update HISTORY.md and version.h for 9.4.1

Fix file deletions in DestroyDB not rate limited (#12891 )
Summary: Make `DestroyDB` slowly delete files if it's configured and enabled via `SstFileManager`. It's currently not available mainly because of DeleteScheduler's logic related to tracked total_size_ and total_trash_size_. These accounting and logic should not be applied to `DestroyDB`. This PR adds a `DeleteUnaccountedDBFile` util for this purpose which deletes files without accounting it. This util also supports assigning a file to a specified trash bucket so that user can later wait for a specific trash bucket to be empty. For `DestroyDB`, files with more than 1 hard links will be deleted immediately. Pull Request resolved: https://github.com/facebook/rocksdb/pull/12891 Test Plan: Added unit tests, existing tests. Reviewed By: anand1976 Differential Revision: D60300220 Pulled By: jowlyzhang fbshipit-source-id: 8b18109a177a3a9532f6dc2e40e08310c08ca3c7
2026-11-06 11:02:36 -05:00 · 2024-08-05 14:44:17 -07:00 · 2024-08-05 14:42:33 -07:00 · 2024-06-24 16:48:34 -07:00 · 2024-06-23 22:20:59 -07:00 · 2024-06-23 21:54:27 -07:00
340 changed files with 18663 additions and 4448 deletions
@@ -96,3 +96,13 @@ jobs:
    steps:
    - uses: actions/checkout@v4.1.0
    - uses: "./.github/actions/windows-build-steps"
+  build-linux-arm-test-full:
+    if: ${{ github.repository_owner == 'facebook' }}
+    runs-on:
+      labels: 4-core-ubuntu-arm
+    steps:
+      - uses: actions/checkout@v4.1.0
+      - uses: "./.github/actions/pre-steps"
+      - run: sudo apt-get update && sudo apt-get install -y build-essential libgflags-dev
+      - run: make V=1 J=4 -j4 check
+      - uses: "./.github/actions/post-steps"
@@ -90,6 +90,9 @@ jobs:
    container:
      image: zjay437/rocksdb:0.6
      options: --shm-size=16gb
+    env:
+      CC: gcc-10
+      CXX: g++-10
    steps:
    - uses: actions/checkout@v4.1.0
    - uses: "./.github/actions/pre-steps"
@@ -104,38 +107,47 @@ jobs:
    container:
      image: zjay437/rocksdb:0.6
      options: --shm-size=16gb
+    env:
+      CC: gcc-10
+      CXX: g++-10
    steps:
    - uses: actions/checkout@v4.1.0
    - uses: "./.github/actions/pre-steps"
    - uses: "./.github/actions/setup-folly"
    - run: "(mkdir build && cd build && cmake -DUSE_FOLLY_LITE=1 -DWITH_GFLAGS=1 .. && make V=1 -j20)"
    - uses: "./.github/actions/post-steps"
-  build-linux-gcc-7-with-folly:
+  build-linux-make-with-folly:
    if: ${{ github.repository_owner == 'facebook' }}
    runs-on:
      labels: 16-core-ubuntu
    container:
      image: zjay437/rocksdb:0.6
      options: --shm-size=16gb
+    env:
+      CC: gcc-10
+      CXX: g++-10
    steps:
    - uses: actions/checkout@v4.1.0
    - uses: "./.github/actions/pre-steps"
    - uses: "./.github/actions/setup-folly"
    - uses: "./.github/actions/build-folly"
-    - run: USE_FOLLY=1 LIB_MODE=static CC=gcc-7 CXX=g++-7 V=1 make -j32 check
+    - run: USE_FOLLY=1 LIB_MODE=static V=1 make -j32 check
    - uses: "./.github/actions/post-steps"
-  build-linux-gcc-7-with-folly-lite-no-test:
+  build-linux-make-with-folly-lite-no-test:
    if: ${{ github.repository_owner == 'facebook' }}
    runs-on:
      labels: 16-core-ubuntu
    container:
      image: zjay437/rocksdb:0.6
      options: --shm-size=16gb
+    env:
+      CC: gcc-10
+      CXX: g++-10
    steps:
    - uses: actions/checkout@v4.1.0
    - uses: "./.github/actions/pre-steps"
    - uses: "./.github/actions/setup-folly"
-    - run: USE_FOLLY_LITE=1 CC=gcc-7 CXX=g++-7 V=1 make -j32 all
+    - run: USE_FOLLY_LITE=1 V=1 make -j32 all
    - uses: "./.github/actions/post-steps"
  build-linux-cmake-with-folly-coroutines:
    if: ${{ github.repository_owner == 'facebook' }}
@@ -607,3 +619,13 @@ jobs:
      with:
        name: maven-site
        path: "${{ github.workspace }}/java/target/site"
+  build-linux-arm:
+    if: ${{ github.repository_owner == 'facebook' }}
+    runs-on:
+      labels: 4-core-ubuntu-arm
+    steps:
+      - uses: actions/checkout@v4.1.0
+      - uses: "./.github/actions/pre-steps"
+      - run: sudo apt-get update && sudo apt-get install -y build-essential
+      - run: ROCKSDBTESTS_PLATFORM_DEPENDENT=only make V=1 J=4 -j4 all_but_some_tests check_some
+      - uses: "./.github/actions/post-steps"
@@ -47,6 +47,9 @@ package/
 unity.a
 tags
 etags
+GPATH
+GRTAGS
+GTAGS
 rocksdb_dump
 rocksdb_undump
 db_test2
@@ -267,8 +267,8 @@ endif(CMAKE_SYSTEM_PROCESSOR MATCHES "s390x")
 if(CMAKE_SYSTEM_PROCESSOR MATCHES "loongarch64")
  CHECK_C_COMPILER_FLAG("-march=loongarch64" HAS_LOONGARCH64)
  if(HAS_LOONGARCH64)
-    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mcpu=loongarch64 -mtune=loongarch64")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mcpu=loongarch64 -mtune=loongarch64")
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=loongarch64 -mtune=loongarch64")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=loongarch64 -mtune=loongarch64")
  endif(HAS_LOONGARCH64)
 endif(CMAKE_SYSTEM_PROCESSOR MATCHES "loongarch64")

@@ -488,6 +488,8 @@ if(CMAKE_SYSTEM_NAME MATCHES "Cygwin")
  add_definitions(-fno-builtin-memcmp -DCYGWIN)
 elseif(CMAKE_SYSTEM_NAME MATCHES "Darwin")
  add_definitions(-DOS_MACOSX)
+elseif(CMAKE_SYSTEM_NAME MATCHES "iOS")
+  add_definitions(-DOS_MACOSX -DIOS_CROSS_COMPILE)
 elseif(CMAKE_SYSTEM_NAME MATCHES "Linux")
  add_definitions(-DOS_LINUX)
 elseif(CMAKE_SYSTEM_NAME MATCHES "SunOS")
@@ -615,7 +617,7 @@ if(USE_FOLLY)
    FMT_INST_PATH)
    exec_program(ls ARGS -d ${FOLLY_INST_PATH}/../gflags* OUTPUT_VARIABLE
    GFLAGS_INST_PATH)
-    set(Boost_DIR ${BOOST_INST_PATH}/lib/cmake/Boost-1.78.0)
+    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)
    else()
@@ -653,6 +655,7 @@ set(SOURCES
        cache/sharded_cache.cc
        cache/tiered_secondary_cache.cc
        db/arena_wrapped_db_iter.cc
+        db/attribute_group_iterator_impl.cc
        db/blob/blob_contents.cc
        db/blob/blob_fetcher.cc
        db/blob/blob_file_addition.cc
@@ -669,6 +672,7 @@ set(SOURCES
        db/blob/prefetch_buffer_collection.cc
        db/builder.cc
        db/c.cc
+        db/coalescing_iterator.cc
        db/column_family.cc
        db/compaction/compaction.cc
        db/compaction/compaction_iterator.cc
@@ -689,6 +693,7 @@ set(SOURCES
        db/db_impl/db_impl_write.cc
        db/db_impl/db_impl_compaction_flush.cc
        db/db_impl/db_impl_files.cc
+        db/db_impl/db_impl_follower.cc
        db/db_impl/db_impl_open.cc
        db/db_impl/db_impl_debug.cc
        db/db_impl/db_impl_experimental.cc
@@ -715,7 +720,6 @@ set(SOURCES
        db/memtable_list.cc
        db/merge_helper.cc
        db/merge_operator.cc
-        db/multi_cf_iterator.cc
        db/output_validator.cc
        db/periodic_task_scheduler.cc
        db/range_del_aggregator.cc
@@ -747,6 +751,7 @@ set(SOURCES
        env/env_encryption.cc
        env/file_system.cc
        env/file_system_tracer.cc
+        env/fs_on_demand.cc
        env/fs_remap.cc
        env/mock_env.cc
        env/unique_id_gen.cc
@@ -934,6 +939,7 @@ set(SOURCES
        utilities/persistent_cache/volatile_tier_impl.cc
        utilities/simulator_cache/cache_simulator.cc
        utilities/simulator_cache/sim_cache.cc
+        utilities/table_properties_collectors/compact_for_tiering_collector.cc
        utilities/table_properties_collectors/compact_on_deletion_collector.cc
        utilities/trace/file_trace_reader_writer.cc
        utilities/trace/replayer_impl.cc
@@ -954,6 +960,7 @@ set(SOURCES
        utilities/transactions/write_prepared_txn_db.cc
        utilities/transactions/write_unprepared_txn.cc
        utilities/transactions/write_unprepared_txn_db.cc
+        utilities/types_util.cc
        utilities/ttl/db_ttl_impl.cc
        utilities/wal_filter.cc
        utilities/write_batch_with_index/write_batch_with_index.cc
@@ -1045,6 +1052,7 @@ if(USE_FOLLY_LITE)
  list(APPEND SOURCES
    third-party/folly/folly/container/detail/F14Table.cpp
    third-party/folly/folly/detail/Futex.cpp
+    third-party/folly/folly/lang/Exception.cpp
    third-party/folly/folly/lang/SafeAssert.cpp
    third-party/folly/folly/lang/ToAscii.cpp
    third-party/folly/folly/ScopeGuard.cpp
@@ -1052,6 +1060,12 @@ if(USE_FOLLY_LITE)
    third-party/folly/folly/synchronization/DistributedMutex.cpp
    third-party/folly/folly/synchronization/ParkingLot.cpp)
  include_directories(${PROJECT_SOURCE_DIR}/third-party/folly)
+  exec_program(python3 ${PROJECT_SOURCE_DIR}/third-party/folly ARGS
+  build/fbcode_builder/getdeps.py show-source-dir boost OUTPUT_VARIABLE
+  BOOST_SOURCE_PATH)
+  exec_program(ls ARGS -d ${BOOST_SOURCE_PATH}/boost* OUTPUT_VARIABLE
+  BOOST_INCLUDE_DIR)
+  include_directories(${BOOST_INCLUDE_DIR})
  add_definitions(-DUSE_FOLLY -DFOLLY_NO_CONFIG)
  list(APPEND THIRDPARTY_LIBS glog)
 endif()
@@ -1361,6 +1375,7 @@ if(WITH_TESTS)
        db/file_indexer_test.cc
        db/filename_test.cc
        db/flush_job_test.cc
+        db/db_follower_test.cc
        db/import_column_family_test.cc
        db/listener_test.cc
        db/log_test.cc
@@ -1474,6 +1489,7 @@ if(WITH_TESTS)
        utilities/persistent_cache/persistent_cache_test.cc
        utilities/simulator_cache/cache_simulator_test.cc
        utilities/simulator_cache/sim_cache_test.cc
+        utilities/table_properties_collectors/compact_for_tiering_collector_test.cc
        utilities/table_properties_collectors/compact_on_deletion_collector_test.cc
        utilities/transactions/optimistic_transaction_test.cc
        utilities/transactions/transaction_test.cc
@@ -1484,6 +1500,7 @@ if(WITH_TESTS)
        utilities/transactions/lock/range/range_locking_test.cc
        utilities/transactions/timestamped_snapshot_test.cc
        utilities/ttl/ttl_test.cc
+        utilities/types_util_test.cc
        utilities/util_merge_operators_test.cc
        utilities/write_batch_with_index/write_batch_with_index_test.cc
 	${PLUGIN_TESTS}
@@ -1,6 +1,142 @@
 # Rocksdb Change Log
 > NOTE: Entries for next release do not go here. Follow instructions in `unreleased_history/README.txt`

+## 9.4.1 (08/02/2024)
+### Bug Fixes
+* *Make DestroyDB supports slow deletion when it's configured in `SstFileManager`. The slow deletion is subject to the configured `rate_bytes_per_sec`, but not subject to the `max_trash_db_ratio`.
+
+## 9.4.0 (06/23/2024)
+### New Features
+* Added a `CompactForTieringCollectorFactory` to auto trigger compaction for tiering use case.
+* Optimistic transactions and pessimistic transactions with the WriteCommitted policy now support the `GetEntityForUpdate` API.
+* Added a new "count" command to the ldb repl shell. By default, it prints a count of keys in the database from start to end. The options --from=<key> and/or --to=<key> can be specified to limit the range.
+* Add `rocksdb_writebatch_update_timestamps`, `rocksdb_writebatch_wi_update_timestamps` in C API.
+* Add `rocksdb_iter_refresh` in C API.
+* Add `rocksdb_writebatch_create_with_params`, `rocksdb_writebatch_wi_create_with_params` to create WB and WBWI with all options in C API
+
+### Public API Changes
+* Deprecated names `LogFile` and `VectorLogPtr` in favor of new names `WalFile` and `VectorWalPtr`.
+* Introduce a new universal compaction option CompactionOptionsUniversal::max_read_amp which allows user to define the limit on the number of sorted runs separately from the trigger for compaction (`level0_file_num_compaction_trigger`) #12477.
+
+### Behavior Changes
+* Inactive WALs are immediately closed upon being fully sync-ed rather than in a background thread. This is to ensure LinkFile() is not called on files still open for write, which might not be supported by some FileSystem implementations. This should not be a performance issue, but an opt-out is available with with new DB option `background_close_inactive_wals`.
+
+### Bug Fixes
+* Fix a rare case in which a hard-linked WAL in a Checkpoint is not fully synced (so might lose data on power loss).
+* Fixed the output of the `ldb dump_wal` command for `PutEntity` records so it prints the key and correctly resets the hexadecimal formatting flag after printing the wide-column entity.
+* Fixed an issue where `PutEntity` records were handled incorrectly while rebuilding transactions during recovery.
+* Various read operations could ignore various ReadOptions that might be relevant. Fixed many such cases, which can result in behavior change but a better reflection of specified options.
+
+### Performance Improvements
+* Improved write throughput to memtable when there's a large number of concurrent writers and allow_concurrent_memtable_write=true(#12545)
+
+## 9.3.0 (05/17/2024)
+### New Features
+* Optimistic transactions and pessimistic transactions with the WriteCommitted policy now support the `GetEntity` API.
+* Added new `Iterator` property, "rocksdb.iterator.is-value-pinned", for checking whether the `Slice` returned by `Iterator::value()` can be used until the `Iterator` is destroyed.
+* Optimistic transactions and WriteCommitted pessimistic transactions now support the `MultiGetEntity` API.
+* Optimistic transactions and pessimistic transactions with the WriteCommitted policy now support the `PutEntity` API. Support for read APIs and other write policies (WritePrepared, WriteUnprepared) will be added later.
+
+### Public API Changes
+* Exposed block based metadata cache options via C API
+* Exposed compaction pri via c api.
+* Add a kAdmPolicyAllowAll option to TieredAdmissionPolicy that admits all blocks evicted from the primary block cache into the compressed secondary cache.
+
+### Behavior Changes
+* CompactRange() with change_level=true on a CF with FIFO compaction will return Status::NotSupported().
+* External file ingestion with FIFO compaction will always ingest to L0.
+
+### Bug Fixes
+* Fixed a bug for databases using `DBOptions::allow_2pc == true` (all `TransactionDB`s except `OptimisticTransactionDB`) that have exactly one column family. Due to a missing WAL sync, attempting to open the DB could have returned a `Status::Corruption` with a message like "SST file is ahead of WALs".
+* Fix a bug in CreateColumnFamilyWithImport() where if multiple CFs are imported, we were not resetting files' epoch number and L0 files can have overlapping key range but the same epoch number.
+* Fixed race conditions when `ColumnFamilyOptions::inplace_update_support == true` between user overwrites and reads on the same key.
+* Fix a bug where `CompactFiles()` can compact files of range conflict with other ongoing compactions' when `preclude_last_level_data_seconds > 0` is used
+* Fixed a false positive `Status::Corruption` reported when reopening a DB that used `DBOptions::recycle_log_file_num > 0` and `DBOptions::wal_compression != kNoCompression`.
+* While WAL is locked with LockWAL(), some operations like Flush() and IngestExternalFile() are now blocked as they should have been.
+* Fixed a bug causing stale memory access when using the TieredSecondaryCache with an NVM secondary cache, and a file system that supports return an FS allocated buffer for MultiRead (FSSupportedOps::kFSBuffer is set).
+
+## 9.2.0 (05/01/2024)
+### New Features
+* Added two options `deadline` and `max_size_bytes` for CacheDumper to exit early
+* Added a new API `GetEntityFromBatchAndDB` to `WriteBatchWithIndex` that can be used for wide-column point lookups with read-your-own-writes consistency. Similarly to `GetFromBatchAndDB`, the API can combine data from the write batch with data from the underlying database if needed. See the API comments for more details.
+* [Experimental] Introduce two new cross-column-family iterators - CoalescingIterator and AttributeGroupIterator. The CoalescingIterator enables users to iterate over multiple column families and access their values and columns. During this iteration, if the same key exists in more than one column family, the keys in the later column family will overshadow the previous ones. The AttributeGroupIterator allows users to gather wide columns per Column Family and create attribute groups while iterating over keys across all CFs.
+* Added a new API `MultiGetEntityFromBatchAndDB` to `WriteBatchWithIndex` that can be used for batched wide-column point lookups with read-your-own-writes consistency. Similarly to `MultiGetFromBatchAndDB`, the API can combine data from the write batch with data from the underlying database if needed. See the API comments for more details.
+* *Adds a `SstFileReader::NewTableIterator` API to support programmatically read a SST file as a raw table file.
+* Add an option to `WaitForCompactOptions` - `wait_for_purge` to make `WaitForCompact()` API wait for background purge to complete
+
+### Public API Changes
+* DeleteRange() will return NotSupported() if row_cache is configured since they don't work together in some cases.
+* Deprecated `CompactionOptions::compression` since `CompactionOptions`'s API for configuring compression was incomplete, unsafe, and likely unnecessary
+* Using `OptionChangeMigration()` to migrate from non-FIFO to FIFO compaction
+with `Options::compaction_options_fifo.max_table_files_size` > 0 can cause
+the whole DB to be dropped right after migration if the migrated data is larger than
+`max_table_files_size`
+
+### Behavior Changes
+* Enabling `BlockBasedTableOptions::block_align` is now incompatible (i.e., APIs will return `Status::InvalidArgument`) with more ways of enabling compression: `CompactionOptions::compression`, `ColumnFamilyOptions::compression_per_level`, and `ColumnFamilyOptions::bottommost_compression`.
+* Changed the default value of `CompactionOptions::compression` to `kDisableCompressionOption`, which means the compression type is determined by the `ColumnFamilyOptions`.
+* `BlockBasedTableOptions::optimize_filters_for_memory` is now set to true by default. When `partition_filters=false`, this could lead to somewhat increased average RSS memory usage by the block cache, but this "extra" usage is within the allowed memory budget and should make memory usage more consistent (by minimizing internal fragmentation for more kinds of blocks).
+* Dump all keys for cache dumper impl if `SetDumpFilter()` is not called
+* `CompactRange()` with `CompactRangeOptions::change_level = true` and `CompactRangeOptions::target_level = 0` that ends up moving more than 1 file from non-L0 to L0 will return `Status::Aborted()`.
+* On distributed file systems that support file system level checksum verification and reconstruction reads, RocksDB will now retry a file read if the initial read fails RocksDB block level or record level checksum verification. This applies to MANIFEST file reads when the DB is opened, and to SST file reads at all times.
+
+### Bug Fixes
+* Fix a bug causing `VerifyFileChecksums()` to return false-positive corruption under `BlockBasedTableOptions::block_align=true`
+* Provide consistent view of the database across the column families for `NewIterators()` API.
+* Fixed feature interaction bug for `DeleteRange()` together with `ColumnFamilyOptions::memtable_insert_with_hint_prefix_extractor`. The impact of this bug would likely be corruption or crashing.
+* Fixed hang in `DisableManualCompactions()` where compactions waiting to be scheduled due to conflicts would not be canceled promptly
+* Fixed a regression when `ColumnFamilyOptions::max_successive_merges > 0` where the CPU overhead for deciding whether to merge could have increased unless the user had set the option `ColumnFamilyOptions::strict_max_successive_merges`
+* Fixed a bug in `MultiGet()` and `MultiGetEntity()` together with blob files (`ColumnFamilyOptions::enable_blob_files == true`). An error looking up one of the keys could cause the results to be wrong for other keys for which the statuses were `Status::OK`.
+* Fixed a bug where wrong padded bytes are used to generate file checksum and `DataVerificationInfo::checksum` upon file creation
+* Correctly implemented the move semantics of `PinnableWideColumns`.
+* Fixed a bug when the recycle_log_file_num in DBOptions is changed from 0 to non-zero when a DB is reopened. On a subsequent reopen, if a log file created when recycle_log_file_num==0 was reused previously, is alive and is empty, we could end up inserting stale WAL records into the memtable.
+* *Fix a bug where obsolete files' deletion during DB::Open are not rate limited with `SstFilemManager`'s slow deletion feature even if it's configured.
+
+## 9.1.0 (03/22/2024)
+### New Features
+* Added an option, `GetMergeOperandsOptions::continue_cb`, to give users the ability to end `GetMergeOperands()`'s lookup process before all merge operands were found.
+* *Add sanity checks for ingesting external files that currently checks if the user key comparator used to create the file is compatible with the column family's user key comparator.
+*Support ingesting external files for column family that has user-defined timestamps in memtable only enabled.
+* On file systems that support storage level data checksum and reconstruction, retry SST block reads for point lookups, scans, and flush and compaction if there's a checksum mismatch on the initial read.
+* Some enhancements and fixes to experimental Temperature handling features, including new `default_write_temperature` CF option and opening an `SstFileWriter` with a temperature.
+* `WriteBatchWithIndex` now supports wide-column point lookups via the `GetEntityFromBatch` API. See the API comments for more details.
+* *Implement experimental features: API `Iterator::GetProperty("rocksdb.iterator.write-time")` to allow users to get data's approximate write unix time and write data with a specific write time via `WriteBatch::TimedPut` API.
+
+### Public API Changes
+* Best-effort recovery (`best_efforts_recovery == true`) may now be used together with atomic flush (`atomic_flush == true`). The all-or-nothing recovery guarantee for atomically flushed data will be upheld.
+* Remove deprecated option `bottommost_temperature`, already replaced by `last_level_temperature`
+* Added new PerfContext counters for block cache bytes read - block_cache_index_read_byte, block_cache_filter_read_byte, block_cache_compression_dict_read_byte, and block_cache_read_byte.
+* Deprecate experimental Remote Compaction APIs - StartV2() and WaitForCompleteV2() and introduce Schedule() and Wait(). The new APIs essentially does the same thing as the old APIs. They allow taking externally generated unique id to wait for remote compaction to complete.
+* *For API `WriteCommittedTransaction::GetForUpdate`, if the column family enables user-defined timestamp, it was mandated that argument `do_validate` cannot be false, and UDT based validation has to be done with a user set read timestamp. It's updated to make the UDT based validation optional if user sets `do_validate` to false and does not set a read timestamp. With this, `GetForUpdate` skips UDT based validation and it's users' responsibility to enforce the UDT invariant. SO DO NOT skip this UDT-based validation if users do not have ways to enforce the UDT invariant. Ways to enforce the invariant on the users side include manage a monotonically increasing timestamp, commit transactions in a single thread etc.
+* Defined a new PerfLevel `kEnableWait` to measure time spent by user threads blocked in RocksDB other than mutex, such as a write thread waiting to be added to a write group, a write thread delayed or stalled etc.
+* `RateLimiter`'s API no longer requires the burst size to be the refill size. Users of `NewGenericRateLimiter()` can now provide burst size in `single_burst_bytes`. Implementors of `RateLimiter::SetSingleBurstBytes()` need to adapt their implementations to match the changed API doc.
+* Add `write_memtable_time` to the newly introduced PerfLevel `kEnableWait`.
+
+### Behavior Changes
+* `RateLimiter`s created by `NewGenericRateLimiter()` no longer modify the refill period when `SetSingleBurstBytes()` is called.
+* Merge writes will only keep merge operand count within `ColumnFamilyOptions::max_successive_merges` when the key's merge operands are all found in memory, unless `strict_max_successive_merges` is explicitly set.
+
+### Bug Fixes
+* Fixed `kBlockCacheTier` reads to return `Status::Incomplete` when I/O is needed to fetch a merge chain's base value from a blob file.
+* Fixed `kBlockCacheTier` reads to return `Status::Incomplete` on table cache miss rather than incorrectly returning an empty value.
+* Fixed a data race in WalManager that may affect how frequent PurgeObsoleteWALFiles() runs.
+* Re-enable the recycle_log_file_num option in DBOptions for kPointInTimeRecovery WAL recovery mode, which was previously disabled due to a bug in the recovery logic. This option is incompatible with WriteOptions::disableWAL. A Status::InvalidArgument() will be returned if disableWAL is specified.
+
+### Performance Improvements
+* Java API `multiGet()` variants now take advantage of the underlying batched `multiGet()` performance improvements.
+Before
+```
+Benchmark (columnFamilyTestType) (keyCount) (keySize) (multiGetSize) (valueSize) Mode Cnt Score Error Units
+MultiGetBenchmarks.multiGetList10 no_column_family 10000 16 100 64 thrpt 25 6315.541 ± 8.106 ops/s
+MultiGetBenchmarks.multiGetList10 no_column_family 10000 16 100 1024 thrpt 25 6975.468 ± 68.964 ops/s
+```
+After
+```
+Benchmark (columnFamilyTestType) (keyCount) (keySize) (multiGetSize) (valueSize) Mode Cnt Score Error Units
+MultiGetBenchmarks.multiGetList10 no_column_family 10000 16 100 64 thrpt 25 7046.739 ± 13.299 ops/s
+MultiGetBenchmarks.multiGetList10 no_column_family 10000 16 100 1024 thrpt 25 7654.521 ± 60.121 ops/s
+```
+
 ## 9.0.0 (02/16/2024)
 ### New Features
 * Provide support for FSBuffer for point lookups. Also added support for scans and compactions that don't go through prefetching.
@@ -103,6 +103,7 @@ dummy := $(shell (export ROCKSDB_ROOT="$(CURDIR)"; \
                  export LIB_MODE="$(LIB_MODE)"; \
 		  export ROCKSDB_CXX_STANDARD="$(ROCKSDB_CXX_STANDARD)"; \
 		  export USE_FOLLY="$(USE_FOLLY)"; \
+		  export USE_FOLLY_LITE="$(USE_FOLLY_LITE)"; \
                  "$(CURDIR)/build_tools/build_detect_platform" "$(CURDIR)/make_config.mk"))
 # this file is generated by the previous line to set build flags and sources
 include make_config.mk
@@ -500,6 +501,17 @@ endif
 ifeq ($(USE_FOLLY_LITE),1)
 	# Path to the Folly source code and include files
 	FOLLY_DIR = ./third-party/folly
+ifneq ($(strip $(BOOST_SOURCE_PATH)),)
+	BOOST_INCLUDE = $(shell (ls -d $(BOOST_SOURCE_PATH)/boost*/))
+	# AIX: pre-defined system headers are surrounded by an extern "C" block
+	ifeq ($(PLATFORM), OS_AIX)
+		PLATFORM_CCFLAGS += -I$(BOOST_INCLUDE)
+		PLATFORM_CXXFLAGS += -I$(BOOST_INCLUDE)
+	else
+		PLATFORM_CCFLAGS += -isystem $(BOOST_INCLUDE)
+		PLATFORM_CXXFLAGS += -isystem $(BOOST_INCLUDE)
+	endif
+endif  # BOOST_SOURCE_PATH
 	# AIX: pre-defined system headers are surrounded by an extern "C" block
 	ifeq ($(PLATFORM), OS_AIX)
 		PLATFORM_CCFLAGS += -I$(FOLLY_DIR)
@@ -636,26 +648,37 @@ TESTS = $(patsubst %.cc, %, $(notdir $(TEST_MAIN_SOURCES)))
 TESTS += $(patsubst %.c, %, $(notdir $(TEST_MAIN_SOURCES_C)))
 TESTS += $(PLUGIN_TESTS)

-# `make check-headers` to very that each header file includes its own
-# dependencies
+# `make check-headers` to verify that each header file includes its own deps
+# and that public headers do not depend on internal headers
 ifneq ($(filter check-headers, $(MAKECMDGOALS)),)
 # TODO: add/support JNI headers
 	DEV_HEADER_DIRS := $(sort include/ $(dir $(ALL_SOURCES)))
 # Some headers like in port/ are platform-specific
-	DEV_HEADERS := $(shell $(FIND) $(DEV_HEADER_DIRS) -type f -name '*.h' | grep -E -v 'port/|plugin/|lua/|range_tree/')
+	DEV_HEADERS_TO_CHECK := $(shell $(FIND) $(DEV_HEADER_DIRS) -type f -name '*.h' | grep -E -v 'port/|plugin/|lua/|range_tree/')
+	PUBLIC_HEADERS_TO_CHECK := $(shell $(FIND) include/ -type f -name '*.h' | grep -E -v 'lua/')
 else
-	DEV_HEADERS :=
+	DEV_HEADERS_TO_CHECK :=
+	PUBLIC_HEADERS_TO_CHECK :=
 endif
-HEADER_OK_FILES = $(patsubst %.h, %.h.ok, $(DEV_HEADERS))
+HEADER_OK_FILES = $(patsubst %.h, %.h.ok, $(DEV_HEADERS_TO_CHECK)) \
+	$(patsubst %.h, %.h.pub, $(PUBLIC_HEADERS_TO_CHECK))

 AM_V_CCH = $(am__v_CCH_$(V))
 am__v_CCH_ = $(am__v_CCH_$(AM_DEFAULT_VERBOSITY))
 am__v_CCH_0 = @echo "  CC.h    " $<;
 am__v_CCH_1 =

+# verify headers include their own dependencies, under dev build settings
 %.h.ok: %.h # .h.ok not actually created, so re-checked on each invocation
 # -DROCKSDB_NAMESPACE=42 ensures the namespace header is included
-	$(AM_V_CCH) echo '#include "$<"' | $(CXX) $(CXXFLAGS) -DROCKSDB_NAMESPACE=42 -x c++ -c - -o /dev/null
+	$(AM_V_CCH) echo '#include "$<"' | $(CXX) $(CXXFLAGS) \
+	  -DROCKSDB_NAMESPACE=42 -x c++ -c - -o /dev/null
+
+# verify public headers do not depend on internal headers, under typical
+# user build settings
+%.h.pub: %.h # .h.pub not actually created, so re-checked on each invocation
+	$(AM_V_CCH) cd include/ && echo '#include "$(patsubst include/%,%,$<)"' | \
+	  $(CXX) -I. -DROCKSDB_NAMESPACE=42 -x c++ -c - -o /dev/null

 check-headers: $(HEADER_OK_FILES)

@@ -1610,6 +1633,9 @@ object_registry_test: $(OBJ_DIR)/utilities/object_registry_test.o $(TEST_LIBRARY
 ttl_test: $(OBJ_DIR)/utilities/ttl/ttl_test.o $(TEST_LIBRARY) $(LIBRARY)
 	$(AM_LINK)

+types_util_test: $(OBJ_DIR)/utilities/types_util_test.o $(TEST_LIBRARY) $(LIBRARY)
+	$(AM_LINK)
+
 write_batch_with_index_test: $(OBJ_DIR)/utilities/write_batch_with_index/write_batch_with_index_test.o $(TEST_LIBRARY) $(LIBRARY)
 	$(AM_LINK)

@@ -1628,6 +1654,9 @@ compaction_job_stats_test: $(OBJ_DIR)/db/compaction/compaction_job_stats_test.o
 compaction_service_test: $(OBJ_DIR)/db/compaction/compaction_service_test.o $(TEST_LIBRARY) $(LIBRARY)
 	$(AM_LINK)

+compact_for_tiering_collector_test: $(OBJ_DIR)/utilities/table_properties_collectors/compact_for_tiering_collector_test.o $(TEST_LIBRARY) $(LIBRARY)
+	$(AM_LINK)
+
 compact_on_deletion_collector_test: $(OBJ_DIR)/utilities/table_properties_collectors/compact_on_deletion_collector_test.o $(TEST_LIBRARY) $(LIBRARY)
 	$(AM_LINK)

@@ -1919,6 +1948,9 @@ sst_file_reader_test: $(OBJ_DIR)/table/sst_file_reader_test.o $(TEST_LIBRARY) $(
 db_secondary_test: $(OBJ_DIR)/db/db_secondary_test.o $(TEST_LIBRARY) $(LIBRARY)
 	$(AM_LINK)

+db_follower_test: $(OBJ_DIR)/db/db_follower_test.o $(TEST_LIBRARY) $(LIBRARY)
+	$(AM_LINK)
+
 block_cache_tracer_test: $(OBJ_DIR)/trace_replay/block_cache_tracer_test.o $(TEST_LIBRARY) $(LIBRARY)
 	$(AM_LINK)

@@ -2452,14 +2484,15 @@ checkout_folly:
 	fi
 	@# Pin to a particular version for public CI, so that PR authors don't
 	@# need to worry about folly breaking our integration. Update periodically
-	cd third-party/folly && git reset --hard beacd86d63cd71c904632262e6c36f60874d78ba
-	@# A hack to remove boost dependency.
-	@# NOTE: this hack is only needed if building using USE_FOLLY_LITE
-	perl -pi -e 's/^(#include <boost)/\/\/$$1/' third-party/folly/folly/functional/Invoke.h
+	cd third-party/folly && git reset --hard 17054c85c2dee0e84cbee7ef28d7bae7d0b1630e
 	@# NOTE: this hack is required for clang in some cases
 	perl -pi -e 's/int rv = syscall/int rv = (int)syscall/' third-party/folly/folly/detail/Futex.cpp
 	@# NOTE: this hack is required for gcc in some cases
 	perl -pi -e 's/(__has_include.<experimental.memory_resource>.)/__cpp_rtti && $$1/' third-party/folly/folly/memory/MemoryResource.h
+	@# NOTE: this hack is only needed if building target build_folly in the future
+	perl -pi -e 's,https://github.com/jedisct1/libsodium/releases/download/1.0.17/libsodium-1.0.17.tar.gz,https://download.libsodium.org/libsodium/releases/old/unsupported/libsodium-1.0.17.tar.gz,' third-party/folly/build/fbcode_builder/manifests/libsodium
+	@# NOTE: boost source will be needed for any build including `USE_FOLLY_LITE` builds as those depend on boost headers
+	cd third-party/folly && $(PYTHON) build/fbcode_builder/getdeps.py fetch boost

 CXX_M_FLAGS = $(filter -m%, $(CXXFLAGS))

@@ -2471,8 +2504,6 @@ build_folly:
 		echo "Please run checkout_folly first"; \
 		false; \
 	fi
-	# Restore the original version of Invoke.h with boost dependency
-	cd third-party/folly && ${GIT_COMMAND} checkout folly/functional/Invoke.h
 	cd third-party/folly && \
 		CXXFLAGS=" $(CXX_M_FLAGS) -DHAVE_CXX11_ATOMIC " $(PYTHON) build/fbcode_builder/getdeps.py build --no-tests

@@ -4,6 +4,7 @@
 # This file is a Facebook-specific integration for buck builds, so can
 # only be validated by Facebook employees.
 load("//rocks/buckifier:defs.bzl", "cpp_library_wrapper","rocks_cpp_library_wrapper","cpp_binary_wrapper","cpp_unittest_wrapper","fancy_bench_wrapper","add_c_test_wrapper")
+load("@fbcode_macros//build_defs:export_files.bzl", "export_file")


 cpp_library_wrapper(name="rocksdb_lib", srcs=[
@@ -21,6 +22,7 @@ cpp_library_wrapper(name="rocksdb_lib", srcs=[
        "cache/sharded_cache.cc",
        "cache/tiered_secondary_cache.cc",
        "db/arena_wrapped_db_iter.cc",
+        "db/attribute_group_iterator_impl.cc",
        "db/blob/blob_contents.cc",
        "db/blob/blob_fetcher.cc",
        "db/blob/blob_file_addition.cc",
@@ -37,6 +39,7 @@ cpp_library_wrapper(name="rocksdb_lib", srcs=[
        "db/blob/prefetch_buffer_collection.cc",
        "db/builder.cc",
        "db/c.cc",
+        "db/coalescing_iterator.cc",
        "db/column_family.cc",
        "db/compaction/compaction.cc",
        "db/compaction/compaction_iterator.cc",
@@ -58,6 +61,7 @@ cpp_library_wrapper(name="rocksdb_lib", srcs=[
        "db/db_impl/db_impl_debug.cc",
        "db/db_impl/db_impl_experimental.cc",
        "db/db_impl/db_impl_files.cc",
+        "db/db_impl/db_impl_follower.cc",
        "db/db_impl/db_impl_open.cc",
        "db/db_impl/db_impl_readonly.cc",
        "db/db_impl/db_impl_secondary.cc",
@@ -83,7 +87,6 @@ cpp_library_wrapper(name="rocksdb_lib", srcs=[
        "db/memtable_list.cc",
        "db/merge_helper.cc",
        "db/merge_operator.cc",
-        "db/multi_cf_iterator.cc",
        "db/output_validator.cc",
        "db/periodic_task_scheduler.cc",
        "db/range_del_aggregator.cc",
@@ -116,6 +119,7 @@ cpp_library_wrapper(name="rocksdb_lib", srcs=[
        "env/env_posix.cc",
        "env/file_system.cc",
        "env/file_system_tracer.cc",
+        "env/fs_on_demand.cc",
        "env/fs_posix.cc",
        "env/fs_remap.cc",
        "env/io_posix.cc",
@@ -313,6 +317,7 @@ cpp_library_wrapper(name="rocksdb_lib", srcs=[
        "utilities/persistent_cache/volatile_tier_impl.cc",
        "utilities/simulator_cache/cache_simulator.cc",
        "utilities/simulator_cache/sim_cache.cc",
+        "utilities/table_properties_collectors/compact_for_tiering_collector.cc",
        "utilities/table_properties_collectors/compact_on_deletion_collector.cc",
        "utilities/trace/file_trace_reader_writer.cc",
        "utilities/trace/replayer_impl.cc",
@@ -346,6 +351,7 @@ cpp_library_wrapper(name="rocksdb_lib", srcs=[
        "utilities/transactions/write_unprepared_txn.cc",
        "utilities/transactions/write_unprepared_txn_db.cc",
        "utilities/ttl/db_ttl_impl.cc",
+        "utilities/types_util.cc",
        "utilities/wal_filter.cc",
        "utilities/write_batch_with_index/write_batch_with_index.cc",
        "utilities/write_batch_with_index/write_batch_with_index_internal.cc",
@@ -389,6 +395,7 @@ rocks_cpp_library_wrapper(name="rocksdb_stress_lib", srcs=[
        "db_stress_tool/cf_consistency_stress.cc",
        "db_stress_tool/db_stress_common.cc",
        "db_stress_tool/db_stress_driver.cc",
+        "db_stress_tool/db_stress_filters.cc",
        "db_stress_tool/db_stress_gflags.cc",
        "db_stress_tool/db_stress_listener.cc",
        "db_stress_tool/db_stress_shared_state.cc",
@@ -406,6 +413,8 @@ rocks_cpp_library_wrapper(name="rocksdb_stress_lib", srcs=[
    ], headers=None)


+cpp_binary_wrapper(name="ldb", srcs=["tools/ldb.cc"], deps=[":rocksdb_tools_lib"], extra_preprocessor_flags=[], extra_bench_libs=False)
+
 cpp_binary_wrapper(name="db_stress", srcs=["db_stress_tool/db_stress.cc"], deps=[":rocksdb_stress_lib"], extra_preprocessor_flags=[], extra_bench_libs=False)

 cpp_binary_wrapper(name="cache_bench", srcs=["cache/cache_bench.cc"], deps=[":rocksdb_cache_bench_tools_lib"], extra_preprocessor_flags=[], extra_bench_libs=False)
@@ -4619,6 +4628,12 @@ cpp_unittest_wrapper(name="compact_files_test",
            extra_compiler_flags=[])


+cpp_unittest_wrapper(name="compact_for_tiering_collector_test",
+            srcs=["utilities/table_properties_collectors/compact_for_tiering_collector_test.cc"],
+            deps=[":rocksdb_test_lib"],
+            extra_compiler_flags=[])
+
+
 cpp_unittest_wrapper(name="compact_on_deletion_collector_test",
            srcs=["utilities/table_properties_collectors/compact_on_deletion_collector_test.cc"],
            deps=[":rocksdb_test_lib"],
@@ -4793,6 +4808,12 @@ cpp_unittest_wrapper(name="db_flush_test",
            extra_compiler_flags=[])


+cpp_unittest_wrapper(name="db_follower_test",
+            srcs=["db/db_follower_test.cc"],
+            deps=[":rocksdb_test_lib"],
+            extra_compiler_flags=[])
+
+
 cpp_unittest_wrapper(name="db_inplace_update_test",
            srcs=["db/db_inplace_update_test.cc"],
            deps=[":rocksdb_test_lib"],
@@ -5527,6 +5548,12 @@ cpp_unittest_wrapper(name="ttl_test",
            extra_compiler_flags=[])


+cpp_unittest_wrapper(name="types_util_test",
+            srcs=["utilities/types_util_test.cc"],
+            deps=[":rocksdb_test_lib"],
+            extra_compiler_flags=[])
+
+
 cpp_unittest_wrapper(name="udt_util_test",
            srcs=["util/udt_util_test.cc"],
            deps=[":rocksdb_test_lib"],
@@ -193,6 +193,10 @@ def generate_targets(repo_path, deps_map):
        + src_mk.get("STRESS_LIB_SOURCES", [])
        + ["test_util/testutil.cc"],
    )
+    # ldb binary
+    TARGETS.add_binary(
+        "ldb", ["tools/ldb.cc"], [":rocksdb_tools_lib"]
+    )
    # db_stress binary
    TARGETS.add_binary(
        "db_stress", ["db_stress_tool/db_stress.cc"], [":rocksdb_stress_lib"]
@@ -7,6 +7,7 @@ rocksdb_target_header_template = """# This file \100generated by:
 # This file is a Facebook-specific integration for buck builds, so can
 # only be validated by Facebook employees.
 load("//rocks/buckifier:defs.bzl", "cpp_library_wrapper","rocks_cpp_library_wrapper","cpp_binary_wrapper","cpp_unittest_wrapper","fancy_bench_wrapper","add_c_test_wrapper")
+load("@fbcode_macros//build_defs:export_files.bzl", "export_file")

 """

@@ -163,24 +163,6 @@ case "$TARGET_OS" in
            PLATFORM_LDFLAGS="$PLATFORM_LDFLAGS -latomic"
        fi
        PLATFORM_LDFLAGS="$PLATFORM_LDFLAGS -lpthread -lrt -ldl"
-        if test -z "$ROCKSDB_USE_IO_URING"; then
-            ROCKSDB_USE_IO_URING=1
-        fi
-        if test "$ROCKSDB_USE_IO_URING" -ne 0; then
-            # check for liburing
-            $CXX $PLATFORM_CXXFLAGS -x c++ - -luring -o test.o 2>/dev/null  <<EOF
-              #include <liburing.h>
-              int main() {
-                struct io_uring ring;
-                io_uring_queue_init(1, &ring, 0);
-                return 0;
-              }
-EOF
-            if [ "$?" = 0 ]; then
-                PLATFORM_LDFLAGS="$PLATFORM_LDFLAGS -luring"
-                COMMON_FLAGS="$COMMON_FLAGS -DROCKSDB_IOURING_PRESENT"
-            fi
-        fi
        # PORT_FILES=port/linux/linux_specific.cc
        ;;
    SunOS)
@@ -603,7 +585,7 @@ EOF
            PLATFORM_LDFLAGS="$PLATFORM_LDFLAGS -lbenchmark"
        fi
    fi
-    if test $USE_FOLLY; then
+    if test $USE_FOLLY || test $USE_FOLLY_LITE; then
        # Test whether libfolly library is installed
        $CXX $PLATFORM_CXXFLAGS $COMMON_FLAGS -x c++ - -o /dev/null 2>/dev/null  <<EOF
          #include <folly/synchronization/DistributedMutex.h>
@@ -614,6 +596,24 @@ EOF
        fi
    fi

