  // Initialize shared resources (hot-key generator, block cache, WBM,
+  // rate limiter) once so that all DB instances share them.
   InitializeHotKeyGenerator(FLAGS_hot_key_alpha);
-  shared.reset(new SharedState(db_stress_env, stress.get()));
-  bool run_stress_test = RunStressTest(shared.get());
-  // Close DB in CleanUp() before destructor to prevent race between destructor
-  // and operations in listener callbacks (e.g. MultiOpsTxnsStressListener).
-  stress->CleanUp();
-  return run_stress_test ? 0 : 1;
+  block_cache =
+      StressTest::NewCache(FLAGS_cache_size, FLAGS_cache_numshardbits);
+  if (FLAGS_use_write_buffer_manager) {
+    wbm = std::make_shared<WriteBufferManager>(FLAGS_db_write_buffer_size,
+                                               block_cache);
+  }
+  if (FLAGS_rate_limiter_bytes_per_sec > 0) {
+    rate_limiter.reset(NewGenericRateLimiter(
+        FLAGS_rate_limiter_bytes_per_sec, 1000 /* refill_period_us */,
+        10 /* fairness */,
+        FLAGS_rate_limit_bg_reads ? RateLimiter::Mode::kReadsOnly
+                                  : RateLimiter::Mode::kWritesOnly));
+  }
+
+  // Phase 1: Create StressTest instances (one per DB).
+  std::vector<std::unique_ptr<StressTest>> stress_tests(num_dbs);
+  for (int i = 0; i < num_dbs; i++) {
+    if (FLAGS_test_cf_consistency) {
+      stress_tests[i].reset(CreateCfConsistencyStressTest(
+          i, db_paths[i], ev_paths[i], sec_paths[i]));
+    } else if (FLAGS_test_batches_snapshots) {
+      stress_tests[i].reset(CreateBatchedOpsStressTest(
+          i, db_paths[i], ev_paths[i], sec_paths[i]));
+    } else if (FLAGS_test_multi_ops_txns) {
+      stress_tests[i].reset(CreateMultiOpsTxnsStressTest(
+          i, db_paths[i], ev_paths[i], sec_paths[i]));
+    } else {
+      stress_tests[i].reset(CreateNonBatchedOpsStressTest(
+          i, db_paths[i], ev_paths[i], sec_paths[i]));
+    }
+  }
+
+  RegisterCrashCallbacks(stress_tests, num_dbs);
+
+  // Phase 2: Create SharedState for every DB before any worker thread starts.
+  std::vector<std::unique_ptr<SharedState>> shared_states(num_dbs);
+  for (int i = 0; i < num_dbs; i++) {
+    shared_states[i].reset(new SharedState(raw_env, stress_tests[i].get()));
+  }
+
+  // Phase 3: Launch each DB's stress test on its own thread.
+  std::vector<int> results(num_dbs, 0);
+  std::vector<std::thread> stress_test_runners;
+  stress_test_runners.reserve(num_dbs);
+  for (int i = 0; i < num_dbs; i++) {
+    stress_test_runners.emplace_back([i, &results, &shared_states]() {
+      results[i] = RunStressTest(shared_states[i].get()) ? 1 : 0;
+    });
+  }
+
+  // Phase 4: Wait for all DB threads, collect results, clean up.
+  bool all_passed = true;
+  for (int i = 0; i < num_dbs; i++) {
+    stress_test_runners[i].join();
+    if (!results[i]) {
+      all_passed = false;
+    }
+    stress_tests[i]->CleanUp();
+  }
+  for (auto& fs : fault_fs_for_crash_report) {
+    fs = nullptr;
+  }
+  return all_passed ? 0 : 1;
 }
 
 }  // namespace ROCKSDB_NAMESPACE

db_stress_tool/multi_ops_txns_stress.cc

@@ -40,7 +40,7 @@ DEFINE_int32(clear_wp_commit_cache_one_in, 0,
              "write-unprepared transactions.");
 
 extern "C" bool rocksdb_write_prepared_TEST_ShouldClearCommitCache(void) {
-  static Random rand(static_cast<uint32_t>(db_stress_env->NowMicros()));
+  static Random rand(static_cast<uint32_t>(raw_env->NowMicros()));
   return FLAGS_clear_wp_commit_cache_one_in > 0 &&
          rand.OneIn(FLAGS_clear_wp_commit_cache_one_in);
 }
@@ -1413,7 +1413,7 @@ Status MultiOpsTxnsStressTest::CommitAndCreateTimestampedSnapshotIfNeeded(
   Status s;
   if (FLAGS_create_timestamped_snapshot_one_in > 0 &&
       thread->rand.OneInOpt(FLAGS_create_timestamped_snapshot_one_in)) {
-    uint64_t ts = db_stress_env->NowNanos();
+    uint64_t ts = raw_env->NowNanos();
     std::shared_ptr<const Snapshot> snapshot;
     s = txn.CommitAndTryCreateSnapshot(/*notifier=*/nullptr, ts, &snapshot);
   } else {
@@ -1428,7 +1428,7 @@ Status MultiOpsTxnsStressTest::CommitAndCreateTimestampedSnapshotIfNeeded(
   assert(txn_db_);
   if (FLAGS_create_timestamped_snapshot_one_in > 0 &&
       thread->rand.OneInOpt(50000)) {
-    uint64_t now = db_stress_env->NowNanos();
+    uint64_t now = raw_env->NowNanos();
     constexpr uint64_t time_diff = static_cast<uint64_t>(1000) * 1000 * 1000;
     txn_db_->ReleaseTimestampedSnapshotsOlderThan(now - time_diff);
   }
@@ -1767,8 +1767,11 @@ void MultiOpsTxnsStressTest::ScanExistingDb(SharedState* shared, int threads) {
   }
 }
 