+    if test -z "$ROCKSDB_USE_IO_URING"; then
+        ROCKSDB_USE_IO_URING=1
+    fi
+    if [ "$ROCKSDB_USE_IO_URING" -ne 0 -a "$PLATFORM" = OS_LINUX ]; then
+        # check for liburing
+        $CXX $PLATFORM_CXXFLAGS -x c++ - -luring -o test.o 2>/dev/null  <<EOF
+          #include <liburing.h>
+          int main() {
+            struct io_uring ring;
+            io_uring_queue_init(1, &ring, 0);
+            return 0;
+          }
+EOF
+        if [ "$?" = 0 ]; then
+            PLATFORM_LDFLAGS="$PLATFORM_LDFLAGS -luring"
+            COMMON_FLAGS="$COMMON_FLAGS -DROCKSDB_IOURING_PRESENT"
+        fi
+    fi
 fi

 # TODO(tec): Fix -Wshorten-64-to-32 errors on FreeBSD and enable the warning.
@@ -751,6 +751,11 @@ if [ "$USE_FOLLY" ]; then
    FOLLY_PATH=`cd $FOLLY_DIR && $PYTHON build/fbcode_builder/getdeps.py show-inst-dir folly`
  fi
 fi
+if [ "$USE_FOLLY_LITE" ]; then
+  if [ "$FOLLY_DIR" ]; then
+    BOOST_SOURCE_PATH=`cd $FOLLY_DIR && $PYTHON build/fbcode_builder/getdeps.py show-source-dir boost`
+  fi
+fi

 PLATFORM_CCFLAGS="$PLATFORM_CCFLAGS $COMMON_FLAGS"
 PLATFORM_CXXFLAGS="$PLATFORM_CXXFLAGS $COMMON_FLAGS"
@@ -792,6 +797,7 @@ echo "PROFILING_FLAGS=$PROFILING_FLAGS" >> "$OUTPUT"
 echo "FIND=$FIND" >> "$OUTPUT"
 echo "WATCH=$WATCH" >> "$OUTPUT"
 echo "FOLLY_PATH=$FOLLY_PATH" >> "$OUTPUT"
+echo "BOOST_SOURCE_PATH=$BOOST_SOURCE_PATH" >> "$OUTPUT"

 # This will enable some related identifiers for the preprocessor
 if test -n "$JEMALLOC"; then
@@ -113,7 +113,7 @@ CLANG_LIB="$CLANG_BASE/lib"
 CLANG_SRC="$CLANG_BASE/../../src"

 CLANG_ANALYZER="$CLANG_BIN/clang++"
-CLANG_SCAN_BUILD="$CLANG_SRC/llvm/tools/clang/tools/scan-build/bin/scan-build"
+CLANG_SCAN_BUILD="$CLANG_BIN/scan-build

 if [ -z "$USE_CLANG" ]; then
  # gcc
@@ -110,7 +110,7 @@ CLANG_LIB="$CLANG_BASE/lib"
 CLANG_SRC="$CLANG_BASE/../../src"

 CLANG_ANALYZER="$CLANG_BIN/clang++"
-CLANG_SCAN_BUILD="$CLANG_SRC/llvm/clang/tools/scan-build/bin/scan-build"
+CLANG_SCAN_BUILD="$CLANG_BIN/scan-build"

 if [ -z "$USE_CLANG" ]; then
  # gcc
@@ -129,7 +129,6 @@ TEST_F(CacheReservationManagerTest,

 TEST(CacheReservationManagerIncreaseReservcationOnFullCacheTest,
     IncreaseCacheReservationOnFullCache) {
-  ;
  constexpr std::size_t kSizeDummyEntry =
      CacheReservationManagerImpl<CacheEntryRole::kMisc>::GetDummyEntrySize();
  constexpr std::size_t kSmallCacheCapacity = 4 * kSizeDummyEntry;
@@ -18,6 +18,7 @@
 #include "cache/lru_cache.h"
 #include "cache/typed_cache.h"
 #include "port/stack_trace.h"
+#include "table/block_based/block_cache.h"
 #include "test_util/secondary_cache_test_util.h"
 #include "test_util/testharness.h"
 #include "util/coding.h"
@@ -1017,6 +1018,63 @@ INSTANTIATE_TEST_CASE_P(CacheTestInstance, CacheTest,
 INSTANTIATE_TEST_CASE_P(CacheTestInstance, LRUCacheTest,
                        testing::Values(secondary_cache_test_util::kLRU));

+TEST(MiscBlockCacheTest, UncacheAggressivenessAdvisor) {
+  // Aggressiveness to a sequence of Report() calls (as string of 0s and 1s)
+  // exactly until the first ShouldContinue() == false.
+  const std::vector<std::pair<uint32_t, Slice>> expectedTraces{
+      // Aggressiveness 1 aborts on first unsuccessful erasure.
+      {1, "0"},
+      {1, "11111111111111111111110"},
+      // For sufficient evidence, aggressiveness 2 requires a minimum of two
+      // unsuccessful erasures.
+      {2, "00"},
+      {2, "0110"},
+      {2, "1100"},
+      {2, "011111111111111111111111111111111111111111111111111111111111111100"},
+      {2, "0111111111111111111111111111111111110"},
+      // For sufficient evidence, aggressiveness 3 and higher require a minimum
+      // of three unsuccessful erasures.
+      {3, "000"},
+      {3, "01010"},
+      {3, "111000"},
+      {3, "00111111111111111111111111111111111100"},
+      {3, "00111111111111111111110"},
+
+      {4, "000"},
+      {4, "01010"},
+      {4, "111000"},
+      {4, "001111111111111111111100"},
+      {4, "0011111111111110"},
+
+      {6, "000"},
+      {6, "01010"},
+      {6, "111000"},
+      {6, "00111111111111100"},
+      {6, "0011111110"},
+
+      // 69 -> 50% threshold, now up to minimum of 4
+      {69, "0000"},
+      {69, "010000"},
+      {69, "01010000"},
+      {69, "101010100010101000"},
+
+      // 230 -> 10% threshold, appropriately higher minimum
+      {230, "000000000000"},
+      {230, "0000000000010000000000"},
+      {230, "00000000000100000000010000000000"}};
+  for (const auto& [aggressiveness, t] : expectedTraces) {
+    SCOPED_TRACE("aggressiveness=" + std::to_string(aggressiveness) + " with " +
+                 t.ToString());
+    UncacheAggressivenessAdvisor uaa(aggressiveness);
+    for (size_t i = 0; i < t.size(); ++i) {
+      SCOPED_TRACE("i=" + std::to_string(i));
+      ASSERT_TRUE(uaa.ShouldContinue());
+      uaa.Report(t[i] & 1);
+    }
+    ASSERT_FALSE(uaa.ShouldContinue());
+  }
+}
+
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
@@ -134,12 +134,14 @@ bool CacheWithSecondaryAdapter::EvictionHandler(const Slice& key,
    auto obj = target_->Value(handle);
    // Ignore dummy entry
    if (obj != kDummyObj) {
-      bool hit = false;
+      bool force = false;
      if (adm_policy_ == TieredAdmissionPolicy::kAdmPolicyAllowCacheHits) {
-        hit = was_hit;
+        force = was_hit;
+      } else if (adm_policy_ == TieredAdmissionPolicy::kAdmPolicyAllowAll) {
+        force = true;
      }
      // Spill into secondary cache.
-      secondary_cache_->Insert(key, obj, helper, hit).PermitUncheckedError();
+      secondary_cache_->Insert(key, obj, helper, force).PermitUncheckedError();
    }
  }
  // Never takes ownership of obj
@@ -661,6 +663,7 @@ std::shared_ptr<Cache> NewTieredCache(const TieredCacheOptions& _opts) {
        break;
      case TieredAdmissionPolicy::kAdmPolicyPlaceholder:
      case TieredAdmissionPolicy::kAdmPolicyAllowCacheHits:
+      case TieredAdmissionPolicy::kAdmPolicyAllowAll:
        if (opts.nvm_sec_cache) {
          valid_adm_policy = false;
        }
@@ -816,11 +816,174 @@ TEST_P(DBTieredAdmPolicyTest, CompressedOnlyTest) {
  Destroy(options);
 }

+TEST_P(DBTieredAdmPolicyTest, CompressedCacheAdmission) {
+  if (!LZ4_Supported()) {
+    ROCKSDB_GTEST_SKIP("This test requires LZ4 support.");
+    return;
+  }
+
+  BlockBasedTableOptions table_options;
+  // We want a block cache of size 5KB, and a compressed secondary cache of
+  // size 5KB. However, we specify a block cache size of 256KB here in order
+  // to take into account the cache reservation in the block cache on
+  // behalf of the compressed cache. The unit of cache reservation is 256KB.
+  // The effective block cache capacity will be calculated as 256 + 5 = 261KB,
+  // and 256KB will be reserved for the compressed cache, leaving 10KB for
+  // the primary block cache. We only have to worry about this here because
+  // the cache size is so small.
+  table_options.block_cache = NewCache(256 * 1024, 5 * 1024, 0, GetParam());
+  table_options.block_size = 4 * 1024;
+  table_options.cache_index_and_filter_blocks = false;
+  Options options = GetDefaultOptions();
+  options.create_if_missing = true;
+  options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+
+  size_t comp_cache_usage = compressed_secondary_cache()->TEST_GetUsage();
+  // Disable paranoid_file_checks so that flush will not read back the newly
+  // written file
+  options.paranoid_file_checks = false;
+  DestroyAndReopen(options);
+  Random rnd(301);
+  const int N = 256;
+  for (int i = 0; i < N; i++) {
+    std::string p_v;
+    test::CompressibleString(&rnd, 0.5, 1007, &p_v);
+    ASSERT_OK(Put(Key(i), p_v));
+  }
+
+  ASSERT_OK(Flush());
+
+  // The second Get (for 5) will evict the data block loaded by the first
+  // Get, which will be admitted into the compressed secondary cache only
+  // for the kAdmPolicyAllowAll policy
+  std::string v = Get(Key(0));
+  ASSERT_EQ(1007, v.size());
+
+  v = Get(Key(5));
+  ASSERT_EQ(1007, v.size());
+
+  if (GetParam() == TieredAdmissionPolicy::kAdmPolicyAllowAll) {
+    ASSERT_GT(compressed_secondary_cache()->TEST_GetUsage(),
+              comp_cache_usage + 128);
+  } else {
+    ASSERT_LT(compressed_secondary_cache()->TEST_GetUsage(),
+              comp_cache_usage + 128);
+  }
+
+  Destroy(options);
+}
+
+TEST_F(DBTieredSecondaryCacheTest, FSBufferTest) {
+  class WrapFS : public FileSystemWrapper {
+   public:
+    explicit WrapFS(const std::shared_ptr<FileSystem>& _target)
+        : FileSystemWrapper(_target) {}
+    ~WrapFS() override {}
+    const char* Name() const override { return "WrapFS"; }
+
+    IOStatus NewRandomAccessFile(const std::string& fname,
+                                 const FileOptions& opts,
+                                 std::unique_ptr<FSRandomAccessFile>* result,
+                                 IODebugContext* dbg) override {
+      class WrappedRandomAccessFile : public FSRandomAccessFileOwnerWrapper {
+       public:
+        explicit WrappedRandomAccessFile(
+            std::unique_ptr<FSRandomAccessFile>& file)
+            : FSRandomAccessFileOwnerWrapper(std::move(file)) {}
+
+        IOStatus MultiRead(FSReadRequest* reqs, size_t num_reqs,
+                           const IOOptions& options,
+                           IODebugContext* dbg) override {
+          for (size_t i = 0; i < num_reqs; ++i) {
+            FSReadRequest& req = reqs[i];
+            FSAllocationPtr buffer(new char[req.len], [](void* ptr) {
+              delete[] static_cast<char*>(ptr);
+            });
+            req.fs_scratch = std::move(buffer);
+            req.status = Read(req.offset, req.len, options, &req.result,
+                              static_cast<char*>(req.fs_scratch.get()), dbg);
+          }
+          return IOStatus::OK();
+        }
+      };
+
+      std::unique_ptr<FSRandomAccessFile> file;
+      IOStatus s = target()->NewRandomAccessFile(fname, opts, &file, dbg);
+      EXPECT_OK(s);
+      result->reset(new WrappedRandomAccessFile(file));
+
+      return s;
+    }
+
+    void SupportedOps(int64_t& supported_ops) override {
+      supported_ops = 1 << FSSupportedOps::kAsyncIO;
+      supported_ops |= 1 << FSSupportedOps::kFSBuffer;
+    }
+  };
+
+  if (!LZ4_Supported()) {
+    ROCKSDB_GTEST_SKIP("This test requires LZ4 support.");
+    return;
+  }
+
+  std::shared_ptr<WrapFS> wrap_fs =
+      std::make_shared<WrapFS>(env_->GetFileSystem());
+  std::unique_ptr<Env> wrap_env(new CompositeEnvWrapper(env_, wrap_fs));
+  BlockBasedTableOptions table_options;
+  table_options.block_cache = NewCache(250 * 1024, 20 * 1024, 256 * 1024,
+                                       TieredAdmissionPolicy::kAdmPolicyAuto,
+                                       /*ready_before_wait=*/true);
+  table_options.block_size = 4 * 1024;
+  table_options.cache_index_and_filter_blocks = false;
+  Options options = GetDefaultOptions();
+  options.create_if_missing = true;
+  options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+  options.statistics = CreateDBStatistics();
+  options.env = wrap_env.get();
+
+  options.paranoid_file_checks = false;
+  DestroyAndReopen(options);
+  Random rnd(301);
+  const int N = 256;
+  for (int i = 0; i < N; i++) {
+    std::string p_v;
+    test::CompressibleString(&rnd, 0.5, 1007, &p_v);
+    ASSERT_OK(Put(Key(i), p_v));
+  }
+
+  ASSERT_OK(Flush());
+
+  std::vector<std::string> keys;
+  std::vector<std::string> values;
+
+  keys.push_back(Key(0));
+  keys.push_back(Key(4));
+  keys.push_back(Key(8));
+  values = MultiGet(keys, /*snapshot=*/nullptr, /*async=*/true);
+  ASSERT_EQ(values.size(), keys.size());
+  for (const auto& value : values) {
+    ASSERT_EQ(1007, value.size());
+  }
+  ASSERT_EQ(nvm_sec_cache()->num_insert_saved(), 3u);
+  ASSERT_EQ(nvm_sec_cache()->num_misses(), 3u);
+  ASSERT_EQ(nvm_sec_cache()->num_hits(), 0u);
+
+  std::string v = Get(Key(12));
+  ASSERT_EQ(1007, v.size());
+  ASSERT_EQ(nvm_sec_cache()->num_insert_saved(), 4u);
+  ASSERT_EQ(nvm_sec_cache()->num_misses(), 4u);
+  ASSERT_EQ(options.statistics->getTickerCount(BLOCK_CACHE_MISS), 4u);
+
+  Close();
+  Destroy(options);
+}
+
 INSTANTIATE_TEST_CASE_P(
    DBTieredAdmPolicyTest, DBTieredAdmPolicyTest,
    ::testing::Values(TieredAdmissionPolicy::kAdmPolicyAuto,
                      TieredAdmissionPolicy::kAdmPolicyPlaceholder,
-                      TieredAdmissionPolicy::kAdmPolicyAllowCacheHits));
+                      TieredAdmissionPolicy::kAdmPolicyAllowCacheHits,
+                      TieredAdmissionPolicy::kAdmPolicyAllowAll));

 }  // namespace ROCKSDB_NAMESPACE

@@ -0,0 +1,20 @@
+//  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/attribute_group_iterator_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+const AttributeGroups kNoAttributeGroups;
+const IteratorAttributeGroups kNoIteratorAttributeGroups;
+
+void AttributeGroupIteratorImpl::AddToAttributeGroups(
+    const autovector<MultiCfIteratorInfo>& items) {
+  for (const auto& item : items) {
+    attribute_groups_.emplace_back(item.cfh, &item.iterator->columns());
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
@@ -0,0 +1,83 @@
+//  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).
+
+#pragma once
+
+#include "db/multi_cf_iterator_impl.h"
+#include "rocksdb/attribute_groups.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class AttributeGroupIteratorImpl : public AttributeGroupIterator {
+ public:
+  AttributeGroupIteratorImpl(
+      const Comparator* comparator,
+      const std::vector<ColumnFamilyHandle*>& column_families,
+      const std::vector<Iterator*>& child_iterators)
+      : impl_(
+            comparator, column_families, child_iterators, [this]() { Reset(); },
+            [this](const autovector<MultiCfIteratorInfo>& items) {
+              AddToAttributeGroups(items);
+            }) {}
+  ~AttributeGroupIteratorImpl() override {}
+
+  // No copy allowed
+  AttributeGroupIteratorImpl(const AttributeGroupIteratorImpl&) = delete;
+  AttributeGroupIteratorImpl& operator=(const AttributeGroupIteratorImpl&) =
+      delete;
+
+  bool Valid() const override { return impl_.Valid(); }
+  void SeekToFirst() override { impl_.SeekToFirst(); }
+  void SeekToLast() override { impl_.SeekToLast(); }
+  void Seek(const Slice& target) override { impl_.Seek(target); }
+  void SeekForPrev(const Slice& target) override { impl_.SeekForPrev(target); }
+  void Next() override { impl_.Next(); }
+  void Prev() override { impl_.Prev(); }
+  Slice key() const override { return impl_.key(); }
+  Status status() const override { return impl_.status(); }
+
+  const IteratorAttributeGroups& attribute_groups() const override {
+    assert(Valid());
+    return attribute_groups_;
+  }
+
+  void Reset() { attribute_groups_.clear(); }
+
+ private:
+  MultiCfIteratorImpl impl_;
+  IteratorAttributeGroups attribute_groups_;
+  void AddToAttributeGroups(const autovector<MultiCfIteratorInfo>& items);
+};
+
+class EmptyAttributeGroupIterator : public AttributeGroupIterator {
+ public:
+  explicit EmptyAttributeGroupIterator(const Status& s) : status_(s) {}
+  bool Valid() const override { return false; }
+  void Seek(const Slice& /*target*/) override {}
+  void SeekForPrev(const Slice& /*target*/) override {}
+  void SeekToFirst() override {}
+  void SeekToLast() override {}
+  void Next() override { assert(false); }
+  void Prev() override { assert(false); }
+  Slice key() const override {
+    assert(false);
+    return Slice();
+  }
+  Status status() const override { return status_; }
+
+  const IteratorAttributeGroups& attribute_groups() const override {
+    return kNoIteratorAttributeGroups;
+  }
+
+ private:
+  Status status_;
+};
+
+inline std::unique_ptr<AttributeGroupIterator> NewAttributeGroupErrorIterator(
+    const Status& status) {
+  return std::make_unique<EmptyAttributeGroupIterator>(status);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
@@ -1551,6 +1551,57 @@ TEST_P(DBBlobBasicIOErrorMultiGetTest, MultipleBlobFiles) {
  ASSERT_TRUE(statuses[1].IsIOError());
 }

+TEST_F(DBBlobBasicTest, MultiGetFindTable_IOError) {
+  // Repro test for a specific bug where `MultiGet()` would fail to open a table
+  // in `FindTable()` and then proceed to return raw blob handles for the other
+  // keys.
+  Options options = GetDefaultOptions();
+  options.enable_blob_files = true;
+  options.min_blob_size = 0;
+
+  Reopen(options);
+
+  // Force no table cache so every read will preload the SST file.
+  dbfull()->TEST_table_cache()->SetCapacity(0);
+
+  constexpr size_t num_keys = 2;
+
+  constexpr char key1[] = "key1";
+  constexpr char value1[] = "blob1";
+
+  ASSERT_OK(Put(key1, value1));
+  ASSERT_OK(Flush());
+
+  constexpr char key2[] = "key2";
+  constexpr char value2[] = "blob2";
+
+  ASSERT_OK(Put(key2, value2));
+  ASSERT_OK(Flush());
+
+  std::atomic<int> num_files_opened = 0;
+  // This test would be more realistic if we injected an `IOError` from the
+  // `FileSystem`
+  SyncPoint::GetInstance()->SetCallBack(
+      "TableCache::MultiGet:FindTable", [&](void* status) {
+        num_files_opened++;
+        if (num_files_opened == 2) {
+          Status* s = static_cast<Status*>(status);
+          *s = Status::IOError();
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  std::array<Slice, num_keys> keys{{key1, key2}};
+  std::array<PinnableSlice, num_keys> values;
+  std::array<Status, num_keys> statuses;
+  db_->MultiGet(ReadOptions(), db_->DefaultColumnFamily(), num_keys,
+                keys.data(), values.data(), statuses.data());
+
+  ASSERT_TRUE(statuses[0].IsIOError());
+  ASSERT_OK(statuses[1]);
+  ASSERT_EQ(value2, values[1]);
+}
+
 namespace {

 class ReadBlobCompactionFilter : public CompactionFilter {
@@ -64,7 +64,7 @@ Status BuildTable(
    std::vector<std::unique_ptr<FragmentedRangeTombstoneIterator>>
        range_del_iters,
    FileMetaData* meta, std::vector<BlobFileAddition>* blob_file_additions,
-    std::vector<SequenceNumber> snapshots,
+    std::vector<SequenceNumber> snapshots, SequenceNumber earliest_snapshot,
    SequenceNumber earliest_write_conflict_snapshot,
    SequenceNumber job_snapshot, SnapshotChecker* snapshot_checker,
    bool paranoid_file_checks, InternalStats* internal_stats,
@@ -195,7 +195,7 @@ Status BuildTable(

    const std::atomic<bool> kManualCompactionCanceledFalse{false};
    CompactionIterator c_iter(
-        iter, ucmp, &merge, kMaxSequenceNumber, &snapshots,
+        iter, ucmp, &merge, kMaxSequenceNumber, &snapshots, earliest_snapshot,
        earliest_write_conflict_snapshot, job_snapshot, snapshot_checker, env,
        ShouldReportDetailedTime(env, ioptions.stats),
        true /* internal key corruption is not ok */, range_del_agg.get(),
@@ -210,6 +210,7 @@ Status BuildTable(
    const bool logical_strip_timestamp =
        ts_sz > 0 && !ioptions.persist_user_defined_timestamps;

+    SequenceNumber smallest_preferred_seqno = kMaxSequenceNumber;
    std::string key_after_flush_buf;
    std::string value_buf;
    c_iter.SeekToFirst();
@@ -242,6 +243,8 @@ Status BuildTable(
        if (preferred_seqno < ikey.sequence) {
          value_after_flush =
              PackValueAndSeqno(unpacked_value, preferred_seqno, &value_buf);
+          smallest_preferred_seqno =
+              std::min(smallest_preferred_seqno, preferred_seqno);
        } else {
          // Cannot get a useful preferred seqno, convert it to a kTypeValue.
          UpdateInternalKey(&key_after_flush_buf, ikey.sequence, kTypeValue);
@@ -326,9 +329,10 @@ Status BuildTable(
    } else {
      SeqnoToTimeMapping relevant_mapping;
      if (seqno_to_time_mapping) {
-        relevant_mapping.CopyFromSeqnoRange(*seqno_to_time_mapping,
-                                            meta->fd.smallest_seqno,
-                                            meta->fd.largest_seqno);
+        relevant_mapping.CopyFromSeqnoRange(
+            *seqno_to_time_mapping,
+            std::min(meta->fd.smallest_seqno, smallest_preferred_seqno),
+            meta->fd.largest_seqno);
        relevant_mapping.SetCapacity(kMaxSeqnoTimePairsPerSST);
        relevant_mapping.Enforce(tboptions.file_creation_time);
      }
@@ -57,7 +57,7 @@ Status BuildTable(
    std::vector<std::unique_ptr<FragmentedRangeTombstoneIterator>>
        range_del_iters,
    FileMetaData* meta, std::vector<BlobFileAddition>* blob_file_additions,
-    std::vector<SequenceNumber> snapshots,
+    std::vector<SequenceNumber> snapshots, SequenceNumber earliest_snapshot,
    SequenceNumber earliest_write_conflict_snapshot,
    SequenceNumber job_snapshot, SnapshotChecker* snapshot_checker,
    bool paranoid_file_checks, InternalStats* internal_stats,
@@ -231,6 +231,7 @@ struct rocksdb_livefiles_t {
 };
 struct rocksdb_column_family_handle_t {
  ColumnFamilyHandle* rep;
+  bool immortal;  /* only true for default cf */
 };
 struct rocksdb_column_family_metadata_t {
  ColumnFamilyMetaData rep;
@@ -546,6 +547,7 @@ static bool SaveError(char** errptr, const Status& s) {
  return true;
 }

+// Copies str to a new malloc()-ed buffer. The buffer is not NUL terminated.
 static char* CopyString(const std::string& str) {
  char* result = reinterpret_cast<char*>(malloc(sizeof(char) * str.size()));
  memcpy(result, str.data(), sizeof(char) * str.size());
@@ -906,6 +908,7 @@ rocksdb_t* rocksdb_open_and_trim_history(
    rocksdb_column_family_handle_t* c_handle =
        new rocksdb_column_family_handle_t;
    c_handle->rep = handles[i];
+    c_handle->immortal = false;
    column_family_handles[i] = c_handle;
  }
  rocksdb_t* result = new rocksdb_t;
@@ -936,6 +939,7 @@ rocksdb_t* rocksdb_open_column_families(
    rocksdb_column_family_handle_t* c_handle =
        new rocksdb_column_family_handle_t;
    c_handle->rep = handles[i];
+    c_handle->immortal = false;
    column_family_handles[i] = c_handle;
  }
  rocksdb_t* result = new rocksdb_t;
@@ -971,6 +975,7 @@ rocksdb_t* rocksdb_open_column_families_with_ttl(
    rocksdb_column_family_handle_t* c_handle =
        new rocksdb_column_family_handle_t;
    c_handle->rep = handles[i];
+    c_handle->immortal = false;
    column_family_handles[i] = c_handle;
  }
  rocksdb_t* result = new rocksdb_t;
@@ -1004,6 +1009,7 @@ rocksdb_t* rocksdb_open_for_read_only_column_families(
    rocksdb_column_family_handle_t* c_handle =
        new rocksdb_column_family_handle_t;
    c_handle->rep = handles[i];
+    c_handle->immortal = false;
    column_family_handles[i] = c_handle;
  }
  rocksdb_t* result = new rocksdb_t;
@@ -1035,6 +1041,7 @@ rocksdb_t* rocksdb_open_as_secondary_column_families(
    rocksdb_column_family_handle_t* c_handle =
        new rocksdb_column_family_handle_t;
    c_handle->rep = handles[i];
+    c_handle->immortal = false;
    column_family_handles[i] = c_handle;
  }
  rocksdb_t* result = new rocksdb_t;
@@ -1072,6 +1079,7 @@ rocksdb_column_family_handle_t* rocksdb_create_column_family(
  SaveError(errptr, db->rep->CreateColumnFamily(
                        ColumnFamilyOptions(column_family_options->rep),
                        std::string(column_family_name), &(handle->rep)));
+  handle->immortal = false;
  return handle;
 }

@@ -1095,6 +1103,7 @@ rocksdb_column_family_handle_t** rocksdb_create_column_families(
  for (size_t i = 0; i != handles.size(); ++i) {
    c_handles[i] = new rocksdb_column_family_handle_t;
    c_handles[i]->rep = handles[i];
+    c_handles[i]->immortal = false;
  }

  return c_handles;
@@ -1114,6 +1123,7 @@ rocksdb_column_family_handle_t* rocksdb_create_column_family_with_ttl(
  SaveError(errptr, db_with_ttl->CreateColumnFamilyWithTtl(
                        ColumnFamilyOptions(column_family_options->rep),
                        std::string(column_family_name), &(handle->rep), ttl));
+  handle->immortal = false;
  return handle;
 }

@@ -1135,9 +1145,19 @@ char* rocksdb_column_family_handle_get_name(
  return CopyString(name);
 }

+rocksdb_column_family_handle_t* rocksdb_get_default_column_family_handle(
+    rocksdb_t* db) {
+  rocksdb_column_family_handle_t* handle = new rocksdb_column_family_handle_t;
+  handle->rep = db->rep->DefaultColumnFamily();
+  handle->immortal = true;
+  return handle;
+}
+
 void rocksdb_column_family_handle_destroy(
    rocksdb_column_family_handle_t* handle) {
-  delete handle->rep;
+  if (!handle->immortal) {
+    delete handle->rep;
+  }
  delete handle;
 }

@@ -1373,6 +1393,18 @@ char* rocksdb_get_cf_with_ts(rocksdb_t* db,
  return result;
 }

+char* rocksdb_get_db_identity(rocksdb_t* db, size_t* id_len) {
+  std::string identity_tmp;
+  Status s = db->rep->GetDbIdentity(identity_tmp);
+  if (!s.ok()) {
+    *id_len = 0;
+    return nullptr;
+  }
+
+  *id_len = identity_tmp.size();
+  return CopyString(identity_tmp);
+}
+
 void rocksdb_multi_get(rocksdb_t* db, const rocksdb_readoptions_t* options,
                       size_t num_keys, const char* const* keys_list,
                       const size_t* keys_list_sizes, char** values_list,
@@ -1947,6 +1979,10 @@ void rocksdb_iter_get_error(const rocksdb_iterator_t* iter, char** errptr) {
  SaveError(errptr, iter->rep->status());
 }

+void rocksdb_iter_refresh(const rocksdb_iterator_t* iter, char** errptr) {
+  SaveError(errptr, iter->rep->Refresh());
+}
+
 rocksdb_writebatch_t* rocksdb_writebatch_create() {
  return new rocksdb_writebatch_t;
 }
@@ -1958,6 +1994,15 @@ rocksdb_writebatch_t* rocksdb_writebatch_create_from(const char* rep,
  return b;
 }

+rocksdb_writebatch_t* rocksdb_writebatch_create_with_params(
+    size_t reserved_bytes, size_t max_bytes, size_t protection_bytes_per_key,
+    size_t default_cf_ts_sz) {
+  rocksdb_writebatch_t* b = new rocksdb_writebatch_t;
+  b->rep = WriteBatch(reserved_bytes, max_bytes, protection_bytes_per_key,
+                      default_cf_ts_sz);
+  return b;
+}
+
 void rocksdb_writebatch_destroy(rocksdb_writebatch_t* b) { delete b; }

 void rocksdb_writebatch_clear(rocksdb_writebatch_t* b) { b->rep.Clear(); }
@@ -2223,6 +2268,35 @@ rocksdb_writebatch_wi_t* rocksdb_writebatch_wi_create(
  return b;
 }

+rocksdb_writebatch_wi_t* rocksdb_writebatch_wi_create_with_params(
+    rocksdb_comparator_t* backup_index_comparator, size_t reserved_bytes,
+    unsigned char overwrite_key, size_t max_bytes,
+    size_t protection_bytes_per_key) {
+  rocksdb_writebatch_wi_t* b = new rocksdb_writebatch_wi_t;
+  b->rep = new WriteBatchWithIndex(backup_index_comparator, reserved_bytes,
+                                   overwrite_key, max_bytes,
+                                   protection_bytes_per_key);
+  return b;
+}
+
+void rocksdb_writebatch_update_timestamps(
+    rocksdb_writebatch_t* wb, const char* ts, size_t tslen, void* state,
+    size_t (*get_ts_size)(void*, uint32_t), char** errptr) {
+  SaveError(errptr, wb->rep.UpdateTimestamps(
+                        Slice(ts, tslen), [&get_ts_size, &state](uint32_t cf) {
+                          return (*get_ts_size)(state, cf);
+                        }));
+}
+
+void rocksdb_writebatch_wi_update_timestamps(
+    rocksdb_writebatch_wi_t* wb, const char* ts, size_t tslen, void* state,
+    size_t (*get_ts_size)(void*, uint32_t), char** errptr) {
+  SaveError(errptr, wb->rep->GetWriteBatch()->UpdateTimestamps(
+                        Slice(ts, tslen), [&get_ts_size, &state](uint32_t cf) {
+                          return (*get_ts_size)(state, cf);
+                        }));
+}
+
 void rocksdb_writebatch_wi_destroy(rocksdb_writebatch_wi_t* b) {
  if (b->rep) {
    delete b->rep;
@@ -2745,6 +2819,24 @@ void rocksdb_options_set_block_based_table_factory(
  }
 }

+void rocksdb_block_based_options_set_top_level_index_pinning_tier(
+    rocksdb_block_based_table_options_t* options, int v) {
+  options->rep.metadata_cache_options.top_level_index_pinning =
+      static_cast<ROCKSDB_NAMESPACE::PinningTier>(v);
+}
+
+void rocksdb_block_based_options_set_partition_pinning_tier(
+    rocksdb_block_based_table_options_t* options, int v) {
+  options->rep.metadata_cache_options.partition_pinning =
+      static_cast<ROCKSDB_NAMESPACE::PinningTier>(v);
+}
+
+void rocksdb_block_based_options_set_unpartitioned_pinning_tier(
+    rocksdb_block_based_table_options_t* options, int v) {
+  options->rep.metadata_cache_options.unpartitioned_pinning =
+      static_cast<ROCKSDB_NAMESPACE::PinningTier>(v);
+}
+
 rocksdb_cuckoo_table_options_t* rocksdb_cuckoo_options_create() {
  return new rocksdb_cuckoo_table_options_t;
 }
@@ -3867,6 +3959,27 @@ void rocksdb_options_set_plain_table_factory(
  opt->rep.table_factory.reset(factory);
 }

+unsigned char rocksdb_options_get_write_dbid_to_manifest(
+    rocksdb_options_t* opt) {
+  return opt->rep.write_dbid_to_manifest;
+}
+
+void rocksdb_options_set_write_dbid_to_manifest(
+    rocksdb_options_t* opt, unsigned char write_dbid_to_manifest) {
+  opt->rep.write_dbid_to_manifest = write_dbid_to_manifest;
+}
+
+unsigned char rocksdb_options_get_track_and_verify_wals_in_manifest(
+    rocksdb_options_t* opt) {
+  return opt->rep.track_and_verify_wals_in_manifest;
+}
+
+void rocksdb_options_set_track_and_verify_wals_in_manifest(
+    rocksdb_options_t* opt, unsigned char track_and_verify_wals_in_manifest) {
+  opt->rep.track_and_verify_wals_in_manifest =
+      track_and_verify_wals_in_manifest;
+}
+
 void rocksdb_options_set_max_successive_merges(rocksdb_options_t* opt,
                                               size_t v) {
  opt->rep.max_successive_merges = v;
@@ -3930,6 +4043,14 @@ void rocksdb_options_set_fifo_compaction_options(
  opt->rep.compaction_options_fifo = fifo->rep;
 }

+void rocksdb_options_set_compaction_pri(rocksdb_options_t* opt, int pri) {
+  opt->rep.compaction_pri = static_cast<ROCKSDB_NAMESPACE::CompactionPri>(pri);
+}
+
+int rocksdb_options_get_compaction_pri(rocksdb_options_t* opt) {
+  return opt->rep.compaction_pri;
+}
+
 char* rocksdb_options_statistics_get_string(rocksdb_options_t* opt) {
  ROCKSDB_NAMESPACE::Statistics* statistics = opt->rep.statistics.get();
  if (statistics) {
@@ -4238,13 +4359,8 @@ void rocksdb_perfcontext_destroy(rocksdb_perfcontext_t* context) {

 /*
 TODO:
-DB::OpenForReadOnly
-DB::KeyMayExist
 DB::GetOptions
 DB::GetSortedWalFiles
-DB::GetLatestSequenceNumber
-DB::GetUpdatesSince
-DB::GetDbIdentity
 DB::RunManualCompaction
 custom cache
 table_properties_collectors
@@ -5721,6 +5837,7 @@ rocksdb_column_family_handle_t* rocksdb_transactiondb_create_column_family(
  SaveError(errptr, txn_db->rep->CreateColumnFamily(
                        ColumnFamilyOptions(column_family_options->rep),
                        std::string(column_family_name), &(handle->rep)));
+  handle->immortal = false;
  return handle;
 }

@@ -5763,6 +5880,7 @@ rocksdb_transactiondb_t* rocksdb_transactiondb_open_column_families(
    rocksdb_column_family_handle_t* c_handle =
        new rocksdb_column_family_handle_t;
    c_handle->rep = handles[i];
+    c_handle->immortal = false;
    column_family_handles[i] = c_handle;
  }
  rocksdb_transactiondb_t* result = new rocksdb_transactiondb_t;
@@ -6557,6 +6675,7 @@ rocksdb_optimistictransactiondb_open_column_families(
    rocksdb_column_family_handle_t* c_handle =
        new rocksdb_column_family_handle_t;
    c_handle->rep = handles[i];
+    c_handle->immortal = false;
    column_family_handles[i] = c_handle;
  }
  rocksdb_optimistictransactiondb_t* result =
@@ -704,6 +704,10 @@ 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_top_level_index_pinning_tier(table_options,
+                                                               1);
+  rocksdb_block_based_options_set_partition_pinning_tier(table_options, 2);
+  rocksdb_block_based_options_set_unpartitioned_pinning_tier(table_options, 3);
  rocksdb_options_set_block_based_table_factory(options, table_options);

  rocksdb_options_set_compression(options, rocksdb_no_compression);
@@ -794,6 +798,11 @@ int main(int argc, char** argv) {
    rocksdb_delete(db, woptions, "foo", 3, &err);
    CheckNoError(err);

+    // get the identity before the backup
+    size_t before_db_id_len = 0;
+    char* before_db_id = rocksdb_get_db_identity(db, &before_db_id_len);
+    CheckCondition(before_db_id_len == 36);
+
    rocksdb_close(db);

    rocksdb_destroy_db(options, dbname, &err);
@@ -814,6 +823,14 @@ int main(int argc, char** argv) {

    CheckGet(db, roptions, "foo", "hello");

+    // the db_identity after the backup is different
+    size_t after_db_id_len = 0;
+    char* after_db_id = rocksdb_get_db_identity(db, &after_db_id_len);
+    CheckCondition(after_db_id_len == 36);
+    CheckCondition(memcmp(after_db_id, before_db_id, after_db_id_len) != 0);
+    Free(&before_db_id);
+    Free(&after_db_id);
+
    rocksdb_backup_engine_close(be);
  }

@@ -829,6 +846,14 @@ int main(int argc, char** argv) {
    rocksdb_checkpoint_create(checkpoint, dbcheckpointname, 0, &err);
    CheckNoError(err);

+    rocksdb_checkpoint_object_destroy(checkpoint);
+    checkpoint = NULL;
+
+    // get the identity before the checkpoint
+    size_t before_db_id_len = 0;
+    char* before_db_id = rocksdb_get_db_identity(db, &before_db_id_len);
+    CheckCondition(before_db_id_len == 36);
+
    // start a new database from the checkpoint
    rocksdb_close(db);
    rocksdb_options_set_error_if_exists(options, 0);
@@ -837,7 +862,13 @@ int main(int argc, char** argv) {

    CheckGet(db, roptions, "foo", "hello");

-    rocksdb_checkpoint_object_destroy(checkpoint);
+    // the db_identity after the checkpoint is different
+    size_t after_db_id_len = 0;
+    char* after_db_id = rocksdb_get_db_identity(db, &after_db_id_len);
+    CheckCondition(after_db_id_len == 36);
+    CheckCondition(memcmp(after_db_id, before_db_id, after_db_id_len) != 0);
+    Free(&before_db_id);
+    Free(&after_db_id);

    rocksdb_close(db);
    rocksdb_destroy_db(options, dbcheckpointname, &err);
@@ -848,6 +879,73 @@ int main(int argc, char** argv) {
    rocksdb_options_set_error_if_exists(options, 1);
  }

+  StartPhase("checkpoint_db_id_in_manifest");
+  {
+    // create new DB with set_write_dbid_to_manifest=true
+    // db_identity is now the same across checkpoints
+    rocksdb_close(db);
+    db = NULL;
+
+    rocksdb_options_t* options_dbid_in_manifest = rocksdb_options_create();
+    rocksdb_options_set_create_if_missing(options_dbid_in_manifest, 1);
+
+    unsigned char write_to_manifest =
+        rocksdb_options_get_write_dbid_to_manifest(options_dbid_in_manifest);
+    CheckCondition(!write_to_manifest);
+    rocksdb_options_set_write_dbid_to_manifest(options_dbid_in_manifest, true);
+    CheckCondition(!write_to_manifest);
+    write_to_manifest =
+        rocksdb_options_get_write_dbid_to_manifest(options_dbid_in_manifest);
+    CheckCondition(write_to_manifest);
+
+    db = rocksdb_open(options_dbid_in_manifest, dbbackupname, &err);
+    CheckNoError(err);
+
+    rocksdb_checkpoint_t* checkpoint =
+        rocksdb_checkpoint_object_create(db, &err);
+    CheckNoError(err);
+
+    rocksdb_checkpoint_create(checkpoint, dbcheckpointname, 0, &err);
+    CheckNoError(err);
+
+    rocksdb_checkpoint_object_destroy(checkpoint);
+    checkpoint = NULL;
+
+    // get the identity before the backup
+    size_t before_db_id_len = 0;
+    char* before_db_id = rocksdb_get_db_identity(db, &before_db_id_len);
+    CheckCondition(before_db_id_len == 36);
+
+    // open the checkpoint
+    rocksdb_close(db);
+    rocksdb_destroy_db(options_dbid_in_manifest, dbbackupname, &err);
+    CheckNoError(err);
+
+    rocksdb_options_set_error_if_exists(options_dbid_in_manifest, 0);
+    db = rocksdb_open(options_dbid_in_manifest, dbcheckpointname, &err);
+    CheckNoError(err);
+
+    // the db_identity after the checkpoint is the same
+    size_t after_db_id_len = 0;
+    char* after_db_id = rocksdb_get_db_identity(db, &after_db_id_len);
+    CheckCondition(after_db_id_len == 36);
+    CheckCondition(memcmp(after_db_id, before_db_id, after_db_id_len) == 0);
+    Free(&before_db_id);
+    Free(&after_db_id);
+
+    rocksdb_close(db);
+    rocksdb_destroy_db(options_dbid_in_manifest, dbcheckpointname, &err);
+    CheckNoError(err);
+    rocksdb_options_destroy(options_dbid_in_manifest);
+    options_dbid_in_manifest = NULL;
+
+    // re-open the default database
+    rocksdb_options_set_error_if_exists(options, 0);
+    db = rocksdb_open(options, dbname, &err);
+    CheckNoError(err);
+    rocksdb_options_set_error_if_exists(options, 1);
+  }
+
  StartPhase("compactall");
  rocksdb_compact_range(db, NULL, 0, NULL, 0);
  CheckGet(db, roptions, "foo", "hello");
@@ -1318,6 +1416,8 @@ int main(int argc, char** argv) {
      policy = rocksdb_filterpolicy_create_ribbon_hybrid(8.0, 1);
    }
    rocksdb_block_based_options_set_filter_policy(table_options, policy);
+    rocksdb_block_based_options_set_optimize_filters_for_memory(table_options,
+                                                                0);

    // Create new database
    rocksdb_close(db);
@@ -2039,7 +2139,6 @@ int main(int argc, char** argv) {
    rocksdb_options_set_advise_random_on_open(o, 1);
    CheckCondition(1 == rocksdb_options_get_advise_random_on_open(o));

-
    rocksdb_options_set_use_adaptive_mutex(o, 1);
    CheckCondition(1 == rocksdb_options_get_use_adaptive_mutex(o));

@@ -2110,6 +2209,9 @@ int main(int argc, char** argv) {
    rocksdb_options_set_compaction_style(o, 2);
    CheckCondition(2 == rocksdb_options_get_compaction_style(o));

+    rocksdb_options_set_compaction_pri(o, 4);
+    CheckCondition(4 == rocksdb_options_get_compaction_pri(o));
+
    rocksdb_options_set_atomic_flush(o, 1);
    CheckCondition(1 == rocksdb_options_get_atomic_flush(o));

@@ -2132,6 +2234,12 @@ int main(int argc, char** argv) {
    CheckCondition(rocksdb_statistics_level_all ==
                   rocksdb_options_get_statistics_level(o));

+    CheckCondition(0 ==
+                   rocksdb_options_get_track_and_verify_wals_in_manifest(o));
+    rocksdb_options_set_track_and_verify_wals_in_manifest(o, 42);
+    CheckCondition(1 ==
+                   rocksdb_options_get_track_and_verify_wals_in_manifest(o));
+
    /* Blob Options */
    rocksdb_options_set_enable_blob_files(o, 1);
    CheckCondition(1 == rocksdb_options_get_enable_blob_files(o));
@@ -2614,6 +2722,10 @@ int main(int argc, char** argv) {
    CheckCondition(1 == rocksdb_options_get_compaction_style(copy));
    CheckCondition(2 == rocksdb_options_get_compaction_style(o));

+    rocksdb_options_set_compaction_pri(copy, 1);
+    CheckCondition(1 == rocksdb_options_get_compaction_pri(copy));
+    CheckCondition(4 == rocksdb_options_get_compaction_pri(o));
+
    rocksdb_options_set_atomic_flush(copy, 0);
    CheckCondition(0 == rocksdb_options_get_atomic_flush(copy));
    CheckCondition(1 == rocksdb_options_get_atomic_flush(o));
@@ -0,0 +1,47 @@
+//  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/coalescing_iterator.h"
+
+#include "db/wide/wide_columns_helper.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+void CoalescingIterator::Coalesce(
+    const autovector<MultiCfIteratorInfo>& items) {
+  assert(wide_columns_.empty());
+  MinHeap heap;
+  for (const auto& item : items) {
+    assert(item.iterator);
+    for (auto& column : item.iterator->columns()) {
+      heap.push(WideColumnWithOrder{&column, item.order});
+    }
+  }
+  if (heap.empty()) {
+    return;
+  }
+  wide_columns_.reserve(heap.size());
+  auto current = heap.top();
+  heap.pop();
+  while (!heap.empty()) {
+    int comparison = current.column->name().compare(heap.top().column->name());
+    if (comparison < 0) {
+      wide_columns_.push_back(*current.column);
+    } else if (comparison > 0) {
+      // Shouldn't reach here.
+      // Current item in the heap is greater than the top item in the min heap
+      assert(false);
+    }
+    current = heap.top();
+    heap.pop();
+  }
+  wide_columns_.push_back(*current.column);
+
+  if (WideColumnsHelper::HasDefaultColumn(wide_columns_)) {
+    value_ = WideColumnsHelper::GetDefaultColumn(wide_columns_);
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
@@ -0,0 +1,79 @@
+//  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).
+
+#pragma once
+
+#include "db/multi_cf_iterator_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// EXPERIMENTAL
+class CoalescingIterator : public Iterator {
+ public:
+  CoalescingIterator(const Comparator* comparator,
+                     const std::vector<ColumnFamilyHandle*>& column_families,
+                     const std::vector<Iterator*>& child_iterators)
+      : impl_(
+            comparator, column_families, child_iterators, [this]() { Reset(); },
+            [this](const autovector<MultiCfIteratorInfo>& items) {
+              Coalesce(items);
+            }) {}
+  ~CoalescingIterator() override {}
+
+  // No copy allowed
+  CoalescingIterator(const CoalescingIterator&) = delete;
+  CoalescingIterator& operator=(const CoalescingIterator&) = delete;
+
+  bool Valid() const override { return impl_.Valid(); }
+  void SeekToFirst() override { impl_.SeekToFirst(); }
+  void SeekToLast() override { impl_.SeekToLast(); }
+  void Seek(const Slice& target) override { impl_.Seek(target); }
+  void SeekForPrev(const Slice& target) override { impl_.SeekForPrev(target); }
+  void Next() override { impl_.Next(); }
+  void Prev() override { impl_.Prev(); }
+  Slice key() const override { return impl_.key(); }
+  Status status() const override { return impl_.status(); }
+
+  Slice value() const override {
+    assert(Valid());
+    return value_;
+  }
+  const WideColumns& columns() const override {
+    assert(Valid());
+    return wide_columns_;
+  }
+
+  void Reset() {
+    value_.clear();
+    wide_columns_.clear();
+  }
+
+ private:
+  MultiCfIteratorImpl impl_;
+  Slice value_;
+  WideColumns wide_columns_;
+
+  struct WideColumnWithOrder {
+    const WideColumn* column;
+    int order;
+  };
+
+  class WideColumnWithOrderComparator {
+   public:
+    explicit WideColumnWithOrderComparator() {}
+    bool operator()(const WideColumnWithOrder& a,
+                    const WideColumnWithOrder& b) const {
+      int c = a.column->name().compare(b.column->name());
+      return c == 0 ? a.order - b.order > 0 : c > 0;
+    }
+  };
+
+  using MinHeap =
+      BinaryHeap<WideColumnWithOrder, WideColumnWithOrderComparator>;
+
+  void Coalesce(const autovector<MultiCfIteratorInfo>& items);
+};
+
+}  // namespace ROCKSDB_NAMESPACE
@@ -538,6 +538,7 @@ ColumnFamilyData::ColumnFamilyData(
      refs_(0),
      initialized_(false),
      dropped_(false),
+      flush_skip_reschedule_(false),
      internal_comparator_(cf_options.comparator),
      initial_cf_options_(SanitizeOptions(db_options, cf_options)),
      ioptions_(db_options, initial_cf_options_),
@@ -1528,6 +1529,20 @@ Status ColumnFamilyData::ValidateOptions(
      }
    }
  }
+
+  if (cf_options.compaction_style == kCompactionStyleUniversal) {
+    int max_read_amp = cf_options.compaction_options_universal.max_read_amp;
+    if (max_read_amp < -1) {
+      return Status::NotSupported(
+          "CompactionOptionsUniversal::max_read_amp should be at least -1.");
+    } else if (0 < max_read_amp &&
+               max_read_amp < cf_options.level0_file_num_compaction_trigger) {
+      return Status::NotSupported(
+          "CompactionOptionsUniversal::max_read_amp limits the number of sorted"
+          " runs but is smaller than the compaction trigger "
+          "level0_file_num_compaction_trigger.");
+    }
+  }
  return s;
 }

@@ -1607,6 +1622,19 @@ FSDirectory* ColumnFamilyData::GetDataDir(size_t path_id) const {
  return data_dirs_[path_id].get();
 }

+void ColumnFamilyData::SetFlushSkipReschedule() {
+  const Comparator* ucmp = user_comparator();
+  const size_t ts_sz = ucmp->timestamp_size();
+  if (ts_sz == 0 || ioptions_.persist_user_defined_timestamps) {
+    return;
+  }
+  flush_skip_reschedule_.store(true);
+}
+
+bool ColumnFamilyData::GetAndClearFlushSkipReschedule() {
+  return flush_skip_reschedule_.exchange(false);
+}
+
 bool ColumnFamilyData::ShouldPostponeFlushToRetainUDT(
    uint64_t max_memtable_id) {
  const Comparator* ucmp = user_comparator();
@@ -329,6 +329,10 @@ class ColumnFamilyData {
  void SetDropped();
  bool IsDropped() const { return dropped_.load(std::memory_order_relaxed); }

+  void SetFlushSkipReschedule();
+
+  bool GetAndClearFlushSkipReschedule();
+
  // thread-safe
  int NumberLevels() const { return ioptions_.num_levels; }

@@ -568,6 +572,10 @@ class ColumnFamilyData {
  // of its files (if missing)
  void RecoverEpochNumbers();

+  int GetUnflushedMemTableCountForWriteStallCheck() const {
+    return (mem_->IsEmpty() ? 0 : 1) + imm_.NumNotFlushed();
+  }
+
 private:
  friend class ColumnFamilySet;
  ColumnFamilyData(uint32_t id, const std::string& name,
@@ -592,6 +600,15 @@ class ColumnFamilyData {
  std::atomic<bool> initialized_;
  std::atomic<bool> dropped_;  // true if client dropped it

+  // When user-defined timestamps in memtable only feature is enabled, this
+  // flag indicates a successfully requested flush that should
+  // skip being rescheduled and haven't undergone the rescheduling check yet.
+  // This flag is cleared when a check skips rescheduling a FlushRequest.
+  // With this flag, automatic flushes in regular cases can continue to
+  // retain UDTs by getting rescheduled as usual while manual flushes and
+  // error recovery flushes will proceed without getting rescheduled.
+  std::atomic<bool> flush_skip_reschedule_;
+
  const InternalKeyComparator internal_comparator_;
  InternalTblPropCollFactories internal_tbl_prop_coll_factories_;

@@ -23,6 +23,7 @@
 #include "rocksdb/db.h"
 #include "rocksdb/env.h"
 #include "rocksdb/iterator.h"
+#include "rocksdb/listener.h"
 #include "rocksdb/utilities/object_registry.h"
 #include "test_util/sync_point.h"
 #include "test_util/testharness.h"
@@ -34,6 +35,13 @@
 #include "utilities/merge_operators.h"

 namespace ROCKSDB_NAMESPACE {
+namespace {
+std::string EncodeAsUint64(uint64_t number) {
+  std::string result;
+  PutFixed64(&result, number);
+  return result;
+}
+}  // namespace

 static const int kValueSize = 1000;

@@ -3170,6 +3178,8 @@ TEST_P(ColumnFamilyTest, IteratorCloseWALFile1) {
  SpecialEnv env(Env::Default());
  db_options_.env = &env;
  db_options_.max_background_flushes = 1;
+  // When this option is removed, the test will need re-engineering
+  db_options_.background_close_inactive_wals = true;
  column_family_options_.memtable_factory.reset(
      test::NewSpecialSkipListFactory(2));
  Open();
@@ -3222,6 +3232,8 @@ TEST_P(ColumnFamilyTest, IteratorCloseWALFile2) {
  env.SetBackgroundThreads(2, Env::HIGH);
  db_options_.env = &env;
  db_options_.max_background_flushes = 1;
+  // When this option is removed, the test will need re-engineering
+  db_options_.background_close_inactive_wals = true;
  column_family_options_.memtable_factory.reset(
      test::NewSpecialSkipListFactory(2));
  Open();
@@ -3279,6 +3291,8 @@ TEST_P(ColumnFamilyTest, ForwardIteratorCloseWALFile) {
  env.SetBackgroundThreads(2, Env::HIGH);
  db_options_.env = &env;
  db_options_.max_background_flushes = 1;
+  // When this option is removed, the test will need re-engineering
+  db_options_.background_close_inactive_wals = true;
  column_family_options_.memtable_factory.reset(
      test::NewSpecialSkipListFactory(3));
  column_family_options_.level0_file_num_compaction_trigger = 2;
@@ -3603,7 +3617,9 @@ TEST(ColumnFamilyTest, ValidateMemtableKVChecksumOption) {
 }

 // Tests the flushing behavior of a column family to retain user-defined
-// timestamp when `persist_user_defined_timestamp` is false.
+// timestamp when `persist_user_defined_timestamp` is false. The behavior of
+// auto flush is it makes some effort to retain user-defined timestamps while
+// the behavior of manual flush is that it skips retaining UDTs.
 class ColumnFamilyRetainUDTTest : public ColumnFamilyTestBase {
 public:
  ColumnFamilyRetainUDTTest() : ColumnFamilyTestBase(kLatestFormatVersion) {}
@@ -3621,6 +3637,27 @@ class ColumnFamilyRetainUDTTest : public ColumnFamilyTestBase {
    return db_->Put(WriteOptions(), handles_[cf], Slice(key), Slice(ts),
                    Slice(value));
  }
+
+  std::string Get(int cf, const std::string& key, const std::string& read_ts) {
+    ReadOptions ropts;
+    Slice timestamp = read_ts;
+    ropts.timestamp = &timestamp;
+    std::string value;
+    Status s = db_->Get(ropts, handles_[cf], Slice(key), &value);
+    if (s.IsNotFound()) {
+      return "NOT_FOUND";
+    } else if (s.ok()) {
+      return value;
+    }
+    return "";
+  }
+
+  void CheckEffectiveCutoffTime(uint64_t expected_cutoff) {
+    std::string effective_full_history_ts_low;
+    EXPECT_OK(
+        db_->GetFullHistoryTsLow(handles_[0], &effective_full_history_ts_low));
+    EXPECT_EQ(EncodeAsUint64(expected_cutoff), effective_full_history_ts_low);
+  }
 };

 class TestTsComparator : public Comparator {
@@ -3664,7 +3701,9 @@ TEST_F(ColumnFamilyRetainUDTTest, SanityCheck) {
  Close();
 }

-TEST_F(ColumnFamilyRetainUDTTest, FullHistoryTsLowNotSet) {
+class AutoFlushRetainUDTTest : public ColumnFamilyRetainUDTTest {};
+
+TEST_F(AutoFlushRetainUDTTest, FullHistoryTsLowNotSet) {
  SyncPoint::GetInstance()->SetCallBack(
      "DBImpl::BackgroundFlush:CheckFlushRequest:cb", [&](void* arg) {
        ASSERT_NE(nullptr, arg);
@@ -3674,28 +3713,22 @@ TEST_F(ColumnFamilyRetainUDTTest, FullHistoryTsLowNotSet) {

  SyncPoint::GetInstance()->EnableProcessing();
  Open();
-  std::string write_ts;
-  PutFixed64(&write_ts, 1);
-  ASSERT_OK(Put(0, "foo", write_ts, "v1"));
-  // No `full_history_ts_low` explicitly set by user, flush is continued
+  ASSERT_OK(Put(0, "foo", EncodeAsUint64(1), "v1"));
+  // No `full_history_ts_low` explicitly set by user, auto flush is continued
  // without checking if its UDTs expired.
-  ASSERT_OK(Flush(0));
+  ASSERT_OK(dbfull()->TEST_SwitchWAL());
+  ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());

  // After flush, `full_history_ts_low` should be automatically advanced to
  // the effective cutoff timestamp: write_ts + 1
-  std::string cutoff_ts;
-  PutFixed64(&cutoff_ts, 2);
-  std::string effective_full_history_ts_low;
-  ASSERT_OK(
-      db_->GetFullHistoryTsLow(handles_[0], &effective_full_history_ts_low));
-  ASSERT_EQ(cutoff_ts, effective_full_history_ts_low);
+  CheckEffectiveCutoffTime(2);
  Close();

  SyncPoint::GetInstance()->DisableProcessing();
  SyncPoint::GetInstance()->ClearAllCallBacks();
 }

-TEST_F(ColumnFamilyRetainUDTTest, AllKeysExpired) {
+TEST_F(AutoFlushRetainUDTTest, AllKeysExpired) {
  SyncPoint::GetInstance()->SetCallBack(
      "DBImpl::BackgroundFlush:CheckFlushRequest:cb", [&](void* arg) {
        ASSERT_NE(nullptr, arg);
@@ -3705,27 +3738,22 @@ TEST_F(ColumnFamilyRetainUDTTest, AllKeysExpired) {

  SyncPoint::GetInstance()->EnableProcessing();
  Open();
-  std::string write_ts;
-  PutFixed64(&write_ts, 1);
-  ASSERT_OK(Put(0, "foo", write_ts, "v1"));
-  std::string cutoff_ts;
-  PutFixed64(&cutoff_ts, 3);
-  ASSERT_OK(db_->IncreaseFullHistoryTsLow(handles_[0], cutoff_ts));
-  // All keys expired w.r.t the configured `full_history_ts_low`, flush continue
-  // without the need for a re-schedule.
-  ASSERT_OK(Flush(0));
+  ASSERT_OK(Put(0, "foo", EncodeAsUint64(1), "v1"));
+  ASSERT_OK(db_->IncreaseFullHistoryTsLow(handles_[0], EncodeAsUint64(3)));
+  // All keys expired w.r.t the configured `full_history_ts_low`, auto flush
+  // continue without the need for a re-schedule.
+  ASSERT_OK(dbfull()->TEST_SwitchWAL());
+  ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());

  // `full_history_ts_low` stays unchanged after flush.
-  std::string effective_full_history_ts_low;
-  ASSERT_OK(
-      db_->GetFullHistoryTsLow(handles_[0], &effective_full_history_ts_low));
-  ASSERT_EQ(cutoff_ts, effective_full_history_ts_low);
+  CheckEffectiveCutoffTime(3);
  Close();

  SyncPoint::GetInstance()->DisableProcessing();
  SyncPoint::GetInstance()->ClearAllCallBacks();
 }
-TEST_F(ColumnFamilyRetainUDTTest, NotAllKeysExpiredFlushToAvoidWriteStall) {
+
+TEST_F(AutoFlushRetainUDTTest, NotAllKeysExpiredFlushToAvoidWriteStall) {
  SyncPoint::GetInstance()->SetCallBack(
      "DBImpl::BackgroundFlush:CheckFlushRequest:cb", [&](void* arg) {
        ASSERT_NE(nullptr, arg);
@@ -3735,72 +3763,218 @@ TEST_F(ColumnFamilyRetainUDTTest, NotAllKeysExpiredFlushToAvoidWriteStall) {

  SyncPoint::GetInstance()->EnableProcessing();
  Open();
-  std::string cutoff_ts;
-  std::string write_ts;
-  PutFixed64(&write_ts, 1);
-  ASSERT_OK(Put(0, "foo", write_ts, "v1"));
-  PutFixed64(&cutoff_ts, 1);
-  ASSERT_OK(db_->IncreaseFullHistoryTsLow(handles_[0], cutoff_ts));
+  ASSERT_OK(Put(0, "foo", EncodeAsUint64(1), "v1"));
+  ASSERT_OK(db_->IncreaseFullHistoryTsLow(handles_[0], EncodeAsUint64(1)));
  ASSERT_OK(db_->SetOptions(handles_[0], {{"max_write_buffer_number", "1"}}));
-  // Not all keys expired, but flush is continued without a re-schedule because
-  // of risk of write stall.
-  ASSERT_OK(Flush(0));
+  // Not all keys expired, but auto flush is continued without a re-schedule
+  // because of risk of write stall.
+  ASSERT_OK(dbfull()->TEST_SwitchWAL());
+  ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());

  // After flush, `full_history_ts_low` should be automatically advanced to
  // the effective cutoff timestamp: write_ts + 1
-  std::string effective_full_history_ts_low;
-  ASSERT_OK(
-      db_->GetFullHistoryTsLow(handles_[0], &effective_full_history_ts_low));
-
-  cutoff_ts.clear();
-  PutFixed64(&cutoff_ts, 2);
-  ASSERT_EQ(cutoff_ts, effective_full_history_ts_low);
+  CheckEffectiveCutoffTime(2);
  Close();

  SyncPoint::GetInstance()->DisableProcessing();
  SyncPoint::GetInstance()->ClearAllCallBacks();
 }

-TEST_F(ColumnFamilyRetainUDTTest, NotAllKeysExpiredFlushRescheduled) {
-  std::string cutoff_ts;
+TEST_F(AutoFlushRetainUDTTest, NotAllKeysExpiredFlushRescheduled) {
+  std::atomic<int> local_counter{1};
  SyncPoint::GetInstance()->SetCallBack(
      "DBImpl::AfterRetainUDTReschedule:cb", [&](void* /*arg*/) {
        // Increasing full_history_ts_low so all keys expired after the initial
        // FlushRequest is rescheduled
-        cutoff_ts.clear();
-        PutFixed64(&cutoff_ts, 3);
-        ASSERT_OK(db_->IncreaseFullHistoryTsLow(handles_[0], cutoff_ts));
+        ASSERT_OK(
+            db_->IncreaseFullHistoryTsLow(handles_[0], EncodeAsUint64(3)));
      });
  SyncPoint::GetInstance()->SetCallBack(
      "DBImpl::BackgroundFlush:CheckFlushRequest:cb", [&](void* arg) {
        ASSERT_NE(nullptr, arg);
        auto reschedule_count = *static_cast<int*>(arg);
        ASSERT_EQ(2, reschedule_count);
+        local_counter.fetch_add(1);
      });
  SyncPoint::GetInstance()->EnableProcessing();

  Open();
-  std::string write_ts;
-  PutFixed64(&write_ts, 1);
-  ASSERT_OK(Put(0, "foo", write_ts, "v1"));
-  PutFixed64(&cutoff_ts, 1);
-  ASSERT_OK(db_->IncreaseFullHistoryTsLow(handles_[0], cutoff_ts));
+  ASSERT_OK(Put(0, "foo", EncodeAsUint64(1), "v1"));
+  ASSERT_OK(db_->IncreaseFullHistoryTsLow(handles_[0], EncodeAsUint64(1)));
  // Not all keys expired, and there is no risk of write stall. Flush is
  // rescheduled. The actual flush happens after `full_history_ts_low` is
  // increased to mark all keys expired.
-  ASSERT_OK(Flush(0));
+  ASSERT_OK(dbfull()->TEST_SwitchWAL());
+  ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
+  // Make sure callback is not skipped.
+  ASSERT_EQ(2, local_counter);

-  std::string effective_full_history_ts_low;
-  ASSERT_OK(
-      db_->GetFullHistoryTsLow(handles_[0], &effective_full_history_ts_low));
-  // `full_history_ts_low` stays unchanged.
-  ASSERT_EQ(cutoff_ts, effective_full_history_ts_low);
+  CheckEffectiveCutoffTime(3);
  Close();

  SyncPoint::GetInstance()->DisableProcessing();
  SyncPoint::GetInstance()->ClearAllCallBacks();
 }

+class ManualFlushSkipRetainUDTTest : public ColumnFamilyRetainUDTTest {
+ public:
+  // Write an entry with timestamp that is not expired w.r.t cutoff timestamp,
+  // and make sure automatic flush would be rescheduled to retain UDT.
+  void CheckAutomaticFlushRetainUDT(uint64_t write_ts) {
+    std::atomic<int> local_counter{1};
+    SyncPoint::GetInstance()->SetCallBack(
+        "DBImpl::AfterRetainUDTReschedule:cb", [&](void* /*arg*/) {
+          // Increasing full_history_ts_low so all keys expired after the
+          // initial FlushRequest is rescheduled
+          ASSERT_OK(db_->IncreaseFullHistoryTsLow(
+              handles_[0], EncodeAsUint64(write_ts + 1)));
+        });
+    SyncPoint::GetInstance()->SetCallBack(
+        "DBImpl::BackgroundFlush:CheckFlushRequest:cb", [&](void* arg) {
+          ASSERT_NE(nullptr, arg);
+          auto reschedule_count = *static_cast<int*>(arg);
+          ASSERT_EQ(2, reschedule_count);
+          local_counter.fetch_add(1);
+        });
+    SyncPoint::GetInstance()->EnableProcessing();
+    EXPECT_OK(Put(0, "foo", EncodeAsUint64(write_ts),
+                  "foo" + std::to_string(write_ts)));
+    EXPECT_OK(dbfull()->TEST_SwitchWAL());
+    EXPECT_OK(dbfull()->TEST_WaitForFlushMemTable());
+    // Make sure callback is not skipped.
+    EXPECT_EQ(2, local_counter);
+
+    SyncPoint::GetInstance()->DisableProcessing();
+    SyncPoint::GetInstance()->ClearAllCallBacks();
+  }
+};
+
+TEST_F(ManualFlushSkipRetainUDTTest, ManualFlush) {
+  Open();
+  ASSERT_OK(db_->IncreaseFullHistoryTsLow(handles_[0], EncodeAsUint64(0)));
+
+  // Manual flush proceeds without trying to retain UDT.
+  ASSERT_OK(Put(0, "foo", EncodeAsUint64(1), "v1"));
+  ASSERT_OK(Flush(0));
+  CheckEffectiveCutoffTime(2);
+  CheckAutomaticFlushRetainUDT(3);
+
+  Close();
+}
+
+TEST_F(ManualFlushSkipRetainUDTTest, ManualCompaction) {
+  Open();
+  ASSERT_OK(db_->IncreaseFullHistoryTsLow(handles_[0], EncodeAsUint64(0)));
+
+  // Manual compaction proceeds without trying to retain UDT.
+  ASSERT_OK(Put(0, "foo", EncodeAsUint64(1), "v2"));
+  ASSERT_OK(
+      db_->CompactRange(CompactRangeOptions(), handles_[0], nullptr, nullptr));
+  CheckEffectiveCutoffTime(2);
+  CheckAutomaticFlushRetainUDT(3);
+
+  Close();
+}
+
+TEST_F(ManualFlushSkipRetainUDTTest, BulkLoading) {
+  Open();
+  ASSERT_OK(db_->IncreaseFullHistoryTsLow(handles_[0], EncodeAsUint64(0)));
+  ASSERT_OK(Put(0, "foo", EncodeAsUint64(1), "v1"));
+
+  // Test flush behavior in bulk loading scenarios.
+  Options options(db_options_, column_family_options_);
+  std::string sst_files_dir = dbname_ + "/sst_files/";
+  ASSERT_OK(DestroyDir(env_, sst_files_dir));
+  ASSERT_OK(env_->CreateDir(sst_files_dir));
+  SstFileWriter sst_file_writer(EnvOptions(), options);
+  std::string file1 = sst_files_dir + "file1.sst";
+  ASSERT_OK(sst_file_writer.Open(file1));
+  ASSERT_OK(sst_file_writer.Put("foo", EncodeAsUint64(0), "v2"));
+  ExternalSstFileInfo file1_info;
+  ASSERT_OK(sst_file_writer.Finish(&file1_info));
+
+  // Bulk loading in UDT mode doesn't support external file key range overlap
+  // with DB key range.
+  ASSERT_TRUE(db_->IngestExternalFile({file1}, IngestExternalFileOptions())
+                  .IsInvalidArgument());
+
+  std::string file2 = sst_files_dir + "file2.sst";
+  ASSERT_OK(sst_file_writer.Open(file2));
+  ASSERT_OK(sst_file_writer.Put("bar", EncodeAsUint64(0), "val"));
+  ExternalSstFileInfo file2_info;
+  ASSERT_OK(sst_file_writer.Finish(&file2_info));
+  // A successful bulk loading, and it doesn't trigger any flush. As a result
+  // the effective cutoff timestamp is also unchanged.
+  ASSERT_OK(db_->IngestExternalFile({file2}, IngestExternalFileOptions()));
+
+  ASSERT_EQ(Get(0, "foo", EncodeAsUint64(1)), "v1");
+  ASSERT_EQ(Get(0, "bar", EncodeAsUint64(0)), "val");
+  CheckEffectiveCutoffTime(0);
+  CheckAutomaticFlushRetainUDT(1);
+
+  Close();
+}
+
+TEST_F(ManualFlushSkipRetainUDTTest, AutomaticFlushQueued) {
+  Open();
+  ASSERT_OK(db_->IncreaseFullHistoryTsLow(handles_[0], EncodeAsUint64(0)));
+
+  ASSERT_OK(Put(0, "foo", EncodeAsUint64(1), "v1"));
+  ASSERT_OK(dbfull()->TEST_SwitchWAL());
+  CheckEffectiveCutoffTime(0);
+
+  // Default `max_write_buffer_number=2` used, writing another memtable can get
+  // automatic flush to proceed because of memory pressure. Not doing that so
+  // we can test automatic flush gets to proceed because of an ongoing manual
+  // flush attempt.
+  ASSERT_OK(Flush(0));
+  CheckEffectiveCutoffTime(2);
+  CheckAutomaticFlushRetainUDT(3);
+
+  Close();
+}
+
+TEST_F(ManualFlushSkipRetainUDTTest, ConcurrentManualFlushes) {
+  Open();
+  ASSERT_OK(db_->IncreaseFullHistoryTsLow(handles_[0], EncodeAsUint64(0)));
+
+  std::vector<ROCKSDB_NAMESPACE::port::Thread> manual_flush_tds;
+  std::atomic<int> next_ts{0};
+  std::mutex mtx;
+  std::condition_variable cv;
+
+  auto manual_flush = [&](int write_ts) {
+    {
+      std::unique_lock<std::mutex> lock(mtx);
+      cv.wait(lock,
+              [&write_ts, &next_ts] { return write_ts == next_ts.load(); });
+      ASSERT_OK(Put(0, "foo" + std::to_string(write_ts),
+                    EncodeAsUint64(write_ts),
+                    "val_" + std::to_string(write_ts)));
+      next_ts.fetch_add(1);
+      cv.notify_all();
+    }
+    if (write_ts % 2 == 0) {
+      ASSERT_OK(Flush(0));
+    } else {
+      ASSERT_OK(db_->CompactRange(CompactRangeOptions(), handles_[0], nullptr,
+                                  nullptr));
+    }
+  };
+
+  for (int write_ts = 0; write_ts < 10; write_ts++) {
+    manual_flush_tds.emplace_back(manual_flush, write_ts);
+  }
+
+  for (auto& td : manual_flush_tds) {
+    td.join();
+  }
+
+  CheckEffectiveCutoffTime(10);
+  CheckAutomaticFlushRetainUDT(11);
+  Close();
+}
+
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
@@ -441,6 +441,50 @@ TEST_F(CompactFilesTest, SentinelCompressionType) {
  }
 }

+TEST_F(CompactFilesTest, CompressionWithBlockAlign) {
+  Options options;
+  options.compression = CompressionType::kNoCompression;
+  options.create_if_missing = true;
+  options.disable_auto_compactions = true;
+
+  std::shared_ptr<FlushedFileCollector> collector =
+      std::make_shared<FlushedFileCollector>();
+  options.listeners.push_back(collector);
+
+  {
+    BlockBasedTableOptions bbto;
+    bbto.block_align = true;
+    options.table_factory.reset(NewBlockBasedTableFactory(bbto));
+  }
+
+  std::unique_ptr<DB> db;
+  {
+    DB* _db = nullptr;
+    ASSERT_OK(DB::Open(options, db_name_, &_db));
+    db.reset(_db);
+  }
+
+  ASSERT_OK(db->Put(WriteOptions(), "key", "val"));
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  // Ensure background work is fully finished including listener callbacks
+  // before accessing listener state.
+  ASSERT_OK(
+      static_cast_with_check<DBImpl>(db.get())->TEST_WaitForBackgroundWork());
+  auto l0_files = collector->GetFlushedFiles();
+  ASSERT_EQ(1, l0_files.size());
+
+  // We can run this test even without Snappy support because we expect the
+  // `CompactFiles()` to fail before actually invoking Snappy compression.
+  CompactionOptions compaction_opts;
+  compaction_opts.compression = CompressionType::kSnappyCompression;
+  ASSERT_TRUE(db->CompactFiles(compaction_opts, l0_files, 1 /* output_level */)
+                  .IsInvalidArgument());
+
+  compaction_opts.compression = CompressionType::kDisableCompressionOption;
+  ASSERT_OK(db->CompactFiles(compaction_opts, l0_files, 1 /* output_level */));
+}
+
 TEST_F(CompactFilesTest, GetCompactionJobInfo) {
  Options options;
  options.create_if_missing = true;
@@ -25,6 +25,7 @@ namespace ROCKSDB_NAMESPACE {
 CompactionIterator::CompactionIterator(
    InternalIterator* input, const Comparator* cmp, MergeHelper* merge_helper,
    SequenceNumber last_sequence, std::vector<SequenceNumber>* snapshots,
+    SequenceNumber earliest_snapshot,
    SequenceNumber earliest_write_conflict_snapshot,
    SequenceNumber job_snapshot, const SnapshotChecker* snapshot_checker,
    Env* env, bool report_detailed_time, bool expect_valid_internal_key,
@@ -40,7 +41,7 @@ CompactionIterator::CompactionIterator(
    const SequenceNumber preserve_time_min_seqno,
    const SequenceNumber preclude_last_level_min_seqno)
    : CompactionIterator(
-          input, cmp, merge_helper, last_sequence, snapshots,
+          input, cmp, merge_helper, last_sequence, snapshots, earliest_snapshot,
          earliest_write_conflict_snapshot, job_snapshot, snapshot_checker, env,
          report_detailed_time, expect_valid_internal_key, range_del_agg,
          blob_file_builder, allow_data_in_errors, enforce_single_del_contracts,
@@ -54,6 +55,7 @@ CompactionIterator::CompactionIterator(
 CompactionIterator::CompactionIterator(
    InternalIterator* input, const Comparator* cmp, MergeHelper* merge_helper,
    SequenceNumber /*last_sequence*/, std::vector<SequenceNumber>* snapshots,
+    SequenceNumber earliest_snapshot,
    SequenceNumber earliest_write_conflict_snapshot,
    SequenceNumber job_snapshot, const SnapshotChecker* snapshot_checker,
    Env* env, bool report_detailed_time, bool expect_valid_internal_key,
@@ -91,9 +93,7 @@ CompactionIterator::CompactionIterator(
      // snapshots_ cannot be nullptr, but we will assert later in the body of
      // the constructor.
      visible_at_tip_(snapshots_ ? snapshots_->empty() : false),
-      earliest_snapshot_(!snapshots_ || snapshots_->empty()
-                             ? kMaxSequenceNumber
-                             : snapshots_->at(0)),
+      earliest_snapshot_(earliest_snapshot),
      info_log_(info_log),
      allow_data_in_errors_(allow_data_in_errors),
      enforce_single_del_contracts_(enforce_single_del_contracts),
@@ -997,25 +997,37 @@ void CompactionIterator::NextFromInput() {
      // A special case involving range deletion is handled separately below.
      auto [unpacked_value, preferred_seqno] =
          ParsePackedValueWithSeqno(value_);
-      assert(preferred_seqno < ikey_.sequence);
-      InternalKey ikey_after_swap(ikey_.user_key, preferred_seqno, kTypeValue);
-      Slice ikey_after_swap_slice(*ikey_after_swap.rep());
+      assert(preferred_seqno < ikey_.sequence || ikey_.sequence == 0);
      if (range_del_agg_->ShouldDelete(
-              ikey_after_swap_slice,
-              RangeDelPositioningMode::kForwardTraversal)) {
-        // A range tombstone that doesn't cover this kTypeValuePreferredSeqno
-        // entry may end up covering the entry, so it's not safe to swap
-        // preferred sequence number. In this case, we output the entry as is.
-        validity_info_.SetValid(ValidContext::kNewUserKey);
+              key_, RangeDelPositioningMode::kForwardTraversal)) {
+        ++iter_stats_.num_record_drop_hidden;
+        ++iter_stats_.num_record_drop_range_del;
+        AdvanceInputIter();
      } else {
-        iter_stats_.num_timed_put_swap_preferred_seqno++;
-        ikey_.sequence = preferred_seqno;
-        ikey_.type = kTypeValue;
-        current_key_.UpdateInternalKey(ikey_.sequence, ikey_.type);
-        key_ = current_key_.GetInternalKey();
-        ikey_.user_key = current_key_.GetUserKey();
-        value_ = unpacked_value;
-        validity_info_.SetValid(ValidContext::kSwapPreferredSeqno);
+        InternalKey ikey_after_swap(ikey_.user_key,
+                                    std::min(preferred_seqno, ikey_.sequence),
+                                    kTypeValue);
+        Slice ikey_after_swap_slice(*ikey_after_swap.rep());
+        if (range_del_agg_->ShouldDelete(
+                ikey_after_swap_slice,
+                RangeDelPositioningMode::kForwardTraversal)) {
+          // A range tombstone that doesn't cover this kTypeValuePreferredSeqno
+          // entry will end up covering the entry, so it's not safe to swap
+          // preferred sequence number. In this case, we output the entry as is.
+          validity_info_.SetValid(ValidContext::kNewUserKey);
+        } else {
+          if (ikey_.sequence != 0) {
+            iter_stats_.num_timed_put_swap_preferred_seqno++;
+            saved_seq_for_penul_check_ = ikey_.sequence;
+            ikey_.sequence = preferred_seqno;
+          }
+          ikey_.type = kTypeValue;
+          current_key_.UpdateInternalKey(ikey_.sequence, ikey_.type);
+          key_ = current_key_.GetInternalKey();
+          ikey_.user_key = current_key_.GetUserKey();
+          value_ = unpacked_value;
+          validity_info_.SetValid(ValidContext::kSwapPreferredSeqno);
+        }
      }
    } else if (ikey_.type == kTypeMerge) {
      if (!merge_helper_->HasOperator()) {
@@ -1268,6 +1280,21 @@ void CompactionIterator::DecideOutputLevel() {
  }
 #endif  // NDEBUG

+  // saved_seq_for_penul_check_ is populated in `NextFromInput` when the
+  // entry's sequence number is non zero and validity context for output this
+  // entry is kSwapPreferredSeqno for use in `DecideOutputLevel`. It should be
+  // cleared out here unconditionally. Otherwise, it may end up getting consumed
+  // incorrectly by a different entry.
+  SequenceNumber seq_for_range_check =
+      (saved_seq_for_penul_check_.has_value() &&
+       saved_seq_for_penul_check_.value() != kMaxSequenceNumber)
+          ? saved_seq_for_penul_check_.value()
+          : ikey_.sequence;
+  saved_seq_for_penul_check_ = std::nullopt;
+  ParsedInternalKey ikey_for_range_check = ikey_;
+  if (seq_for_range_check != ikey_.sequence) {
+    ikey_for_range_check.sequence = seq_for_range_check;
+  }
  if (output_to_penultimate_level_) {
    // If it's decided to output to the penultimate level, but unsafe to do so,
    // still output to the last level. For example, moving the data from a lower
@@ -1276,7 +1303,7 @@ void CompactionIterator::DecideOutputLevel() {
    // not from this compaction.
    // TODO: add statistic for declined output_to_penultimate_level
    bool safe_to_penultimate_level =
-        compaction_->WithinPenultimateLevelOutputRange(ikey_);
+        compaction_->WithinPenultimateLevelOutputRange(ikey_for_range_check);
    if (!safe_to_penultimate_level) {
      output_to_penultimate_level_ = false;
      // It could happen when disable/enable `last_level_temperature` while
@@ -1288,7 +1315,7 @@ void CompactionIterator::DecideOutputLevel() {
      // snapshot is released before enabling `last_level_temperature` feature
      // We will migrate the feature to `last_level_temperature` and maybe make
      // it not dynamically changeable.
-      if (ikey_.sequence > earliest_snapshot_) {
+      if (seq_for_range_check > earliest_snapshot_) {
        status_ = Status::Corruption(
            "Unsafe to store Seq later than snapshot in the last level if "
            "per_key_placement is enabled");
@@ -203,6 +203,7 @@ class CompactionIterator {
  CompactionIterator(
      InternalIterator* input, const Comparator* cmp, MergeHelper* merge_helper,
      SequenceNumber last_sequence, std::vector<SequenceNumber>* snapshots,
+      SequenceNumber earliest_snapshot,
      SequenceNumber earliest_write_conflict_snapshot,
      SequenceNumber job_snapshot, const SnapshotChecker* snapshot_checker,
      Env* env, bool report_detailed_time, bool expect_valid_internal_key,
@@ -222,6 +223,7 @@ class CompactionIterator {
  CompactionIterator(
      InternalIterator* input, const Comparator* cmp, MergeHelper* merge_helper,
      SequenceNumber last_sequence, std::vector<SequenceNumber>* snapshots,
+      SequenceNumber earliest_snapshot,
      SequenceNumber earliest_write_conflict_snapshot,
      SequenceNumber job_snapshot, const SnapshotChecker* snapshot_checker,
      Env* env, bool report_detailed_time, bool expect_valid_internal_key,
@@ -437,6 +439,17 @@ class CompactionIterator {
  // iterator output (or current key in the underlying iterator during
  // NextFromInput()).
  ParsedInternalKey ikey_;
+
+  // When a kTypeValuePreferredSeqno entry's preferred seqno is safely swapped
+  // in in this compaction, this field saves its original sequence number for
+  // range checking whether it's safe to be placed on the penultimate level.
+  // This is to ensure when such an entry happens to be the right boundary of
+  // penultimate safe range, it won't get excluded because with the preferred
+  // seqno swapped in, it's now larger than the right boundary (itself before
+  // the swap). This is safe to do, because preferred seqno is swapped in only
+  // when no entries with the same user key exist on lower levels and this entry
+  // is already visible in the earliest snapshot.
+  std::optional<SequenceNumber> saved_seq_for_penul_check_ = kMaxSequenceNumber;
  // Stores whether ikey_.user_key is valid. If set to false, the user key is
  // not compared against the current key in the underlying iterator.
  bool has_current_user_key_ = false;
@@ -296,6 +296,7 @@ class CompactionIteratorTest : public testing::TestWithParam<bool> {
    iter_->SeekToFirst();
    c_iter_.reset(new CompactionIterator(
        iter_.get(), cmp_, merge_helper_.get(), last_sequence, &snapshots_,
+        snapshots_.empty() ? kMaxSequenceNumber : snapshots_.at(0),
        earliest_write_conflict_snapshot, kMaxSequenceNumber,
        snapshot_checker_.get(), Env::Default(),
        false /* report_detailed_time */, false, range_del_agg_.get(),
@@ -1260,6 +1261,28 @@ TEST_F(CompactionIteratorWithSnapshotCheckerTest,
          false /*key_not_exists_beyond_output_level*/);
 }

+TEST_F(CompactionIteratorWithSnapshotCheckerTest,
+       TimedPut_ShouldBeCoverredByRangeDeletionBeforeSwap_NoOutput) {
+  InitIterators({test::KeyStr("morning", 5, kTypeValuePreferredSeqno),
+                 test::KeyStr("morning", 2, kTypeValuePreferredSeqno),
+                 test::KeyStr("night", 6, kTypeValue)},
+                {ValueWithPreferredSeqno("zao", 3),
+                 ValueWithPreferredSeqno("zao", 1), "wan"},
+                {test::KeyStr("ma", 6, kTypeRangeDeletion)}, {"mz"}, 6,
+                kMaxSequenceNumber /*last_committed_sequence*/,
+                nullptr /*merge_op*/, nullptr /*filter*/,
+                false /*bottommost_level*/,
+                kMaxSequenceNumber /*earliest_write_conflict_snapshot*/,
+                true /*key_not_exists_beyond_output_level*/);
+  c_iter_->SeekToFirst();
+  ASSERT_TRUE(c_iter_->Valid());
+  ASSERT_EQ(test::KeyStr("night", 6, kTypeValue), c_iter_->key().ToString());
+  ASSERT_EQ("wan", c_iter_->value().ToString());
+  c_iter_->Next();
+  ASSERT_FALSE(c_iter_->Valid());
+  ASSERT_OK(c_iter_->status());
+}
+
 TEST_F(CompactionIteratorWithSnapshotCheckerTest,
       TimedPut_WillBeHiddenByRangeDeletionAfterSwap_NoSwap) {
  InitIterators({test::KeyStr("morning", 5, kTypeValuePreferredSeqno),
@@ -1316,6 +1339,19 @@ TEST_F(
      true /*key_not_exists_beyond_output_level*/);
 }

+TEST_F(CompactionIteratorWithSnapshotCheckerTest,
+       TimedPut_SequenceNumberAlreadyZeroedOut_ChangeType) {
+  RunTest(
+      {test::KeyStr("bar", 0, kTypeValuePreferredSeqno),
+       test::KeyStr("bar", 0, kTypeValuePreferredSeqno),
+       test::KeyStr("foo", 0, kTypeValue)},
+      {ValueWithPreferredSeqno("bv2", 2), ValueWithPreferredSeqno("bv1", 1),
+       "fv1"},
+      {test::KeyStr("bar", 0, kTypeValue), test::KeyStr("foo", 0, kTypeValue)},
+      {"bv2", "fv1"}, 6 /*last_committed_seq*/, nullptr /*merge_operator*/,
+      nullptr /*compaction_filter*/, true /*bottommost_level*/);
+}
+
 // Compaction filter should keep uncommitted key as-is, and
 //   * Convert the latest value to deletion, and/or
 //   * if latest value is a merge, apply filter to all subsequent merges.
@@ -174,6 +174,9 @@ CompactionJob::CompactionJob(
      db_mutex_(db_mutex),
      db_error_handler_(db_error_handler),
      existing_snapshots_(std::move(existing_snapshots)),
+      earliest_snapshot_(existing_snapshots_.empty()
+                             ? kMaxSequenceNumber
+                             : existing_snapshots_.at(0)),
      earliest_write_conflict_snapshot_(earliest_write_conflict_snapshot),
      snapshot_checker_(snapshot_checker),
      job_context_(job_context),
@@ -282,6 +285,7 @@ void CompactionJob::Prepare() {