-StressTest* CreateMultiOpsTxnsStressTest() {
-  return new MultiOpsTxnsStressTest();
+StressTest* CreateMultiOpsTxnsStressTest(int db_index,
+                                         const std::string& db_path,
+                                         const std::string& ev_path,
+                                         const std::string& sec_path) {
+  return new MultiOpsTxnsStressTest(db_index, db_path, ev_path, sec_path);
 }
 
 void CheckAndSetOptionsForMultiOpsTxnStressTest() {

db_stress_tool/multi_ops_txns_stress.h

@@ -9,6 +9,7 @@
 
 #ifdef GFLAGS
 #include "db_stress_tool/db_stress_common.h"
+#include "util/atomic.h"
 
 namespace ROCKSDB_NAMESPACE {
 
@@ -191,7 +192,10 @@ class MultiOpsTxnsStressTest : public StressTest {
     uint32_t c_{0};
   };
 
-  MultiOpsTxnsStressTest() {}
+  MultiOpsTxnsStressTest(int db_index, const std::string& db_path,
+                         const std::string& ev_path,
+                         const std::string& sec_path)
+      : StressTest(db_index, db_path, ev_path, sec_path) {}
 
   ~MultiOpsTxnsStressTest() override {}
 
@@ -423,11 +427,16 @@ class MultiOpsTxnsStressListener : public EventListener {
 
   ~MultiOpsTxnsStressListener() override {}
 
+  void OnDBShutdownBegin(DB* /*db*/) override { shutting_down_.Store(true); }
+
   void OnFlushCompleted(DB* db, const FlushJobInfo& info) override {
     assert(db);
 #ifdef NDEBUG
     (void)db;
 #endif
+    if (shutting_down_.Load()) {
+      return;
+    }
     assert(info.cf_id == 0);
     const ReadOptions read_options(Env::IOActivity::kFlush);
     stress_test_->VerifyPkSkFast(read_options, info.job_id);
@@ -438,6 +447,9 @@ class MultiOpsTxnsStressListener : public EventListener {
 #ifdef NDEBUG
     (void)db;
 #endif
+    if (shutting_down_.Load()) {
+      return;
+    }
     assert(info.cf_id == 0);
     const ReadOptions read_options(Env::IOActivity::kCompaction);
     stress_test_->VerifyPkSkFast(read_options, info.job_id);
@@ -445,6 +457,7 @@ class MultiOpsTxnsStressListener : public EventListener {
 
  private:
   MultiOpsTxnsStressTest* const stress_test_ = nullptr;
+  Atomic<bool> shutting_down_{false};
 };
 
 }  // namespace ROCKSDB_NAMESPACE

db_stress_tool/no_batched_ops_stress.cc

@@ -14,6 +14,7 @@
 #include "rocksdb/status.h"
 #ifdef GFLAGS
 #include <cinttypes>
+#include <deque>
 #include <unordered_map>
 
 #include "db/wide/wide_columns_helper.h"
@@ -24,7 +25,10 @@
 namespace ROCKSDB_NAMESPACE {
 class NonBatchedOpsStressTest : public StressTest {
  public:
-  NonBatchedOpsStressTest() = default;
+  NonBatchedOpsStressTest(int db_index, const std::string& db_path,
+                          const std::string& ev_path,
+                          const std::string& sec_path)
+      : StressTest(db_index, db_path, ev_path, sec_path) {}
 
   virtual ~NonBatchedOpsStressTest() = default;
 
@@ -265,20 +269,20 @@ class NonBatchedOpsStressTest : public StressTest {
             std::string from_db;
 
             // Temporarily disable error injection to verify the secondary
-            if (fault_fs_guard) {
-              fault_fs_guard->DisableThreadLocalErrorInjection(
+            if (db_fault_injection_fs_) {
+              db_fault_injection_fs_->DisableThreadLocalErrorInjection(
                   FaultInjectionIOType::kRead);
-              fault_fs_guard->DisableThreadLocalErrorInjection(
+              db_fault_injection_fs_->DisableThreadLocalErrorInjection(
                   FaultInjectionIOType::kMetadataRead);
             }
 
             s = secondary_db_->Get(options, secondary_cfhs_[cf], key, &from_db);
 
             // Re-enable error injection after verifying the secondary
-            if (fault_fs_guard) {
-              fault_fs_guard->EnableThreadLocalErrorInjection(
+            if (db_fault_injection_fs_) {
+              db_fault_injection_fs_->EnableThreadLocalErrorInjection(
                   FaultInjectionIOType::kRead);
-              fault_fs_guard->EnableThreadLocalErrorInjection(
+              db_fault_injection_fs_->EnableThreadLocalErrorInjection(
                   FaultInjectionIOType::kMetadataRead);
             }
 
@@ -688,10 +692,10 @@ class NonBatchedOpsStressTest : public StressTest {
     bool read_older_ts = MaybeUseOlderTimestampForPointLookup(
         thread, read_ts_str, read_ts_slice, read_opts_copy);
 
-    if (fault_fs_guard) {
-      fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+    if (db_fault_injection_fs_) {
+      db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
           FaultInjectionIOType::kRead);
-      fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+      db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
           FaultInjectionIOType::kMetadataRead);
       SharedState::ignore_read_error = false;
     }
@@ -703,11 +707,11 @@ class NonBatchedOpsStressTest : public StressTest {
         thread->shared->Get(rand_column_families[0], rand_keys[0]);
 