  // collect all seqno->time information from the input files which will be used
  // to encode seqno->time to the output files.
+
  uint64_t preserve_time_duration =
      std::max(c->immutable_options()->preserve_internal_time_seconds,
               c->immutable_options()->preclude_last_level_data_seconds);
@@ -319,28 +323,11 @@ void CompactionJob::Prepare() {
      seqno_to_time_mapping_.Enforce();
    } else {
      seqno_to_time_mapping_.Enforce(_current_time);
-      uint64_t preserve_time =
-          static_cast<uint64_t>(_current_time) > preserve_time_duration
-              ? _current_time - preserve_time_duration
-              : 0;
-      // GetProximalSeqnoBeforeTime tells us the last seqno known to have been
-      // written at or before the given time. + 1 to get the minimum we should
-      // preserve without excluding anything that might have been written on or
-      // after the given time.
-      preserve_time_min_seqno_ =
-          seqno_to_time_mapping_.GetProximalSeqnoBeforeTime(preserve_time) + 1;
-      if (c->immutable_options()->preclude_last_level_data_seconds > 0) {
-        uint64_t preclude_last_level_time =
-            static_cast<uint64_t>(_current_time) >
-                    c->immutable_options()->preclude_last_level_data_seconds
-                ? _current_time -
-                      c->immutable_options()->preclude_last_level_data_seconds
-                : 0;
-        preclude_last_level_min_seqno_ =
-            seqno_to_time_mapping_.GetProximalSeqnoBeforeTime(
-                preclude_last_level_time) +
-            1;
-      }
+      seqno_to_time_mapping_.GetCurrentTieringCutoffSeqnos(
+          static_cast<uint64_t>(_current_time),
+          c->immutable_options()->preserve_internal_time_seconds,
+          c->immutable_options()->preclude_last_level_data_seconds,
+          &preserve_time_min_seqno_, &preclude_last_level_min_seqno_);
    }
    // For accuracy of the GetProximalSeqnoBeforeTime queries above, we only
    // limit the capacity after them.
@@ -1295,8 +1282,9 @@ void CompactionJob::ProcessKeyValueCompaction(SubcompactionState* sub_compact) {

  auto c_iter = std::make_unique<CompactionIterator>(
      input, cfd->user_comparator(), &merge, versions_->LastSequence(),
-      &existing_snapshots_, earliest_write_conflict_snapshot_, job_snapshot_seq,
-      snapshot_checker_, env_, ShouldReportDetailedTime(env_, stats_),
+      &existing_snapshots_, earliest_snapshot_,
+      earliest_write_conflict_snapshot_, job_snapshot_seq, snapshot_checker_,
+      env_, ShouldReportDetailedTime(env_, stats_),
      /*expect_valid_internal_key=*/true, range_del_agg.get(),
      blob_file_builder.get(), db_options_.allow_data_in_errors,
      db_options_.enforce_single_del_contracts, manual_compaction_canceled_,
@@ -1969,7 +1957,10 @@ Status CompactionJob::OpenCompactionOutputFile(SubcompactionState* sub_compact,
      cfd->GetName(), sub_compact->compaction->output_level(),
      bottommost_level_, TableFileCreationReason::kCompaction,
      0 /* oldest_key_time */, current_time, db_id_, db_session_id_,
-      sub_compact->compaction->max_output_file_size(), file_number);
+      sub_compact->compaction->max_output_file_size(), file_number,
+      preclude_last_level_min_seqno_ == kMaxSequenceNumber
+          ? preclude_last_level_min_seqno_
+          : std::min(earliest_snapshot_, preclude_last_level_min_seqno_));

  outputs.NewBuilder(tboptions);

@@ -314,6 +314,8 @@ class CompactionJob {
  // deleted because that version is not visible in any snapshot.
  std::vector<SequenceNumber> existing_snapshots_;

+  SequenceNumber earliest_snapshot_;
+
  // This is the earliest snapshot that could be used for write-conflict
  // checking by a transaction.  For any user-key newer than this snapshot, we
  // should make sure not to remove evidence that a write occurred.
@@ -27,7 +27,9 @@ Status CompactionOutputs::Finish(
  if (s.ok()) {
    SeqnoToTimeMapping relevant_mapping;
    relevant_mapping.CopyFromSeqnoRange(
-        seqno_to_time_mapping, meta->fd.smallest_seqno, meta->fd.largest_seqno);
+        seqno_to_time_mapping,
+        std::min(smallest_preferred_seqno_, meta->fd.smallest_seqno),
+        meta->fd.largest_seqno);
    relevant_mapping.SetCapacity(kMaxSeqnoTimePairsPerSST);
    builder_->SetSeqnoTimeTableProperties(relevant_mapping,
                                          meta->oldest_ancester_time);
@@ -422,6 +424,11 @@ Status CompactionOutputs::AddToOutput(
  }

  const ParsedInternalKey& ikey = c_iter.ikey();
+  if (ikey.type == kTypeValuePreferredSeqno) {
+    SequenceNumber preferred_seqno = ParsePackedValueForSeqno(value);
+    smallest_preferred_seqno_ =
+        std::min(smallest_preferred_seqno_, preferred_seqno);
+  }
  s = current_output().meta.UpdateBoundaries(key, value, ikey.sequence,
                                             ikey.type);

@@ -297,6 +297,7 @@ class CompactionOutputs {
  std::unique_ptr<TableBuilder> builder_;
  std::unique_ptr<WritableFileWriter> file_writer_;
  uint64_t current_output_file_size_ = 0;
+  SequenceNumber smallest_preferred_seqno_ = kMaxSequenceNumber;

  // all the compaction outputs so far
  std::vector<Output> outputs_;
@@ -340,8 +340,8 @@ Compaction* CompactionPicker::CompactFiles(
 #ifndef NDEBUG
  assert(input_files.size());
  // This compaction output should not overlap with a running compaction as
-  // `SanitizeCompactionInputFiles` should've checked earlier and db mutex
-  // shouldn't have been released since.
+  // `SanitizeAndConvertCompactionInputFiles` should've checked earlier and db
+  // mutex shouldn't have been released since.
  int start_level = Compaction::kInvalidLevel;
  for (const auto& in : input_files) {
    // input_files should already be sorted by level
@@ -457,9 +457,10 @@ bool CompactionPicker::IsRangeInCompaction(VersionStorageInfo* vstorage,
 // Returns false if files on parent level are currently in compaction, which
 // means that we can't compact them
 bool CompactionPicker::SetupOtherInputs(
-    const std::string& cf_name, VersionStorageInfo* vstorage,
-    CompactionInputFiles* inputs, CompactionInputFiles* output_level_inputs,
-    int* parent_index, int base_index, bool only_expand_towards_right) {
+    const std::string& cf_name, const MutableCFOptions& mutable_cf_options,
+    VersionStorageInfo* vstorage, CompactionInputFiles* inputs,
+    CompactionInputFiles* output_level_inputs, int* parent_index,
+    int base_index, bool only_expand_towards_right) {
  assert(!inputs->empty());
  assert(output_level_inputs->empty());
  const int input_level = inputs->level;
@@ -525,8 +526,15 @@ bool CompactionPicker::SetupOtherInputs(
    if (!ExpandInputsToCleanCut(cf_name, vstorage, &expanded_inputs)) {
      try_overlapping_inputs = false;
    }
+    // It helps to reduce write amp and avoid a further separate compaction
+    // to include more input level files without expanding output level files.
+    // So we apply a softer limit. We still need a limit to avoid overly large
+    // compactions and potential high space amp spikes.
+    const uint64_t limit =
+        MultiplyCheckOverflow(mutable_cf_options.max_compaction_bytes, 2.0);
    if (try_overlapping_inputs && expanded_inputs.size() > inputs->size() &&
-        !AreFilesInCompaction(expanded_inputs.files)) {
+        !AreFilesInCompaction(expanded_inputs.files) &&
+        output_level_inputs_size + expanded_inputs_size < limit) {
      InternalKey new_start, new_limit;
      GetRange(expanded_inputs, &new_start, &new_limit);
      CompactionInputFiles expanded_output_level_inputs;
@@ -548,7 +556,8 @@ bool CompactionPicker::SetupOtherInputs(
                                             base_index, nullptr);
      expanded_inputs_size = TotalFileSize(expanded_inputs.files);
      if (expanded_inputs.size() > inputs->size() &&
-          !AreFilesInCompaction(expanded_inputs.files)) {
+          !AreFilesInCompaction(expanded_inputs.files) &&
+          (output_level_inputs_size + expanded_inputs_size) < limit) {
        expand_inputs = true;
      }
    }
@@ -808,8 +817,8 @@ Compaction* CompactionPicker::CompactRange(
  output_level_inputs.level = output_level;
  if (input_level != output_level) {
    int parent_index = -1;
-    if (!SetupOtherInputs(cf_name, vstorage, &inputs, &output_level_inputs,
-                          &parent_index, -1)) {
+    if (!SetupOtherInputs(cf_name, mutable_cf_options, vstorage, &inputs,
+                          &output_level_inputs, &parent_index, -1)) {
      // manual compaction is now multi-threaded, so it can
      // happen that SetupOtherInputs fails
      // we handle it higher in RunManualCompaction
@@ -1031,19 +1040,18 @@ Status CompactionPicker::SanitizeCompactionInputFilesForAllLevels(
      }
    }
  }
-  if (RangeOverlapWithCompaction(smallestkey, largestkey, output_level)) {
-    return Status::Aborted(
-        "A running compaction is writing to the same output level in an "
-        "overlapping key range");
-  }
  return Status::OK();
 }

-Status CompactionPicker::SanitizeCompactionInputFiles(
+Status CompactionPicker::SanitizeAndConvertCompactionInputFiles(
    std::unordered_set<uint64_t>* input_files,
-    const ColumnFamilyMetaData& cf_meta, const int output_level) const {
+    const ColumnFamilyMetaData& cf_meta, const int output_level,
+    const VersionStorageInfo* vstorage,
+    std::vector<CompactionInputFiles>* converted_input_files) const {
  assert(static_cast<int>(cf_meta.levels.size()) - 1 ==
         cf_meta.levels[cf_meta.levels.size() - 1].level);
+  assert(converted_input_files);
+
  if (output_level >= static_cast<int>(cf_meta.levels.size())) {
    return Status::InvalidArgument(
        "Output level for column family " + cf_meta.name +
@@ -1069,7 +1077,6 @@ Status CompactionPicker::SanitizeCompactionInputFiles(

  Status s = SanitizeCompactionInputFilesForAllLevels(input_files, cf_meta,
                                                      output_level);
-
  if (!s.ok()) {
    return s;
  }
@@ -1110,6 +1117,22 @@ Status CompactionPicker::SanitizeCompactionInputFiles(
    }
  }

+  s = GetCompactionInputsFromFileNumbers(converted_input_files, input_files,
+                                         vstorage, CompactionOptions());
+  if (!s.ok()) {
+    return s;
+  }
+  assert(converted_input_files->size() > 0);
+  if (output_level != 0 &&
+      FilesRangeOverlapWithCompaction(
+          *converted_input_files, output_level,
+          Compaction::EvaluatePenultimateLevel(
+              vstorage, ioptions_, (*converted_input_files)[0].level,
+              output_level))) {
+    return Status::Aborted(
+        "A running compaction is writing to the same output level(s) in an "
+        "overlapping key range");
+  }
  return Status::OK();
 }

@@ -86,16 +86,20 @@ class CompactionPicker {

  virtual bool NeedsCompaction(const VersionStorageInfo* vstorage) const = 0;

-// Sanitize the input set of compaction input files.
-// When the input parameters do not describe a valid compaction, the
-// function will try to fix the input_files by adding necessary
-// files.  If it's not possible to conver an invalid input_files
-// into a valid one by adding more files, the function will return a
-// non-ok status with specific reason.
-//
-  Status SanitizeCompactionInputFiles(std::unordered_set<uint64_t>* input_files,
-                                      const ColumnFamilyMetaData& cf_meta,
-                                      const int output_level) const;
+  // Sanitize the input set of compaction input files and convert it to
+  // `std::vector<CompactionInputFiles>` in the output parameter
+  // `converted_input_files`.
+  // When the input parameters do not describe a valid
+  // compaction, the function will try to fix the input_files by adding
+  // necessary files.  If it's not possible to convert an invalid input_files
+  // into a valid one by adding more files, the function will return a
+  // non-ok status with specific reason.
+  //
+  Status SanitizeAndConvertCompactionInputFiles(
+      std::unordered_set<uint64_t>* input_files,
+      const ColumnFamilyMetaData& cf_meta, const int output_level,
+      const VersionStorageInfo* vstorage,
+      std::vector<CompactionInputFiles>* converted_input_files) const;

  // Free up the files that participated in a compaction
  //
@@ -109,8 +113,8 @@ class CompactionPicker {
  // object.
  //
  // Caller must provide a set of input files that has been passed through
-  // `SanitizeCompactionInputFiles` earlier. The lock should not be released
-  // between that call and this one.
+  // `SanitizeAndConvertCompactionInputFiles` earlier. The lock should not be
+  // released between that call and this one.
  Compaction* CompactFiles(const CompactionOptions& compact_options,
                           const std::vector<CompactionInputFiles>& input_files,
                           int output_level, VersionStorageInfo* vstorage,
@@ -120,6 +124,7 @@ class CompactionPicker {

  // Converts a set of compaction input file numbers into
  // a list of CompactionInputFiles.
+  // TODO(hx235): remove the unused paramter `compact_options`
  Status GetCompactionInputsFromFileNumbers(
      std::vector<CompactionInputFiles>* input_files,
      std::unordered_set<uint64_t>* input_set,
@@ -186,6 +191,7 @@ class CompactionPicker {
      int penultimate_level) const;

  bool SetupOtherInputs(const std::string& cf_name,
+                        const MutableCFOptions& mutable_cf_options,
                        VersionStorageInfo* vstorage,
                        CompactionInputFiles* inputs,
                        CompactionInputFiles* output_level_inputs,
@@ -224,8 +230,8 @@ class CompactionPicker {
 protected:
  const ImmutableOptions& ioptions_;

-// A helper function to SanitizeCompactionInputFiles() that
-// sanitizes "input_files" by adding necessary files.
+  // A helper function to SanitizeAndConvertCompactionInputFiles() that
+  // sanitizes "input_files" by adding necessary files.
  virtual Status SanitizeCompactionInputFilesForAllLevels(
      std::unordered_set<uint64_t>* input_files,
      const ColumnFamilyMetaData& cf_meta, const int output_level) const;
@@ -467,8 +467,9 @@ bool LevelCompactionBuilder::SetupOtherInputsIfNeeded() {
    }
    if (!is_l0_trivial_move_ &&
        !compaction_picker_->SetupOtherInputs(
-            cf_name_, vstorage_, &start_level_inputs_, &output_level_inputs_,
-            &parent_index_, base_index_, round_robin_expanding)) {
+            cf_name_, mutable_cf_options_, vstorage_, &start_level_inputs_,
+            &output_level_inputs_, &parent_index_, base_index_,
+            round_robin_expanding)) {
      return false;
    }

@@ -2465,7 +2465,7 @@ TEST_F(CompactionPickerTest, IsBottommostLevelTest) {
  DeleteVersionStorage();
 }

-TEST_F(CompactionPickerTest, IgnoreCompactionLimitWhenAddFileFromInputLevel) {
+TEST_F(CompactionPickerTest, CompactionLimitWhenAddFileFromInputLevel) {
  mutable_cf_options_.max_bytes_for_level_base = 1000000u;
  mutable_cf_options_.max_compaction_bytes = 800000u;
  ioptions_.level_compaction_dynamic_level_bytes = false;
@@ -2473,8 +2473,10 @@ TEST_F(CompactionPickerTest, IgnoreCompactionLimitWhenAddFileFromInputLevel) {
  // A compaction should be triggered and pick file 2 and 5.
  // It pulls in other compaction input file from the input level L1
  // without pulling in more output level files.
-  // Files 1, 3, 4 will be included in the compaction.
+  // Files 1, 3, 4 are eligible.
  // File 6 is excluded since it overlaps with file 7.
+  // It can expand input level since in this case, the limit on compaction size
+  // is 2 * max_compaction_bytes.
  Add(1, 1U, "100", "150", 300000U);
  Add(1, 2U, "151", "200", 300001U, 0, 0);
  Add(1, 3U, "201", "250", 300000U, 0, 0);
@@ -2498,6 +2500,38 @@ TEST_F(CompactionPickerTest, IgnoreCompactionLimitWhenAddFileFromInputLevel) {
  ASSERT_EQ(5U, compaction->input(1, 0)->fd.GetNumber());
 }

+TEST_F(CompactionPickerTest, HitCompactionLimitWhenAddFileFromInputLevel) {
+  mutable_cf_options_.max_bytes_for_level_base = 1000000u;
+  mutable_cf_options_.max_compaction_bytes = 800000u;
+  ioptions_.level_compaction_dynamic_level_bytes = false;
+  NewVersionStorage(6, kCompactionStyleLevel);
+  // A compaction should be triggered and pick file 2 and 5.
+  // It pulls in other compaction input file from the input level L1
+  // without pulling in more output level files.
+  // Files 1, 3, 4 are eligible.
+  // File 6 is excluded since it overlaps with file 7.
+  // It can not expand input level since total compaction size hit the limit
+  // 2 * max_compaction_bytes.
+  Add(1, 1U, "100", "150", 400000U);
+  Add(1, 2U, "151", "200", 400001U, 0, 0);
+  Add(1, 3U, "201", "250", 400000U, 0, 0);
+  Add(1, 4U, "251", "300", 400000U, 0, 0);
+  Add(1, 6U, "325", "400", 400000U, 0, 0);
+  Add(2, 5U, "100", "350", 1U);
+  Add(2, 7U, "375", "425", 1U);
+  UpdateVersionStorageInfo();
+
+  std::unique_ptr<Compaction> compaction(level_compaction_picker.PickCompaction(
+      cf_name_, mutable_cf_options_, mutable_db_options_, vstorage_.get(),
+      &log_buffer_));
+  ASSERT_TRUE(compaction.get() != nullptr);
+  ASSERT_EQ(2U, compaction->num_input_levels());
+  ASSERT_EQ(1U, compaction->num_input_files(0));
+  ASSERT_EQ(1U, compaction->num_input_files(1));
+  ASSERT_EQ(2U, compaction->input(0, 0)->fd.GetNumber());
+  ASSERT_EQ(5U, compaction->input(1, 0)->fd.GetNumber());
+}
+
 TEST_F(CompactionPickerTest, IsTrivialMoveOn) {
  mutable_cf_options_.max_bytes_for_level_base = 10000u;
  mutable_cf_options_.max_compaction_bytes = 10001u;
@@ -4296,6 +4330,118 @@ TEST_F(CompactionPickerTest, IntraL0WhenL0IsSmall) {
  }
 }

+TEST_F(CompactionPickerTest, UniversalMaxReadAmpLargeDB) {
+  ioptions_.compaction_style = kCompactionStyleUniversal;
+  ioptions_.num_levels = 50;
+  mutable_cf_options_.RefreshDerivedOptions(ioptions_);
+  mutable_cf_options_.compaction_options_universal.size_ratio = 10;
+  mutable_cf_options_.write_buffer_size = 256 << 20;
+  // Avoid space amp compaction
+  mutable_cf_options_.compaction_options_universal
+      .max_size_amplification_percent = 200;
+  const int kMaxRuns = 8;
+  for (int max_read_amp : {kMaxRuns, 0, -1}) {
+    SCOPED_TRACE("max_read_amp = " + std::to_string(max_read_amp));
+    if (max_read_amp == -1) {
+      mutable_cf_options_.level0_file_num_compaction_trigger = kMaxRuns;
+    } else {
+      mutable_cf_options_.level0_file_num_compaction_trigger = 4;
+    }
+    mutable_cf_options_.compaction_options_universal.max_read_amp =
+        max_read_amp;
+    UniversalCompactionPicker universal_compaction_picker(ioptions_, &icmp_);
+    uint64_t max_run_size = 20ull << 30;
+    // When max_read_amp = 0, we estimate the number of levels needed based on
+    // size_ratio and write_buffer_size. See more in
+    // UniversalCompactionBuilder::PickCompaction().
+    // With a 20GB last level, we estimate that 8 levels are needed:
+    // L0 256MB
+    // L1 256MB * 1.1 (size_ratio) = 282MB
+    // L2 (256MB + 282MB) * 1.1 = 592MB
+    // L3 1243MB
+    // L4 2610MB
+    // L5 5481MB
+    // L6 11510MB
+    // L7 24171MB > 20GB
+    for (int i = 0; i <= kMaxRuns; ++i) {
+      SCOPED_TRACE("i = " + std::to_string(i));
+      NewVersionStorage(/*num_levels=*/50, kCompactionStyleUniversal);
+      Add(/*level=*/49, /*file_number=*/10, /*smallest=*/"100",
+          /*largest=*/"200", /*file_size=*/max_run_size, /*path_id=*/0,
+          /*smallest_seq=*/0, /*largest_seq=*/0,
+          /*compensated_file_size=*/max_run_size);
+      // Besides the last sorted run, we add additional `i` sorted runs
+      // without triggering space-amp or size-amp compactions.
+      uint64_t file_size = 1 << 20;
+      for (int j = 0; j < i; ++j) {
+        Add(/*level=*/j, /*file_number=*/100 - j, /*smallest=*/"100",
+            /*largest=*/"200", /*file_size=*/file_size, /*path_id=*/0,
+            /*smallest_seq=*/100 - j, /*largest_seq=*/100 - j,
+            /*compensated_file_size=*/file_size);
+        // to avoid space-amp and size-amp compaction
+        file_size *= 2;
+      }
+      UpdateVersionStorageInfo();
+      // level0_file_num_compaction_trigger is still used as trigger to
+      // check potential compactions
+      ASSERT_EQ(
+          universal_compaction_picker.NeedsCompaction(vstorage_.get()),
+          i + 1 >= mutable_cf_options_.level0_file_num_compaction_trigger);
+      std::unique_ptr<Compaction> compaction(
+          universal_compaction_picker.PickCompaction(
+              cf_name_, mutable_cf_options_, mutable_db_options_,
+              vstorage_.get(), &log_buffer_));
+      if (i == kMaxRuns) {
+        // There are in total i + 1 > kMaxRuns sorted runs.
+        // This triggers compaction ignoring size_ratio.
+        ASSERT_NE(nullptr, compaction);
+        ASSERT_EQ(CompactionReason::kUniversalSortedRunNum,
+                  compaction->compaction_reason());
+        // First two runs are compacted
+        ASSERT_EQ(0, compaction->start_level());
+        ASSERT_EQ(1, compaction->output_level());
+        ASSERT_EQ(1U, compaction->num_input_files(0));
+        ASSERT_EQ(1U, compaction->num_input_files(1));
+      } else {
+        ASSERT_EQ(nullptr, compaction);
+      }
+    }
+  }
+}
+
+TEST_F(CompactionPickerTest, UniversalMaxReadAmpSmallDB) {
+  ioptions_.compaction_style = kCompactionStyleUniversal;
+  ioptions_.num_levels = 50;
+  mutable_cf_options_.RefreshDerivedOptions(ioptions_);
+  mutable_cf_options_.level0_file_num_compaction_trigger = 1;
+  mutable_cf_options_.compaction_options_universal.size_ratio = 10;
+  mutable_cf_options_.write_buffer_size = 256 << 20;
+  mutable_cf_options_.compaction_options_universal
+      .max_size_amplification_percent = 200;
+  const int kMaxRuns = 1;
+  for (int max_read_amp : {-1, kMaxRuns, 0}) {
+    SCOPED_TRACE("max_read_amp = " + std::to_string(max_read_amp));
+    mutable_cf_options_.compaction_options_universal.max_read_amp =
+        max_read_amp;
+    UniversalCompactionPicker universal_compaction_picker(ioptions_, &icmp_);
+    NewVersionStorage(/*num_levels=*/50, kCompactionStyleUniversal);
+    // max_run_size is much smaller than write_buffer_size,
+    // only 1 level is needed.
+    uint64_t max_run_size = 8 << 10;
+    Add(/*level=*/49, /*file_number=*/10, /*smallest=*/"100",
+        /*largest=*/"200", /*file_size=*/max_run_size, /*path_id=*/0,
+        /*smallest_seq=*/0, /*largest_seq=*/0,
+        /*compensated_file_size=*/max_run_size);
+    UpdateVersionStorageInfo();
+    ASSERT_TRUE(universal_compaction_picker.NeedsCompaction(vstorage_.get()));
+    std::unique_ptr<Compaction> compaction(
+        universal_compaction_picker.PickCompaction(
+            cf_name_, mutable_cf_options_, mutable_db_options_, vstorage_.get(),
+            &log_buffer_));
+    ASSERT_EQ(nullptr, compaction);
+  }
+}
+
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
@@ -227,6 +227,7 @@ class UniversalCompactionBuilder {
  const InternalKeyComparator* icmp_;
  double score_;
  std::vector<SortedRun> sorted_runs_;
+  uint64_t max_run_size_;
  const std::string& cf_name_;
  const MutableCFOptions& mutable_cf_options_;
  const MutableDBOptions& mutable_db_options_;
@@ -235,7 +236,8 @@ class UniversalCompactionBuilder {
  LogBuffer* log_buffer_;

  static std::vector<UniversalCompactionBuilder::SortedRun> CalculateSortedRuns(
-      const VersionStorageInfo& vstorage, int last_level);
+      const VersionStorageInfo& vstorage, int last_level,
+      uint64_t* max_run_size);

  // Pick a path ID to place a newly generated file, with its estimated file
  // size.
@@ -440,11 +442,15 @@ void UniversalCompactionBuilder::SortedRun::DumpSizeInfo(

 std::vector<UniversalCompactionBuilder::SortedRun>
 UniversalCompactionBuilder::CalculateSortedRuns(
-    const VersionStorageInfo& vstorage, int last_level) {
+    const VersionStorageInfo& vstorage, int last_level,
+    uint64_t* max_run_size) {
+  assert(max_run_size);
+  *max_run_size = 0;
  std::vector<UniversalCompactionBuilder::SortedRun> ret;
  for (FileMetaData* f : vstorage.LevelFiles(0)) {
    ret.emplace_back(0, f, f->fd.GetFileSize(), f->compensated_file_size,
                     f->being_compacted);
+    *max_run_size = std::max(*max_run_size, f->fd.GetFileSize());
  }
  for (int level = 1; level <= last_level; level++) {
    uint64_t total_compensated_size = 0U;
@@ -466,6 +472,7 @@ UniversalCompactionBuilder::CalculateSortedRuns(
      ret.emplace_back(level, nullptr, total_size, total_compensated_size,
                       being_compacted);
    }
+    *max_run_size = std::max(*max_run_size, total_size);
  }
  return ret;
 }
@@ -477,13 +484,16 @@ Compaction* UniversalCompactionBuilder::PickCompaction() {
  score_ = vstorage_->CompactionScore(kLevel0);
  int max_output_level =
      vstorage_->MaxOutputLevel(ioptions_.allow_ingest_behind);
-  sorted_runs_ = CalculateSortedRuns(*vstorage_, max_output_level);
+  max_run_size_ = 0;
+  sorted_runs_ =
+      CalculateSortedRuns(*vstorage_, max_output_level, &max_run_size_);
+  int file_num_compaction_trigger =
+      mutable_cf_options_.level0_file_num_compaction_trigger;

  if (sorted_runs_.size() == 0 ||
      (vstorage_->FilesMarkedForPeriodicCompaction().empty() &&
       vstorage_->FilesMarkedForCompaction().empty() &&
-       sorted_runs_.size() < (unsigned int)mutable_cf_options_
-                                 .level0_file_num_compaction_trigger)) {
+       sorted_runs_.size() < (unsigned int)file_num_compaction_trigger)) {
    ROCKS_LOG_BUFFER(log_buffer_, "[%s] Universal: nothing to do\n",
                     cf_name_.c_str());
    TEST_SYNC_POINT_CALLBACK(
@@ -505,11 +515,9 @@ Compaction* UniversalCompactionBuilder::PickCompaction() {
    TEST_SYNC_POINT_CALLBACK("PostPickPeriodicCompaction", c);
  }

-  // Check for size amplification.
  if (c == nullptr &&
-      sorted_runs_.size() >=
-          static_cast<size_t>(
-              mutable_cf_options_.level0_file_num_compaction_trigger)) {
+      sorted_runs_.size() >= static_cast<size_t>(file_num_compaction_trigger)) {
+    // Check for size amplification.
    if ((c = PickCompactionToReduceSizeAmp()) != nullptr) {
      TEST_SYNC_POINT("PickCompactionToReduceSizeAmpReturnNonnullptr");
      ROCKS_LOG_BUFFER(log_buffer_, "[%s] Universal: compacting for size amp\n",
@@ -527,13 +535,48 @@ Compaction* UniversalCompactionBuilder::PickCompaction() {
                         cf_name_.c_str());
      } else {
        // Size amplification and file size ratios are within configured limits.
-        // If max read amplification is exceeding configured limits, then force
-        // compaction without looking at filesize ratios and try to reduce
-        // the number of files to fewer than level0_file_num_compaction_trigger.
+        // If max read amplification exceeds configured limits, then force
+        // compaction to reduce the number sorted runs without looking at file
+        // size ratios.
+
        // This is guaranteed by NeedsCompaction()
        assert(sorted_runs_.size() >=
-               static_cast<size_t>(
-                   mutable_cf_options_.level0_file_num_compaction_trigger));
+               static_cast<size_t>(file_num_compaction_trigger));
+        int max_num_runs =
+            mutable_cf_options_.compaction_options_universal.max_read_amp;
+        if (max_num_runs < 0) {
+          // any value < -1 is not valid
+          assert(max_num_runs == -1);
+          // By default, fall back to `level0_file_num_compaction_trigger`
+          max_num_runs = file_num_compaction_trigger;
+        } else if (max_num_runs == 0) {
+          if (mutable_cf_options_.compaction_options_universal.stop_style ==
+              kCompactionStopStyleTotalSize) {
+            // 0 means auto-tuning by RocksDB. We estimate max num run based on
+            // max_run_size, size_ratio and write buffer size:
+            // Assume the size of the lowest level size is equal to
+            // write_buffer_size. Each subsequent level is the max size without
+            // triggering size_ratio compaction. `max_num_runs` is the minimum
+            // number of levels required such that the target size of the
+            // largest level is at least `max_run_size_`.
+            max_num_runs = 1;
+            double cur_level_max_size =
+                static_cast<double>(mutable_cf_options_.write_buffer_size);
+            double total_run_size = 0;
+            while (cur_level_max_size < static_cast<double>(max_run_size_)) {
+              // This loop should not take too many iterations since
+              // cur_level_max_size at least doubles each iteration.
+              total_run_size += cur_level_max_size;
+              cur_level_max_size = (100.0 + ratio) / 100.0 * total_run_size;
+              ++max_num_runs;
+            }
+          } else {
+            // TODO: implement the auto-tune logic for this stop style
+            max_num_runs = file_num_compaction_trigger;
+          }
+        } else {
+          // max_num_runs > 0, it's the limit on the number of sorted run
+        }
        // Get the total number of sorted runs that are not being compacted
        int num_sr_not_compacted = 0;
        for (size_t i = 0; i < sorted_runs_.size(); i++) {
@@ -544,17 +587,25 @@ Compaction* UniversalCompactionBuilder::PickCompaction() {

        // The number of sorted runs that are not being compacted is greater
        // than the maximum allowed number of sorted runs
-        if (num_sr_not_compacted >
-            mutable_cf_options_.level0_file_num_compaction_trigger) {
-          unsigned int num_files =
-              num_sr_not_compacted -
-              mutable_cf_options_.level0_file_num_compaction_trigger + 1;
+        if (num_sr_not_compacted > max_num_runs) {
+          unsigned int num_files = num_sr_not_compacted - max_num_runs + 1;
          if ((c = PickCompactionToReduceSortedRuns(UINT_MAX, num_files)) !=
              nullptr) {
            ROCKS_LOG_BUFFER(log_buffer_,
-                             "[%s] Universal: compacting for file num -- %u\n",
-                             cf_name_.c_str(), num_files);
+                             "[%s] Universal: compacting for file num, to "
+                             "compact file num -- %u, max num runs allowed"
+                             "-- %d, max_run_size -- %" PRIu64 "\n",
+                             cf_name_.c_str(), num_files, max_num_runs,
+                             max_run_size_);
          }
+        } else {
+          ROCKS_LOG_BUFFER(
+              log_buffer_,
+              "[%s] Universal: skipping compaction for file num, num runs not "
+              "being compacted -- %u, max num runs allowed -- %d, max_run_size "
+              "-- %" PRIu64 "\n",
+              cf_name_.c_str(), num_sr_not_compacted, max_num_runs,
+              max_run_size_);
        }
      }
    }
@@ -1139,7 +1190,8 @@ Compaction* UniversalCompactionBuilder::PickIncrementalForReduceSizeAmp(
  // from bottom_start_idx and bottom_end_idx, but for now, we use
  // SetupOtherInputs() for simplicity.
  int parent_index = -1;  // Create and use bottom_start_idx?
-  if (!picker_->SetupOtherInputs(cf_name_, vstorage_, &second_last_level_inputs,
+  if (!picker_->SetupOtherInputs(cf_name_, mutable_cf_options_, vstorage_,
+                                 &second_last_level_inputs,
                                 &bottom_level_inputs, &parent_index,
                                 /*base_index=*/-1)) {
    return nullptr;
@@ -1310,8 +1362,9 @@ Compaction* UniversalCompactionBuilder::PickDeleteTriggeredCompaction() {
      int parent_index = -1;

      output_level_inputs.level = output_level;
-      if (!picker_->SetupOtherInputs(cf_name_, vstorage_, &start_level_inputs,
-                                     &output_level_inputs, &parent_index, -1)) {
+      if (!picker_->SetupOtherInputs(cf_name_, mutable_cf_options_, vstorage_,
+                                     &start_level_inputs, &output_level_inputs,
+                                     &parent_index, -1)) {
        return nullptr;
      }
      inputs.push_back(start_level_inputs);
@@ -13,6 +13,7 @@
 #include "rocksdb/iostats_context.h"
 #include "rocksdb/listener.h"
 #include "rocksdb/utilities/debug.h"
+#include "rocksdb/utilities/table_properties_collectors.h"
 #include "test_util/mock_time_env.h"
 #include "utilities/merge_operators.h"

@@ -1584,7 +1585,16 @@ TEST_F(PrecludeLastLevelTest, SmallPrecludeTime) {
  Close();
 }

-TEST_F(PrecludeLastLevelTest, FastTrackTimedPutToLastLevel) {
+// Test Param: protection_bytes_per_key for WriteBatch
+class TimedPutPrecludeLastLevelTest
+    : public PrecludeLastLevelTest,
+      public testing::WithParamInterface<size_t> {
+ public:
+  TimedPutPrecludeLastLevelTest()
+      : PrecludeLastLevelTest("timed_put_preclude_last_level_test") {}
+};
+
+TEST_P(TimedPutPrecludeLastLevelTest, FastTrackTimedPutToLastLevel) {
  const int kNumTrigger = 4;
  const int kNumLevels = 7;
  const int kNumKeys = 100;
@@ -1598,6 +1608,8 @@ TEST_F(PrecludeLastLevelTest, FastTrackTimedPutToLastLevel) {
  options.num_levels = kNumLevels;
  options.last_level_temperature = Temperature::kCold;
  DestroyAndReopen(options);
+  WriteOptions wo;
+  wo.protection_bytes_per_key = GetParam();

  Random rnd(301);

@@ -1606,7 +1618,7 @@ TEST_F(PrecludeLastLevelTest, FastTrackTimedPutToLastLevel) {
  });

  for (int i = 0; i < kNumKeys / 2; i++) {
-    ASSERT_OK(Put(Key(i), rnd.RandomString(100)));
+    ASSERT_OK(Put(Key(i), rnd.RandomString(100), wo));
    dbfull()->TEST_WaitForPeriodicTaskRun([&] {
      mock_clock_->MockSleepForSeconds(static_cast<int>(rnd.Uniform(2)));
    });
@@ -1620,7 +1632,7 @@ TEST_F(PrecludeLastLevelTest, FastTrackTimedPutToLastLevel) {
  // These data are eligible to be put on the last level once written to db
  // and compaction will fast track them to the last level.
  for (int i = kNumKeys / 2; i < kNumKeys; i++) {
-    ASSERT_OK(TimedPut(0, Key(i), rnd.RandomString(100), 50));
+    ASSERT_OK(TimedPut(0, Key(i), rnd.RandomString(100), 50, wo));
  }
  ASSERT_OK(Flush());

@@ -1640,6 +1652,168 @@ TEST_F(PrecludeLastLevelTest, FastTrackTimedPutToLastLevel) {
  Close();
 }

+TEST_P(TimedPutPrecludeLastLevelTest, InterleavedTimedPutAndPut) {
+  Options options = CurrentOptions();
+  options.compaction_style = kCompactionStyleUniversal;
+  options.disable_auto_compactions = true;
+  options.preclude_last_level_data_seconds = 1 * 24 * 60 * 60;
+  options.env = mock_env_.get();
+  options.num_levels = 7;
+  options.last_level_temperature = Temperature::kCold;
+  options.default_write_temperature = Temperature::kHot;
+  DestroyAndReopen(options);
+  WriteOptions wo;
+  wo.protection_bytes_per_key = GetParam();
+
+  // Start time: kMockStartTime = 10000000;
+  ASSERT_OK(TimedPut(0, Key(0), "v0", kMockStartTime - 1 * 24 * 60 * 60, wo));
+  ASSERT_OK(Put(Key(1), "v1", wo));
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  ASSERT_EQ("0,0,0,0,0,1,1", FilesPerLevel());
+  ASSERT_GT(GetSstSizeHelper(Temperature::kHot), 0);
+  ASSERT_GT(GetSstSizeHelper(Temperature::kCold), 0);
+  Close();
+}
+
+TEST_P(TimedPutPrecludeLastLevelTest, PreserveTimedPutOnPenultimateLevel) {
+  Options options = CurrentOptions();
+  options.compaction_style = kCompactionStyleUniversal;
+  options.disable_auto_compactions = true;
+  options.preclude_last_level_data_seconds = 3 * 24 * 60 * 60;
+  int seconds_between_recording = (3 * 24 * 60 * 60) / kMaxSeqnoTimePairsPerCF;
+  options.env = mock_env_.get();
+  options.num_levels = 7;
+  options.last_level_temperature = Temperature::kCold;
+  options.default_write_temperature = Temperature::kHot;
+  DestroyAndReopen(options);
+  WriteOptions wo;
+  wo.protection_bytes_per_key = GetParam();
+
+  // Creating a snapshot to manually control when preferred sequence number is
+  // swapped in. An entry's preferred seqno won't get swapped in until it's
+  // visible to the earliest snapshot. With this, we can test relevant seqno to
+  // time mapping recorded in SST file also covers preferred seqno, not just
+  // the seqno in the internal keys.
+  auto* snap1 = db_->GetSnapshot();
+  // Start time: kMockStartTime = 10000000;
+  ASSERT_OK(TimedPut(0, Key(0), "v0", kMockStartTime - 1 * 24 * 60 * 60, wo));
+  ASSERT_OK(TimedPut(0, Key(1), "v1", kMockStartTime - 1 * 24 * 60 * 60, wo));
+  ASSERT_OK(TimedPut(0, Key(2), "v2", kMockStartTime - 1 * 24 * 60 * 60, wo));
+  ASSERT_OK(Flush());
+
+  // Should still be in penultimate level.
+  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  ASSERT_EQ("0,0,0,0,0,1", FilesPerLevel());
+  ASSERT_GT(GetSstSizeHelper(Temperature::kHot), 0);
+  ASSERT_EQ(GetSstSizeHelper(Temperature::kCold), 0);
+
+  // Wait one more day and release snapshot. Data's preferred seqno should be
+  // swapped in, but data should still stay in penultimate level. SST file's
+  // seqno to time mapping should continue to cover preferred seqno after
+  // compaction.
+  db_->ReleaseSnapshot(snap1);
+  mock_clock_->MockSleepForSeconds(1 * 24 * 60 * 60);
+  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  ASSERT_EQ("0,0,0,0,0,1", FilesPerLevel());
+  ASSERT_GT(GetSstSizeHelper(Temperature::kHot), 0);
+  ASSERT_EQ(GetSstSizeHelper(Temperature::kCold), 0);
+
+  // Wait one more day and data are eligible to be placed on last level.
+  // Instead of waiting exactly one more day, here we waited
+  // `seconds_between_recording` less seconds to show that it's not precise.
+  // Data could start to be placed on cold tier one recording interval before
+  // they exactly become cold based on the setting. For this one column family
+  // setting preserving 3 days of recording, it's about 43 minutes.
+  mock_clock_->MockSleepForSeconds(1 * 24 * 60 * 60 -
+                                   seconds_between_recording);
+  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  ASSERT_EQ("0,0,0,0,0,0,1", FilesPerLevel());
+  ASSERT_EQ(GetSstSizeHelper(Temperature::kHot), 0);
+  ASSERT_GT(GetSstSizeHelper(Temperature::kCold), 0);
+  Close();
+}
+
+TEST_P(TimedPutPrecludeLastLevelTest, AutoTriggerCompaction) {
+  const int kNumTrigger = 10;
+  const int kNumLevels = 7;
+  const int kNumKeys = 200;
+
+  Options options = CurrentOptions();
+  options.compaction_style = kCompactionStyleUniversal;
+  options.preclude_last_level_data_seconds = 60;
+  options.preserve_internal_time_seconds = 0;
+  options.env = mock_env_.get();
+  options.level0_file_num_compaction_trigger = kNumTrigger;
+  options.num_levels = kNumLevels;
+  options.last_level_temperature = Temperature::kCold;
+  ConfigOptions config_options;
+  config_options.ignore_unsupported_options = false;
+  std::shared_ptr<TablePropertiesCollectorFactory> factory;
+  std::string id = CompactForTieringCollectorFactory::kClassName();
+  ASSERT_OK(TablePropertiesCollectorFactory::CreateFromString(
+      config_options, "compaction_trigger_ratio=0.4; id=" + id, &factory));
+  auto collector_factory =
+      factory->CheckedCast<CompactForTieringCollectorFactory>();
+  options.table_properties_collector_factories.push_back(factory);
+  DestroyAndReopen(options);
+  WriteOptions wo;
+  wo.protection_bytes_per_key = GetParam();
+
+  Random rnd(301);
+
+  dbfull()->TEST_WaitForPeriodicTaskRun([&] {
+    mock_clock_->MockSleepForSeconds(static_cast<int>(rnd.Uniform(10) + 1));
+  });
+
+  for (int i = 0; i < kNumKeys / 4; i++) {
+    ASSERT_OK(Put(Key(i), rnd.RandomString(100), wo));
+    dbfull()->TEST_WaitForPeriodicTaskRun([&] {
+      mock_clock_->MockSleepForSeconds(static_cast<int>(rnd.Uniform(2)));
+    });
+  }
+  // Create one file with regular Put.
+  ASSERT_OK(Flush());
+
+  // Create one file with TimedPut.
+  // These data are eligible to be put on the last level once written to db
+  // and compaction will fast track them to the last level.
+  for (int i = kNumKeys / 4; i < kNumKeys / 2; i++) {
+    ASSERT_OK(TimedPut(0, Key(i), rnd.RandomString(100), 50, wo));
+  }
+  ASSERT_OK(Flush());
+
+  // TimedPut file moved to the last level via auto triggered compaction.
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+  ASSERT_EQ("1,0,0,0,0,0,1", FilesPerLevel());
+  ASSERT_GT(GetSstSizeHelper(Temperature::kUnknown), 0);
+  ASSERT_GT(GetSstSizeHelper(Temperature::kCold), 0);
+
+  collector_factory->SetCompactionTriggerRatio(1.1);
+  for (int i = kNumKeys / 2; i < kNumKeys * 3 / 4; i++) {
+    ASSERT_OK(TimedPut(0, Key(i), rnd.RandomString(100), 50, wo));
+  }
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+  ASSERT_EQ("2,0,0,0,0,0,1", FilesPerLevel());
+
+  collector_factory->SetCompactionTriggerRatio(0);
+  for (int i = kNumKeys * 3 / 4; i < kNumKeys; i++) {
+    ASSERT_OK(TimedPut(0, Key(i), rnd.RandomString(100), 50, wo));
+  }
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+  ASSERT_EQ("3,0,0,0,0,0,1", FilesPerLevel());
+
+  Close();
+}
+
+INSTANTIATE_TEST_CASE_P(TimedPutPrecludeLastLevelTest,
+                        TimedPutPrecludeLastLevelTest, ::testing::Values(0, 8));
+
 TEST_F(PrecludeLastLevelTest, LastLevelOnlyCompactionPartial) {
  const int kNumTrigger = 4;
  const int kNumLevels = 7;
@@ -71,6 +71,9 @@ Status VerifySstFileChecksumInternal(const Options& options,
  } else {
    return s;
  }
+  if (!s.ok()) {
+    return s;
+  }
  std::unique_ptr<TableReader> table_reader;
  std::unique_ptr<RandomAccessFileReader> file_reader(
      new RandomAccessFileReader(
@@ -464,12 +464,14 @@ TEST_F(DBBasicTest, TimedPutBasic) {
    ASSERT_EQ("bv1", Get(1, "bar"));
    ASSERT_OK(TimedPut(1, "baz", "bzv1", /*write_unix_time=*/0));
    ASSERT_EQ("bzv1", Get(1, "baz"));
-    std::string range_del_begin = "b";
-    std::string range_del_end = "baz";
-    Slice begin_rdel = range_del_begin, end_rdel = range_del_end;
-    ASSERT_OK(
-        db_->DeleteRange(WriteOptions(), handles_[1], begin_rdel, end_rdel));
-    ASSERT_EQ("NOT_FOUND", Get(1, "bar"));
+    if (option_config_ != kRowCache) {
+      std::string range_del_begin = "b";
+      std::string range_del_end = "baz";
+      Slice begin_rdel = range_del_begin, end_rdel = range_del_end;
+      ASSERT_OK(
+          db_->DeleteRange(WriteOptions(), handles_[1], begin_rdel, end_rdel));
+      ASSERT_EQ("NOT_FOUND", Get(1, "bar"));
+    }