     int injected_error_count = 0;
-    if (fault_fs_guard) {
+    if (db_fault_injection_fs_) {
       injected_error_count = GetMinInjectedErrorCount(
-          fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+          db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
               FaultInjectionIOType::kRead),
-          fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+          db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
               FaultInjectionIOType::kMetadataRead));
       if (!SharedState::ignore_read_error && injected_error_count > 0 &&
           (s.ok() || s.IsNotFound())) {
@@ -716,9 +720,9 @@ class NonBatchedOpsStressTest : public StressTest {
         MutexLock l(thread->shared->GetMutex());
         fprintf(stderr, "Didn't get expected error from Get\n");
         fprintf(stderr, "Callstack that injected the fault\n");
-        fault_fs_guard->PrintInjectedThreadLocalErrorBacktrace(
+        db_fault_injection_fs_->PrintInjectedThreadLocalErrorBacktrace(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->PrintInjectedThreadLocalErrorBacktrace(
+        db_fault_injection_fs_->PrintInjectedThreadLocalErrorBacktrace(
             FaultInjectionIOType::kMetadataRead);
         std::terminate();
       }
@@ -821,10 +825,10 @@ class NonBatchedOpsStressTest : public StressTest {
     std::unique_ptr<Transaction> txn;
     if (use_txn) {
       // TODO(hx235): test fault injection with MultiGet() with transactions
-      if (fault_fs_guard) {
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kMetadataRead);
       }
       WriteOptions wo;
@@ -850,20 +854,20 @@ class NonBatchedOpsStressTest : public StressTest {
     int injected_error_count = 0;
 
     if (!use_txn) {
-      if (fault_fs_guard) {
-        fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+        db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
             FaultInjectionIOType::kMetadataRead);
         SharedState::ignore_read_error = false;
       }
       db_->MultiGet(readoptionscopy, cfh, num_keys, keys.data(), values.data(),
                     statuses.data());
-      if (fault_fs_guard) {
+      if (db_fault_injection_fs_) {
         injected_error_count = GetMinInjectedErrorCount(
-            fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+            db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
                 FaultInjectionIOType::kRead),
-            fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+            db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
                 FaultInjectionIOType::kMetadataRead));
 
         if (injected_error_count > 0) {
@@ -883,9 +887,9 @@ class NonBatchedOpsStressTest : public StressTest {
                     "num_keys %zu Expected %d errors, seen at least %d\n",
                     num_keys, injected_error_count, stat_nok_nfound);
             fprintf(stderr, "Callstack that injected the fault\n");
-            fault_fs_guard->PrintInjectedThreadLocalErrorBacktrace(
+            db_fault_injection_fs_->PrintInjectedThreadLocalErrorBacktrace(
                 FaultInjectionIOType::kRead);
-            fault_fs_guard->PrintInjectedThreadLocalErrorBacktrace(
+            db_fault_injection_fs_->PrintInjectedThreadLocalErrorBacktrace(
                 FaultInjectionIOType::kMetadataRead);
             std::terminate();
           }
@@ -952,10 +956,10 @@ class NonBatchedOpsStressTest : public StressTest {
             const Status& s,
             const std::optional<ExpectedValue>& ryw_expected_value) -> bool {
       //  Temporarily disable error injection for verification
-      if (fault_fs_guard) {
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kMetadataRead);
       }
 
@@ -1049,10 +1053,10 @@ class NonBatchedOpsStressTest : public StressTest {
       }
 
       // Enable back error injection disbled for checking results
-      if (fault_fs_guard) {
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kMetadataRead);
       }
       return check_multiget_res;
@@ -1091,10 +1095,10 @@ class NonBatchedOpsStressTest : public StressTest {
     if (use_txn) {
       txn->Rollback().PermitUncheckedError();
       // Enable back error injection disbled for transactions
-      if (fault_fs_guard) {
-        fault_fs_guard->EnableThreadLocalErrorInjection(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->EnableThreadLocalErrorInjection(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->EnableThreadLocalErrorInjection(
+        db_fault_injection_fs_->EnableThreadLocalErrorInjection(
             FaultInjectionIOType::kMetadataRead);
       }
     }
@@ -1141,10 +1145,10 @@ class NonBatchedOpsStressTest : public StressTest {
     const ExpectedValue pre_read_expected_value =
         thread->shared->Get(column_family, key);
 
-    if (fault_fs_guard) {
-      fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+    if (db_fault_injection_fs_) {
+      db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
           FaultInjectionIOType::kRead);
-      fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+      db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
           FaultInjectionIOType::kMetadataRead);
       SharedState::ignore_read_error = false;
     }
@@ -1164,11 +1168,11 @@ class NonBatchedOpsStressTest : public StressTest {
         thread->shared->Get(column_family, key);
 
     int injected_error_count = 0;
-    if (fault_fs_guard) {
+    if (db_fault_injection_fs_) {
       injected_error_count = GetMinInjectedErrorCount(
-          fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+          db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
               FaultInjectionIOType::kRead),
-          fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+          db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
               FaultInjectionIOType::kMetadataRead));
       if (!SharedState::ignore_read_error && injected_error_count > 0 &&
           (s.ok() || s.IsNotFound())) {
@@ -1177,9 +1181,9 @@ class NonBatchedOpsStressTest : public StressTest {
         MutexLock l(thread->shared->GetMutex());
         fprintf(stderr, "Didn't get expected error from GetEntity\n");
         fprintf(stderr, "Callstack that injected the fault\n");
-        fault_fs_guard->PrintInjectedThreadLocalErrorBacktrace(
+        db_fault_injection_fs_->PrintInjectedThreadLocalErrorBacktrace(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->PrintInjectedThreadLocalErrorBacktrace(
+        db_fault_injection_fs_->PrintInjectedThreadLocalErrorBacktrace(
             FaultInjectionIOType::kMetadataRead);
         std::terminate();
       }
@@ -1281,10 +1285,10 @@ class NonBatchedOpsStressTest : public StressTest {
     if (FLAGS_use_txn) {
       // TODO(hx235): test fault injection with MultiGetEntity() with
       // transactions
-      if (fault_fs_guard) {
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kMetadataRead);
       }
       WriteOptions write_options;
@@ -1318,11 +1322,11 @@ class NonBatchedOpsStressTest : public StressTest {
     int injected_error_count = 0;
 