    ASSERT_EQ("bzv1", Get(1, "baz"));
    ASSERT_OK(SingleDelete(1, "baz"));
@@ -1413,7 +1415,7 @@ TEST_P(DBMultiGetTestWithParam, MultiGetMultiCF) {
  int get_sv_count = 0;
  ROCKSDB_NAMESPACE::DBImpl* db = static_cast_with_check<DBImpl>(db_);
  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
-      "DBImpl::MultiGet::AfterRefSV", [&](void* /*arg*/) {
+      "DBImpl::MultiCFSnapshot::AfterRefSV", [&](void* /*arg*/) {
        if (++get_sv_count == 2) {
          // After MultiGet refs a couple of CFs, flush all CFs so MultiGet
          // is forced to repeat the process
@@ -1513,9 +1515,10 @@ TEST_P(DBMultiGetTestWithParam, MultiGetMultiCFMutex) {
  int retries = 0;
  bool last_try = false;
  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
-      "DBImpl::MultiGet::LastTry", [&](void* /*arg*/) { last_try = true; });
+      "DBImpl::MultiCFSnapshot::LastTry",
+      [&](void* /*arg*/) { last_try = true; });
  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
-      "DBImpl::MultiGet::AfterRefSV", [&](void* /*arg*/) {
+      "DBImpl::MultiCFSnapshot::AfterRefSV", [&](void* /*arg*/) {
        if (last_try) {
          return;
        }
@@ -1531,10 +1534,10 @@ TEST_P(DBMultiGetTestWithParam, MultiGetMultiCFMutex) {
        }
      });
  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({
-      {"DBImpl::MultiGet::AfterLastTryRefSV",
+      {"DBImpl::MultiCFSnapshot::AfterLastTryRefSV",
       "DBMultiGetTestWithParam::MultiGetMultiCFMutex:BeforeCreateSV"},
      {"DBMultiGetTestWithParam::MultiGetMultiCFMutex:AfterCreateSV",
-       "DBImpl::MultiGet::BeforeLastTryUnRefSV"},
+       "DBImpl::MultiCFSnapshot::BeforeLastTryUnRefSV"},
  });
  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();

@@ -1600,7 +1603,7 @@ TEST_P(DBMultiGetTestWithParam, MultiGetMultiCFSnapshot) {
  int get_sv_count = 0;
  ROCKSDB_NAMESPACE::DBImpl* db = static_cast_with_check<DBImpl>(db_);
  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
-      "DBImpl::MultiGet::AfterRefSV", [&](void* /*arg*/) {
+      "DBImpl::MultiCFSnapshot::AfterRefSV", [&](void* /*arg*/) {
        if (++get_sv_count == 2) {
          for (int i = 0; i < 8; ++i) {
            ASSERT_OK(Flush(i));
@@ -26,6 +26,7 @@
 #include "rocksdb/table_properties.h"
 #include "table/block_based/block_based_table_reader.h"
 #include "table/unique_id_impl.h"
+#include "test_util/secondary_cache_test_util.h"
 #include "util/compression.h"
 #include "util/defer.h"
 #include "util/hash.h"
@@ -740,118 +741,6 @@ class LookupLiarCache : public CacheWrapper {

 }  // anonymous namespace

-TEST_F(DBBlockCacheTest, AddRedundantStats) {
-  const size_t capacity = size_t{1} << 25;
-  const int num_shard_bits = 0;  // 1 shard
-  int iterations_tested = 0;
-  for (const std::shared_ptr<Cache>& base_cache :
-       {NewLRUCache(capacity, num_shard_bits),
-        // FixedHyperClockCache
-        HyperClockCacheOptions(
-            capacity,
-            BlockBasedTableOptions().block_size /*estimated_value_size*/,
-            num_shard_bits)
-            .MakeSharedCache(),
-        // AutoHyperClockCache
-        HyperClockCacheOptions(capacity, 0 /*estimated_value_size*/,
-                               num_shard_bits)
-            .MakeSharedCache()}) {
-    if (!base_cache) {
-      // Skip clock cache when not supported
-      continue;
-    }
-    ++iterations_tested;
-    Options options = CurrentOptions();
-    options.create_if_missing = true;
-    options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
-
-    std::shared_ptr<LookupLiarCache> cache =
-        std::make_shared<LookupLiarCache>(base_cache);
-
-    BlockBasedTableOptions table_options;
-    table_options.cache_index_and_filter_blocks = true;
-    table_options.block_cache = cache;
-    table_options.filter_policy.reset(NewBloomFilterPolicy(50));
-    options.table_factory.reset(NewBlockBasedTableFactory(table_options));
-    DestroyAndReopen(options);
-
-    // Create a new table.
-    ASSERT_OK(Put("foo", "value"));
-    ASSERT_OK(Put("bar", "value"));
-    ASSERT_OK(Flush());
-    ASSERT_EQ(1, NumTableFilesAtLevel(0));
-
-    // Normal access filter+index+data.
-    ASSERT_EQ("value", Get("foo"));
-
-    ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD));
-    ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD));
-    ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD));
-    // --------
-    ASSERT_EQ(3, TestGetTickerCount(options, BLOCK_CACHE_ADD));
-
-    ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD_REDUNDANT));
-    ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD_REDUNDANT));
-    ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD_REDUNDANT));
-    // --------
-    ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_ADD_REDUNDANT));
-
-    // Againt access filter+index+data, but force redundant load+insert on index
-    cache->SetNthLookupNotFound(2);
-    ASSERT_EQ("value", Get("bar"));
-
-    ASSERT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD));
-    ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD));
-    ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD));
-    // --------
-    ASSERT_EQ(4, TestGetTickerCount(options, BLOCK_CACHE_ADD));
-
-    ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD_REDUNDANT));
-    ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD_REDUNDANT));
-    ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD_REDUNDANT));
-    // --------
-    ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_ADD_REDUNDANT));
-
-    // Access just filter (with high probability), and force redundant
-    // load+insert
-    cache->SetNthLookupNotFound(1);
-    ASSERT_EQ("NOT_FOUND", Get("this key was not added"));
-
-    EXPECT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD));
-    EXPECT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD));
-    EXPECT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD));
-    // --------
-    EXPECT_EQ(5, TestGetTickerCount(options, BLOCK_CACHE_ADD));
-
-    EXPECT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD_REDUNDANT));
-    EXPECT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD_REDUNDANT));
-    EXPECT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD_REDUNDANT));
-    // --------
-    EXPECT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_ADD_REDUNDANT));
-
-    // Access just data, forcing redundant load+insert
-    ReadOptions read_options;
-    std::unique_ptr<Iterator> iter{db_->NewIterator(read_options)};
-    cache->SetNthLookupNotFound(1);
-    iter->SeekToFirst();
-    ASSERT_TRUE(iter->Valid());
-    ASSERT_EQ(iter->key(), "bar");
-
-    EXPECT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD));
-    EXPECT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD));
-    EXPECT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD));
-    // --------
-    EXPECT_EQ(6, TestGetTickerCount(options, BLOCK_CACHE_ADD));
-
-    EXPECT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD_REDUNDANT));
-    EXPECT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD_REDUNDANT));
-    EXPECT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD_REDUNDANT));
-    // --------
-    EXPECT_EQ(3, TestGetTickerCount(options, BLOCK_CACHE_ADD_REDUNDANT));
-  }
-  EXPECT_GE(iterations_tested, 1);
-}
-
 TEST_F(DBBlockCacheTest, ParanoidFileChecks) {
  Options options = CurrentOptions();
  options.create_if_missing = true;
@@ -1347,6 +1236,274 @@ TEST_F(DBBlockCacheTest, HyperClockCacheReportProblems) {
  EXPECT_EQ(logger->PopCounts(), (std::array<int, 3>{{0, 1, 0}}));
 }

+class DBBlockCacheTypeTest
+    : public DBBlockCacheTest,
+      public secondary_cache_test_util::WithCacheTypeParam {};
+
+INSTANTIATE_TEST_CASE_P(DBBlockCacheTypeTestInstance, DBBlockCacheTypeTest,
+                        secondary_cache_test_util::GetTestingCacheTypes());
+
+TEST_P(DBBlockCacheTypeTest, AddRedundantStats) {
+  BlockBasedTableOptions table_options;
+
+  const size_t capacity = size_t{1} << 25;
+  const int num_shard_bits = 0;  // 1 shard
+  estimated_value_size_ = table_options.block_size;
+  std::shared_ptr<Cache> base_cache =
+      NewCache(capacity, num_shard_bits, /*strict_capacity_limit=*/false);
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
+
+  std::shared_ptr<LookupLiarCache> cache =
+      std::make_shared<LookupLiarCache>(base_cache);
+
+  table_options.cache_index_and_filter_blocks = true;
+  table_options.block_cache = cache;
+  table_options.filter_policy.reset(NewBloomFilterPolicy(50));
+  options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+  DestroyAndReopen(options);
+
+  // Create a new table.
+  ASSERT_OK(Put("foo", "value"));
+  ASSERT_OK(Put("bar", "value"));
+  ASSERT_OK(Flush());
+  ASSERT_EQ(1, NumTableFilesAtLevel(0));
+
+  // Normal access filter+index+data.
+  ASSERT_EQ("value", Get("foo"));
+
+  ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD));
+  ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD));
+  ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD));
+  // --------
+  ASSERT_EQ(3, TestGetTickerCount(options, BLOCK_CACHE_ADD));
+
+  ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD_REDUNDANT));
+  ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD_REDUNDANT));
+  ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD_REDUNDANT));
+  // --------
+  ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_ADD_REDUNDANT));
+
+  // Againt access filter+index+data, but force redundant load+insert on index
+  cache->SetNthLookupNotFound(2);
+  ASSERT_EQ("value", Get("bar"));
+
+  ASSERT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD));
+  ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD));
+  ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD));
+  // --------
+  ASSERT_EQ(4, TestGetTickerCount(options, BLOCK_CACHE_ADD));
+
+  ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD_REDUNDANT));
+  ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD_REDUNDANT));
+  ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD_REDUNDANT));
+  // --------
+  ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_ADD_REDUNDANT));
+
+  // Access just filter (with high probability), and force redundant
+  // load+insert
+  cache->SetNthLookupNotFound(1);
+  ASSERT_EQ("NOT_FOUND", Get("this key was not added"));
+
+  EXPECT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD));
+  EXPECT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD));
+  EXPECT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD));
+  // --------
+  EXPECT_EQ(5, TestGetTickerCount(options, BLOCK_CACHE_ADD));
+
+  EXPECT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD_REDUNDANT));
+  EXPECT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD_REDUNDANT));
+  EXPECT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD_REDUNDANT));
+  // --------
+  EXPECT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_ADD_REDUNDANT));
+
+  // Access just data, forcing redundant load+insert
+  ReadOptions read_options;
+  std::unique_ptr<Iterator> iter{db_->NewIterator(read_options)};
+  cache->SetNthLookupNotFound(1);
+  iter->SeekToFirst();
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ(iter->key(), "bar");
+
+  EXPECT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD));
+  EXPECT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD));
+  EXPECT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD));
+  // --------
+  EXPECT_EQ(6, TestGetTickerCount(options, BLOCK_CACHE_ADD));
+
+  EXPECT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_ADD_REDUNDANT));
+  EXPECT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_ADD_REDUNDANT));
+  EXPECT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_DATA_ADD_REDUNDANT));
+  // --------
+  EXPECT_EQ(3, TestGetTickerCount(options, BLOCK_CACHE_ADD_REDUNDANT));
+}
+
+namespace {
+std::string AltKey(int i) {
+  char buf[100];
+  snprintf(buf, sizeof(buf), "altkey%06d", i);
+  return std::string(buf);
+}
+}  // namespace
+
+TEST_P(DBBlockCacheTypeTest, Uncache) {
+  for (bool partitioned : {false, true}) {
+    SCOPED_TRACE("partitioned=" + std::to_string(partitioned));
+    for (uint32_t ua : {0, 1, 2, 10000}) {
+      SCOPED_TRACE("ua=" + std::to_string(ua));
+
+      BlockBasedTableOptions table_options;
+      Options options = CurrentOptions();
+      options.uncache_aggressiveness = ua;
+      options.create_if_missing = true;
+      // Don't allow background operations to keep Versions referenced
+      options.stats_dump_period_sec = 0;
+      options.stats_persist_period_sec = 0;
+      auto stats = ROCKSDB_NAMESPACE::CreateDBStatistics();
+      options.statistics = stats;
+
+      const size_t capacity = size_t{1} << 25;
+      const int num_shard_bits = 0;  // 1 shard
+      estimated_value_size_ = table_options.block_size;
+      std::shared_ptr<Cache> cache =
+          NewCache(capacity, num_shard_bits, /*strict_capacity_limit=*/false);
+
+      table_options.cache_index_and_filter_blocks = true;
+      table_options.block_cache = cache;
+      table_options.filter_policy.reset(NewBloomFilterPolicy(10));
+      table_options.partition_filters = partitioned;
+      table_options.index_type =
+          partitioned ? BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch
+                      : BlockBasedTableOptions::IndexType::kBinarySearch;
+      options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+      DestroyAndReopen(options);
+
+      size_t kBaselineCount = 1;  // Because of entry stats collector
+
+      ASSERT_EQ(kBaselineCount, cache->GetOccupancyCount());
+      ASSERT_EQ(0U, cache->GetUsage());
+
+      constexpr uint8_t kNumDataBlocks = 10;
+      constexpr uint8_t kNumFiles = 3;
+      for (int i = 0; i < kNumDataBlocks; i++) {
+        // Force some overlap with ordering
+        ASSERT_OK(Put(Key((i * 7) % kNumDataBlocks),
+                      Random::GetTLSInstance()->RandomBinaryString(
+                          static_cast<int>(table_options.block_size))));
+        if (i >= kNumDataBlocks - kNumFiles) {
+          ASSERT_OK(Flush());
+        }
+      }
+      ASSERT_EQ(int{kNumFiles}, NumTableFilesAtLevel(0));
+
+      for (int i = 0; i < kNumDataBlocks; i++) {
+        ASSERT_NE(Get(Key(i)), "NOT_FOUND");
+      }
+
+      size_t meta_blocks_per_file = /*index & filter*/ 2U * (1U + partitioned);
+      ASSERT_EQ(
+          cache->GetOccupancyCount(),
+          kBaselineCount + kNumDataBlocks + meta_blocks_per_file * kNumFiles);
+      ASSERT_GE(cache->GetUsage(), kNumDataBlocks * table_options.block_size);
+
+      // Combine into one file, making the originals obsolete
+      ASSERT_OK(db_->CompactRange({}, nullptr, nullptr));
+      ASSERT_OK(dbfull()->TEST_WaitForBackgroundWork());
+
+      ASSERT_EQ(1, NumTableFilesAtLevel(1));
+
+      for (int i = 0; i < kNumDataBlocks; i++) {
+        ASSERT_NE(Get(Key(i)), "NOT_FOUND");
+      }
+
+      if (ua == 0) {
+        // Expect to see cache entries for new file and obsolete files
+        EXPECT_EQ(cache->GetOccupancyCount(),
+                  kBaselineCount + kNumDataBlocks * 2U +
+                      meta_blocks_per_file * (kNumFiles + 1));
+        EXPECT_GE(cache->GetUsage(),
+                  kNumDataBlocks * table_options.block_size * 2U);
+      } else {
+        // Expect only to see cache entries for new file
+        EXPECT_EQ(cache->GetOccupancyCount(),
+                  kBaselineCount + kNumDataBlocks + meta_blocks_per_file);
+        EXPECT_GE(cache->GetUsage(), kNumDataBlocks * table_options.block_size);
+        EXPECT_LT(cache->GetUsage(),
+                  kNumDataBlocks * table_options.block_size * 2U);
+      }
+
+      size_t alt_baseline_count = cache->GetOccupancyCount();
+      size_t alt_baseline_usage = cache->GetUsage();
+      ASSERT_OK(stats->Reset());
+      // We aren't generally cleaning up cache entries on DB::Close, especially
+      // because someone might just re-open the same DB.
+      Reopen(options);
+      for (int i = 0; i < kNumDataBlocks; i++) {
+        ASSERT_NE(Get(Key(i)), "NOT_FOUND");
+      }
+
+      EXPECT_EQ(cache->GetOccupancyCount(), alt_baseline_count);
+      EXPECT_EQ(cache->GetUsage(), alt_baseline_usage);
+
+      // Check for unnecessary unncessary cache churn
+      ASSERT_EQ(stats->getTickerCount(BLOCK_CACHE_ADD), 0U);
+      ASSERT_EQ(stats->getTickerCount(BLOCK_CACHE_MISS), 0U);
+      ASSERT_GT(stats->getTickerCount(BLOCK_CACHE_HIT), 0U);
+
+      // And now do a similar test as above except with trivial moves, making
+      // sure that we aren't falsely uncaching in that case, which would cause
+      // unnecessary cache misses. Using AltKey instead of Key to avoid
+      // interference.
+      for (int i = 0; i < kNumDataBlocks; i++) {
+        // No overlap
+        ASSERT_OK(
+            Put(AltKey(i), Random::GetTLSInstance()->RandomBinaryString(
+                               static_cast<int>(table_options.block_size))));
+        if (i >= kNumDataBlocks - kNumFiles) {
+          ASSERT_OK(Flush());
+        }
+      }
+      ASSERT_EQ(int{kNumFiles}, NumTableFilesAtLevel(0));
+
+      for (int i = 0; i < kNumDataBlocks; i++) {
+        ASSERT_NE(Get(AltKey(i)), "NOT_FOUND");
+      }
+
+      ASSERT_EQ(cache->GetOccupancyCount(),
+                alt_baseline_count + kNumDataBlocks +
+                    meta_blocks_per_file * kNumFiles);
+      ASSERT_GE(cache->GetUsage(),
+                alt_baseline_usage + kNumDataBlocks * table_options.block_size);
+
+      ASSERT_OK(stats->Reset());
+
+      // Make trivial move
+      {
+        auto a = AltKey(0);
+        auto b = AltKey(kNumDataBlocks);
+        Slice slice_a{a};
+        Slice slice_b{b};
+        ASSERT_OK(db_->CompactRange({}, &slice_a, &slice_b));
+      }
+      ASSERT_EQ(/*old*/ 1 + /*new*/ int{kNumFiles}, NumTableFilesAtLevel(1));
+
+      for (int i = 0; i < kNumDataBlocks; i++) {
+        ASSERT_NE(Get(AltKey(i)), "NOT_FOUND");
+      }
+
+      // Should be the same if trivial move
+      ASSERT_EQ(cache->GetOccupancyCount(),
+                alt_baseline_count + kNumDataBlocks +
+                    meta_blocks_per_file * kNumFiles);
+
+      // Check for unnecessary unncessary cache churn
+      ASSERT_EQ(stats->getTickerCount(BLOCK_CACHE_ADD), 0U);
+      ASSERT_EQ(stats->getTickerCount(BLOCK_CACHE_MISS), 0U);
+      ASSERT_GT(stats->getTickerCount(BLOCK_CACHE_HIT), 0U);
+    }
+  }
+}

 class DBBlockCacheKeyTest
    : public DBTestBase,
@@ -529,6 +529,7 @@ TEST_P(DBBloomFilterTestWithParam, BloomFilter) {
    BlockBasedTableOptions table_options;
    table_options.no_block_cache = true;
    table_options.filter_policy = Create(10, bfp_impl_);
+    table_options.optimize_filters_for_memory = false;
    table_options.partition_filters = partition_filters_;
    if (partition_filters_) {
      table_options.index_type =
@@ -1695,6 +1696,7 @@ TEST_F(DBBloomFilterTest, ContextCustomFilterPolicy) {
    BlockBasedTableOptions table_options;
    table_options.filter_policy = policy;
    table_options.format_version = 5;
+    table_options.optimize_filters_for_memory = false;
    options.table_factory.reset(NewBlockBasedTableFactory(table_options));

    ASSERT_OK(TryReopen(options));
@@ -16,6 +16,7 @@
 #include "env/mock_env.h"
 #include "port/port.h"
 #include "port/stack_trace.h"
+#include "rocksdb/advanced_options.h"
 #include "rocksdb/concurrent_task_limiter.h"
 #include "rocksdb/experimental.h"
 #include "rocksdb/sst_file_writer.h"
@@ -3997,7 +3998,7 @@ TEST_P(DBCompactionTestWithParam, FullCompactionInBottomPriThreadPool) {
  Env::Default()->SetBackgroundThreads(0, Env::Priority::BOTTOM);
 }

-TEST_F(DBCompactionTest, CancelCompactionWaitingOnConflict) {
+TEST_F(DBCompactionTest, CancelCompactionWaitingOnRunningConflict) {
  // This test verifies cancellation of a compaction waiting to be scheduled due
  // to conflict with a running compaction.
  //
@@ -4036,7 +4037,7 @@ TEST_F(DBCompactionTest, CancelCompactionWaitingOnConflict) {
  // Make sure the manual compaction has seen the conflict before being canceled
  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
      {{"ColumnFamilyData::CompactRange:Return",
-        "DBCompactionTest::CancelCompactionWaitingOnConflict:"
+        "DBCompactionTest::CancelCompactionWaitingOnRunningConflict:"
        "PreDisableManualCompaction"}});
  auto manual_compaction_thread = port::Thread([this]() {
    ASSERT_TRUE(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr)
@@ -4047,12 +4048,73 @@ TEST_F(DBCompactionTest, CancelCompactionWaitingOnConflict) {
  // despite finding a conflict with an automatic compaction that is still
  // running
  TEST_SYNC_POINT(
-      "DBCompactionTest::CancelCompactionWaitingOnConflict:"
+      "DBCompactionTest::CancelCompactionWaitingOnRunningConflict:"
      "PreDisableManualCompaction");
  db_->DisableManualCompaction();
  manual_compaction_thread.join();
 }

+TEST_F(DBCompactionTest, CancelCompactionWaitingOnScheduledConflict) {
+  // This test verifies cancellation of a compaction waiting to be scheduled due
+  // to conflict with a scheduled (but not running) compaction.
+  //
+  // A `CompactRange()` in universal compacts all files, waiting for files to
+  // become available if they are locked for another compaction. This test
+  // blocks the compaction thread pool and then calls `CompactRange()` twice.
+  // The first call to `CompactRange()` schedules a compaction that is queued
+  // in the thread pool. The second call to `CompactRange()` blocks on the first
+  // call due to the conflict in file picking. The test verifies that
+  // `DisableManualCompaction()` can cancel both while the thread pool remains
+  // blocked.
+  const int kNumSortedRuns = 4;
+
+  Options options = CurrentOptions();
+  options.compaction_style = kCompactionStyleUniversal;
+  options.disable_auto_compactions = true;
+  options.memtable_factory.reset(
+      test::NewSpecialSkipListFactory(KNumKeysByGenerateNewFile - 1));
+  Reopen(options);
+
+  test::SleepingBackgroundTask sleeping_task_low;
+  env_->Schedule(&test::SleepingBackgroundTask::DoSleepTask, &sleeping_task_low,
+                 Env::Priority::LOW);
+
+  // Fill overlapping files in L0
+  Random rnd(301);
+  for (int i = 0; i < kNumSortedRuns; ++i) {
+    int key_idx = 0;
+    GenerateNewFile(&rnd, &key_idx, false /* nowait */);
+  }
+
+  std::atomic<int> num_compact_range_calls{0};
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "ColumnFamilyData::CompactRange:Return",
+      [&](void* /* arg */) { num_compact_range_calls++; });
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  const int kNumManualCompactions = 2;
+  port::Thread manual_compaction_threads[kNumManualCompactions];
+  for (int i = 0; i < kNumManualCompactions; i++) {
+    manual_compaction_threads[i] = port::Thread([this]() {
+      ASSERT_TRUE(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr)
+                      .IsIncomplete());
+    });
+  }
+  while (num_compact_range_calls < kNumManualCompactions) {
+  }
+
+  // Cancel it. Threads should be joinable, i.e., both the scheduled and blocked
+  // manual compactions were canceled despite no compaction could have ever run.
+  db_->DisableManualCompaction();
+  for (int i = 0; i < kNumManualCompactions; i++) {
+    manual_compaction_threads[i].join();
+  }
+
+  sleeping_task_low.WakeUp();
+  sleeping_task_low.WaitUntilDone();
+}
+
 TEST_F(DBCompactionTest, OptimizedDeletionObsoleting) {
  // Deletions can be dropped when compacted to non-last level if they fall
  // outside the lower-level files' key-ranges.
@@ -6082,6 +6144,14 @@ TEST_F(DBCompactionTest, CompactionLimiter) {
        }
      });

+  std::vector<std::string> pending_compaction_cfs;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "SchedulePendingCompaction::cfd", [&](void* arg) {
+        const std::string& cf_name =
+            static_cast<ColumnFamilyData*>(arg)->GetName();
+        pending_compaction_cfs.emplace_back(cf_name);
+      });
+
  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();

  // Block all compact threads in thread pool.
@@ -6135,8 +6205,15 @@ TEST_F(DBCompactionTest, CompactionLimiter) {
              NumTableFilesAtLevel(0, 0));
  }

-  // All CFs are pending compaction
-  ASSERT_EQ(cf_count, env_->GetThreadPoolQueueLen(Env::LOW));
+  // Wait until all CFs are pending compaction. WaitForFlushMemtable() can
+  // return before the next compaction is scheduled, so we need to do some
+  // waiting here.
+  unsigned int tp_len = env_->GetThreadPoolQueueLen(Env::LOW);
+  for (int i = 0; i < 10000 && tp_len < cf_count; i++) {
+    env_->SleepForMicroseconds(1000);
+    tp_len = env_->GetThreadPoolQueueLen(Env::LOW);
+  }
+  ASSERT_EQ(cf_count, tp_len);

  // Unblock all compaction threads
  for (size_t i = 0; i < kTotalCompactTasks; i++) {
@@ -7044,7 +7121,8 @@ class DBCompactionTestWithOngoingFileIngestionParam
      SyncPoint::GetInstance()->SetCallBack(
          "ExternalSstFileIngestionJob::Run", [&](void*) {
            SyncPoint::GetInstance()->LoadDependency(
-                {{"DBImpl::CompactFilesImpl::PostSanitizeCompactionInputFiles",
+                {{"DBImpl::CompactFilesImpl::"
+                  "PostSanitizeAndConvertCompactionInputFiles",
                  "VersionSet::LogAndApply:WriteManifest"}});
          });
    } else {
@@ -7526,6 +7604,72 @@ class DBCompactionTestL0FilesMisorderCorruption : public DBCompactionTest {
  std::shared_ptr<test::SleepingBackgroundTask> sleeping_task_;
 };

+TEST_F(DBCompactionTest, CompactFilesSupportKeyPlacementRangeConflict) {
+  Options options;
+  options.create_if_missing = true;
+  options.compaction_style = kCompactionStyleUniversal;
+  options.num_levels = 3;
+  DestroyAndReopen(options);
+
+  // To create LSM of below shape:
+  // L0: [k2]
+  // L1: [k3],[k4]
+  // L2: [k1,  k5]
+  ASSERT_OK(Put("k1", "v"));
+  ASSERT_OK(Put("k5", "v"));
+  ASSERT_OK(Flush());
+  CompactRangeOptions cro;
+  cro.bottommost_level_compaction = BottommostLevelCompaction::kForceOptimized;
+  ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
+  ASSERT_EQ("0,0,1", FilesPerLevel());
+
+  ASSERT_OK(Put("k3", "v"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("k4", "v"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(experimental::PromoteL0(db_, db_->DefaultColumnFamily(), 1));
+  ASSERT_EQ("0,2,1", FilesPerLevel());
+
+  ASSERT_OK(Put("k2", "v"));
+  ASSERT_OK(Flush());
+  ASSERT_EQ("1,2,1", FilesPerLevel());
+
+  Close();
+
+  // To force below two CompactFiles() in order to coerce range conflict on L1
+  // upon (2)
+  // (1): Compact [k2] at L0 and [k3] at L1 with output to L1
+  // (2): Compact [k4] at L1 and [k1, k5] at L2 and output to L1 and L2
+  options.preclude_last_level_data_seconds = 1;
+  Reopen(options);
+
+  ColumnFamilyMetaData cf_meta_data;
+  db_->GetColumnFamilyMetaData(&cf_meta_data);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "CompactFilesImpl:0", [&](void* /*arg*/) {
+        std::vector<std::string> c2_input_files;
+        c2_input_files.push_back(cf_meta_data.levels[1].files[1].name);
+        c2_input_files.push_back(cf_meta_data.levels[2].files[0].name);
+        // To verify CompactFiles() is aborted upon range conflict instead
+        // of crashing upon internal assertion
+        Status s = db_->CompactFiles(CompactionOptions(), c2_input_files,
+                                     2 /* output_level */);
+        ASSERT_TRUE(s.IsAborted());
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  std::vector<std::string> c1_input_files;
+  c1_input_files.push_back(cf_meta_data.levels[0].files[0].name);
+  c1_input_files.push_back(cf_meta_data.levels[1].files[0].name);
+  Status s = db_->CompactFiles(CompactionOptions(), c1_input_files,
+                               1 /* output_level */);
+  ASSERT_OK(s);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
 TEST_F(DBCompactionTestL0FilesMisorderCorruption,
       FlushAfterIntraL0LevelCompactionWithIngestedFile) {
  SetupOptions(CompactionStyle::kCompactionStyleLevel, "");
@@ -9890,6 +10034,62 @@ TEST_F(DBCompactionTest, TurnOnLevelCompactionDynamicLevelBytesUCToLC) {
  ASSERT_EQ(expected_lsm, FilesPerLevel(1));
 }

+TEST_F(DBCompactionTest, DisallowRefitFilesFromNonL0ToL02) {
+  Options options = CurrentOptions();
+  options.compaction_style = CompactionStyle::kCompactionStyleLevel;
+  options.num_levels = 3;
+  DestroyAndReopen(options);
+
+  // To set up LSM shape:
+  // L0
+  // L1
+  // L2:[a@1, k@3], [k@2, z@4] (sorted by ascending smallest key)
+  // Both of these 2 files have epoch number = 1
+  const Snapshot* s1 = db_->GetSnapshot();
+  ASSERT_OK(Put("a", "@1"));
+  ASSERT_OK(Put("k", "@2"));
+  const Snapshot* s2 = db_->GetSnapshot();
+  ASSERT_OK(Put("k", "@3"));
+  ASSERT_OK(Put("z", "v3"));
+  ASSERT_OK(Flush());
+  // Cut file between k@3 and k@2
+  SyncPoint::GetInstance()->SetCallBack(
+      "CompactionOutputs::ShouldStopBefore::manual_decision",
+      [options](void* p) {
+        auto* pair = (std::pair<bool*, const Slice>*)p;
+        if ((options.comparator->Compare(ExtractUserKey(pair->second), "k") ==
+             0) &&
+            (GetInternalKeySeqno(pair->second) == 2)) {
+          *(pair->first) = true;
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+  CompactRangeOptions cro;
+  cro.bottommost_level_compaction = BottommostLevelCompaction::kForceOptimized;
+  cro.change_level = true;
+  cro.target_level = 2;
+  Status s = dbfull()->CompactRange(cro, nullptr, nullptr);
+  ASSERT_OK(s);
+  ASSERT_EQ("0,0,2", FilesPerLevel());
+  std::vector<LiveFileMetaData> files;
+  dbfull()->GetLiveFilesMetaData(&files);
+  ASSERT_EQ(files.size(), 2);
+  ASSERT_EQ(files[0].smallestkey, "a");
+  ASSERT_EQ(files[0].largestkey, "k");
+  ASSERT_EQ(files[1].smallestkey, "k");
+  ASSERT_EQ(files[1].largestkey, "z");
+
+  // Disallow moving 2 non-L0 files to L0
+  CompactRangeOptions cro2;
+  cro2.change_level = true;
+  cro2.target_level = 0;
+  s = dbfull()->CompactRange(cro2, nullptr, nullptr);
+  ASSERT_TRUE(s.IsAborted());
+
+  db_->ReleaseSnapshot(s1);
+  db_->ReleaseSnapshot(s2);
+}
+
 TEST_F(DBCompactionTest, DrainUnnecessaryLevelsAfterMultiplierChanged) {
  // When the level size multiplier increases such that fewer levels become
  // necessary, unnecessary levels should to be drained.
@@ -91,7 +91,7 @@ Status DBImpl::GetLiveFiles(std::vector<std::string>& ret,
  return Status::OK();
 }

-Status DBImpl::GetSortedWalFiles(VectorLogPtr& files) {
+Status DBImpl::GetSortedWalFiles(VectorWalPtr& files) {
  // Record tracked WALs as a (minimum) cross-check for directory scan
  std::vector<uint64_t> required_by_manifest;

@@ -151,10 +151,33 @@ Status DBImpl::GetSortedWalFiles(VectorLogPtr& files) {
    }
  }

+  if (s.ok()) {
+    size_t wal_count = files.size();
+    ROCKS_LOG_INFO(immutable_db_options_.info_log,
+                   "Number of WAL files %" ROCKSDB_PRIszt " (%" ROCKSDB_PRIszt
+                   " required by manifest)",
+                   wal_count, required_by_manifest.size());
+#ifndef NDEBUG
+    std::ostringstream wal_names;
+    for (const auto& wal : files) {
+      wal_names << wal->PathName() << " ";
+    }
+
+    std::ostringstream wal_required_by_manifest_names;
+    for (const auto& wal : required_by_manifest) {
+      wal_required_by_manifest_names << wal << ".log ";
+    }
+
+    ROCKS_LOG_INFO(immutable_db_options_.info_log,
+                   "Log files : %s .Log files required by manifest: %s.",
+                   wal_names.str().c_str(),
+                   wal_required_by_manifest_names.str().c_str());
+#endif  // NDEBUG
+  }
  return s;
 }

-Status DBImpl::GetCurrentWalFile(std::unique_ptr<LogFile>* current_log_file) {
+Status DBImpl::GetCurrentWalFile(std::unique_ptr<WalFile>* current_log_file) {
  uint64_t current_logfile_number;
  {
    InstrumentedMutexLock l(&mutex_);
@@ -175,13 +198,14 @@ Status DBImpl::GetLiveFilesStorageInfo(
  // NOTE: This implementation was largely migrated from Checkpoint.

  Status s;
-  VectorLogPtr live_wal_files;
+  VectorWalPtr live_wal_files;
  bool flush_memtable = true;
  if (!immutable_db_options_.allow_2pc) {
    if (opts.wal_size_for_flush == std::numeric_limits<uint64_t>::max()) {
      flush_memtable = false;
    } else if (opts.wal_size_for_flush > 0) {
-      // If the outstanding log files are small, we skip the flush.
+      // FIXME: avoid querying the filesystem for current WAL state
+      // If the outstanding WAL files are small, we skip the flush.
      s = GetSortedWalFiles(live_wal_files);

      if (!s.ok()) {
@@ -206,12 +230,18 @@ Status DBImpl::GetLiveFilesStorageInfo(
  // metadata.
  mutex_.Lock();
  if (flush_memtable) {
-    Status status = FlushForGetLiveFiles();
-    if (!status.ok()) {
-      mutex_.Unlock();
-      ROCKS_LOG_ERROR(immutable_db_options_.info_log, "Cannot Flush data %s\n",
-                      status.ToString().c_str());
-      return status;
+    bool wal_locked = lock_wal_count_ > 0;
+    if (wal_locked) {
+      ROCKS_LOG_INFO(immutable_db_options_.info_log,
+                     "Can't FlushForGetLiveFiles while WAL is locked");
+    } else {
+      Status status = FlushForGetLiveFiles();
+      if (!status.ok()) {
+        mutex_.Unlock();
+        ROCKS_LOG_ERROR(immutable_db_options_.info_log,
+                        "Cannot Flush data %s\n", status.ToString().c_str());
+        return status;
+      }
    }
  }

@@ -287,6 +317,8 @@ Status DBImpl::GetLiveFilesStorageInfo(
  const uint64_t options_number = versions_->options_file_number();
  const uint64_t options_size = versions_->options_file_size_;
  const uint64_t min_log_num = MinLogNumberToKeep();
+  // Ensure consistency with manifest for track_and_verify_wals_in_manifest
+  const uint64_t max_log_num = logfile_number_;

  mutex_.Unlock();

@@ -350,10 +382,13 @@ Status DBImpl::GetLiveFilesStorageInfo(
  TEST_SYNC_POINT("CheckpointImpl::CreateCheckpoint:SavedLiveFiles2");

  if (s.ok()) {
-    // To maximize the effectiveness of track_and_verify_wals_in_manifest,
-    // sync WAL when it is enabled.
-    s = FlushWAL(
-        immutable_db_options_.track_and_verify_wals_in_manifest /* sync */);
+    // FlushWAL is required to ensure we can physically copy everything
+    // logically written to the WAL. (Sync not strictly required for
+    // active WAL to be copied rather than hard linked, even when
+    // Checkpoint guarantees that the copied-to file is sync-ed. Plus we can't
+    // help track_and_verify_wals_in_manifest after manifest_size is
+    // already determined.)
+    s = FlushWAL(/*sync=*/false);
    if (s.IsNotSupported()) {  // read-only DB or similar
      s = Status::OK();
    }
@@ -362,25 +397,51 @@ Status DBImpl::GetLiveFilesStorageInfo(
  TEST_SYNC_POINT("CheckpointImpl::CreateCustomCheckpoint:AfterGetLive1");
  TEST_SYNC_POINT("CheckpointImpl::CreateCustomCheckpoint:AfterGetLive2");

-  // If we have more than one column family, we also need to get WAL files.
+  // Even after WAL flush, there could be multiple WALs that are not
+  // fully synced. Although the output DB of a Checkpoint or Backup needs
+  // to be fully synced on return, we don't strictly need to sync this
+  // DB (the input DB). If we allow Checkpoint to hard link an inactive
+  // WAL that isn't fully synced, that could result in an insufficiently
+  // sync-ed Checkpoint. Here we get the set of WALs that are potentially
+  // unsynced or still being written to, to prevent them from being hard
+  // linked. Enforcing max_log_num from above ensures any new WALs after
+  // GetOpenWalSizes() and before GetSortedWalFiles() are not included in
+  // the results.
+  // NOTE: we might still hard link a file that is open for writing, even
+  // if we don't do any more writes to it.
+  //
+  // In a step toward reducing unnecessary file metadata queries, we also
+  // get and use our known flushed sizes for those WALs.
+  // FIXME: eventually we should not be using filesystem queries at all for
+  // the required set of WAL files.
+  //
+  // However for recycled log files, we just copy the whole file,
+  // for better or worse.
+  //
+  std::map<uint64_t, uint64_t> open_wal_number_to_size;
+  bool recycling_log_files = immutable_db_options_.recycle_log_file_num > 0;
+  if (s.ok() && !recycling_log_files) {
+    s = GetOpenWalSizes(open_wal_number_to_size);
+  }
+
+  // [old comment] If we have more than one column family, we also need to get
+  // WAL files.
  if (s.ok()) {
+    // FIXME: avoid querying the filesystem for current WAL state
    s = GetSortedWalFiles(live_wal_files);
  }
  if (!s.ok()) {
    return s;
  }

-  size_t wal_size = live_wal_files.size();
-
-  ROCKS_LOG_INFO(immutable_db_options_.info_log,
-                 "Number of log files %" ROCKSDB_PRIszt, live_wal_files.size());
-
+  size_t wal_count = live_wal_files.size();
  // Link WAL files. Copy exact size of last one because it is the only one
  // that has changes after the last flush.
  auto wal_dir = immutable_db_options_.GetWalDir();
-  for (size_t i = 0; s.ok() && i < wal_size; ++i) {
+  for (size_t i = 0; s.ok() && i < wal_count; ++i) {
    if ((live_wal_files[i]->Type() == kAliveLogFile) &&
-        (!flush_memtable || live_wal_files[i]->LogNumber() >= min_log_num)) {
+        (!flush_memtable || live_wal_files[i]->LogNumber() >= min_log_num) &&
+        live_wal_files[i]->LogNumber() <= max_log_num) {
      results.emplace_back();
      LiveFileStorageInfo& info = results.back();
      auto f = live_wal_files[i]->PathName();
@@ -389,12 +450,29 @@ Status DBImpl::GetLiveFilesStorageInfo(
      info.directory = wal_dir;
      info.file_number = live_wal_files[i]->LogNumber();
      info.file_type = kWalFile;
-      info.size = live_wal_files[i]->SizeFileBytes();
-      // Trim the log either if its the last one, or log file recycling is
-      // enabled. In the latter case, a hard link doesn't prevent the file
-      // from being renamed and recycled. So we need to copy it instead.
-      info.trim_to_size = (i + 1 == wal_size) ||
-                          (immutable_db_options_.recycle_log_file_num > 0);
+      if (recycling_log_files) {
+        info.size = live_wal_files[i]->SizeFileBytes();
+        // Recyclable WAL files must be copied instead of hard linked
+        info.trim_to_size = true;
+      } else {
+        auto it = open_wal_number_to_size.find(info.file_number);
+        if (it == open_wal_number_to_size.end()) {
+          // Known fully synced and no future writes (in part from
+          // max_log_num check). Ok to hard link
+          info.size = live_wal_files[i]->SizeFileBytes();
+          assert(!info.trim_to_size);
+        } else {
+          // Marked as (possibly) still open -> use our known flushed size
+          // and force file copy instead of hard link
+          info.size = it->second;
+          info.trim_to_size = true;
+          // FIXME: this is needed as long as db_stress uses
+          // SetReadUnsyncedData(false), because it will only be able to
+          // copy the synced portion of the WAL, which under
+          // SetReadUnsyncedData(false) is given by the reported file size.
+          info.size = std::min(info.size, live_wal_files[i]->SizeFileBytes());
+        }
+      }
      if (opts.include_checksum_info) {
        info.file_checksum_func_name = kUnknownFileChecksumFuncName;
        info.file_checksum = kUnknownFileChecksum;
@@ -86,8 +86,8 @@ TEST_F(DBFlushTest, SyncFail) {
  options.env = fault_injection_env.get();

  SyncPoint::GetInstance()->LoadDependency(
-      {{"DBFlushTest::SyncFail:1", "DBImpl::SyncClosedLogs:Start"},
-       {"DBImpl::SyncClosedLogs:Failed", "DBFlushTest::SyncFail:2"}});
+      {{"DBFlushTest::SyncFail:1", "DBImpl::SyncClosedWals:Start"},
+       {"DBImpl::SyncClosedWals:Failed", "DBFlushTest::SyncFail:2"}});
  SyncPoint::GetInstance()->EnableProcessing();

  CreateAndReopenWithCF({"pikachu"}, options);
@@ -111,8 +111,8 @@ TEST_F(DBFlushTest, SyncSkip) {
  Options options = CurrentOptions();

  SyncPoint::GetInstance()->LoadDependency(
-      {{"DBFlushTest::SyncSkip:1", "DBImpl::SyncClosedLogs:Skip"},
-       {"DBImpl::SyncClosedLogs:Skip", "DBFlushTest::SyncSkip:2"}});
+      {{"DBFlushTest::SyncSkip:1", "DBImpl::SyncClosedWals:Skip"},
+       {"DBImpl::SyncClosedWals:Skip", "DBFlushTest::SyncSkip:2"}});
  SyncPoint::GetInstance()->EnableProcessing();