     auto verify_expected_errors = [&](auto get_status) {
-      assert(fault_fs_guard);
+      assert(db_fault_injection_fs_);
       injected_error_count = GetMinInjectedErrorCount(
-          fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+          db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
               FaultInjectionIOType::kRead),
-          fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+          db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
               FaultInjectionIOType::kMetadataRead));
       if (injected_error_count) {
         int stat_nok_nfound = 0;
@@ -1344,9 +1348,9 @@ class NonBatchedOpsStressTest : public StressTest {
           fprintf(stderr, "num_keys %zu Expected %d errors, seen %d\n",
                   num_keys, injected_error_count, stat_nok_nfound);
           fprintf(stderr, "Call stack that injected the fault\n");
-          fault_fs_guard->PrintInjectedThreadLocalErrorBacktrace(
+          db_fault_injection_fs_->PrintInjectedThreadLocalErrorBacktrace(
               FaultInjectionIOType::kRead);
-          fault_fs_guard->PrintInjectedThreadLocalErrorBacktrace(
+          db_fault_injection_fs_->PrintInjectedThreadLocalErrorBacktrace(
               FaultInjectionIOType::kMetadataRead);
           std::terminate();
         }
@@ -1356,10 +1360,10 @@ class NonBatchedOpsStressTest : public StressTest {
     auto check_results = [&](auto get_columns, auto get_status,
                              auto do_extra_check, auto call_get_entity) {
       // Temporarily disable error injection for checking results
-      if (fault_fs_guard) {
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kMetadataRead);
       }
       const bool check_get_entity =
@@ -1538,10 +1542,10 @@ class NonBatchedOpsStressTest : public StressTest {
         }
       }
       // Enable back error injection disbled for checking results
-      if (fault_fs_guard) {
-        fault_fs_guard->EnableThreadLocalErrorInjection(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->EnableThreadLocalErrorInjection(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->EnableThreadLocalErrorInjection(
+        db_fault_injection_fs_->EnableThreadLocalErrorInjection(
             FaultInjectionIOType::kMetadataRead);
       }
     };
@@ -1624,19 +1628,19 @@ class NonBatchedOpsStressTest : public StressTest {
 
       txn->Rollback().PermitUncheckedError();
       // Enable back error injection disbled for transactions
-      if (fault_fs_guard) {
-        fault_fs_guard->EnableThreadLocalErrorInjection(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->EnableThreadLocalErrorInjection(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->EnableThreadLocalErrorInjection(
+        db_fault_injection_fs_->EnableThreadLocalErrorInjection(
             FaultInjectionIOType::kMetadataRead);
       }
     } else if (FLAGS_use_attribute_group) {
       // AttributeGroup MultiGetEntity verification
 
-      if (fault_fs_guard) {
-        fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+        db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
             FaultInjectionIOType::kMetadataRead);
         SharedState::ignore_read_error = false;
       }
@@ -1652,7 +1656,7 @@ class NonBatchedOpsStressTest : public StressTest {
       db_->MultiGetEntity(read_opts_copy, num_keys, key_slices.data(),
                           results.data());
 
-      if (fault_fs_guard) {
+      if (db_fault_injection_fs_) {
         verify_expected_errors(
             [&](size_t i) { return results[i][0].status(); });
       }
@@ -1668,10 +1672,10 @@ class NonBatchedOpsStressTest : public StressTest {
     } else {
       // Non-AttributeGroup MultiGetEntity verification
 
-      if (fault_fs_guard) {
-        fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+        db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
             FaultInjectionIOType::kMetadataRead);
         SharedState::ignore_read_error = false;
       }
@@ -1682,7 +1686,7 @@ class NonBatchedOpsStressTest : public StressTest {
       db_->MultiGetEntity(read_opts_copy, cfh, num_keys, key_slices.data(),
                           results.data(), statuses.data());
 
-      if (fault_fs_guard) {
+      if (db_fault_injection_fs_) {
         verify_expected_errors([&](size_t i) { return statuses[i]; });
       }
 
@@ -1733,10 +1737,10 @@ class NonBatchedOpsStressTest : public StressTest {
     MaybeUseOlderTimestampForRangeScan(thread, read_ts_str, read_ts_slice,
                                        ro_copy);
 
-    if (fault_fs_guard) {
-      fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+    if (db_fault_injection_fs_) {
+      db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
           FaultInjectionIOType::kRead);
-      fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+      db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
           FaultInjectionIOType::kMetadataRead);
       SharedState::ignore_read_error = false;
     }
@@ -1798,11 +1802,11 @@ class NonBatchedOpsStressTest : public StressTest {
     }
 
     int injected_error_count = 0;
-    if (fault_fs_guard) {
+    if (db_fault_injection_fs_) {
       injected_error_count = GetMinInjectedErrorCount(
-          fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+          db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
               FaultInjectionIOType::kRead),
-          fault_fs_guard->GetAndResetInjectedThreadLocalErrorCount(
+          db_fault_injection_fs_->GetAndResetInjectedThreadLocalErrorCount(
               FaultInjectionIOType::kMetadataRead));
       if (!SharedState::ignore_read_error && injected_error_count > 0 &&
           s.ok()) {
@@ -1811,9 +1815,9 @@ class NonBatchedOpsStressTest : public StressTest {
         MutexLock l(thread->shared->GetMutex());
         fprintf(stderr, "Didn't get expected error from PrefixScan\n");
         fprintf(stderr, "Callstack that injected the fault\n");
-        fault_fs_guard->PrintInjectedThreadLocalErrorBacktrace(
+        db_fault_injection_fs_->PrintInjectedThreadLocalErrorBacktrace(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->PrintInjectedThreadLocalErrorBacktrace(
+        db_fault_injection_fs_->PrintInjectedThreadLocalErrorBacktrace(
             FaultInjectionIOType::kMetadataRead);
         std::terminate();
       }
@@ -1880,10 +1884,10 @@ class NonBatchedOpsStressTest : public StressTest {
 
     if (FLAGS_verify_before_write) {
       // Temporarily disable error injection for preparation
-      if (fault_fs_guard) {
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->DisableThreadLocalErrorInjection(
+        db_fault_injection_fs_->DisableThreadLocalErrorInjection(
             FaultInjectionIOType::kMetadataRead);
       }
 
@@ -1894,10 +1898,10 @@ class NonBatchedOpsStressTest : public StressTest {
           /* msg_prefix */ "Pre-Put Get verification", from_db, s);
 
       // Enable back error injection disabled for preparation
-      if (fault_fs_guard) {
-        fault_fs_guard->EnableThreadLocalErrorInjection(
+      if (db_fault_injection_fs_) {
+        db_fault_injection_fs_->EnableThreadLocalErrorInjection(
             FaultInjectionIOType::kRead);
-        fault_fs_guard->EnableThreadLocalErrorInjection(
+        db_fault_injection_fs_->EnableThreadLocalErrorInjection(
             FaultInjectionIOType::kMetadataRead);
       }
       if (!res) {
@@ -2000,7 +2004,7 @@ class NonBatchedOpsStressTest : public StressTest {
               &commit_bypass_memtable);
         }
       }
-      UpdateIfInitialWriteFails(db_stress_env, s, &initial_write_s,
+      UpdateIfInitialWriteFails(raw_env, s, &initial_write_s,
                                 &initial_wal_write_may_succeed,
                                 &wait_for_recover_start_time);
 
@@ -2031,7 +2035,7 @@ class NonBatchedOpsStressTest : public StressTest {
       }
     } else {
       PrintWriteRecoveryWaitTimeIfNeeded(
-          db_stress_env, initial_write_s, initial_wal_write_may_succeed,
+          raw_env, initial_write_s, initial_wal_write_may_succeed,
           wait_for_recover_start_time, "TestPut");
       pending_expected_value.Commit();
       thread->stats.AddBytesForWrites(1, sz);
@@ -2107,7 +2111,7 @@ class NonBatchedOpsStressTest : public StressTest {
               &commit_bypass_memtable);
         }
         UpdateIfInitialWriteFails(
-            db_stress_env, s, &initial_write_s, &initial_wal_write_may_succeed,
+            raw_env, s, &initial_write_s, &initial_wal_write_may_succeed,
             &wait_for_recover_start_time, commit_bypass_memtable);
       } while (!s.ok() && IsErrorInjectedAndRetryable(s) &&
                initial_wal_write_may_succeed);
@@ -2137,7 +2141,7 @@ class NonBatchedOpsStressTest : public StressTest {
         }
       } else {
         PrintWriteRecoveryWaitTimeIfNeeded(
-            db_stress_env, initial_write_s, initial_wal_write_may_succeed,
+            raw_env, initial_write_s, initial_wal_write_may_succeed,
             wait_for_recover_start_time, "TestDelete");
         pending_expected_value.Commit();
         thread->stats.AddDeletes(1);
@@ -2179,7 +2183,7 @@ class NonBatchedOpsStressTest : public StressTest {
               &commit_bypass_memtable);
         }
         UpdateIfInitialWriteFails(
-            db_stress_env, s, &initial_write_s, &initial_wal_write_may_succeed,
+            raw_env, s, &initial_write_s, &initial_wal_write_may_succeed,
             &wait_for_recover_start_time, commit_bypass_memtable);
       } while (!s.ok() && IsErrorInjectedAndRetryable(s) &&
                initial_wal_write_may_succeed);
@@ -2209,7 +2213,7 @@ class NonBatchedOpsStressTest : public StressTest {
         }
       } else {
         PrintWriteRecoveryWaitTimeIfNeeded(
-            db_stress_env, initial_write_s, initial_wal_write_may_succeed,
+            raw_env, initial_write_s, initial_wal_write_may_succeed,
             wait_for_recover_start_time, "TestDelete");
         pending_expected_value.Commit();
         thread->stats.AddSingleDeletes(1);
@@ -2274,7 +2278,7 @@ class NonBatchedOpsStressTest : public StressTest {
       } else {
         s = db_->DeleteRange(write_opts, cfh, key, end_key);
       }
-      UpdateIfInitialWriteFails(db_stress_env, s, &initial_write_s,
+      UpdateIfInitialWriteFails(raw_env, s, &initial_write_s,
                                 &initial_wal_write_may_succeed,
                                 &wait_for_recover_start_time);
     } while (!s.ok() && IsErrorInjectedAndRetryable(s) &&
@@ -2302,7 +2306,7 @@ class NonBatchedOpsStressTest : public StressTest {
       }
     } else {
       PrintWriteRecoveryWaitTimeIfNeeded(
-          db_stress_env, initial_write_s, initial_wal_write_may_succeed,
+          raw_env, initial_write_s, initial_wal_write_may_succeed,
           wait_for_recover_start_time, "TestDeleteRange");
       for (PendingExpectedValue& pending_expected_value :
            pending_expected_values) {
@@ -2323,60 +2327,16 @@ class NonBatchedOpsStressTest : public StressTest {
     // deletion file's compaction input optimization.
     bool test_standalone_range_deletion = thread->rand.OneInOpt(
         FLAGS_test_ingest_standalone_range_deletion_one_in);
-    std::vector<std::string> external_files;
-    const std::string sst_filename =
-        FLAGS_db + "/." + std::to_string(thread->tid) + ".sst";
-    external_files.push_back(sst_filename);
-    std::string standalone_rangedel_filename;
-    if (test_standalone_range_deletion) {
-      standalone_rangedel_filename = FLAGS_db + "/." +
-                                     std::to_string(thread->tid) +
-                                     "_standalone_rangedel.sst";
-      external_files.push_back(standalone_rangedel_filename);
-    }
+    // When true, reuse the writer's metadata via IngestExternalFileArg's
+    // file_infos so ingestion skips re-opening and scanning the file. Not
+    // combined with the standalone range deletion mode (a range-del-only file).
+    bool use_file_info =
+        !test_standalone_range_deletion &&
+        thread->rand.OneInOpt(FLAGS_ingest_external_file_use_file_info_one_in);
+
     Status s;
     std::ostringstream ingest_options_oss;
 