  Reopen(options);
@@ -2381,7 +2381,7 @@ TEST_F(DBFlushTest, PickRightMemtables) {
  SyncPoint::GetInstance()->DisableProcessing();
  SyncPoint::GetInstance()->ClearAllCallBacks();
  SyncPoint::GetInstance()->SetCallBack(
-      "DBImpl::SyncClosedLogs:BeforeReLock", [&](void* /*arg*/) {
+      "DBImpl::SyncClosedWals:BeforeReLock", [&](void* /*arg*/) {
        ASSERT_OK(db_->Put(WriteOptions(), handles_[1], "what", "v"));
        auto* cfhi =
            static_cast_with_check<ColumnFamilyHandleImpl>(handles_[1]);
@@ -0,0 +1,527 @@
+//  Copyright (c) 2024-present, Facebook, Inc.  All rights reserved.
+//  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 "port/stack_trace.h"
+#include "test_util/sync_point.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+#ifdef OS_LINUX
+
+class DBFollowerTest : public DBTestBase {
+ public:
+  // Create directories for leader and follower
+  // Create the leader DB object
+  DBFollowerTest() : DBTestBase("/db_follower_test", /*env_do_fsync*/ false) {
+    follower_name_ = dbname_ + "/follower";
+    db_parent_ = dbname_;
+    Close();
+    Destroy(CurrentOptions());
+    EXPECT_EQ(env_->CreateDirIfMissing(dbname_), Status::OK());
+    dbname_ = dbname_ + "/leader";
+    Reopen(CurrentOptions());
+  }
+
+  ~DBFollowerTest() {
+    follower_.reset();
+    EXPECT_EQ(DestroyDB(follower_name_, CurrentOptions()), Status::OK());
+    Destroy(CurrentOptions());
+    dbname_ = db_parent_;
+  }
+
+ protected:
+  class DBFollowerTestFS : public FileSystemWrapper {
+   public:
+    explicit DBFollowerTestFS(const std::shared_ptr<FileSystem>& target)
+        : FileSystemWrapper(target),
+          cv_(&mutex_),
+          barrier_(false),
+          count_(0),
+          reinit_count_(0) {}
+
+    const char* Name() const override { return "DBFollowerTestFS"; }
+
+    IOStatus NewSequentialFile(const std::string& fname,
+                               const FileOptions& file_opts,
+                               std::unique_ptr<FSSequentialFile>* result,
+                               IODebugContext* dbg = nullptr) override {
+      class DBFollowerTestSeqFile : public FSSequentialFileWrapper {
+       public:
+        DBFollowerTestSeqFile(DBFollowerTestFS* fs,
+                              std::unique_ptr<FSSequentialFile>&& file,
+                              uint64_t /*size*/)
+            : FSSequentialFileWrapper(file.get()),
+              fs_(fs),
+              file_(std::move(file)) {}
+
+        IOStatus Read(size_t n, const IOOptions& options, Slice* result,
+                      char* scratch, IODebugContext* dbg) override {
+          fs_->BarrierWait();
+          return target()->Read(n, options, result, scratch, dbg);
+        }
+
+       private:
+        DBFollowerTestFS* fs_;
+        std::unique_ptr<FSSequentialFile> file_;
+      };
+
+      std::unique_ptr<FSSequentialFile> file;
+      IOStatus s = target()->NewSequentialFile(fname, file_opts, &file, dbg);
+
+      if (s.ok() && test::GetFileType(fname) == kDescriptorFile) {
+        uint64_t size = 0;
+        EXPECT_EQ(target()->GetFileSize(fname, IOOptions(), &size, nullptr),
+                  IOStatus::OK());
+        result->reset(new DBFollowerTestSeqFile(this, std::move(file), size));
+      } else {
+        *result = std::move(file);
+      }
+      return s;
+    }
+
+    void BarrierInit(int count) {
+      MutexLock l(&mutex_);
+      barrier_ = true;
+      count_ = count;
+    }
+
+    void BarrierWait() {
+      MutexLock l(&mutex_);
+      if (!barrier_) {
+        return;
+      }
+      if (--count_ == 0) {
+        if (reinit_count_ > 0) {
+          count_ = reinit_count_;
+          reinit_count_ = 0;
+        } else {
+          barrier_ = false;
+        }
+        cv_.SignalAll();
+      } else {
+        cv_.Wait();
+      }
+    }
+
+    void BarrierWaitAndReinit(int count) {
+      MutexLock l(&mutex_);
+      if (!barrier_) {
+        return;
+      }
+      reinit_count_ = count;
+      if (--count_ == 0) {
+        if (reinit_count_ > 0) {
+          count_ = reinit_count_;
+          reinit_count_ = 0;
+        } else {
+          barrier_ = false;
+        }
+        cv_.SignalAll();
+      } else {
+        cv_.Wait();
+      }
+    }
+
+   private:
+    port::Mutex mutex_;
+    port::CondVar cv_;
+    bool barrier_;
+    int count_;
+    int reinit_count_;
+  };
+
+  class DBFollowerTestSstPartitioner : public SstPartitioner {
+   public:
+    explicit DBFollowerTestSstPartitioner(uint64_t max_keys)
+        : max_keys_(max_keys), num_keys_(0) {}
+
+    const char* Name() const override { return "DBFollowerTestSstPartitioner"; }
+
+    PartitionerResult ShouldPartition(
+        const PartitionerRequest& /*request*/) override {
+      if (++num_keys_ > max_keys_) {
+        num_keys_ = 0;
+        return PartitionerResult::kRequired;
+      } else {
+        return PartitionerResult::kNotRequired;
+      }
+    }
+
+    bool CanDoTrivialMove(const Slice& /*smallest_user_key*/,
+                          const Slice& /*largest_user_key*/) override {
+      return true;
+    }
+
+   private:
+    uint64_t max_keys_;
+    uint64_t num_keys_;
+  };
+
+  class DBFollowerTestSstPartitionerFactory : public SstPartitionerFactory {
+   public:
+    explicit DBFollowerTestSstPartitionerFactory(uint64_t max_keys)
+        : max_keys_(max_keys) {}
+
+    std::unique_ptr<SstPartitioner> CreatePartitioner(
+        const SstPartitioner::Context& /*context*/) const override {
+      std::unique_ptr<SstPartitioner> partitioner;
+      partitioner.reset(new DBFollowerTestSstPartitioner(max_keys_));
+      return partitioner;
+    }
+
+    const char* Name() const override {
+      return "DBFollowerTestSstPartitionerFactory";
+    }
+
+   private:
+    uint64_t max_keys_;
+  };
+
+  Status OpenAsFollower() {
+    Options opts = CurrentOptions();
+    if (!follower_env_) {
+      follower_env_ = NewCompositeEnv(
+          std::make_shared<DBFollowerTestFS>(env_->GetFileSystem()));
+    }
+    opts.env = follower_env_.get();
+    opts.follower_refresh_catchup_period_ms = 100;
+    return DB::OpenAsFollower(opts, follower_name_, dbname_, &follower_);
+  }
+
+  std::string FollowerGet(const std::string& k) {
+    ReadOptions options;
+    options.verify_checksums = true;
+    std::string result;
+    Status s = follower()->Get(options, k, &result);
+    if (s.IsNotFound()) {
+      result = "NOT_FOUND";
+    } else if (!s.ok()) {
+      result = s.ToString();
+    }
+    return result;
+  }
+
+  DB* follower() { return follower_.get(); }
+  DBFollowerTestFS* follower_fs() {
+    return static_cast<DBFollowerTestFS*>(follower_env_->GetFileSystem().get());
+  }
+
+  void CheckDirs() {
+    std::vector<std::string> db_children;
+    std::vector<std::string> follower_children;
+    EXPECT_OK(env_->GetChildren(dbname_, &db_children));
+    EXPECT_OK(env_->GetChildren(follower_name_, &follower_children));
+
+    std::set<uint64_t> db_filenums;
+    std::set<uint64_t> follower_filenums;
+    for (auto& name : db_children) {
+      if (test::GetFileType(name) != kTableFile) {
+        continue;
+      }
+      db_filenums.insert(test::GetFileNumber(name));
+    }
+    for (auto& name : follower_children) {
+      if (test::GetFileType(name) != kTableFile) {
+        continue;
+      }
+      follower_filenums.insert(test::GetFileNumber(name));
+    }
+    db_filenums.merge(follower_filenums);
+    EXPECT_EQ(follower_filenums.size(), db_filenums.size());
+  }
+
+ private:
+  std::string follower_name_;
+  std::string db_parent_;
+  std::unique_ptr<Env> follower_env_;
+  std::unique_ptr<DB> follower_;
+};
+
+TEST_F(DBFollowerTest, Basic) {
+  ASSERT_OK(Put("k1", "v1"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("k2", "v2"));
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(OpenAsFollower());
+  std::string val;
+  ASSERT_OK(follower()->Get(ReadOptions(), "k1", &val));
+  ASSERT_EQ(val, "v1");
+  CheckDirs();
+}
+
+TEST_F(DBFollowerTest, Flush) {
+  SyncPoint::GetInstance()->LoadDependency({
+      {"DBImplFollower::TryCatchupWithLeader:Begin1", "Leader::Start"},
+      {"Leader::Done", "DBImplFollower::TryCatchupWithLeader:Begin2"},
+      {"DBImplFollower::TryCatchupWithLeader:End", "Follower::WaitForCatchup"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(OpenAsFollower());
+  TEST_SYNC_POINT("Leader::Start");
+  ASSERT_OK(Put("k1", "v1"));
+  ASSERT_OK(Flush());
+  TEST_SYNC_POINT("Leader::Done");
+
+  TEST_SYNC_POINT("Follower::WaitForCatchup");
+  std::string val;
+  ASSERT_OK(follower()->Get(ReadOptions(), "k1", &val));
+  ASSERT_EQ(val, "v1");
+  CheckDirs();
+
+  SyncPoint::GetInstance()->DisableProcessing();
+}
+
+// This test creates 4 L0 files, immediately followed by a compaction to L1.
+// The follower replays the 4 flush records from the MANIFEST unsuccessfully,
+// and then successfully recovers a Version from the compaction record
+TEST_F(DBFollowerTest, RetryCatchup) {
+  Options opts = CurrentOptions();
+  opts.disable_auto_compactions = true;
+  Reopen(opts);
+
+  ASSERT_OK(OpenAsFollower());
+  SyncPoint::GetInstance()->LoadDependency({
+      {"DBImplFollower::TryCatchupWithLeader:Begin1", "Leader::Start"},
+      {"DBImpl::BackgroundCompaction:Start",
+       "DBImplFollower::TryCatchupWithLeader:Begin2"},
+      {"VersionEditHandlerPointInTime::MaybeCreateVersion:Begin1",
+       "DBImpl::BackgroundCompaction:BeforeCompaction"},
+      {"DBImpl::BackgroundCallCompaction:PurgedObsoleteFiles",
+       "VersionEditHandlerPointInTime::MaybeCreateVersion:Begin2"},
+      {"DBImplFollower::TryCatchupWithLeader:End", "Follower::WaitForCatchup"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+  TEST_SYNC_POINT("Leader::Start");
+  ASSERT_OK(Put("k1", "v1"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("k1", "v2"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("k1", "v3"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("k1", "v4"));
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(dbfull()->TEST_CompactRange(0, nullptr, nullptr, nullptr, true));
+
+  TEST_SYNC_POINT("Follower::WaitForCatchup");
+  ASSERT_EQ(FollowerGet("k1"), "v4");
+  CheckDirs();
+
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+  SyncPoint::GetInstance()->DisableProcessing();
+}
+
+// This test validates the same as the previous test, except there is a
+// MANIFEST rollover between the flushes and compaction. The follower
+// does not switch to a new MANIFEST in ReadAndApply. So it would require
+// another round of refresh before catching up.
+TEST_F(DBFollowerTest, RetryCatchupManifestRollover) {
+  Options opts = CurrentOptions();
+  opts.disable_auto_compactions = true;
+  Reopen(opts);
+
+  ASSERT_OK(Put("k1", "v1"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("k1", "v2"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("k1", "v3"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(OpenAsFollower());
+  SyncPoint::GetInstance()->LoadDependency({
+      {"DBImplFollower::TryCatchupWithLeader:Begin1", "Leader::Start"},
+      {"Leader::Flushed", "DBImplFollower::TryCatchupWithLeader:Begin2"},
+      {"VersionEditHandlerPointInTime::MaybeCreateVersion:Begin1",
+       "Leader::Done"},
+      {"DBImpl::BackgroundCallCompaction:PurgedObsoleteFiles",
+       "VersionEditHandlerPointInTime::MaybeCreateVersion:Begin2"},
+      {"DBImplFollower::TryCatchupWithLeader:End",
+       "Follower::WaitForCatchup:1"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+  TEST_SYNC_POINT("Leader::Start");
+  ASSERT_OK(Put("k1", "v4"));
+  ASSERT_OK(Flush());
+
+  TEST_SYNC_POINT("Leader::Flushed");
+  TEST_SYNC_POINT("Leader::Done");
+  Reopen(opts);
+  ASSERT_OK(dbfull()->TEST_CompactRange(0, nullptr, nullptr, nullptr, true));
+
+  TEST_SYNC_POINT("Follower::WaitForCatchup:1");
+  SyncPoint::GetInstance()->LoadDependency({
+      {"DBImplFollower::TryCatchupWithLeader:End",
+       "Follower::WaitForCatchup:2"},
+  });
+  TEST_SYNC_POINT("Follower::WaitForCatchup:2");
+  ASSERT_EQ(FollowerGet("k1"), "v4");
+
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+  SyncPoint::GetInstance()->DisableProcessing();
+}
+
+// This test creates 4 L0 files and compacts them. The follower, during catchup,
+// successfully instantiates 4 Versions corresponding to the 4 files (but
+// donesn't install them yet), followed by deleting those 4 and adding a new
+// file from compaction. The test verifies that the 4 L0 files are deleted
+// correctly by the follower.
+// We use teh Barrier* functions to ensure that the follower first sees the 4
+// L0 files and is able to link them, and then sees the compaction that
+// obsoletes those L0 files (so those L0 files are intermediates that it has
+// to explicitly delete). Suppose we don't have any barriers, its possible
+// the follower reads the L0 records and compaction records from the MANIFEST
+// in one read, which means those L0 files would have already been deleted
+// by the leader and the follower cannot link to them.
+TEST_F(DBFollowerTest, IntermediateObsoleteFiles) {
+  Options opts = CurrentOptions();
+  opts.disable_auto_compactions = true;
+  Reopen(opts);
+  ASSERT_OK(OpenAsFollower());
+
+  follower_fs()->BarrierInit(2);
+  ASSERT_OK(Put("k1", "v1"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("k1", "v2"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("k1", "v3"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("k1", "v4"));
+  ASSERT_OK(Flush());
+  follower_fs()->BarrierWaitAndReinit(2);
+
+  ASSERT_OK(dbfull()->TEST_CompactRange(0, nullptr, nullptr, nullptr, true));
+  follower_fs()->BarrierWait();
+
+  SyncPoint::GetInstance()->LoadDependency({
+      {"DBImplFollower::TryCatchupWithLeader:End",
+       "Follower::WaitForCatchup:1"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+  TEST_SYNC_POINT("Follower::WaitForCatchup:1");
+  CheckDirs();
+  ASSERT_EQ(FollowerGet("k1"), "v4");
+}
+
+// This test verifies a scenario where the follower can recover a Version
+// partially (i.e some of the additions cannot be found), and the files
+// that are found are obsoleted by a subsequent VersionEdit.
+TEST_F(DBFollowerTest, PartialVersionRecovery) {
+  Options opts = CurrentOptions();
+  opts.disable_auto_compactions = true;
+  opts.sst_partitioner_factory =
+      std::make_shared<DBFollowerTestSstPartitionerFactory>(1);
+  Reopen(opts);
+
+  ASSERT_OK(Put("k1", "v1"));
+  ASSERT_OK(Put("k2", "v1"));
+  ASSERT_OK(Put("k3", "v1"));
+  ASSERT_OK(Flush());
+  MoveFilesToLevel(2);
+
+  ASSERT_OK(Put("k1", "v2"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("k3", "v2"));
+  ASSERT_OK(Flush());
+  MoveFilesToLevel(1);
+
+  ASSERT_OK(OpenAsFollower());
+  ASSERT_OK(dbfull()->SetOptions(dbfull()->DefaultColumnFamily(),
+                                 {{"max_compaction_bytes", "1"}}));
+
+  follower_fs()->BarrierInit(2);
+  Slice key("k1");
+  ASSERT_OK(dbfull()->TEST_CompactRange(1, &key, &key, nullptr, true));
+
+  follower_fs()->BarrierWaitAndReinit(2);
+
+  // The second compaction input overlaps the previous compaction outputs
+  // by one file. This file is never added to  VersionStorageInfo since it
+  // was added and deleted before the catch up completes. We later verify that
+  // the follower correctly deleted this file.
+  key = Slice("k3");
+  ASSERT_OK(dbfull()->TEST_CompactRange(1, &key, &key, nullptr, true));
+  follower_fs()->BarrierWait();
+
+  SyncPoint::GetInstance()->LoadDependency({
+      {"DBImplFollower::TryCatchupWithLeader:End",
+       "Follower::WaitForCatchup:1"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+  TEST_SYNC_POINT("Follower::WaitForCatchup:1");
+  CheckDirs();
+  ASSERT_EQ(FollowerGet("k1"), "v2");
+  ASSERT_EQ(FollowerGet("k2"), "v1");
+  ASSERT_EQ(FollowerGet("k3"), "v2");
+  SyncPoint::GetInstance()->DisableProcessing();
+}
+
+// This test verifies a scenario similar to the PartialVersionRecovery, except
+// with a MANIFEST rollover in between. When there is a rollover, the
+// follower's attempt ends without installing a new Version. The next catch up
+// attempt will recover a full Version.
+TEST_F(DBFollowerTest, PartialVersionRecoveryWithRollover) {
+  Options opts = CurrentOptions();
+  opts.disable_auto_compactions = true;
+  opts.sst_partitioner_factory =
+      std::make_shared<DBFollowerTestSstPartitionerFactory>(1);
+  Reopen(opts);
+
+  ASSERT_OK(Put("k1", "v1"));
+  ASSERT_OK(Put("k2", "v1"));
+  ASSERT_OK(Put("k3", "v1"));
+  ASSERT_OK(Flush());
+  MoveFilesToLevel(2);
+
+  ASSERT_OK(Put("k1", "v2"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put("k3", "v2"));
+  ASSERT_OK(Flush());
+  MoveFilesToLevel(1);
+
+  opts.max_compaction_bytes = 1;
+  Reopen(opts);
+
+  ASSERT_OK(OpenAsFollower());
+
+  follower_fs()->BarrierInit(2);
+  Slice key("k1");
+  ASSERT_OK(dbfull()->TEST_CompactRange(1, &key, &key, nullptr, true));
+
+  follower_fs()->BarrierWaitAndReinit(2);
+  Reopen(opts);
+  key = Slice("k3");
+  ASSERT_OK(dbfull()->TEST_CompactRange(1, &key, &key, nullptr, true));
+  follower_fs()->BarrierWait();
+
+  SyncPoint::GetInstance()->LoadDependency({
+      {"DBImplFollower::TryCatchupWithLeader:Begin1",
+       "Follower::WaitForCatchup:1"},
+      {"Follower::WaitForCatchup:2",
+       "DBImplFollower::TryCatchupWithLeader:Begin2"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+  TEST_SYNC_POINT("Follower::WaitForCatchup:1");
+  TEST_SYNC_POINT("Follower::WaitForCatchup:2");
+  SyncPoint::GetInstance()->LoadDependency({
+      {"DBImplFollower::TryCatchupWithLeader:End",
+       "Follower::WaitForCatchup:3"},
+  });
+  TEST_SYNC_POINT("Follower::WaitForCatchup:3");
+  CheckDirs();
+  ASSERT_EQ(FollowerGet("k1"), "v2");
+  ASSERT_EQ(FollowerGet("k2"), "v1");
+  ASSERT_EQ(FollowerGet("k3"), "v2");
+  SyncPoint::GetInstance()->DisableProcessing();
+}
+#endif
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
@@ -9,6 +9,7 @@
 #pragma once

 #include <atomic>
+#include <cstdint>
 #include <deque>
 #include <functional>
 #include <limits>
@@ -50,12 +51,14 @@
 #include "monitoring/instrumented_mutex.h"
 #include "options/db_options.h"
 #include "port/port.h"
+#include "rocksdb/attribute_groups.h"
 #include "rocksdb/db.h"
 #include "rocksdb/env.h"
 #include "rocksdb/memtablerep.h"
 #include "rocksdb/status.h"
 #include "rocksdb/trace_reader_writer.h"
 #include "rocksdb/transaction_log.h"
+#include "rocksdb/user_write_callback.h"
 #include "rocksdb/utilities/replayer.h"
 #include "rocksdb/write_buffer_manager.h"
 #include "table/merging_iterator.h"
@@ -230,6 +233,10 @@ class DBImpl : public DB {
  using DB::Write;
  Status Write(const WriteOptions& options, WriteBatch* updates) override;

+  using DB::WriteWithCallback;
+  Status WriteWithCallback(const WriteOptions& options, WriteBatch* updates,
+                           UserWriteCallback* user_write_cb) override;
+
  using DB::Get;
  Status Get(const ReadOptions& _read_options,
             ColumnFamilyHandle* column_family, const Slice& key,
@@ -290,6 +297,11 @@ class DBImpl : public DB {
                      const Slice* keys,
                      PinnableAttributeGroups* results) override;

+  void MultiGetEntityWithCallback(
+      const ReadOptions& read_options, ColumnFamilyHandle* column_family,
+      ReadCallback* callback,
+      autovector<KeyContext*, MultiGetContext::MAX_BATCH_SIZE>* sorted_keys);
+
  Status CreateColumnFamily(const ColumnFamilyOptions& cf_options,
                            const std::string& column_family,
                            ColumnFamilyHandle** handle) override {
@@ -351,9 +363,13 @@ class DBImpl : public DB {
  const Snapshot* GetSnapshot() override;
  void ReleaseSnapshot(const Snapshot* snapshot) override;

-  // UNDER CONSTRUCTION - DO NOT USE
-  // Return a cross-column-family iterator from a consistent database state.
-  std::unique_ptr<Iterator> NewMultiCfIterator(
+  // EXPERIMENTAL
+  std::unique_ptr<Iterator> NewCoalescingIterator(
+      const ReadOptions& options,
+      const std::vector<ColumnFamilyHandle*>& column_families) override;
+
+  // EXPERIMENTAL
+  std::unique_ptr<AttributeGroupIterator> NewAttributeGroupIterator(
      const ReadOptions& options,
      const std::vector<ColumnFamilyHandle*>& column_families) override;

@@ -491,8 +507,14 @@ class DBImpl : public DB {
  // All the returned filenames start with "/"
  Status GetLiveFiles(std::vector<std::string>&, uint64_t* manifest_file_size,
                      bool flush_memtable = true) override;
-  Status GetSortedWalFiles(VectorLogPtr& files) override;
-  Status GetCurrentWalFile(std::unique_ptr<LogFile>* current_log_file) override;
+  Status GetSortedWalFiles(VectorWalPtr& files) override;
+  // Get the known flushed sizes of WALs that might still be written to
+  // or have pending sync.
+  // NOTE: unlike alive_log_files_, this function includes WALs that might
+  // be obsolete (but not obsolete to a pending Checkpoint) and not yet fully
+  // synced.
+  Status GetOpenWalSizes(std::map<uint64_t, uint64_t>& number_to_size);
+  Status GetCurrentWalFile(std::unique_ptr<WalFile>* current_log_file) override;
  Status GetCreationTimeOfOldestFile(uint64_t* creation_time) override;

  Status GetUpdatesSince(
@@ -678,7 +700,8 @@ class DBImpl : public DB {
  // thread to determine whether it is safe to perform the write.
  virtual Status WriteWithCallback(const WriteOptions& write_options,
                                   WriteBatch* my_batch,
-                                   WriteCallback* callback);
+                                   WriteCallback* callback,
+                                   UserWriteCallback* user_write_cb = nullptr);

  // Returns the sequence number that is guaranteed to be smaller than or equal
  // to the sequence number of any key that could be inserted into the current
@@ -816,6 +839,8 @@ class DBImpl : public DB {

  uint64_t MinLogNumberToKeep();

+  uint64_t MinLogNumberToRecycle();
+
  // Returns the lower bound file number for SSTs that won't be deleted, even if
  // they're obsolete. This lower bound is used internally to prevent newly
  // created flush/compaction output files from being deleted before they're
@@ -1052,7 +1077,8 @@ class DBImpl : public DB {
  static Status Open(const DBOptions& db_options, const std::string& name,
                     const std::vector<ColumnFamilyDescriptor>& column_families,
                     std::vector<ColumnFamilyHandle*>* handles, DB** dbptr,
-                     const bool seq_per_batch, const bool batch_per_txn);
+                     const bool seq_per_batch, const bool batch_per_txn,
+                     const bool is_retry, bool* can_retry);

  static IOStatus CreateAndNewDirectory(
      FileSystem* fs, const std::string& dirname,
@@ -1198,6 +1224,8 @@ class DBImpl : public DB {
    return logs_.back().number;
  }

+  void TEST_DeleteObsoleteFiles();
+
  const std::unordered_set<uint64_t>& TEST_GetFilesGrabbedForPurge() const {
    return files_grabbed_for_purge_;
  }
@@ -1413,9 +1441,8 @@ class DBImpl : public DB {
    autovector<ColumnFamilyData*> cfds_;
    autovector<const MutableCFOptions*> mutable_cf_opts_;
    autovector<autovector<VersionEdit*>> edit_lists_;
-    // Stale SST files to delete found upon recovery. This stores a mapping from
-    // such a file's absolute path to its parent directory.
-    std::unordered_map<std::string, std::string> files_to_delete_;
+    // All existing data files (SST files and Blob files) found during DB::Open.
+    std::vector<std::string> existing_data_files_;
    bool is_new_db_ = false;
  };

@@ -1435,6 +1462,9 @@ class DBImpl : public DB {
  Status RenameTempFileToOptionsFile(const std::string& file_name);
  Status DeleteObsoleteOptionsFiles();

+  void NotifyOnManualFlushScheduled(autovector<ColumnFamilyData*> cfds,
+                                    FlushReason flush_reason);
+
  void NotifyOnFlushBegin(ColumnFamilyData* cfd, FileMetaData* file_meta,
                          const MutableCFOptions& mutable_cf_options,
                          int job_id, FlushReason flush_reason);
@@ -1485,6 +1515,7 @@ class DBImpl : public DB {
  // batch that does not have duplicate keys.
  Status WriteImpl(const WriteOptions& options, WriteBatch* updates,
                   WriteCallback* callback = nullptr,
+                   UserWriteCallback* user_write_cb = nullptr,
                   uint64_t* log_used = nullptr, uint64_t log_ref = 0,
                   bool disable_memtable = false, uint64_t* seq_used = nullptr,
                   size_t batch_cnt = 0,
@@ -1493,6 +1524,7 @@ class DBImpl : public DB {

  Status PipelinedWriteImpl(const WriteOptions& options, WriteBatch* updates,
                            WriteCallback* callback = nullptr,
+                            UserWriteCallback* user_write_cb = nullptr,
                            uint64_t* log_used = nullptr, uint64_t log_ref = 0,
                            bool disable_memtable = false,
                            uint64_t* seq_used = nullptr);
@@ -1519,7 +1551,8 @@ class DBImpl : public DB {
  // marks start of a new sub-batch.
  Status WriteImplWALOnly(
      WriteThread* write_thread, const WriteOptions& options,
-      WriteBatch* updates, WriteCallback* callback, uint64_t* log_used,
+      WriteBatch* updates, WriteCallback* callback,
+      UserWriteCallback* user_write_cb, uint64_t* log_used,
      const uint64_t log_ref, uint64_t* seq_used, const size_t sub_batch_cnt,
      PreReleaseCallback* pre_release_callback, const AssignOrder assign_order,
      const PublishLastSeq publish_last_seq, const bool disable_memtable);
@@ -1529,7 +1562,7 @@ class DBImpl : public DB {
  Status WriteRecoverableState();

  // Actual implementation of Close()
-  Status CloseImpl();
+  virtual Status CloseImpl();

  // Recover the descriptor from persistent storage.  May do a significant
  // amount of work to recover recently logged updates.  Any changes to
@@ -1541,9 +1574,9 @@ class DBImpl : public DB {
  virtual Status Recover(
      const std::vector<ColumnFamilyDescriptor>& column_families,
      bool read_only = false, bool error_if_wal_file_exists = false,
-      bool error_if_data_exists_in_wals = false,
+      bool error_if_data_exists_in_wals = false, bool is_retry = false,
      uint64_t* recovered_seq = nullptr,
-      RecoveryContext* recovery_ctx = nullptr);
+      RecoveryContext* recovery_ctx = nullptr, bool* can_retry = nullptr);

  virtual bool OwnTablesAndLogs() const { return true; }

@@ -1553,22 +1586,36 @@ class DBImpl : public DB {
  // Assign db_id_ and write DB ID to manifest if necessary.
  void SetDBId(std::string&& id, bool read_only, RecoveryContext* recovery_ctx);

+  // Collect a deduplicated collection of paths used by this DB, including
+  // dbname_, DBOptions.db_paths, ColumnFamilyOptions.cf_paths.
+  std::set<std::string> CollectAllDBPaths();
+
  // REQUIRES: db mutex held when calling this function, but the db mutex can
  // be released and re-acquired. Db mutex will be held when the function
  // returns.
-  // After recovery, there may be SST files in db/cf paths that are
-  // not referenced in the MANIFEST (e.g.
+  // It stores all existing data files (SST and Blob) in RecoveryContext. In
+  // the meantime, we find out the largest file number present in the paths, and
+  // bump up the version set's next_file_number_ to be 1 + largest_file_number.
+  // recovery_ctx stores the context about version edits. All those edits are
+  // persisted to new Manifest after successfully syncing the new WAL.
+  Status MaybeUpdateNextFileNumber(RecoveryContext* recovery_ctx);
+
+  // Track existing data files, including both referenced and unreferenced SST
+  // and Blob files in SstFileManager. This is only called during DB::Open and
+  // it's called before any file deletion start so that their deletion can be
+  // properly rate limited.
+  // Files may not be referenced in the MANIFEST because (e.g.
  // 1. It's best effort recovery;
  // 2. The VersionEdits referencing the SST files are appended to
  // RecoveryContext, DB crashes when syncing the MANIFEST, the VersionEdits are
  // still not synced to MANIFEST during recovery.)
-  // It stores the SST files to be deleted in RecoveryContext. In the
-  // meantime, we find out the largest file number present in the paths, and
-  // bump up the version set's next_file_number_ to be 1 + largest_file_number.
-  // recovery_ctx stores the context about version edits and files to be
-  // deleted. All those edits are persisted to new Manifest after successfully
-  // syncing the new WAL.
-  Status DeleteUnreferencedSstFiles(RecoveryContext* recovery_ctx);
+  //
+  // If the file is referenced in Manifest (typically that's the
+  // vast majority of all files), since it already has the file size
+  // on record, we don't need to query the file system. Otherwise, we query the
+  // file system for the size of an unreferenced file.
+  void TrackExistingDataFiles(
+      const std::vector<std::string>& existing_data_files);

  // SetDbSessionId() should be called in the constuctor DBImpl()
  // to ensure that db_session_id_ gets updated every time the DB is opened
@@ -1622,6 +1669,7 @@ class DBImpl : public DB {
  friend class ForwardIterator;
  friend struct SuperVersion;
  friend class CompactedDBImpl;
+  friend class DBImplFollower;
 #ifndef NDEBUG
  friend class DBTest_ConcurrentFlushWAL_Test;
  friend class DBTest_MixedSlowdownOptionsStop_Test;
@@ -1676,8 +1724,11 @@ class DBImpl : public DB {
      return w;
    }
    Status ClearWriter() {
-      // TODO: plumb Env::IOActivity, Env::IOPriority
-      Status s = writer->WriteBuffer(WriteOptions());
+      Status s;
+      if (writer->file()) {
+        // TODO: plumb Env::IOActivity, Env::IOPriority
+        s = writer->WriteBuffer(WriteOptions());
+      }
      delete writer;
      writer = nullptr;
      return s;
@@ -1692,10 +1743,16 @@ class DBImpl : public DB {

    void PrepareForSync() {
      assert(!getting_synced);
-      // Size is expected to be monotonically increasing.
-      assert(writer->file()->GetFlushedSize() >= pre_sync_size);
+      // Ensure the head of logs_ is marked as getting_synced if any is.
      getting_synced = true;
-      pre_sync_size = writer->file()->GetFlushedSize();
+      // If last sync failed on a later WAL, this could be a fully synced
+      // and closed WAL that just needs to be recorded as synced in the
+      // manifest.
+      if (writer->file()) {
+        // Size is expected to be monotonically increasing.
+        assert(writer->file()->GetFlushedSize() >= pre_sync_size);
+        pre_sync_size = writer->file()->GetFlushedSize();
+      }
    }

    void FinishSync() {
@@ -1893,7 +1950,7 @@ class DBImpl : public DB {
  void ReleaseFileNumberFromPendingOutputs(
      std::unique_ptr<std::list<uint64_t>::iterator>& v);

-  IOStatus SyncClosedLogs(const WriteOptions& write_options,
+  IOStatus SyncClosedWals(const WriteOptions& write_options,
                          JobContext* job_context, VersionEdit* synced_wals,
                          bool error_recovery_in_prog);

@@ -1923,7 +1980,7 @@ class DBImpl : public DB {
  // corrupted_log_found is set to true if we recover from a corrupted log file.
  Status RecoverLogFiles(const std::vector<uint64_t>& log_numbers,
                         SequenceNumber* next_sequence, bool read_only,
-                         bool* corrupted_log_found,
+                         bool is_retry, bool* corrupted_log_found,
                         RecoveryContext* recovery_ctx);

  // The following two methods are used to flush a memtable to
@@ -2027,17 +2084,22 @@ class DBImpl : public DB {
      mutex_.Lock();
    }

-    if (!immutable_db_options_.unordered_write) {
-      // Then the writes are finished before the next write group starts
-      return;
+    if (immutable_db_options_.unordered_write) {
+      // Wait for the ones who already wrote to the WAL to finish their
+      // memtable write.
+      if (pending_memtable_writes_.load() != 0) {
+        // XXX: suspicious wait while holding DB mutex?
+        std::unique_lock<std::mutex> guard(switch_mutex_);
+        switch_cv_.wait(guard,
+                        [&] { return pending_memtable_writes_.load() == 0; });
+      }
+    } else {
+      // (Writes are finished before the next write group starts.)
    }

-    // Wait for the ones who already wrote to the WAL to finish their
-    // memtable write.
-    if (pending_memtable_writes_.load() != 0) {
-      std::unique_lock<std::mutex> guard(switch_mutex_);
-      switch_cv_.wait(guard,
-                      [&] { return pending_memtable_writes_.load() == 0; });
+    // Wait for any LockWAL to clear
+    while (lock_wal_count_ > 0) {
+      bg_cv_.Wait();
    }
  }

@@ -2147,7 +2209,8 @@ class DBImpl : public DB {
  void GenerateFlushRequest(const autovector<ColumnFamilyData*>& cfds,
                            FlushReason flush_reason, FlushRequest* req);

-  void SchedulePendingFlush(const FlushRequest& req);
+  // Returns true if `req` is successfully enqueued.
+  bool SchedulePendingFlush(const FlushRequest& req);

  void SchedulePendingCompaction(ColumnFamilyData* cfd);
  void SchedulePendingPurge(std::string fname, std::string dir_to_sync,
@@ -2223,6 +2286,11 @@ class DBImpl : public DB {
  ColumnFamilyData* PickCompactionFromQueue(
      std::unique_ptr<TaskLimiterToken>* token, LogBuffer* log_buffer);

+  IOStatus SyncWalImpl(bool include_current_wal,
+                       const WriteOptions& write_options,
+                       JobContext* job_context, VersionEdit* synced_wals,
+                       bool error_recovery_in_prog);
+
  // helper function to call after some of the logs_ were synced
  void MarkLogsSynced(uint64_t up_to, bool synced_dir, VersionEdit* edit);
  Status ApplyWALToManifest(const ReadOptions& read_options,
@@ -2331,10 +2399,7 @@ class DBImpl : public DB {
  // A structure to hold the information required to process MultiGet of keys
  // belonging to one column family. For a multi column family MultiGet, there
  // will be a container of these objects.
-  struct MultiGetColumnFamilyData {
-    ColumnFamilyHandle* cf;
-    ColumnFamilyData* cfd;
-
+  struct MultiGetKeyRangePerCf {
    // For the batched MultiGet which relies on sorted keys, start specifies
    // the index of first key belonging to this column family in the sorted
    // list.
@@ -2344,31 +2409,33 @@ class DBImpl : public DB {
    // belonging to this column family in the sorted list
    size_t num_keys;

+    MultiGetKeyRangePerCf() : start(0), num_keys(0) {}
+
+    MultiGetKeyRangePerCf(size_t first, size_t count)
+        : start(first), num_keys(count) {}
+  };
+
+  // A structure to contain ColumnFamilyData and the SuperVersion obtained for
+  // the consistent view of DB
+  struct ColumnFamilySuperVersionPair {
+    ColumnFamilyHandleImpl* cfh;
+    ColumnFamilyData* cfd;
+
    // SuperVersion for the column family obtained in a manner that ensures a
    // consistent view across all column families in the DB
    SuperVersion* super_version;
-    MultiGetColumnFamilyData(ColumnFamilyHandle* column_family,
-                             SuperVersion* sv)
-        : cf(column_family),
-          cfd(static_cast<ColumnFamilyHandleImpl*>(cf)->cfd()),
-          start(0),
-          num_keys(0),
+    ColumnFamilySuperVersionPair(ColumnFamilyHandle* column_family,
+                                 SuperVersion* sv)
+        : cfh(static_cast<ColumnFamilyHandleImpl*>(column_family)),
+          cfd(cfh->cfd()),
          super_version(sv) {}

-    MultiGetColumnFamilyData(ColumnFamilyHandle* column_family, size_t first,
-                             size_t count, SuperVersion* sv)
-        : cf(column_family),
-          cfd(static_cast<ColumnFamilyHandleImpl*>(cf)->cfd()),
-          start(first),
-          num_keys(count),
-          super_version(sv) {}
-
-    MultiGetColumnFamilyData() = default;
+    ColumnFamilySuperVersionPair() = default;
  };

  // A common function to obtain a consistent snapshot, which can be implicit
  // if the user doesn't specify a snapshot in read_options, across
-  // multiple column families for MultiGet. It will attempt to get an implicit
+  // multiple column families. It will attempt to get an implicit
  // snapshot without acquiring the db_mutes, but will give up after a few
  // tries and acquire the mutex if a memtable flush happens. The template
  // allows both the batched and non-batched MultiGet to call this with
@@ -2377,18 +2444,26 @@ class DBImpl : public DB {
  // If callback is non-null, the callback is refreshed with the snapshot
  // sequence number
  //
+  // `extra_sv_ref` is used to indicate whether thread-local SuperVersion
+  // should be obtained with an extra ref (by GetReferencedSuperVersion()) or
+  // not (by GetAndRefSuperVersion()). For instance, point lookup like MultiGet
+  // does not require SuperVersion to be re-acquired throughout the entire
+  // invocation (no need extra ref), while MultiCfIterators may need the
+  // SuperVersion to be updated during Refresh() (requires extra ref).
+  //
  // `sv_from_thread_local` being set to false indicates that the SuperVersion
  // obtained from the ColumnFamilyData, whereas true indicates they are thread
  // local.
+  //
  // A non-OK status will be returned if for a column family that enables
  // user-defined timestamp feature, the specified `ReadOptions.timestamp`
  // attemps to read collapsed history.
-  template <class T>
-  Status MultiCFSnapshot(
-      const ReadOptions& read_options, ReadCallback* callback,
-      std::function<MultiGetColumnFamilyData*(typename T::iterator&)>&
-          iter_deref_func,
-      T* cf_list, SequenceNumber* snapshot, bool* sv_from_thread_local);
+  template <class T, typename IterDerefFuncType>
+  Status MultiCFSnapshot(const ReadOptions& read_options,
+                         ReadCallback* callback,
+                         IterDerefFuncType iter_deref_func, T* cf_list,
+                         bool extra_sv_ref, SequenceNumber* snapshot,
+                         bool* sv_from_thread_local);

  // The actual implementation of the batching MultiGet. The caller is expected
  // to have acquired the SuperVersion and pass in a snapshot sequence number
@@ -2411,6 +2486,13 @@ class DBImpl : public DB {

  bool ShouldReferenceSuperVersion(const MergeContext& merge_context);

+  template <typename IterType, typename ImplType,
+            typename ErrorIteratorFuncType>
+  std::unique_ptr<IterType> NewMultiCfIterator(
+      const ReadOptions& _read_options,
+      const std::vector<ColumnFamilyHandle*>& column_families,
+      ErrorIteratorFuncType error_iterator_func);
+
  // Lock over the persistent DB state.  Non-nullptr iff successfully acquired.
  FileLock* db_lock_;

@@ -2462,6 +2544,11 @@ class DBImpl : public DB {
  uint64_t logfile_number_;
  // Log files that we can recycle. Must be protected by db mutex_.
  std::deque<uint64_t> log_recycle_files_;
+  // The minimum log file number taht can be recycled, if log recycling is
+  // enabled. This is used to ensure that log files created by previous
+  // instances of the database are not recycled, as we cannot be sure they
+  // were created in the recyclable format.
+  uint64_t min_log_number_to_recycle_;
  // Protected by log_write_mutex_.
  bool log_dir_synced_;
  // Without two_write_queues, read and writes to log_empty_ are protected by
@@ -2476,8 +2563,10 @@ class DBImpl : public DB {

  bool persistent_stats_cfd_exists_ = true;

-  // alive_log_files_ is protected by mutex_ and log_write_mutex_ with details
-  // as follows:
+  // The current WAL file and those that have not been found obsolete from
+  // memtable flushes. A WAL not on this list might still be pending writer
+  // flush and/or sync and close and might still be in logs_. alive_log_files_
+  // is protected by mutex_ and log_write_mutex_ with details as follows:
  // 1. read by FindObsoleteFiles() which can be called in either application
  //    thread or RocksDB bg threads, both mutex_ and log_write_mutex_ are
  //    held.
@@ -2528,7 +2617,7 @@ class DBImpl : public DB {
  // 8. read by MarkLogsNotSynced() and MarkLogsSynced() are protected by
  //    log_write_mutex_.
  // 9. erase() by MarkLogsSynced() protected by log_write_mutex_.
-  // 10. read by SyncClosedLogs() protected by only log_write_mutex_. This can
+  // 10. read by SyncClosedWals() protected by only log_write_mutex_. This can
  //     happen in bg flush threads after DB::Open() returns success to
  //     applications.
  // 11. reads, e.g. front(), iteration, and back() called by PreprocessWrite()
@@ -2541,7 +2630,7 @@ class DBImpl : public DB {
  // 13. emplace_back() by SwitchMemtable() hold both mutex_ and
  //     log_write_mutex_. This happens in the write group leader. Can conflict
  //     with bg threads calling FindObsoleteFiles(), MarkLogsSynced(),
-  //     SyncClosedLogs(), etc. as well as application threads calling
+  //     SyncClosedWals(), etc. as well as application threads calling
  //     FlushWAL(), SyncWAL(), LockWAL(). This is fine because all parties
  //     require at least log_write_mutex_.
  // 14. iteration called in WriteToWAL(write_group) protected by
@@ -2947,9 +3036,9 @@ inline Status DBImpl::FailIfReadCollapsedHistory(const ColumnFamilyData* cfd,
  if (!full_history_ts_low.empty() &&
      ucmp->CompareTimestamp(ts, full_history_ts_low) < 0) {
    std::stringstream oss;
-    oss << "Read timestamp: " << ts.ToString(true)
+    oss << "Read timestamp: " << ucmp->TimestampToString(ts)
        << " is smaller than full_history_ts_low: "
-        << Slice(full_history_ts_low).ToString(true) << std::endl;
+        << ucmp->TimestampToString(full_history_ts_low) << std::endl;
    return Status::InvalidArgument(oss.str());
  }
  return Status::OK();
@@ -87,6 +87,9 @@ bool DBImpl::ShouldRescheduleFlushRequestToRetainUDT(
  mutex_.AssertHeld();
  assert(flush_req.cfd_to_max_mem_id_to_persist.size() == 1);
  ColumnFamilyData* cfd = flush_req.cfd_to_max_mem_id_to_persist.begin()->first;
+  if (cfd->GetAndClearFlushSkipReschedule()) {
+    return false;
+  }
  uint64_t max_memtable_id =
      flush_req.cfd_to_max_mem_id_to_persist.begin()->second;
  if (cfd->IsDropped() ||
@@ -98,15 +101,20 @@ bool DBImpl::ShouldRescheduleFlushRequestToRetainUDT(
  // alleviated if we continue with the flush instead of postponing it.
  const auto& mutable_cf_options = *cfd->GetLatestMutableCFOptions();

-  // Taking the status of the active Memtable into consideration so that we are
-  // not just checking if DB is currently already in write stall mode.
-  int mem_to_flush = cfd->mem()->ApproximateMemoryUsageFast() >=
-                             cfd->mem()->write_buffer_size() / 2
-                         ? 1
-                         : 0;
+  // Use the same criteria as WaitUntilFlushWouldNotStallWrites does w.r.t
+  // defining what a write stall is about to happen means. If this uses a
+  // stricter criteria, for example, a write stall is about to happen if the
+  // last memtable is 10% full, there is a possibility that manual flush could
+  // be waiting in `WaitUntilFlushWouldNotStallWrites` with the incorrect
+  // expectation that others will clear up the excessive memtables and
+  // eventually let it proceed. The others in this case won't start clearing
+  // until the last memtable is 10% full. To avoid that scenario, the criteria
+  // this uses should be the same or less strict than
+  // `WaitUntilFlushWouldNotStallWrites` does.
  WriteStallCondition write_stall =
      ColumnFamilyData::GetWriteStallConditionAndCause(
-          cfd->imm()->NumNotFlushed() + mem_to_flush, /*num_l0_files=*/0,
+          cfd->GetUnflushedMemTableCountForWriteStallCheck(),
+          /*num_l0_files=*/0,
          /*num_compaction_needed_bytes=*/0, mutable_cf_options,
          *cfd->ioptions())
          .first;
@@ -116,89 +124,19 @@ bool DBImpl::ShouldRescheduleFlushRequestToRetainUDT(
  return true;
 }