-    // Temporarily disable error injection for preparation
-    if (fault_fs_guard) {
-      fault_fs_guard->DisableThreadLocalErrorInjection(
-          FaultInjectionIOType::kMetadataRead);
-      fault_fs_guard->DisableThreadLocalErrorInjection(
-          FaultInjectionIOType::kMetadataWrite);
-    }
-
-    for (const auto& filename : external_files) {
-      if (db_stress_env->FileExists(filename).ok()) {
-        // Maybe we terminated abnormally before, so cleanup to give this file
-        // ingestion a clean slate
-        s = db_stress_env->DeleteFile(filename);
-      }
-      if (!s.ok()) {
-        return;
-      }
-    }
-
-    if (fault_fs_guard) {
-      fault_fs_guard->EnableThreadLocalErrorInjection(
-          FaultInjectionIOType::kMetadataRead);
-      fault_fs_guard->EnableThreadLocalErrorInjection(
-          FaultInjectionIOType::kMetadataWrite);
-    }
-
-    SstFileWriter sst_file_writer(EnvOptions(options_), options_);
-    SstFileWriter standalone_rangedel_sst_file_writer(EnvOptions(options_),
-                                                      options_);
-    if (s.ok()) {
-      s = sst_file_writer.Open(sst_filename);
-    }
-    if (s.ok() && test_standalone_range_deletion) {
-      s = standalone_rangedel_sst_file_writer.Open(
-          standalone_rangedel_filename);
-    }
-    if (!s.ok()) {
-      return;
-    }
-
     int64_t key_base = rand_keys[0];
     int column_family = rand_column_families[0];
     std::vector<std::unique_ptr<MutexLock>> range_locks;
@@ -2425,13 +2385,76 @@ class NonBatchedOpsStressTest : public StressTest {
       }
     }
 
-    if (s.ok() && keys.empty()) {
+    if (keys.empty()) {
+      return;
+    }
+    size_t total_keys = keys.size();
+
+    const size_t data_file_count =
+        std::min<size_t>(1 + thread->rand.Uniform(3), total_keys);
+    std::vector<std::string> external_files;
+    std::vector<std::string> data_filenames;
+    data_filenames.reserve(data_file_count);
+    for (size_t file_idx = 0; file_idx < data_file_count; ++file_idx) {
+      data_filenames.push_back(GetDbPath() + "/." +
+                               std::to_string(thread->tid) + "_" +
+                               std::to_string(file_idx) + ".sst");
+      external_files.push_back(data_filenames.back());
+    }
+    std::string standalone_rangedel_filename;
+    if (test_standalone_range_deletion) {
+      standalone_rangedel_filename = GetDbPath() + "/." +
+                                     std::to_string(thread->tid) +
+                                     "_standalone_rangedel.sst";
+      external_files.push_back(standalone_rangedel_filename);
+    }
+
+    // Temporarily disable error injection for preparation
+    if (db_fault_injection_fs_) {
+      db_fault_injection_fs_->DisableThreadLocalErrorInjection(
+          FaultInjectionIOType::kMetadataRead);
+      db_fault_injection_fs_->DisableThreadLocalErrorInjection(
+          FaultInjectionIOType::kMetadataWrite);
+    }
+
+    for (const auto& filename : external_files) {
+      if (raw_env->FileExists(filename).ok()) {
+        // Maybe we terminated abnormally before, so cleanup to give this file
+        // ingestion a clean slate
+        s = raw_env->DeleteFile(filename);
+      }
+      if (!s.ok()) {
+        return;
+      }
+    }
+
+    if (db_fault_injection_fs_) {
+      db_fault_injection_fs_->EnableThreadLocalErrorInjection(
+          FaultInjectionIOType::kMetadataRead);
+      db_fault_injection_fs_->EnableThreadLocalErrorInjection(
+          FaultInjectionIOType::kMetadataWrite);
+    }
+
+    std::vector<ExternalSstFileInfo> file_infos(data_file_count);
+    std::deque<SstFileWriter> sst_file_writers;
+    for (size_t file_idx = 0; s.ok() && file_idx < data_file_count;
+         ++file_idx) {
+      sst_file_writers.emplace_back(EnvOptions(options_), options_);
+      s = sst_file_writers.back().Open(data_filenames[file_idx]);
+    }
+    SstFileWriter standalone_rangedel_sst_file_writer(EnvOptions(options_),
+                                                      options_);
+    if (s.ok() && test_standalone_range_deletion) {
+      s = standalone_rangedel_sst_file_writer.Open(
+          standalone_rangedel_filename);
+    }
+    if (!s.ok()) {
       return;
     }
 
     // set pending state on expected values, create and ingest files.
-    size_t total_keys = keys.size();
     for (size_t i = 0; s.ok() && i < total_keys; i++) {
+      auto& sst_file_writer = sst_file_writers[i % sst_file_writers.size()];
       int64_t key = keys.at(i);
       char value[100];
       auto key_str = Key(key);
@@ -2459,7 +2482,9 @@ class NonBatchedOpsStressTest : public StressTest {
       }
     }
     if (s.ok() && !keys.empty()) {
-      s = sst_file_writer.Finish();
+      for (size_t i = 0; s.ok() && i < sst_file_writers.size(); i++) {
+        s = sst_file_writers[i].Finish(&file_infos[i]);
+      }
     }
 
     if (s.ok() && total_keys != 0 && test_standalone_range_deletion) {
@@ -2477,6 +2502,7 @@ class NonBatchedOpsStressTest : public StressTest {
         s = standalone_rangedel_sst_file_writer.Finish();
       }
     }
+    bool dropped_without_commit = false;
     if (s.ok()) {
       IngestExternalFileOptions ingest_options;
       ingest_options.move_files = thread->rand.OneInOpt(2);
@@ -2484,24 +2510,95 @@ class NonBatchedOpsStressTest : public StressTest {
       ingest_options.verify_checksums_readahead_size =
           thread->rand.OneInOpt(2) ? 1024 * 1024 : 0;
       ingest_options.fill_cache = thread->rand.OneInOpt(4);
+      ingest_options.file_opening_threads = 1 + thread->rand.Uniform(4);
+      ingest_options.prefetch_lmax_index_and_filter_blocks =
+          !thread->rand.OneInOpt(4);
+      const bool use_prepare_commit = thread->rand.OneInOpt(
+          FLAGS_ingest_external_file_prepare_commit_one_in);
+      const bool use_separate_prepare_calls = use_prepare_commit &&
+                                              external_files.size() > 1 &&
+                                              thread->rand.OneInOpt(2);
       ingest_options_oss << "move_files: " << ingest_options.move_files
                          << ", verify_checksums_before_ingest: "
                          << ingest_options.verify_checksums_before_ingest
                          << ", verify_checksums_readahead_size: "
                          << ingest_options.verify_checksums_readahead_size
                          << ", fill_cache: " << ingest_options.fill_cache
+                         << ", file_opening_threads: "
+                         << ingest_options.file_opening_threads
+                         << ", prefetch_lmax_index_and_filter_blocks: "
+                         << ingest_options.prefetch_lmax_index_and_filter_blocks
+                         << ", ingest_external_file_data_file_count: "
+                         << data_file_count
+                         << ", num_external_files: " << external_files.size()
                          << ", test_standalone_range_deletion: "
-                         << test_standalone_range_deletion;
-      s = db_->IngestExternalFile(column_families_[column_family],
-                                  external_files, ingest_options);
+                         << test_standalone_range_deletion
+                         << ", use_prepare_commit: " << use_prepare_commit
+                         << ", use_separate_prepare_calls: "
+                         << use_separate_prepare_calls
+                         << ", use_file_info: " << use_file_info;
+      IngestExternalFileArg arg;
+      arg.column_family = column_families_[column_family];
+      arg.external_files = external_files;
+      arg.options = ingest_options;
+      if (use_file_info) {
+        for (const auto& file_info : file_infos) {
+          arg.file_infos.push_back(file_info.prepared_file_info.get());
+        }
+      }
+      if (use_prepare_commit) {
+        std::vector<std::unique_ptr<FileIngestionHandle>> handles;
+        handles.reserve(use_separate_prepare_calls ? external_files.size() : 1);
+        if (use_separate_prepare_calls) {
+          for (const auto& external_file : external_files) {
+            IngestExternalFileArg file_arg;
+            file_arg.column_family = column_families_[column_family];
+            file_arg.external_files = {external_file};
+            file_arg.options = ingest_options;
+            std::unique_ptr<FileIngestionHandle> handle;
+            s = db_->PrepareFileIngestion({file_arg}, &handle);
+            if (!s.ok()) {
+              break;
+            }
+            handles.push_back(std::move(handle));
+          }
+        } else {
+          std::unique_ptr<FileIngestionHandle> handle;
+          s = db_->PrepareFileIngestion({arg}, &handle);
+          if (s.ok()) {
+            handles.push_back(std::move(handle));
+          }
+        }
+        if (s.ok()) {
+          // Occasionally cancel instead of committing, covering both rollback
+          // paths.
+          if (thread->rand.OneInOpt(4)) {
+            if (thread->rand.OneInOpt(2)) {
+              for (auto& handle : handles) {
+                Status abort_status = handle->Abort();
+                if (!abort_status.ok() && s.ok()) {
+                  s = abort_status;
+                }
+              }
+            } else {
+              handles.clear();  // RAII rollback via destructors
+            }
+            dropped_without_commit = true;
+          } else {
+            s = db_->CommitFileIngestionHandles(std::move(handles));
+          }
+        }
+      } else {
+        s = db_->IngestExternalFiles({arg});
+      }
     }
-    if (!s.ok()) {
+    if (!s.ok() || dropped_without_commit) {
       for (PendingExpectedValue& pending_expected_value :
            pending_expected_values) {
         pending_expected_value.Rollback();
       }
 
-      if (!IsErrorInjectedAndRetryable(s)) {
+      if (!s.ok() && !IsErrorInjectedAndRetryable(s)) {
         fprintf(stderr,
                 "file ingestion error: %s under specified "
                 "IngestExternalFileOptions: %s (Empty string or "
@@ -3485,8 +3582,11 @@ class NonBatchedOpsStressTest : public StressTest {
   }
 };
 
-StressTest* CreateNonBatchedOpsStressTest() {
-  return new NonBatchedOpsStressTest();
+StressTest* CreateNonBatchedOpsStressTest(int db_index,
+                                          const std::string& db_path,
+                                          const std::string& ev_path,
+                                          const std::string& sec_path) {
+  return new NonBatchedOpsStressTest(db_index, db_path, ev_path, sec_path);
 }
 