-IOStatus DBImpl::SyncClosedLogs(const WriteOptions& write_options,
+IOStatus DBImpl::SyncClosedWals(const WriteOptions& write_options,
                                JobContext* job_context,
                                VersionEdit* synced_wals,
                                bool error_recovery_in_prog) {
-  TEST_SYNC_POINT("DBImpl::SyncClosedLogs:Start");
-  InstrumentedMutexLock l(&log_write_mutex_);
-  autovector<log::Writer*, 1> logs_to_sync;
-  uint64_t current_log_number = logfile_number_;
-  while (logs_.front().number < current_log_number &&
-         logs_.front().IsSyncing()) {
-    log_sync_cv_.Wait();
+  TEST_SYNC_POINT("DBImpl::SyncClosedWals:Start");
+
+  IOStatus io_s = SyncWalImpl(/*include_current_wal*/ false, write_options,
+                              job_context, synced_wals, error_recovery_in_prog);
+  if (!io_s.ok()) {
+    TEST_SYNC_POINT("DBImpl::SyncClosedWals:Failed");
+  } else {
+    TEST_SYNC_POINT("DBImpl::SyncClosedWals:end");
  }
-  for (auto it = logs_.begin();
-       it != logs_.end() && it->number < current_log_number; ++it) {
-    auto& log = *it;
-    log.PrepareForSync();
-    logs_to_sync.push_back(log.writer);
-  }
-
-  IOStatus io_s;
-  if (!logs_to_sync.empty()) {
-    log_write_mutex_.Unlock();
-
-    assert(job_context);
-
-    for (log::Writer* log : logs_to_sync) {
-      ROCKS_LOG_INFO(immutable_db_options_.info_log,
-                     "[JOB %d] Syncing log #%" PRIu64, job_context->job_id,
-                     log->get_log_number());
-      if (error_recovery_in_prog) {
-        log->file()->reset_seen_error();
-      }
-
-      IOOptions io_options;
-      io_s = WritableFileWriter::PrepareIOOptions(write_options, io_options);
-      if (!io_s.ok()) {
-        break;
-      }
-      io_s = log->file()->Sync(io_options, immutable_db_options_.use_fsync);
-      if (!io_s.ok()) {
-        break;
-      }
-
-      if (immutable_db_options_.recycle_log_file_num > 0) {
-        if (error_recovery_in_prog) {
-          log->file()->reset_seen_error();
-        }
-        // Normally the log file is closed when purging obsolete file, but if
-        // log recycling is enabled, the log file is closed here so that it
-        // can be reused.
-        io_s = log->Close(write_options);
-        if (!io_s.ok()) {
-          break;
-        }
-      }
-    }
-    if (io_s.ok()) {
-      IOOptions io_options;
-      io_s = WritableFileWriter::PrepareIOOptions(write_options, io_options);
-      if (io_s.ok()) {
-        io_s = directories_.GetWalDir()->FsyncWithDirOptions(
-            io_options, nullptr,
-            DirFsyncOptions(DirFsyncOptions::FsyncReason::kNewFileSynced));
-      }
-    }
-
-    TEST_SYNC_POINT_CALLBACK("DBImpl::SyncClosedLogs:BeforeReLock",
-                             /*arg=*/nullptr);
-    log_write_mutex_.Lock();
-
-    // "number <= current_log_number - 1" is equivalent to
-    // "number < current_log_number".
-    if (io_s.ok()) {
-      MarkLogsSynced(current_log_number - 1, true, synced_wals);
-    } else {
-      MarkLogsNotSynced(current_log_number - 1);
-    }
-    if (!io_s.ok()) {
-      TEST_SYNC_POINT("DBImpl::SyncClosedLogs:Failed");
-      return io_s;
-    }
-  }
-  TEST_SYNC_POINT("DBImpl::SyncClosedLogs:end");
  return io_s;
 }

@@ -224,19 +162,25 @@ Status DBImpl::FlushMemTableToOutputFile(
  // the host crashes after flushing and before WAL is persistent, the
  // flushed SST may contain data from write batches whose updates to
  // other (unflushed) column families are missing.
+  //
+  // When 2PC is enabled, non-recent WAL(s) may be needed for crash-recovery,
+  // even when there is only one CF in the DB, for prepared transactions that
+  // had not been committed yet. Make sure we sync them to keep the persisted
+  // WAL state at least as new as the persisted SST state.
  const bool needs_to_sync_closed_wals =
      logfile_number_ > 0 &&
-      versions_->GetColumnFamilySet()->NumberOfColumnFamilies() > 1;
+      (versions_->GetColumnFamilySet()->NumberOfColumnFamilies() > 1 ||
+       allow_2pc());

  // If needs_to_sync_closed_wals is true, we need to record the current
  // maximum memtable ID of this column family so that a later PickMemtables()
  // call will not pick memtables whose IDs are higher. This is due to the fact
-  // that SyncClosedLogs() may release the db mutex, and memtable switch can
+  // that SyncClosedWals() may release the db mutex, and memtable switch can
  // happen for this column family in the meantime. The newly created memtables
  // have their data backed by unsynced WALs, thus they cannot be included in
  // this flush job.
  // Another reason why we must record the current maximum memtable ID of this
-  // column family: SyncClosedLogs() may release db mutex, thus it's possible
+  // column family: SyncClosedWals() may release db mutex, thus it's possible
  // for application to continue to insert into memtables increasing db's
  // sequence number. The application may take a snapshot, but this snapshot is
  // not included in `snapshot_seqs` which will be passed to flush job because
@@ -250,7 +194,7 @@ Status DBImpl::FlushMemTableToOutputFile(

  // If needs_to_sync_closed_wals is false, then the flush job will pick ALL
  // existing memtables of the column family when PickMemTable() is called
-  // later. Although we won't call SyncClosedLogs() in this case, we may still
+  // later. Although we won't call SyncClosedWals() in this case, we may still
  // call the callbacks of the listeners, i.e. NotifyOnFlushBegin() which also
  // releases and re-acquires the db mutex. In the meantime, the application
  // can still insert into the memtables and increase the db's sequence number.
@@ -280,12 +224,12 @@ Status DBImpl::FlushMemTableToOutputFile(
  bool need_cancel = false;
  IOStatus log_io_s = IOStatus::OK();
  if (needs_to_sync_closed_wals) {
-    // SyncClosedLogs() may unlock and re-lock the log_write_mutex multiple
+    // SyncClosedWals() may unlock and re-lock the log_write_mutex multiple
    // times.
    VersionEdit synced_wals;
    bool error_recovery_in_prog = error_handler_.IsRecoveryInProgress();
    mutex_.Unlock();
-    log_io_s = SyncClosedLogs(write_options, job_context, &synced_wals,
+    log_io_s = SyncClosedWals(write_options, job_context, &synced_wals,
                              error_recovery_in_prog);
    mutex_.Lock();
    if (log_io_s.ok() && synced_wals.IsWalAddition()) {
@@ -300,7 +244,7 @@ Status DBImpl::FlushMemTableToOutputFile(
      error_handler_.SetBGError(log_io_s, BackgroundErrorReason::kFlush);
    }
  } else {
-    TEST_SYNC_POINT("DBImpl::SyncClosedLogs:Skip");
+    TEST_SYNC_POINT("DBImpl::SyncClosedWals:Skip");
  }
  s = log_io_s;

@@ -574,7 +518,7 @@ Status DBImpl::AtomicFlushMemTablesToOutputFiles(
    VersionEdit synced_wals;
    bool error_recovery_in_prog = error_handler_.IsRecoveryInProgress();
    mutex_.Unlock();
-    log_io_s = SyncClosedLogs(write_options, job_context, &synced_wals,
+    log_io_s = SyncClosedWals(write_options, job_context, &synced_wals,
                              error_recovery_in_prog);
    mutex_.Lock();
    if (log_io_s.ok() && synced_wals.IsWalAddition()) {
@@ -1088,7 +1032,12 @@ Status DBImpl::IncreaseFullHistoryTsLowImpl(ColumnFamilyData* cfd,
  assert(ucmp->timestamp_size() == ts_low.size() && !ts_low.empty());
  if (!current_ts_low.empty() &&
      ucmp->CompareTimestamp(ts_low, current_ts_low) < 0) {
-    return Status::InvalidArgument("Cannot decrease full_history_ts_low");
+    std::stringstream oss;
+    oss << "Current full_history_ts_low: "
+        << ucmp->TimestampToString(current_ts_low)
+        << " is higher than provided ts: " << ucmp->TimestampToString(ts_low)
+        << std::endl;
+    return Status::InvalidArgument(oss.str());
  }

  Status s = versions_->LogAndApply(cfd, *cfd->GetLatestMutableCFOptions(),
@@ -1120,6 +1069,11 @@ Status DBImpl::CompactRangeInternal(const CompactRangeOptions& options,
  if (options.target_path_id >= cfd->ioptions()->cf_paths.size()) {
    return Status::InvalidArgument("Invalid target path ID");
  }
+  if (options.change_level &&
+      cfd->ioptions()->compaction_style == kCompactionStyleFIFO) {
+    return Status::NotSupported(
+        "FIFO compaction does not support change_level.");
+  }

  bool flush_needed = true;

@@ -1180,6 +1134,16 @@ Status DBImpl::CompactRangeInternal(const CompactRangeOptions& options,
                            final_output_level, options, begin, end, exclusive,
                            false /* disable_trivial_move */,
                            std::numeric_limits<uint64_t>::max(), trim_ts);
+  } else if (cfd->ioptions()->compaction_style == kCompactionStyleFIFO) {
+    // FIFOCompactionPicker::CompactRange() will ignore the input key range
+    // [begin, end] and just try to pick compaction based on the configured
+    // option `compaction_options_fifo`. So we skip checking if [begin, end]
+    // overlaps with the DB here.
+    final_output_level = 0;
+    s = RunManualCompaction(cfd, /*input_level=*/0, final_output_level, options,
+                            begin, end, exclusive,
+                            false /* disable_trivial_move */,
+                            std::numeric_limits<uint64_t>::max(), trim_ts);
  } else {
    int first_overlapped_level = kInvalidLevel;
    {
@@ -1264,8 +1228,7 @@ Status DBImpl::CompactRangeInternal(const CompactRangeOptions& options,
      CleanupSuperVersion(super_version);
    }
    if (s.ok() && first_overlapped_level != kInvalidLevel) {
-      if (cfd->ioptions()->compaction_style == kCompactionStyleUniversal ||
-          cfd->ioptions()->compaction_style == kCompactionStyleFIFO) {
+      if (cfd->ioptions()->compaction_style == kCompactionStyleUniversal) {
        assert(first_overlapped_level == 0);
        s = RunManualCompaction(
            cfd, first_overlapped_level, first_overlapped_level, options, begin,
@@ -1420,6 +1383,14 @@ Status DBImpl::CompactFiles(const CompactionOptions& compact_options,
  LogBuffer log_buffer(InfoLogLevel::INFO_LEVEL,
                       immutable_db_options_.info_log.get());

+  if (compact_options.compression !=
+      CompressionType::kDisableCompressionOption) {
+    ROCKS_LOG_WARN(immutable_db_options_.info_log,
+                   "[%s] [JOB %d] Found use of deprecated option "
+                   "`CompactionOptions::compression`",
+                   cfd->GetName().c_str(), job_context.job_id);
+  }
+
  // Perform CompactFiles
  TEST_SYNC_POINT("TestCompactFiles::IngestExternalFile2");
  TEST_SYNC_POINT_CALLBACK("TestCompactFiles:PausingManualCompaction:3",
@@ -1509,16 +1480,11 @@ Status DBImpl::CompactFilesImpl(
        std::to_string(cfd->ioptions()->num_levels - 1));
  }

-  Status s = cfd->compaction_picker()->SanitizeCompactionInputFiles(
-      &input_set, cf_meta, output_level);
-  TEST_SYNC_POINT("DBImpl::CompactFilesImpl::PostSanitizeCompactionInputFiles");
-  if (!s.ok()) {
-    return s;
-  }
-
  std::vector<CompactionInputFiles> input_files;
-  s = cfd->compaction_picker()->GetCompactionInputsFromFileNumbers(
-      &input_files, &input_set, version->storage_info(), compact_options);
+  Status s = cfd->compaction_picker()->SanitizeAndConvertCompactionInputFiles(
+      &input_set, cf_meta, output_level, version->storage_info(), &input_files);
+  TEST_SYNC_POINT(
+      "DBImpl::CompactFilesImpl::PostSanitizeAndConvertCompactionInputFiles");
  if (!s.ok()) {
    return s;
  }
@@ -1774,6 +1740,8 @@ void DBImpl::NotifyOnCompactionCompleted(

 // REQUIREMENT: block all background work by calling PauseBackgroundWork()
 // before calling this function
+// TODO (hx235): Replace Status::NotSupported() with Status::Aborted() for
+// better semantics like CompactFiles()
 Status DBImpl::ReFitLevel(ColumnFamilyData* cfd, int level, int target_level) {
  assert(level < cfd->NumberLevels());
  if (target_level >= cfd->NumberLevels()) {
@@ -1809,6 +1777,8 @@ Status DBImpl::ReFitLevel(ColumnFamilyData* cfd, int level, int target_level) {
  if (to_level != level) {
    std::vector<CompactionInputFiles> input(1);
    input[0].level = level;
+    // TODO (hx235): Only refit the output files in the current manual
+    // compaction instead of all the files in the output level
    for (auto& f : vstorage->LevelFiles(level)) {
      input[0].files.push_back(f);
    }
@@ -1840,6 +1810,12 @@ Status DBImpl::ReFitLevel(ColumnFamilyData* cfd, int level, int target_level) {
      }
    } else {
      // to_level < level
+      if (to_level == 0 && input[0].files.size() > 1) {
+        refitting_level_ = false;
+        return Status::Aborted(
+            "Moving more than 1 file from non-L0 to L0 is not allowed as it "
+            "does not bring any benefit to read nor write throughput.");
+      }
      // Check levels are empty for a trivial move
      for (int l = to_level; l < level; l++) {
        if (vstorage->NumLevelFiles(l) > 0) {
@@ -2154,16 +2130,6 @@ Status DBImpl::RunManualCompaction(
               manual.begin, manual.end, &manual.manual_end, &manual_conflict,
               max_file_num_to_ignore, trim_ts)) == nullptr &&
          manual_conflict))) {
-      if (!scheduled) {
-        // There is a conflicting compaction
-        if (manual_compaction_paused_ > 0 || manual.canceled == true) {
-          // Stop waiting since it was canceled. Pretend the error came from
-          // compaction so the below cleanup/error handling code can process it.
-          manual.done = true;
-          manual.status =
-              Status::Incomplete(Status::SubCode::kManualCompactionPaused);
-        }
-      }
      if (!manual.done) {
        bg_cv_.Wait();
      }
@@ -2238,6 +2204,17 @@ Status DBImpl::RunManualCompaction(
        *final_output_level = compaction->output_level();
      }
    }
+    if (!scheduled) {
+      // There is nothing scheduled to wait on, so any cancellation can end the
+      // manual now.
+      if (manual_compaction_paused_ > 0 || manual.canceled == true) {
+        // Stop waiting since it was canceled. Pretend the error came from
+        // compaction so the below cleanup/error handling code can process it.
+        manual.done = true;
+        manual.status =
+            Status::Incomplete(Status::SubCode::kManualCompactionPaused);
+      }
+    }
  }

  log_buffer.FlushBufferToLog();
@@ -2271,6 +2248,23 @@ void DBImpl::GenerateFlushRequest(const autovector<ColumnFamilyData*>& cfds,
  }
 }

+void DBImpl::NotifyOnManualFlushScheduled(autovector<ColumnFamilyData*> cfds,
+                                          FlushReason flush_reason) {
+  if (immutable_db_options_.listeners.size() == 0U) {
+    return;
+  }
+  if (shutting_down_.load(std::memory_order_acquire)) {
+    return;
+  }
+  std::vector<ManualFlushInfo> info;
+  for (ColumnFamilyData* cfd : cfds) {
+    info.push_back({cfd->GetID(), cfd->GetName(), flush_reason});
+  }
+  for (const auto& listener : immutable_db_options_.listeners) {
+    listener->OnManualFlushScheduled(this, info);
+  }
+}
+
 Status DBImpl::FlushMemTable(ColumnFamilyData* cfd,
                             const FlushOptions& flush_options,
                             FlushReason flush_reason,
@@ -2375,7 +2369,14 @@ Status DBImpl::FlushMemTable(ColumnFamilyData* cfd,
        }
      }
      for (const auto& req : flush_reqs) {
-        SchedulePendingFlush(req);
+        assert(req.cfd_to_max_mem_id_to_persist.size() == 1);
+        ColumnFamilyData* loop_cfd =
+            req.cfd_to_max_mem_id_to_persist.begin()->first;
+        bool already_queued_for_flush = loop_cfd->queued_for_flush();
+        bool flush_req_enqueued = SchedulePendingFlush(req);
+        if (already_queued_for_flush || flush_req_enqueued) {
+          loop_cfd->SetFlushSkipReschedule();
+        }
      }
      MaybeScheduleFlushOrCompaction();
    }
@@ -2387,6 +2388,8 @@ Status DBImpl::FlushMemTable(ColumnFamilyData* cfd,
      }
    }
  }
+
+  NotifyOnManualFlushScheduled({cfd}, flush_reason);
  TEST_SYNC_POINT("DBImpl::FlushMemTable:AfterScheduleFlush");
  TEST_SYNC_POINT("DBImpl::FlushMemTable:BeforeWaitForBgFlush");
  if (s.ok() && flush_options.wait) {
@@ -2531,6 +2534,7 @@ Status DBImpl::AtomicFlushMemTables(
      }
    }
  }
+  NotifyOnManualFlushScheduled(cfds, flush_reason);
  TEST_SYNC_POINT("DBImpl::AtomicFlushMemTables:AfterScheduleFlush");
  TEST_SYNC_POINT("DBImpl::AtomicFlushMemTables:BeforeWaitForBgFlush");
  if (s.ok() && flush_options.wait) {
@@ -2588,7 +2592,9 @@ Status DBImpl::RetryFlushesForErrorRecovery(FlushReason flush_reason,
          flush_reason,
          {{cfd,
            std::numeric_limits<uint64_t>::max() /* max_mem_id_to_persist */}}};
-      SchedulePendingFlush(flush_req);
+      if (SchedulePendingFlush(flush_req)) {
+        cfd->SetFlushSkipReschedule();
+      };
    }
  }
  MaybeScheduleFlushOrCompaction();
@@ -2676,13 +2682,13 @@ Status DBImpl::WaitUntilFlushWouldNotStallWrites(ColumnFamilyData* cfd,
      // mode due to pending compaction bytes, but that's less common
      // No extra immutable Memtable will be created if the current Memtable is
      // empty.
-      int mem_to_flush = cfd->mem()->IsEmpty() ? 0 : 1;
-      write_stall_condition = ColumnFamilyData::GetWriteStallConditionAndCause(
-                                  cfd->imm()->NumNotFlushed() + mem_to_flush,
-                                  vstorage->l0_delay_trigger_count() + 1,
-                                  vstorage->estimated_compaction_needed_bytes(),
-                                  mutable_cf_options, *cfd->ioptions())
-                                  .first;
+      write_stall_condition =
+          ColumnFamilyData::GetWriteStallConditionAndCause(
+              cfd->GetUnflushedMemTableCountForWriteStallCheck(),
+              vstorage->l0_delay_trigger_count() + 1,
+              vstorage->estimated_compaction_needed_bytes(), mutable_cf_options,
+              *cfd->ioptions())
+              .first;
    } while (write_stall_condition != WriteStallCondition::kNormal);
  }
  return Status::OK();
@@ -2994,13 +3000,14 @@ ColumnFamilyData* DBImpl::PickCompactionFromQueue(
  return cfd;
 }

-void DBImpl::SchedulePendingFlush(const FlushRequest& flush_req) {
+bool DBImpl::SchedulePendingFlush(const FlushRequest& flush_req) {
  mutex_.AssertHeld();
+  bool enqueued = false;
  if (reject_new_background_jobs_) {
-    return;
+    return enqueued;
  }
  if (flush_req.cfd_to_max_mem_id_to_persist.empty()) {
-    return;
+    return enqueued;
  }
  if (!immutable_db_options_.atomic_flush) {
    // For the non-atomic flush case, we never schedule multiple column
@@ -3015,6 +3022,7 @@ void DBImpl::SchedulePendingFlush(const FlushRequest& flush_req) {
      cfd->set_queued_for_flush(true);
      ++unscheduled_flushes_;
      flush_queue_.push_back(flush_req);
+      enqueued = true;
    }
  } else {
    for (auto& iter : flush_req.cfd_to_max_mem_id_to_persist) {
@@ -3023,7 +3031,9 @@ void DBImpl::SchedulePendingFlush(const FlushRequest& flush_req) {
    }
    ++unscheduled_flushes_;
    flush_queue_.push_back(flush_req);
+    enqueued = true;
  }
+  return enqueued;
 }

 void DBImpl::SchedulePendingCompaction(ColumnFamilyData* cfd) {
@@ -3032,6 +3042,8 @@ void DBImpl::SchedulePendingCompaction(ColumnFamilyData* cfd) {
    return;
  }
  if (!cfd->queued_for_compaction() && cfd->NeedsCompaction()) {
+    TEST_SYNC_POINT_CALLBACK("SchedulePendingCompaction::cfd",
+                             static_cast<void*>(cfd));
    AddToCompactionQueue(cfd);
    ++unscheduled_compactions_;
  }
@@ -3319,7 +3331,7 @@ void DBImpl::BackgroundCallFlush(Env::Priority thread_pri) {
      bg_cv_.SignalAll();  // In case a waiter can proceed despite the error
      mutex_.Unlock();
      ROCKS_LOG_ERROR(immutable_db_options_.info_log,
-                      "[JOB %d] Waiting after background flush error: %s"
+                      "[JOB %d] Waiting after background flush error: %s, "
                      "Accumulated background error counts: %" PRIu64,
                      job_context.job_id, s.ToString().c_str(), error_cnt);
      log_buffer.FlushBufferToLog();
@@ -4344,6 +4356,7 @@ Status DBImpl::WaitForCompact(
    }
    if ((bg_bottom_compaction_scheduled_ || bg_compaction_scheduled_ ||
         bg_flush_scheduled_ || unscheduled_compactions_ ||
+         (wait_for_compact_options.wait_for_purge && bg_purge_scheduled_) ||
         unscheduled_flushes_ || error_handler_.IsRecoveryInProgress()) &&
        (error_handler_.GetBGError().ok())) {
      if (wait_for_compact_options.timeout.count()) {
@@ -4351,6 +4364,7 @@ Status DBImpl::WaitForCompact(
          return Status::TimedOut();
        }
      } else {
+        TEST_SYNC_POINT("DBImpl::WaitForCompact:InsideLoop");
        bg_cv_.Wait();
      }
    } else if (wait_for_compact_options.close_db) {
@@ -314,6 +314,11 @@ const autovector<uint64_t>& DBImpl::TEST_GetFilesToQuarantine() const {
  return error_handler_.GetFilesToQuarantine();
 }

+void DBImpl::TEST_DeleteObsoleteFiles() {
+  InstrumentedMutexLock l(&mutex_);
+  DeleteObsoleteFiles();
+}
+
 size_t DBImpl::TEST_EstimateInMemoryStatsHistorySize() const {
  InstrumentedMutexLock l(&const_cast<DBImpl*>(this)->stats_history_mutex_);
  return EstimateInMemoryStatsHistorySize();
@@ -28,6 +28,8 @@ uint64_t DBImpl::MinLogNumberToKeep() {
  return versions_->min_log_number_to_keep();
 }

+uint64_t DBImpl::MinLogNumberToRecycle() { return min_log_number_to_recycle_; }
+
 uint64_t DBImpl::MinObsoleteSstNumberToKeep() {
  mutex_.AssertHeld();
  if (!pending_outputs_.empty()) {
@@ -175,31 +177,10 @@ void DBImpl::FindObsoleteFiles(JobContext* job_context, bool force,
    versions_->AddLiveFiles(&job_context->sst_live, &job_context->blob_live);
    InfoLogPrefix info_log_prefix(!immutable_db_options_.db_log_dir.empty(),
                                  dbname_);
-    std::set<std::string> paths;
-    for (size_t path_id = 0; path_id < immutable_db_options_.db_paths.size();
-         path_id++) {
-      paths.insert(immutable_db_options_.db_paths[path_id].path);
-    }
-
-    // Note that if cf_paths is not specified in the ColumnFamilyOptions
-    // of a particular column family, we use db_paths as the cf_paths
-    // setting. Hence, there can be multiple duplicates of files from db_paths
-    // in the following code. The duplicate are removed while identifying
-    // unique files in PurgeObsoleteFiles.
-    for (auto cfd : *versions_->GetColumnFamilySet()) {
-      for (size_t path_id = 0; path_id < cfd->ioptions()->cf_paths.size();
-           path_id++) {
-        auto& path = cfd->ioptions()->cf_paths[path_id].path;
-
-        if (paths.find(path) == paths.end()) {
-          paths.insert(path);
-        }
-      }
-    }
-
+    // PurgeObsoleteFiles will dedupe duplicate files.
    IOOptions io_opts;
    io_opts.do_not_recurse = true;
-    for (auto& path : paths) {
+    for (auto& path : CollectAllDBPaths()) {
      // set of all files in the directory. We'll exclude files that are still
      // alive in the subsequent processings.
      std::vector<std::string> files;
@@ -298,7 +279,8 @@ void DBImpl::FindObsoleteFiles(JobContext* job_context, bool force,
    while (alive_log_files_.begin()->number < min_log_number) {
      auto& earliest = *alive_log_files_.begin();
      if (immutable_db_options_.recycle_log_file_num >
-          log_recycle_files_.size()) {
+              log_recycle_files_.size() &&
+          earliest.number >= MinLogNumberToRecycle()) {
        ROCKS_LOG_INFO(immutable_db_options_.info_log,
                       "adding log %" PRIu64 " to recycle list\n",
                       earliest.number);
@@ -330,6 +312,26 @@ void DBImpl::FindObsoleteFiles(JobContext* job_context, bool force,
        // logs_ could have changed while we were waiting.
        continue;
      }
+      // This WAL file is not live, so it's OK if we never sync the rest of it.
+      // If it's already closed, then it's been fully synced. If
+      // !background_close_inactive_wals then we need to Close it before
+      // removing from logs_ but not blocking while holding log_write_mutex_.
+      if (!immutable_db_options_.background_close_inactive_wals &&
+          log.writer->file()) {
+        // We are taking ownership of and pinning the front entry, so we can
+        // expect it to be the same after releasing and re-acquiring the lock
+        log.PrepareForSync();
+        log_write_mutex_.Unlock();
+        // TODO: maybe check the return value of Close.
+        // TODO: plumb Env::IOActivity, Env::IOPriority
+        auto s = log.writer->file()->Close({});
+        s.PermitUncheckedError();
+        log_write_mutex_.Lock();
+        log.writer->PublishIfClosed();
+        assert(&log == &logs_.front());
+        log.FinishSync();
+        log_sync_cv_.SignalAll();
+      }
      logs_to_free_.push_back(log.ReleaseWriter());
      logs_.pop_front();
    }
@@ -428,12 +430,24 @@ void DBImpl::PurgeObsoleteFiles(JobContext& state, bool schedule_only) {
      state.manifest_delete_files.size());
  // We may ignore the dbname when generating the file names.
  for (auto& file : state.sst_delete_files) {
-    if (!file.only_delete_metadata) {
-      candidate_files.emplace_back(
-          MakeTableFileName(file.metadata->fd.GetNumber()), file.path);
-    }
-    if (file.metadata->table_reader_handle) {
-      table_cache_->Release(file.metadata->table_reader_handle);
+    auto* handle = file.metadata->table_reader_handle;
+    if (file.only_delete_metadata) {
+      if (handle) {
+        // Simply release handle of file that is not being deleted
+        table_cache_->Release(handle);
+      }
+    } else {
+      // File is being deleted (actually obsolete)
+      auto number = file.metadata->fd.GetNumber();
+      candidate_files.emplace_back(MakeTableFileName(number), file.path);
+      if (handle == nullptr) {
+        // For files not "pinned" in table cache
+        handle = TableCache::Lookup(table_cache_.get(), number);
+      }
+      if (handle) {
+        TableCache::ReleaseObsolete(table_cache_.get(), handle,
+                                    file.uncache_aggressiveness);
+      }
    }
    file.DeleteMetadata();
  }
@@ -509,7 +523,7 @@ void DBImpl::PurgeObsoleteFiles(JobContext& state, bool schedule_only) {
  for (const auto w : state.logs_to_free) {
    // TODO: maybe check the return value of Close.
    // TODO: plumb Env::IOActivity, Env::IOPriority
-    auto s = w->Close(WriteOptions());
+    auto s = w->Close({});
    s.PermitUncheckedError();
  }

@@ -595,8 +609,6 @@ void DBImpl::PurgeObsoleteFiles(JobContext& state, bool schedule_only) {
    std::string fname;
    std::string dir_to_sync;
    if (type == kTableFile) {
-      // evict from cache
-      TableCache::Evict(table_cache_.get(), number);
      fname = MakeTableFileName(candidate_file.file_path, number);
      dir_to_sync = candidate_file.file_path;
    } else if (type == kBlobFile) {
@@ -963,28 +975,26 @@ Status DBImpl::SetupDBId(const WriteOptions& write_options, bool read_only,
  return s;
 }

-Status DBImpl::DeleteUnreferencedSstFiles(RecoveryContext* recovery_ctx) {
-  mutex_.AssertHeld();
-  std::vector<std::string> paths;
-  paths.push_back(NormalizePath(dbname_ + std::string(1, kFilePathSeparator)));
+std::set<std::string> DBImpl::CollectAllDBPaths() {
+  std::set<std::string> all_db_paths;
+  all_db_paths.insert(NormalizePath(dbname_));
  for (const auto& db_path : immutable_db_options_.db_paths) {
-    paths.push_back(
-        NormalizePath(db_path.path + std::string(1, kFilePathSeparator)));
+    all_db_paths.insert(NormalizePath(db_path.path));
  }
  for (const auto* cfd : *versions_->GetColumnFamilySet()) {
    for (const auto& cf_path : cfd->ioptions()->cf_paths) {
-      paths.push_back(
-          NormalizePath(cf_path.path + std::string(1, kFilePathSeparator)));
+      all_db_paths.insert(NormalizePath(cf_path.path));
    }
  }
-  // Dedup paths
-  std::sort(paths.begin(), paths.end());
-  paths.erase(std::unique(paths.begin(), paths.end()), paths.end());
+  return all_db_paths;
+}

+Status DBImpl::MaybeUpdateNextFileNumber(RecoveryContext* recovery_ctx) {
+  mutex_.AssertHeld();
  uint64_t next_file_number = versions_->current_next_file_number();
  uint64_t largest_file_number = next_file_number;
  Status s;
-  for (const auto& path : paths) {
+  for (const auto& path : CollectAllDBPaths()) {
    std::vector<std::string> files;
    s = env_->GetChildren(path, &files);
    if (!s.ok()) {
@@ -996,13 +1006,10 @@ Status DBImpl::DeleteUnreferencedSstFiles(RecoveryContext* recovery_ctx) {
      if (!ParseFileName(fname, &number, &type)) {
        continue;
      }
-      // path ends with '/' or '\\'
-      const std::string normalized_fpath = path + fname;
+      const std::string normalized_fpath = path + kFilePathSeparator + fname;
      largest_file_number = std::max(largest_file_number, number);
-      if (type == kTableFile && number >= next_file_number &&
-          recovery_ctx->files_to_delete_.find(normalized_fpath) ==
-              recovery_ctx->files_to_delete_.end()) {
-        recovery_ctx->files_to_delete_.emplace(normalized_fpath, path);
+      if ((type == kTableFile || type == kBlobFile)) {
+        recovery_ctx->existing_data_files_.push_back(normalized_fpath);
      }
    }
  }
@@ -1022,5 +1029,4 @@ Status DBImpl::DeleteUnreferencedSstFiles(RecoveryContext* recovery_ctx) {
  recovery_ctx->UpdateVersionEdits(default_cfd, edit);
  return s;
 }
-
 }  // namespace ROCKSDB_NAMESPACE
@@ -0,0 +1,348 @@
+//  Copyright (c) 2024-present, Facebook, Inc.  All rights reserved.
+//  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_impl/db_impl_follower.h"
+
+#include <algorithm>
+#include <cinttypes>
+
+#include "db/arena_wrapped_db_iter.h"
+#include "db/merge_context.h"
+#include "env/composite_env_wrapper.h"
+#include "env/fs_on_demand.h"
+#include "logging/auto_roll_logger.h"
+#include "logging/logging.h"
+#include "monitoring/perf_context_imp.h"
+#include "rocksdb/configurable.h"
+#include "rocksdb/db.h"
+#include "util/cast_util.h"
+#include "util/write_batch_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+DBImplFollower::DBImplFollower(const DBOptions& db_options,
+                               std::unique_ptr<Env>&& env,
+                               const std::string& dbname, std::string src_path)
+    : DBImplSecondary(db_options, dbname, ""),
+      env_guard_(std::move(env)),
+      stop_requested_(false),
+      src_path_(std::move(src_path)),
+      cv_(&mu_) {
+  ROCKS_LOG_INFO(immutable_db_options_.info_log,
+                 "Opening the db in follower mode");
+  LogFlush(immutable_db_options_.info_log);
+}
+
+DBImplFollower::~DBImplFollower() {
+  Status s = Close();
+  if (!s.ok()) {
+    ROCKS_LOG_INFO(immutable_db_options_.info_log, "Error closing DB : %s",
+                   s.ToString().c_str());
+  }
+}
+
+// Recover a follower DB instance by reading the MANIFEST. The verification
+// as part of the MANIFEST replay will ensure that local links to the
+// leader's files are created, thus ensuring we can continue reading them
+// even if the leader deletes those files due to compaction.
+// TODO:
+//  1. Devise a mechanism to prevent misconfiguration by, for example,
+//     keeping a local copy of the IDENTITY file and cross checking
+//  2. Make the recovery more robust by retrying if the first attempt
+//     fails.
+Status DBImplFollower::Recover(
+    const std::vector<ColumnFamilyDescriptor>& column_families,
+    bool /*readonly*/, bool /*error_if_wal_file_exists*/,
+    bool /*error_if_data_exists_in_wals*/, bool /*is_retry*/, uint64_t*,
+    RecoveryContext* /*recovery_ctx*/, bool* /*can_retry*/) {
+  mutex_.AssertHeld();
+
+  JobContext job_context(0);
+  Status s;
+  s = static_cast<ReactiveVersionSet*>(versions_.get())
+          ->Recover(column_families, &manifest_reader_, &manifest_reporter_,
+                    &manifest_reader_status_);
+  if (!s.ok()) {
+    if (manifest_reader_status_) {
+      manifest_reader_status_->PermitUncheckedError();
+    }
+    return s;
+  }
+  if (immutable_db_options_.paranoid_checks && s.ok()) {
+    s = CheckConsistency();
+  }
+  if (s.ok()) {
+    default_cf_handle_ = new ColumnFamilyHandleImpl(
+        versions_->GetColumnFamilySet()->GetDefault(), this, &mutex_);
+    default_cf_internal_stats_ = default_cf_handle_->cfd()->internal_stats();
+
+    // Start the periodic catch-up thread
+    // TODO: See if it makes sense to have a threadpool, rather than a thread
+    // per follower DB instance
+    catch_up_thread_.reset(
+        new port::Thread(&DBImplFollower::PeriodicRefresh, this));
+  }
+
+  return s;
+}
+
+// Try to catch up by tailing the MANIFEST.
+// TODO:
+//   1. Cleanup obsolete files afterward
+//   2. Add some error notifications and statistics
+Status DBImplFollower::TryCatchUpWithLeader() {
+  assert(versions_.get() != nullptr);
+  assert(manifest_reader_.get() != nullptr);
+  Status s;
+
+  TEST_SYNC_POINT("DBImplFollower::TryCatchupWithLeader:Begin1");
+  TEST_SYNC_POINT("DBImplFollower::TryCatchupWithLeader:Begin2");
+  // read the manifest and apply new changes to the follower instance
+  std::unordered_set<ColumnFamilyData*> cfds_changed;
+  JobContext job_context(0, true /*create_superversion*/);
+  {
+    InstrumentedMutexLock lock_guard(&mutex_);
+    std::vector<std::string> files_to_delete;
+    s = static_cast_with_check<ReactiveVersionSet>(versions_.get())
+            ->ReadAndApply(&mutex_, &manifest_reader_,
+                           manifest_reader_status_.get(), &cfds_changed,
+                           &files_to_delete);
+    ReleaseFileNumberFromPendingOutputs(pending_outputs_inserted_elem_);
+    pending_outputs_inserted_elem_.reset(new std::list<uint64_t>::iterator(
+        CaptureCurrentFileNumberInPendingOutputs()));
+
+    ROCKS_LOG_INFO(immutable_db_options_.info_log, "Last sequence is %" PRIu64,
+                   static_cast<uint64_t>(versions_->LastSequence()));
+    ROCKS_LOG_INFO(
+        immutable_db_options_.info_log, "Next file number is %" PRIu64,
+        static_cast<uint64_t>(versions_->current_next_file_number()));
+    for (ColumnFamilyData* cfd : cfds_changed) {
+      if (cfd->IsDropped()) {
+        ROCKS_LOG_DEBUG(immutable_db_options_.info_log, "[%s] is dropped\n",
+                        cfd->GetName().c_str());
+        continue;
+      }
+      VersionStorageInfo::LevelSummaryStorage tmp;
+      ROCKS_LOG_DEBUG(immutable_db_options_.info_log,
+                      "[%s] Level summary: %s\n", cfd->GetName().c_str(),
+                      cfd->current()->storage_info()->LevelSummary(&tmp));
+    }
+
+    if (s.ok()) {
+      for (auto cfd : cfds_changed) {
+        if (cfd->mem()->GetEarliestSequenceNumber() <
+            versions_->LastSequence()) {
+          // Construct a new memtable with earliest sequence number set to the
+          // last sequence number in the VersionSet. This matters when
+          // DBImpl::MultiCFSnapshot tries to get consistent references
+          // to super versions in a lock free manner, it checks the earliest
+          // sequence number to detect if there was a change in version in
+          // the meantime.
+          const MutableCFOptions mutable_cf_options =
+              *cfd->GetLatestMutableCFOptions();
+          MemTable* new_mem = cfd->ConstructNewMemtable(
+              mutable_cf_options, versions_->LastSequence());
+          cfd->mem()->SetNextLogNumber(cfd->GetLogNumber());
+          cfd->mem()->ConstructFragmentedRangeTombstones();
+          cfd->imm()->Add(cfd->mem(), &job_context.memtables_to_free);
+          new_mem->Ref();
+          cfd->SetMemtable(new_mem);
+        }
+
+        // This will check if the old memtable is still referenced
+        cfd->imm()->RemoveOldMemTables(cfd->GetLogNumber(),
+                                       &job_context.memtables_to_free);
+        auto& sv_context = job_context.superversion_contexts.back();
+        cfd->InstallSuperVersion(&sv_context, &mutex_);
+        sv_context.NewSuperVersion();
+      }
+    }
+
+    for (auto& file : files_to_delete) {
+      IOStatus io_s = fs_->DeleteFile(file, IOOptions(), nullptr);
+      if (!io_s.ok()) {
+        ROCKS_LOG_INFO(immutable_db_options_.info_log,
+                       "Cannot delete file %s: %s", file.c_str(),
+                       io_s.ToString().c_str());
+      }
+    }
+  }
+  job_context.Clean();
+
+  // Cleanup unused, obsolete files.
+  JobContext purge_files_job_context(0);
+  {
+    InstrumentedMutexLock lock_guard(&mutex_);
+    // Currently, follower instance does not create any database files, thus
+    // is unnecessary for the follower to force full scan.
+    FindObsoleteFiles(&purge_files_job_context, /*force=*/false);
+  }
+  if (purge_files_job_context.HaveSomethingToDelete()) {
+    PurgeObsoleteFiles(purge_files_job_context);
+  }
+  purge_files_job_context.Clean();
+
+  TEST_SYNC_POINT("DBImplFollower::TryCatchupWithLeader:End");
+
+  return s;
+}
+
+void DBImplFollower::PeriodicRefresh() {
+  while (!stop_requested_.load()) {
+    MutexLock l(&mu_);
+    int64_t wait_until =
+        immutable_db_options_.clock->NowMicros() +
+        immutable_db_options_.follower_refresh_catchup_period_ms * 1000;
+    immutable_db_options_.clock->TimedWait(
+        &cv_, std::chrono::microseconds(wait_until));
+    if (stop_requested_.load()) {
+      break;
+    }
+    Status s;
+    for (uint64_t i = 0;
+         i < immutable_db_options_.follower_catchup_retry_count &&
+         !stop_requested_.load();
+         ++i) {
+      s = TryCatchUpWithLeader();
+
+      if (s.ok()) {
+        ROCKS_LOG_INFO(immutable_db_options_.info_log,
+                       "Successful catch up on attempt %llu",
+                       static_cast<unsigned long long>(i));
+        break;
+      }
+      wait_until = immutable_db_options_.clock->NowMicros() +
+                   immutable_db_options_.follower_catchup_retry_wait_ms * 1000;
+      immutable_db_options_.clock->TimedWait(
+          &cv_, std::chrono::microseconds(wait_until));
+    }
+    if (!s.ok()) {
+      ROCKS_LOG_INFO(immutable_db_options_.info_log, "Catch up unsuccessful");
+    }
+  }
+}
+
+Status DBImplFollower::Close() {
+  if (catch_up_thread_) {
+    stop_requested_.store(true);
+    {
+      MutexLock l(&mu_);
+      cv_.SignalAll();
+    }
+    catch_up_thread_->join();
+    catch_up_thread_.reset();
+  }
+
+  ReleaseFileNumberFromPendingOutputs(pending_outputs_inserted_elem_);
+
+  return DBImpl::Close();
+}
+
+Status DB::OpenAsFollower(const Options& options, const std::string& dbname,
+                          const std::string& leader_path,
+                          std::unique_ptr<DB>* dbptr) {
+  dbptr->reset();
+
+  DBOptions db_options(options);
+  ColumnFamilyOptions cf_options(options);
+  std::vector<ColumnFamilyDescriptor> column_families;
+  column_families.emplace_back(kDefaultColumnFamilyName, cf_options);
+  std::vector<ColumnFamilyHandle*> handles;
+
+  Status s = DB::OpenAsFollower(db_options, dbname, leader_path,
+                                column_families, &handles, dbptr);
+  if (s.ok()) {
+    assert(handles.size() == 1);
+    delete handles[0];
+  }
+  return s;
+}
+
+Status DB::OpenAsFollower(
+    const DBOptions& db_options, const std::string& dbname,
+    const std::string& src_path,
+    const std::vector<ColumnFamilyDescriptor>& column_families,
+    std::vector<ColumnFamilyHandle*>* handles, std::unique_ptr<DB>* dbptr) {
+  dbptr->reset();
+
+  FileSystem* fs = db_options.env->GetFileSystem().get();
+  {
+    IOStatus io_s;
+    if (db_options.create_if_missing) {
+      io_s = fs->CreateDirIfMissing(dbname, IOOptions(), nullptr);
+    } else {
+      io_s = fs->FileExists(dbname, IOOptions(), nullptr);
+    }
+    if (!io_s.ok()) {
+      return static_cast<Status>(io_s);
+    }
+  }
+  std::unique_ptr<Env> new_env(new CompositeEnvWrapper(
+      db_options.env, NewOnDemandFileSystem(db_options.env->GetFileSystem(),
+                                            src_path, dbname)));
+
+  DBOptions tmp_opts(db_options);
+  Status s;
+  tmp_opts.env = new_env.get();
+  if (nullptr == tmp_opts.info_log) {
+    s = CreateLoggerFromOptions(dbname, tmp_opts, &tmp_opts.info_log);
+    if (!s.ok()) {
+      tmp_opts.info_log = nullptr;
+      return s;
+    }
+  }
+
+  handles->clear();
+  DBImplFollower* impl =
+      new DBImplFollower(tmp_opts, std::move(new_env), dbname, src_path);
+  impl->versions_.reset(new ReactiveVersionSet(
+      dbname, &impl->immutable_db_options_, impl->file_options_,
+      impl->table_cache_.get(), impl->write_buffer_manager_,
+      &impl->write_controller_, impl->io_tracer_));
+  impl->column_family_memtables_.reset(
+      new ColumnFamilyMemTablesImpl(impl->versions_->GetColumnFamilySet()));
+  impl->wal_in_db_path_ = impl->immutable_db_options_.IsWalDirSameAsDBPath();
+
+  impl->mutex_.Lock();
+  s = impl->Recover(column_families, /*read_only=*/true,
+                    /*error_if_wal_file_exists=*/false,
+                    /*error_if_data_exists_in_wals=*/false);
+  if (s.ok()) {
+    for (const auto& cf : column_families) {
+      auto cfd =
+          impl->versions_->GetColumnFamilySet()->GetColumnFamily(cf.name);
+      if (nullptr == cfd) {
+        s = Status::InvalidArgument("Column family not found", cf.name);
+        break;
+      }
+      handles->push_back(new ColumnFamilyHandleImpl(cfd, impl, &impl->mutex_));
+    }
+  }
+  SuperVersionContext sv_context(false /* create_superversion */);
+  if (s.ok()) {
+    for (auto cfd : *impl->versions_->GetColumnFamilySet()) {
+      sv_context.NewSuperVersion();
+      cfd->InstallSuperVersion(&sv_context, &impl->mutex_);
+    }
+  }
+  impl->mutex_.Unlock();
+  sv_context.Clean();
+  if (s.ok()) {
+    dbptr->reset(impl);
+    for (auto h : *handles) {
+      impl->NewThreadStatusCfInfo(
+          static_cast_with_check<ColumnFamilyHandleImpl>(h)->cfd());
+    }
+  } else {
+    for (auto h : *handles) {
+      delete h;
+    }
+    handles->clear();
+    delete impl;
+  }
+  return s;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
@@ -0,0 +1,54 @@
+//  Copyright (c) 2024-present, Facebook, Inc.  All rights reserved.
+//  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).
+
+#pragma once
+
+#include <string>
+#include <vector>
+
+#include "db/db_impl/db_impl.h"
+#include "db/db_impl/db_impl_secondary.h"
+#include "logging/logging.h"
+#include "port/port.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class DBImplFollower : public DBImplSecondary {
+ public:
+  DBImplFollower(const DBOptions& db_options, std::unique_ptr<Env>&& env,
+                 const std::string& dbname, std::string src_path);
+  ~DBImplFollower();
+
+  Status Close() override;
+
+ protected:
+  bool OwnTablesAndLogs() const override {
+    // TODO: Change this to true once we've properly implemented file
+    // deletion for the read scaling case
+    return true;
+  }
+
+  Status Recover(const std::vector<ColumnFamilyDescriptor>& column_families,
+                 bool /*readonly*/, bool /*error_if_wal_file_exists*/,
+                 bool /*error_if_data_exists_in_wals*/,
+                 bool /*is_retry*/ = false, uint64_t* = nullptr,
+                 RecoveryContext* /*recovery_ctx*/ = nullptr,
+                 bool* /*can_retry*/ = nullptr) override;
+
+ private:
+  friend class DB;
+
+  Status TryCatchUpWithLeader();
+  void PeriodicRefresh();
+
+  std::unique_ptr<Env> env_guard_;
+  std::unique_ptr<port::Thread> catch_up_thread_;
+  std::atomic<bool> stop_requested_;
+  std::string src_path_;
+  port::Mutex mu_;
+  port::CondVar cv_;
+  std::unique_ptr<std::list<uint64_t>::iterator> pending_outputs_inserted_elem_;
+};
+}  // namespace ROCKSDB_NAMESPACE
@@ -175,15 +175,6 @@ DBOptions SanitizeOptions(const std::string& dbname, const DBOptions& src,
      }
    }
  }
-  // When the DB is stopped, it's possible that there are some .trash files that
-  // were not deleted yet, when we open the DB we will find these .trash files
-  // and schedule them to be deleted (or delete immediately if SstFileManager
-  // was not used)
-  auto sfm = static_cast<SstFileManagerImpl*>(result.sst_file_manager.get());
-  for (size_t i = 0; i < result.db_paths.size(); i++) {
-    DeleteScheduler::CleanupDirectory(result.env, sfm, result.db_paths[i].path)
-        .PermitUncheckedError();
-  }