 }  // namespace ROCKSDB_NAMESPACE

docs/_posts/2026-05-19-resumable-remote-compaction.markdown

@@ -0,0 +1,55 @@
+---
+title: Resumable Remote Compaction
+layout: post
+author: hx235
+category: blog
+---
+
+## Background
+
+RocksDB can offload compaction work to remote workers through the `CompactionService` API. In this model, the **primary RocksDB instance** selects the input files and sends a serialized `CompactionServiceInput` to a worker; the **remote worker** runs `DB::OpenAndCompact()`, writes output SSTs to `output_directory`, and returns a serialized `CompactionServiceResult` that the primary RocksDB instance installs into its LSM tree. See the [Remote Compaction wiki](https://github.com/facebook/rocksdb/wiki/Remote-Compaction) for the full architecture. This lets operators scale compaction throughput with stateless workers while keeping the primary RocksDB instance's CPU and I/O available for serving reads and writes. However, remote compaction jobs can be long-running—sometimes processing hundreds of gigabytes of input. When a worker crashes, gets preempted, or times out, the entire compaction must restart from scratch, wasting all output produced before the interruption and increasing compaction debt on the primary RocksDB instance.
+
+## How Resumable Remote Compaction Works
+
+Resumable remote compaction introduces a **checkpoint-and-resume** mechanism. During a compaction, the worker periodically saves its progress to the `output_directory`. If the compaction is interrupted, a subsequent call to `OpenAndCompact()` with the same output directory can pick up from the last checkpoint rather than starting over.
+
+### Checkpointing
+
+After each output SST file is completed, the worker persists a progress checkpoint to a **compaction progress file** in the output directory `output_directory`. The checkpoint records which internal key to resume from and the metadata of all completed output files. Progress records use **delta encoding**—each record only contains files completed since the last checkpoint—to keep serialization cost linear.
+
+![Checkpointing overview](/static/images/resumable-remote-compaction/checkpointing-overview.svg)
+
+The worker skips checkpointing at boundaries where resuming could be unsafe or requires complicated handling: when range deletions span the file boundary or when adjacent output files share the same user key. These constraints ensure that resuming produces the same results as if the compaction was not interrupted.
+
+### Resuming
+
+When `OpenAndCompact()` is called with `allow_resumption = true`, it scans the output directory for a valid progress file. If one is found, it loads the checkpointed state, seeks the input iterator to the recorded resume key, restores the output file state, and continues compaction from that point. If the progress file is corrupted or missing, the system falls back to a fresh compaction by cleaning the directory.
+
+![Resume flow](/static/images/resumable-remote-compaction/resume-flow.svg)
+
+## How to Enable It
+
+On the primary RocksDB instance, set a `CompactionService` implementation on the DB options. On the remote worker, pass `allow_resumption = true` in `OpenAndCompactOptions` when calling `DB::OpenAndCompact()`. The `output_directory` must be the same across retries for resumption to work—each retry call with the same directory will automatically detect and resume from the previous checkpoint. The `REMOTE_COMPACT_RESUMED_BYTES` statistics ticker tracks the total bytes of output files reused from a previous interrupted run, giving visibility into how much work resumption saved.
+
+```cpp
+// Primary RocksDB instance
+DBOptions db_options;
+db_options.compaction_service = std::make_shared<MyCompactionService>();
+
+// Remote worker
+OpenAndCompactOptions options;
+options.allow_resumption = true;
+
+std::string result;
+Status s = DB::OpenAndCompact(
+    options,
+    db_path,             // source database path
+    output_directory,    // where output SSTs and progress are stored
+    compaction_input,    // serialized CompactionServiceInput
+    &result,             // serialized CompactionServiceResult
+    override_options);
+```
+
+## Future Work
+
+Today this feature targets remote compaction. The same checkpoint-and-resume mechanism could also support **local compaction** after a crash. The core persistence and resume logic is already in `CompactionJob`; the remaining work is to integrate it with local compaction scheduling and recovery.

docs/components/write_flow/03_wal.md

@@ -94,11 +94,23 @@ Step 3 - **Synced**: `fsync()` is called when `WriteOptions::sync` is true, or w
 
 Step 4 - **Obsolete**: After all memtables referencing this WAL are flushed, the WAL becomes obsolete and is either recycled or deleted.
 
+## Async WAL Precreation
+
+When `DBOptions::async_wal_precreate` is enabled, RocksDB keeps at most one future WAL file precreated in the background. RocksDB reserves a file number before scheduling the background task; the task opens the file/writer, but deliberately leaves the file as reserved empty storage:
+
+- It does not write compression metadata.
+- It does not write predecessor WAL info.
+- It does not add the file to `logs_`, `alive_wal_files_`, or MANIFEST WAL tracking.
+
+The file becomes a logical WAL only when foreground `SwitchMemtable()` consumes it. At that point RocksDB writes the normal WAL metadata, flushes the previous WAL writer buffer, and installs the new WAL in the same order as synchronous WAL creation. If foreground rotation reaches the switch while the background task is still running, the writer waits for the reserved file number instead of allocating a higher numbered WAL. This avoids a late lower-numbered WAL appearing after a newer WAL has become live. If background precreation fails, the background task logs the failure and foreground rotation falls back to normal synchronous WAL creation.
+
+Unused precreated WALs are safe across crashes and clean shutdowns because they are zero-record future WAL files. Recovery processes WAL files by log number, marks every observed WAL number as used, and treats an empty future WAL as EOF. On clean close, RocksDB releases the unpublished WAL writer but does not need to delete the empty file. The option is sanitized to false when `recycle_log_file_num > 0`.
+
 ## WAL Tracking in MANIFEST
 
 `WalSet` (see `db/wal_edit.h`, tracked in `VersionSet`) manages WAL metadata in the MANIFEST:
 
-- `WalAddition` records are written when a WAL is created or when a closed/inactive WAL's synced size is finalized
+- `WalAddition` records are written when a closed/inactive WAL's synced size is finalized
 - `WalDeletion` records are written when a WAL becomes obsolete after flush
 - Live-WAL syncs (via `DB::SyncWAL()` or `WriteOptions::sync`) are intentionally not tracked in MANIFEST
 

docs/components/write_flow/04_memtable_insert.md

@@ -103,7 +103,7 @@ When a flush is triggered, `DBImpl::SwitchMemtable()` (see `db/db_impl/db_impl_w
 
 Step 1 - Writes any recoverable state to the current memtable.
 
-Step 2 - Creates a new WAL file (unless the current WAL is empty) with the next file number from `VersionSet::NewFileNumber()`.
+Step 2 - Creates a new WAL file (unless the current WAL is empty) with the next file number from `VersionSet::NewFileNumber()`. If `DBOptions::async_wal_precreate` has prepared a future WAL, this step consumes that reserved file number and finalizes the WAL metadata instead of creating the file synchronously.
 
 Step 3 - Constructs a new `MemTable` with `earliest_seq` set to the current last sequence.