  // Create a default SstFileManager for purposes of tracking compaction size
  // and facilitating recovery from out of space errors.
@@ -310,7 +301,10 @@ Status DBImpl::NewDB(std::vector<std::string>* new_filenames) {
  }
  if (immutable_db_options_.write_dbid_to_manifest) {
    std::string temp_db_id;
-    GetDbIdentityFromIdentityFile(&temp_db_id);
+    s = GetDbIdentityFromIdentityFile(&temp_db_id);
+    if (!s.ok()) {
+      return s;
+    }
    new_db.SetDBId(temp_db_id);
  }
  new_db.SetLogNumber(0);
@@ -413,7 +407,8 @@ IOStatus Directories::SetDirectories(FileSystem* fs, const std::string& dbname,
 Status DBImpl::Recover(
    const std::vector<ColumnFamilyDescriptor>& column_families, bool read_only,
    bool error_if_wal_file_exists, bool error_if_data_exists_in_wals,
-    uint64_t* recovered_seq, RecoveryContext* recovery_ctx) {
+    bool is_retry, uint64_t* recovered_seq, RecoveryContext* recovery_ctx,
+    bool* can_retry) {
  mutex_.AssertHeld();

  const WriteOptions write_options(Env::IOActivity::kDBOpen);
@@ -529,7 +524,31 @@ Status DBImpl::Recover(
  Status s;
  bool missing_table_file = false;
  if (!immutable_db_options_.best_efforts_recovery) {
-    s = versions_->Recover(column_families, read_only, &db_id_);
+    // Status of reading the descriptor file
+    Status desc_status;
+    s = versions_->Recover(column_families, read_only, &db_id_,
+                           /*no_error_if_files_missing=*/false, is_retry,
+                           &desc_status);
+    desc_status.PermitUncheckedError();
+    if (can_retry) {
+      // If we're opening for the first time and the failure is likely due to
+      // a corrupt MANIFEST file (could result in either the log::Reader
+      // detecting a corrupt record, or SST files not found error due to
+      // discarding badly formed tail records)
+      if (!is_retry &&
+          (desc_status.IsCorruption() || s.IsNotFound() || s.IsCorruption()) &&
+          CheckFSFeatureSupport(fs_.get(),
+                                FSSupportedOps::kVerifyAndReconstructRead)) {
+        *can_retry = true;
+        ROCKS_LOG_ERROR(
+            immutable_db_options_.info_log,
+            "Possible corruption detected while replaying MANIFEST %s, %s. "
+            "Will be retried.",
+            desc_status.ToString().c_str(), s.ToString().c_str());
+      } else {
+        *can_retry = false;
+      }
+    }
  } else {
    assert(!files_in_dbname.empty());
    s = versions_->TryRecover(column_families, read_only, files_in_dbname,
@@ -644,7 +663,7 @@ Status DBImpl::Recover(
  s = SetupDBId(write_options, read_only, recovery_ctx);
  ROCKS_LOG_INFO(immutable_db_options_.info_log, "DB ID: %s\n", db_id_.c_str());
  if (s.ok() && !read_only) {
-    s = DeleteUnreferencedSstFiles(recovery_ctx);
+    s = MaybeUpdateNextFileNumber(recovery_ctx);
  }

  if (immutable_db_options_.paranoid_checks && s.ok()) {
@@ -767,8 +786,8 @@ Status DBImpl::Recover(
      std::sort(wals.begin(), wals.end());

      bool corrupted_wal_found = false;
-      s = RecoverLogFiles(wals, &next_sequence, read_only, &corrupted_wal_found,
-                          recovery_ctx);
+      s = RecoverLogFiles(wals, &next_sequence, read_only, is_retry,
+                          &corrupted_wal_found, recovery_ctx);
      if (corrupted_wal_found && recovered_seq != nullptr) {
        *recovered_seq = next_sequence;
      }
@@ -946,19 +965,6 @@ Status DBImpl::LogAndApplyForRecovery(const RecoveryContext& recovery_ctx) {
                                    recovery_ctx.mutable_cf_opts_, read_options,
                                    write_options, recovery_ctx.edit_lists_,
                                    &mutex_, directories_.GetDbDir());
-  if (s.ok() && !(recovery_ctx.files_to_delete_.empty())) {
-    mutex_.Unlock();
-    for (const auto& stale_sst_file : recovery_ctx.files_to_delete_) {
-      s = DeleteDBFile(&immutable_db_options_, stale_sst_file.first,
-                       stale_sst_file.second,
-                       /*force_bg=*/false,
-                       /*force_fg=*/false);
-      if (!s.ok()) {
-        break;
-      }
-    }
-    mutex_.Lock();
-  }
  return s;
 }

@@ -1078,7 +1084,7 @@ bool DBImpl::InvokeWalFilterIfNeededOnWalRecord(uint64_t wal_number,
 // REQUIRES: wal_numbers are sorted in ascending order
 Status DBImpl::RecoverLogFiles(const std::vector<uint64_t>& wal_numbers,
                               SequenceNumber* next_sequence, bool read_only,
-                               bool* corrupted_wal_found,
+                               bool is_retry, bool* corrupted_wal_found,
                               RecoveryContext* recovery_ctx) {
  struct LogReporter : public log::Reader::Reporter {
    Env* env;
@@ -1189,7 +1195,7 @@ Status DBImpl::RecoverLogFiles(const std::vector<uint64_t>& wal_numbers,
      }
      file_reader.reset(new SequentialFileReader(
          std::move(file), fname, immutable_db_options_.log_readahead_size,
-          io_tracer_));
+          io_tracer_, /*listeners=*/{}, /*rate_limiter=*/nullptr, is_retry));
    }

    // Create the log reader.
@@ -1667,6 +1673,8 @@ Status DBImpl::WriteLevel0TableForRecovery(int job_id, ColumnFamilyData* cfd,
      SequenceNumber earliest_write_conflict_snapshot;
      std::vector<SequenceNumber> snapshot_seqs =
          snapshots_.GetAll(&earliest_write_conflict_snapshot);
+      SequenceNumber earliest_snapshot =
+          (snapshot_seqs.empty() ? kMaxSequenceNumber : snapshot_seqs.at(0));
      auto snapshot_checker = snapshot_checker_.get();
      if (use_custom_gc_ && snapshot_checker == nullptr) {
        snapshot_checker = DisableGCSnapshotChecker::Instance();
@@ -1686,6 +1694,7 @@ Status DBImpl::WriteLevel0TableForRecovery(int job_id, ColumnFamilyData* cfd,
      IOStatus io_s;
      const ReadOptions read_option(Env::IOActivity::kDBOpen);
      const WriteOptions write_option(Env::IO_HIGH, Env::IOActivity::kDBOpen);
+
      TableBuilderOptions tboptions(
          *cfd->ioptions(), mutable_cf_options, read_option, write_option,
          cfd->internal_comparator(), cfd->internal_tbl_prop_coll_factories(),
@@ -1694,21 +1703,22 @@ Status DBImpl::WriteLevel0TableForRecovery(int job_id, ColumnFamilyData* cfd,
          0 /* level */, false /* is_bottommost */,
          TableFileCreationReason::kRecovery, 0 /* oldest_key_time */,
          0 /* file_creation_time */, db_id_, db_session_id_,
-          0 /* target_file_size */, meta.fd.GetNumber());
+          0 /* target_file_size */, meta.fd.GetNumber(), kMaxSequenceNumber);
      Version* version = cfd->current();
      version->Ref();
      uint64_t num_input_entries = 0;
-      s = BuildTable(
-          dbname_, versions_.get(), immutable_db_options_, tboptions,
-          file_options_for_compaction_, cfd->table_cache(), iter.get(),
-          std::move(range_del_iters), &meta, &blob_file_additions,
-          snapshot_seqs, earliest_write_conflict_snapshot, kMaxSequenceNumber,
-          snapshot_checker, paranoid_file_checks, cfd->internal_stats(), &io_s,
-          io_tracer_, BlobFileCreationReason::kRecovery,
-          nullptr /* seqno_to_time_mapping */, &event_logger_, job_id,
-          nullptr /* table_properties */, write_hint,
-          nullptr /*full_history_ts_low*/, &blob_callback_, version,
-          &num_input_entries);
+      s = BuildTable(dbname_, versions_.get(), immutable_db_options_, tboptions,
+                     file_options_for_compaction_, cfd->table_cache(),
+                     iter.get(), std::move(range_del_iters), &meta,
+                     &blob_file_additions, snapshot_seqs, earliest_snapshot,
+                     earliest_write_conflict_snapshot, kMaxSequenceNumber,
+                     snapshot_checker, paranoid_file_checks,
+                     cfd->internal_stats(), &io_s, io_tracer_,
+                     BlobFileCreationReason::kRecovery,
+                     nullptr /* seqno_to_time_mapping */, &event_logger_,
+                     job_id, nullptr /* table_properties */, write_hint,
+                     nullptr /*full_history_ts_low*/, &blob_callback_, version,
+                     &num_input_entries);
      version->Unref();
      LogFlush(immutable_db_options_.info_log);
      ROCKS_LOG_DEBUG(immutable_db_options_.info_log,
@@ -1833,8 +1843,12 @@ Status DB::Open(const DBOptions& db_options, const std::string& dbname,
  const bool kBatchPerTxn = true;
  ThreadStatusUtil::SetEnableTracking(db_options.enable_thread_tracking);
  ThreadStatusUtil::SetThreadOperation(ThreadStatus::OperationType::OP_DBOPEN);
-  Status s = DBImpl::Open(db_options, dbname, column_families, handles, dbptr,
-                          !kSeqPerBatch, kBatchPerTxn);
+  bool can_retry = false;
+  Status s;
+  do {
+    s = DBImpl::Open(db_options, dbname, column_families, handles, dbptr,
+                     !kSeqPerBatch, kBatchPerTxn, can_retry, &can_retry);
+  } while (!s.ok() && can_retry);
  ThreadStatusUtil::ResetThreadStatus();
  return s;
 }
@@ -1953,10 +1967,55 @@ IOStatus DBImpl::CreateWAL(const WriteOptions& write_options,
  return io_s;
 }

+void DBImpl::TrackExistingDataFiles(
+    const std::vector<std::string>& existing_data_files) {
+  auto sfm = static_cast<SstFileManagerImpl*>(
+      immutable_db_options_.sst_file_manager.get());
+  assert(sfm);
+  std::vector<ColumnFamilyMetaData> metadata;
+  GetAllColumnFamilyMetaData(&metadata);
+
+  std::unordered_set<std::string> referenced_files;
+  for (const auto& md : metadata) {
+    for (const auto& lmd : md.levels) {
+      for (const auto& fmd : lmd.files) {
+        // We're assuming that each sst file name exists in at most one of
+        // the paths.
+        std::string file_path =
+            fmd.directory + kFilePathSeparator + fmd.relative_filename;
+        sfm->OnAddFile(file_path, fmd.size).PermitUncheckedError();
+        referenced_files.insert(file_path);
+      }
+    }
+    for (const auto& bmd : md.blob_files) {
+      std::string name = bmd.blob_file_name;
+      // The BlobMetaData.blob_file_name may start with "/".
+      if (!name.empty() && name[0] == kFilePathSeparator) {
+        name = name.substr(1);
+      }
+      // We're assuming that each blob file name exists in at most one of
+      // the paths.
+      std::string file_path = bmd.blob_file_path + kFilePathSeparator + name;
+      sfm->OnAddFile(file_path, bmd.blob_file_size).PermitUncheckedError();
+      referenced_files.insert(file_path);
+    }
+  }
+
+  for (const auto& file_path : existing_data_files) {
+    if (referenced_files.find(file_path) != referenced_files.end()) {
+      continue;
+    }
+    // There shouldn't be any duplicated files. In case there is, SstFileManager
+    // will take care of deduping it.
+    sfm->OnAddFile(file_path).PermitUncheckedError();
+  }
+}
+
 Status DBImpl::Open(const DBOptions& db_options, const std::string& dbname,
                    const std::vector<ColumnFamilyDescriptor>& column_families,
                    std::vector<ColumnFamilyHandle*>* handles, DB** dbptr,
-                    const bool seq_per_batch, const bool batch_per_txn) {
+                    const bool seq_per_batch, const bool batch_per_txn,
+                    const bool is_retry, bool* can_retry) {
  const WriteOptions write_options(Env::IOActivity::kDBOpen);
  const ReadOptions read_options(Env::IOActivity::kDBOpen);

@@ -1999,7 +2058,7 @@ Status DBImpl::Open(const DBOptions& db_options, const std::string& dbname,
        paths.emplace_back(cf_path.path);
      }
    }
-    for (auto& path : paths) {
+    for (const auto& path : paths) {
      s = impl->env_->CreateDirIfMissing(path);
      if (!s.ok()) {
        break;
@@ -2029,8 +2088,8 @@ Status DBImpl::Open(const DBOptions& db_options, const std::string& dbname,
  uint64_t recovered_seq(kMaxSequenceNumber);
  s = impl->Recover(column_families, false /* read_only */,
                    false /* error_if_wal_file_exists */,
-                    false /* error_if_data_exists_in_wals */, &recovered_seq,
-                    &recovery_ctx);
+                    false /* error_if_data_exists_in_wals */, is_retry,
+                    &recovered_seq, &recovery_ctx, can_retry);
  if (s.ok()) {
    uint64_t new_log_number = impl->versions_->NewFileNumber();
    log::Writer* new_log = nullptr;
@@ -2038,6 +2097,11 @@ Status DBImpl::Open(const DBOptions& db_options, const std::string& dbname,
        impl->GetWalPreallocateBlockSize(max_write_buffer_size);
    s = impl->CreateWAL(write_options, new_log_number, 0 /*recycle_log_number*/,
                        preallocate_block_size, &new_log);
+    if (s.ok()) {
+      // Prevent log files created by previous instance from being recycled.
+      // They might be in alive_log_file_, and might get recycled otherwise.
+      impl->min_log_number_to_recycle_ = new_log_number;
+    }
    if (s.ok()) {
      InstrumentedMutexLock wl(&impl->log_write_mutex_);
      impl->logfile_number_ = new_log_number;
@@ -2165,9 +2229,6 @@ Status DBImpl::Open(const DBOptions& db_options, const std::string& dbname,
        impl->WriteOptionsFile(write_options, true /*db_mutex_already_held*/);
    *dbptr = impl;
    impl->opened_successfully_ = true;
-    impl->DeleteObsoleteFiles();
-    TEST_SYNC_POINT("DBImpl::Open:AfterDeleteFiles");
-    impl->MaybeScheduleFlushOrCompaction();
  } else {
    persist_options_status.PermitUncheckedError();
  }
@@ -2182,73 +2243,10 @@ Status DBImpl::Open(const DBOptions& db_options, const std::string& dbname,
    ROCKS_LOG_INFO(impl->immutable_db_options_.info_log,
                   "SstFileManager instance %p", sfm);

-    // Notify SstFileManager about all sst files that already exist in
-    // db_paths[0] and cf_paths[0] when the DB is opened.
-
-    // SstFileManagerImpl needs to know sizes of the files. For files whose size
-    // we already know (sst files that appear in manifest - typically that's the
-    // vast majority of all files), we'll pass the size to SstFileManager.
-    // For all other files SstFileManager will query the size from filesystem.
-
-    std::vector<ColumnFamilyMetaData> metadata;
-    impl->GetAllColumnFamilyMetaData(&metadata);
-
-    std::unordered_map<std::string, uint64_t> known_file_sizes;
-    for (const auto& md : metadata) {
-      for (const auto& lmd : md.levels) {
-        for (const auto& fmd : lmd.files) {
-          known_file_sizes[fmd.relative_filename] = fmd.size;
-        }
-      }
-      for (const auto& bmd : md.blob_files) {
-        std::string name = bmd.blob_file_name;
-        // The BlobMetaData.blob_file_name may start with "/".
-        if (!name.empty() && name[0] == '/') {
-          name = name.substr(1);
-        }
-        known_file_sizes[name] = bmd.blob_file_size;
-      }
-    }
-
-    std::vector<std::string> paths;
-    paths.emplace_back(impl->immutable_db_options_.db_paths[0].path);
-    for (auto& cf : column_families) {
-      if (!cf.options.cf_paths.empty()) {
-        paths.emplace_back(cf.options.cf_paths[0].path);
-      }
-    }
-    // Remove duplicate paths.
-    std::sort(paths.begin(), paths.end());
-    paths.erase(std::unique(paths.begin(), paths.end()), paths.end());
-    IOOptions io_opts;
-    io_opts.do_not_recurse = true;
-    for (auto& path : paths) {
-      std::vector<std::string> existing_files;
-      impl->immutable_db_options_.fs
-          ->GetChildren(path, io_opts, &existing_files,
-                        /*IODebugContext*=*/nullptr)
-          .PermitUncheckedError();  //**TODO: What do to on error?
-      for (auto& file_name : existing_files) {
-        uint64_t file_number;
-        FileType file_type;
-        std::string file_path = path + "/" + file_name;
-        if (ParseFileName(file_name, &file_number, &file_type) &&
-            (file_type == kTableFile || file_type == kBlobFile)) {
-          // TODO: Check for errors from OnAddFile?
-          if (known_file_sizes.count(file_name)) {
-            // We're assuming that each sst file name exists in at most one of
-            // the paths.
-            sfm->OnAddFile(file_path, known_file_sizes.at(file_name))
-                .PermitUncheckedError();
-          } else {
-            sfm->OnAddFile(file_path).PermitUncheckedError();
-          }
-        }
-      }
-    }
+    impl->TrackExistingDataFiles(recovery_ctx.existing_data_files_);

    // Reserve some disk buffer space. This is a heuristic - when we run out
-    // of disk space, this ensures that there is atleast write_buffer_size
+    // of disk space, this ensures that there is at least write_buffer_size
    // amount of free space before we resume DB writes. In low disk space
    // conditions, we want to avoid a lot of small L0 files due to frequent
    // WAL write failures and resultant forced flushes
@@ -2256,6 +2254,27 @@ Status DBImpl::Open(const DBOptions& db_options, const std::string& dbname,
                           impl->immutable_db_options_.db_paths[0].path);
  }

+  if (s.ok()) {
+    // When the DB is stopped, it's possible that there are some .trash files
+    // that were not deleted yet, when we open the DB we will find these .trash
+    // files and schedule them to be deleted (or delete immediately if
+    // SstFileManager was not used).
+    // Note that we only start doing this and below delete obsolete file after
+    // `TrackExistingDataFiles` are called, the `max_trash_db_ratio` is
+    // ineffective otherwise and these files' deletion won't be rate limited
+    // which can cause discard stall.
+    for (const auto& path : impl->CollectAllDBPaths()) {
+      DeleteScheduler::CleanupDirectory(impl->immutable_db_options_.env, sfm,
+                                        path)
+          .PermitUncheckedError();
+    }
+    impl->mutex_.Lock();
+    // This will do a full scan.
+    impl->DeleteObsoleteFiles();
+    TEST_SYNC_POINT("DBImpl::Open:AfterDeleteFiles");
+    impl->MaybeScheduleFlushOrCompaction();
+    impl->mutex_.Unlock();
+  }

  if (s.ok()) {
    ROCKS_LOG_HEADER(impl->immutable_db_options_.info_log, "DB pointer %p",
@@ -33,8 +33,8 @@ DBImplSecondary::~DBImplSecondary() = default;
 Status DBImplSecondary::Recover(
    const std::vector<ColumnFamilyDescriptor>& column_families,
    bool /*readonly*/, bool /*error_if_wal_file_exists*/,
-    bool /*error_if_data_exists_in_wals*/, uint64_t*,
-    RecoveryContext* /*recovery_ctx*/) {
+    bool /*error_if_data_exists_in_wals*/, bool /*is_retry*/, uint64_t*,
+    RecoveryContext* /*recovery_ctx*/, bool* /*can_retry*/) {
  mutex_.AssertHeld();

  JobContext job_context(0);
@@ -680,7 +680,8 @@ Status DBImplSecondary::TryCatchUpWithPrimary() {
    InstrumentedMutexLock lock_guard(&mutex_);
    s = static_cast_with_check<ReactiveVersionSet>(versions_.get())
            ->ReadAndApply(&mutex_, &manifest_reader_,
-                           manifest_reader_status_.get(), &cfds_changed);
+                           manifest_reader_status_.get(), &cfds_changed,
+                           /*files_to_delete=*/nullptr);

    ROCKS_LOG_INFO(immutable_db_options_.info_log, "Last sequence is %" PRIu64,
                   static_cast<uint64_t>(versions_->LastSequence()));
@@ -82,8 +82,9 @@ class DBImplSecondary : public DBImpl {
  // and log_readers_ to facilitate future operations.
  Status Recover(const std::vector<ColumnFamilyDescriptor>& column_families,
                 bool read_only, bool error_if_wal_file_exists,
-                 bool error_if_data_exists_in_wals, uint64_t* = nullptr,
-                 RecoveryContext* recovery_ctx = nullptr) override;
+                 bool error_if_data_exists_in_wals, bool is_retry = false,
+                 uint64_t* = nullptr, RecoveryContext* recovery_ctx = nullptr,
+                 bool* can_retry = nullptr) override;

  // Can return IOError due to files being deleted by the primary. To avoid
  // IOError in this case, application can coordinate between primary and
@@ -276,6 +277,10 @@ class DBImplSecondary : public DBImpl {
    return false;
  }

+  std::unique_ptr<log::FragmentBufferedReader> manifest_reader_;
+  std::unique_ptr<log::Reader::Reporter> manifest_reporter_;
+  std::unique_ptr<Status> manifest_reader_status_;
+
 private:
  friend class DB;

@@ -304,10 +309,6 @@ class DBImplSecondary : public DBImpl {
                                    const CompactionServiceInput& input,
                                    CompactionServiceResult* result);

-  std::unique_ptr<log::FragmentBufferedReader> manifest_reader_;
-  std::unique_ptr<log::Reader::Reporter> manifest_reporter_;
-  std::unique_ptr<Status> manifest_reader_status_;
-
  // Cache log readers for each log number, used for continue WAL replay
  // after recovery
  std::map<uint64_t, std::unique_ptr<LogReaderContainer>> log_readers_;
@@ -155,21 +155,36 @@ Status DBImpl::Write(const WriteOptions& write_options, WriteBatch* my_batch) {
  }
  if (s.ok()) {
    s = WriteImpl(write_options, my_batch, /*callback=*/nullptr,
+                  /*user_write_cb=*/nullptr,
                  /*log_used=*/nullptr);
  }
  return s;
 }

 Status DBImpl::WriteWithCallback(const WriteOptions& write_options,
-                                 WriteBatch* my_batch,
-                                 WriteCallback* callback) {
+                                 WriteBatch* my_batch, WriteCallback* callback,
+                                 UserWriteCallback* user_write_cb) {
  Status s;
  if (write_options.protection_bytes_per_key > 0) {
    s = WriteBatchInternal::UpdateProtectionInfo(
        my_batch, write_options.protection_bytes_per_key);
  }
  if (s.ok()) {
-    s = WriteImpl(write_options, my_batch, callback, nullptr);
+    s = WriteImpl(write_options, my_batch, callback, user_write_cb);
+  }
+  return s;
+}
+
+Status DBImpl::WriteWithCallback(const WriteOptions& write_options,
+                                 WriteBatch* my_batch,
+                                 UserWriteCallback* user_write_cb) {
+  Status s;
+  if (write_options.protection_bytes_per_key > 0) {
+    s = WriteBatchInternal::UpdateProtectionInfo(
+        my_batch, write_options.protection_bytes_per_key);
+  }
+  if (s.ok()) {
+    s = WriteImpl(write_options, my_batch, /*callback=*/nullptr, user_write_cb);
  }
  return s;
 }
@@ -179,9 +194,9 @@ Status DBImpl::WriteWithCallback(const WriteOptions& write_options,
 // published sequence.
 Status DBImpl::WriteImpl(const WriteOptions& write_options,
                         WriteBatch* my_batch, WriteCallback* callback,
-                         uint64_t* log_used, uint64_t log_ref,
-                         bool disable_memtable, uint64_t* seq_used,
-                         size_t batch_cnt,
+                         UserWriteCallback* user_write_cb, uint64_t* log_used,
+                         uint64_t log_ref, bool disable_memtable,
+                         uint64_t* seq_used, size_t batch_cnt,
                         PreReleaseCallback* pre_release_callback,
                         PostMemTableCallback* post_memtable_callback) {
  assert(!seq_per_batch_ || batch_cnt != 0);
@@ -221,7 +236,11 @@ Status DBImpl::WriteImpl(const WriteOptions& write_options,
    return Status::InvalidArgument(
        "`WriteOptions::protection_bytes_per_key` must be zero or eight");
  } else if (write_options.disableWAL &&
-             immutable_db_options_.recycle_log_file_num > 0) {
+             immutable_db_options_.recycle_log_file_num > 0 &&
+             !(two_write_queues_ && disable_memtable)) {
+    // Corruption detection in recycled WALs relies on sequential sequence
+    // numbers, but WritePreparedTxnDB uses disableWAL internally for split
+    // writes
    return Status::InvalidArgument(
        "WriteOptions::disableWAL option is not supported if "
        "DBOptions::recycle_log_file_num > 0");
@@ -264,6 +283,10 @@ Status DBImpl::WriteImpl(const WriteOptions& write_options,
    return Status::NotSupported(
        "seq_per_batch currently does not honor post_memtable_callback");
  }
+  if (my_batch->HasDeleteRange() && immutable_db_options_.row_cache) {
+    return Status::NotSupported(
+        "DeleteRange is not compatible with row cache.");
+  }
  // Otherwise IsLatestPersistentState optimization does not make sense
  assert(!WriteBatchInternal::IsLatestPersistentState(my_batch) ||
         disable_memtable);
@@ -280,10 +303,10 @@ Status DBImpl::WriteImpl(const WriteOptions& write_options,
        seq_per_batch_ ? kDoAssignOrder : kDontAssignOrder;
    // Otherwise it is WAL-only Prepare batches in WriteCommitted policy and
    // they don't consume sequence.
-    return WriteImplWALOnly(&nonmem_write_thread_, write_options, my_batch,
-                            callback, log_used, log_ref, seq_used, batch_cnt,
-                            pre_release_callback, assign_order,
-                            kDontPublishLastSeq, disable_memtable);
+    return WriteImplWALOnly(
+        &nonmem_write_thread_, write_options, my_batch, callback, user_write_cb,
+        log_used, log_ref, seq_used, batch_cnt, pre_release_callback,
+        assign_order, kDontPublishLastSeq, disable_memtable);
  }

  if (immutable_db_options_.unordered_write) {
@@ -295,9 +318,9 @@ Status DBImpl::WriteImpl(const WriteOptions& write_options,
    // Use a write thread to i) optimize for WAL write, ii) publish last
    // sequence in in increasing order, iii) call pre_release_callback serially
    Status status = WriteImplWALOnly(
-        &write_thread_, write_options, my_batch, callback, log_used, log_ref,
-        &seq, sub_batch_cnt, pre_release_callback, kDoAssignOrder,
-        kDoPublishLastSeq, disable_memtable);
+        &write_thread_, write_options, my_batch, callback, user_write_cb,
+        log_used, log_ref, &seq, sub_batch_cnt, pre_release_callback,
+        kDoAssignOrder, kDoPublishLastSeq, disable_memtable);
    TEST_SYNC_POINT("DBImpl::WriteImpl:UnorderedWriteAfterWriteWAL");
    if (!status.ok()) {
      return status;
@@ -314,17 +337,20 @@ Status DBImpl::WriteImpl(const WriteOptions& write_options,
  }

  if (immutable_db_options_.enable_pipelined_write) {
-    return PipelinedWriteImpl(write_options, my_batch, callback, log_used,
-                              log_ref, disable_memtable, seq_used);
+    return PipelinedWriteImpl(write_options, my_batch, callback, user_write_cb,
+                              log_used, log_ref, disable_memtable, seq_used);
  }

  PERF_TIMER_GUARD(write_pre_and_post_process_time);
-  WriteThread::Writer w(write_options, my_batch, callback, log_ref,
-                        disable_memtable, batch_cnt, pre_release_callback,
-                        post_memtable_callback);
+  WriteThread::Writer w(write_options, my_batch, callback, user_write_cb,
+                        log_ref, disable_memtable, batch_cnt,
+                        pre_release_callback, post_memtable_callback);
  StopWatch write_sw(immutable_db_options_.clock, stats_, DB_WRITE);

  write_thread_.JoinBatchGroup(&w);
+  if (w.state == WriteThread::STATE_PARALLEL_MEMTABLE_CALLER) {
+    write_thread_.SetMemWritersEachStride(&w);
+  }
  if (w.state == WriteThread::STATE_PARALLEL_MEMTABLE_WRITER) {
    // we are a non-leader in a parallel group

@@ -675,6 +701,7 @@ Status DBImpl::WriteImpl(const WriteOptions& write_options,

 Status DBImpl::PipelinedWriteImpl(const WriteOptions& write_options,
                                  WriteBatch* my_batch, WriteCallback* callback,
+                                  UserWriteCallback* user_write_cb,
                                  uint64_t* log_used, uint64_t log_ref,
                                  bool disable_memtable, uint64_t* seq_used) {
  PERF_TIMER_GUARD(write_pre_and_post_process_time);
@@ -682,8 +709,8 @@ Status DBImpl::PipelinedWriteImpl(const WriteOptions& write_options,

  WriteContext write_context;

-  WriteThread::Writer w(write_options, my_batch, callback, log_ref,
-                        disable_memtable, /*_batch_cnt=*/0,
+  WriteThread::Writer w(write_options, my_batch, callback, user_write_cb,
+                        log_ref, disable_memtable, /*_batch_cnt=*/0,
                        /*_pre_release_callback=*/nullptr);
  write_thread_.JoinBatchGroup(&w);
  TEST_SYNC_POINT("DBImplWrite::PipelinedWriteImpl:AfterJoinBatchGroup");
@@ -822,7 +849,9 @@ Status DBImpl::PipelinedWriteImpl(const WriteOptions& write_options,
    // so we need to set its status to pass ASSERT_STATUS_CHECKED
    memtable_write_group.status.PermitUncheckedError();
  }
-
+  if (w.state == WriteThread::STATE_PARALLEL_MEMTABLE_CALLER) {
+    write_thread_.SetMemWritersEachStride(&w);
+  }
  if (w.state == WriteThread::STATE_PARALLEL_MEMTABLE_WRITER) {
    PERF_TIMER_STOP(write_pre_and_post_process_time);
    PERF_TIMER_FOR_WAIT_GUARD(write_memtable_time);
@@ -862,7 +891,8 @@ Status DBImpl::UnorderedWriteMemtable(const WriteOptions& write_options,
  PERF_TIMER_GUARD(write_pre_and_post_process_time);
  StopWatch write_sw(immutable_db_options_.clock, stats_, DB_WRITE);

-  WriteThread::Writer w(write_options, my_batch, callback, log_ref,
+  WriteThread::Writer w(write_options, my_batch, callback,
+                        /*user_write_cb=*/nullptr, log_ref,
                        false /*disable_memtable*/);

  if (w.CheckCallback(this) && w.ShouldWriteToMemtable()) {
@@ -912,13 +942,15 @@ Status DBImpl::UnorderedWriteMemtable(const WriteOptions& write_options,
 // applicable in a two-queue setting.
 Status DBImpl::WriteImplWALOnly(
    WriteThread* write_thread, const WriteOptions& write_options,
-    WriteBatch* my_batch, WriteCallback* callback, uint64_t* log_used,
+    WriteBatch* my_batch, WriteCallback* callback,
+    UserWriteCallback* user_write_cb, uint64_t* log_used,
    const uint64_t log_ref, uint64_t* seq_used, const size_t sub_batch_cnt,
    PreReleaseCallback* pre_release_callback, const AssignOrder assign_order,
    const PublishLastSeq publish_last_seq, const bool disable_memtable) {
  PERF_TIMER_GUARD(write_pre_and_post_process_time);
-  WriteThread::Writer w(write_options, my_batch, callback, log_ref,
-                        disable_memtable, sub_batch_cnt, pre_release_callback);
+  WriteThread::Writer w(write_options, my_batch, callback, user_write_cb,
+                        log_ref, disable_memtable, sub_batch_cnt,
+                        pre_release_callback);
  StopWatch write_sw(immutable_db_options_.clock, stats_, DB_WRITE);

  write_thread->JoinBatchGroup(&w);
@@ -1485,6 +1517,11 @@ IOStatus DBImpl::WriteToWAL(const WriteThread::WriteGroup& write_group,
    RecordTick(stats_, WAL_FILE_BYTES, log_size);
    stats->AddDBStats(InternalStats::kIntStatsWriteWithWal, write_with_wal);
    RecordTick(stats_, WRITE_WITH_WAL, write_with_wal);
+    for (auto* writer : write_group) {
+      if (!writer->CallbackFailed()) {
+        writer->CheckPostWalWriteCallback();
+      }
+    }
  }
  return io_s;
 }
@@ -1549,6 +1586,11 @@ IOStatus DBImpl::ConcurrentWriteToWAL(
    stats->AddDBStats(InternalStats::kIntStatsWriteWithWal, write_with_wal,
                      concurrent);
    RecordTick(stats_, WRITE_WITH_WAL, write_with_wal);
+    for (auto* writer : write_group) {
+      if (!writer->CallbackFailed()) {
+        writer->CheckPostWalWriteCallback();
+      }
+    }
  }
  return io_s;
 }
@@ -2152,6 +2194,8 @@ void DBImpl::NotifyOnMemTableSealed(ColumnFamilyData* /*cfd*/,
 // two_write_queues_ is true (This is to simplify the reasoning.)
 Status DBImpl::SwitchMemtable(ColumnFamilyData* cfd, WriteContext* context) {
  mutex_.AssertHeld();
+  assert(lock_wal_count_ == 0);
+
  // TODO: plumb Env::IOActivity, Env::IOPriority
  const ReadOptions read_options;
  const WriteOptions write_options;
@@ -2195,6 +2239,11 @@ Status DBImpl::SwitchMemtable(ColumnFamilyData* cfd, WriteContext* context) {
  memtable_info.earliest_seqno = cfd->mem()->GetEarliestSequenceNumber();
  memtable_info.num_entries = cfd->mem()->num_entries();
  memtable_info.num_deletes = cfd->mem()->num_deletes();
+  if (!cfd->ioptions()->persist_user_defined_timestamps &&
+      cfd->user_comparator()->timestamp_size() > 0) {
+    const Slice& newest_udt = cfd->mem()->GetNewestUDT();
+    memtable_info.newest_udt.assign(newest_udt.data(), newest_udt.size());
+  }
  // 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();
@@ -2214,14 +2263,15 @@ Status DBImpl::SwitchMemtable(ColumnFamilyData* cfd, WriteContext* context) {
    SequenceNumber seq = versions_->LastSequence();
    new_mem = cfd->ConstructNewMemtable(mutable_cf_options, seq);
    context->superversion_context.NewSuperVersion();
+
+    ROCKS_LOG_INFO(immutable_db_options_.info_log,
+                   "[%s] New memtable created with log file: #%" PRIu64
+                   ". Immutable memtables: %d.\n",
+                   cfd->GetName().c_str(), new_log_number, num_imm_unflushed);
+    // There should be no concurrent write as the thread is at the front of
+    // writer queue
+    cfd->mem()->ConstructFragmentedRangeTombstones();
  }
-  ROCKS_LOG_INFO(immutable_db_options_.info_log,
-                 "[%s] New memtable created with log file: #%" PRIu64
-                 ". Immutable memtables: %d.\n",
-                 cfd->GetName().c_str(), new_log_number, num_imm_unflushed);
-  // There should be no concurrent write as the thread is at the front of
-  // writer queue
-  cfd->mem()->ConstructFragmentedRangeTombstones();

  mutex_.Lock();
  if (recycle_log_number != 0) {
@@ -21,14 +21,15 @@ class CorruptionFS : public FileSystemWrapper {
  int num_writable_file_errors_;

  explicit CorruptionFS(const std::shared_ptr<FileSystem>& _target,
-                        bool fs_buffer)
+                        bool fs_buffer, bool verify_read)
      : FileSystemWrapper(_target),
        writable_file_error_(false),
        num_writable_file_errors_(0),
        corruption_trigger_(INT_MAX),
        read_count_(0),
        rnd_(300),
-        fs_buffer_(fs_buffer) {}
+        fs_buffer_(fs_buffer),
+        verify_read_(verify_read) {}
  ~CorruptionFS() override {
    // Assert that the corruption was reset, which means it got triggered
    assert(corruption_trigger_ == INT_MAX);
@@ -113,11 +114,13 @@ class CorruptionFS : public FileSystemWrapper {
  }

  void SupportedOps(int64_t& supported_ops) override {
-    supported_ops = 1 << FSSupportedOps::kVerifyAndReconstructRead |
-                    1 << FSSupportedOps::kAsyncIO;
+    supported_ops = 1 << FSSupportedOps::kAsyncIO;
    if (fs_buffer_) {
      supported_ops |= 1 << FSSupportedOps::kFSBuffer;
    }
+    if (verify_read_) {
+      supported_ops |= 1 << FSSupportedOps::kVerifyAndReconstructRead;
+    }
  }

 private:
@@ -125,6 +128,7 @@ class CorruptionFS : public FileSystemWrapper {
  int read_count_;
  Random rnd_;
  bool fs_buffer_;
+  bool verify_read_;
 };
 }  // anonymous namespace

@@ -696,23 +700,24 @@ TEST_F(DBIOFailureTest, CompactionSstSyncError) {

 class DBIOCorruptionTest
    : public DBIOFailureTest,
-      public testing::WithParamInterface<std::tuple<bool, bool>> {
+      public testing::WithParamInterface<std::tuple<bool, bool, bool>> {
 public:
  DBIOCorruptionTest() : DBIOFailureTest() {
    BlockBasedTableOptions bbto;
-    Options options = CurrentOptions();
+    options_ = CurrentOptions();

    base_env_ = env_;
    EXPECT_NE(base_env_, nullptr);
-    fs_.reset(
-        new CorruptionFS(base_env_->GetFileSystem(), std::get<0>(GetParam())));
+    fs_.reset(new CorruptionFS(base_env_->GetFileSystem(),
+                               std::get<0>(GetParam()),
+                               std::get<2>(GetParam())));
    env_guard_ = NewCompositeEnv(fs_);
-    options.env = env_guard_.get();
+    options_.env = env_guard_.get();
    bbto.num_file_reads_for_auto_readahead = 0;
-    options.table_factory.reset(NewBlockBasedTableFactory(bbto));
-    options.disable_auto_compactions = true;
+    options_.table_factory.reset(NewBlockBasedTableFactory(bbto));
+    options_.disable_auto_compactions = true;

-    Reopen(options);
+    Reopen(options_);
  }

  ~DBIOCorruptionTest() {
@@ -720,10 +725,13 @@ class DBIOCorruptionTest
    db_ = nullptr;
  }

+  Status ReopenDB() { return TryReopen(options_); }
+
 protected:
  std::unique_ptr<Env> env_guard_;
  std::shared_ptr<CorruptionFS> fs_;
  Env* base_env_;
+  Options options_;
 };

 TEST_P(DBIOCorruptionTest, GetReadCorruptionRetry) {
@@ -737,8 +745,13 @@ TEST_P(DBIOCorruptionTest, GetReadCorruptionRetry) {
  std::string val;
  ReadOptions ro;
  ro.async_io = std::get<1>(GetParam());
-  ASSERT_OK(dbfull()->Get(ReadOptions(), "key1", &val));
-  ASSERT_EQ(val, "val1");
+  Status s = dbfull()->Get(ReadOptions(), "key1", &val);
+  if (std::get<2>(GetParam())) {
+    ASSERT_OK(s);
+    ASSERT_EQ(val, "val1");
+  } else {
+    ASSERT_TRUE(s.IsCorruption());
+  }
 }

 TEST_P(DBIOCorruptionTest, IterReadCorruptionRetry) {
@@ -758,7 +771,11 @@ TEST_P(DBIOCorruptionTest, IterReadCorruptionRetry) {
  while (iter->status().ok() && iter->Valid()) {
    iter->Next();
  }
-  ASSERT_OK(iter->status());
+  if (std::get<2>(GetParam())) {
+    ASSERT_OK(iter->status());
+  } else {
+    ASSERT_TRUE(iter->status().IsCorruption());
+  }
  delete iter;
 }

@@ -779,8 +796,13 @@ TEST_P(DBIOCorruptionTest, MultiGetReadCorruptionRetry) {
  ro.async_io = std::get<1>(GetParam());
  dbfull()->MultiGet(ro, dbfull()->DefaultColumnFamily(), keys.size(),
                     keys.data(), values.data(), statuses.data());
-  ASSERT_EQ(values[0].ToString(), "val1");
-  ASSERT_EQ(values[1].ToString(), "val2");
+  if (std::get<2>(GetParam())) {
+    ASSERT_EQ(values[0].ToString(), "val1");
+    ASSERT_EQ(values[1].ToString(), "val2");
+  } else {
+    ASSERT_TRUE(statuses[0].IsCorruption());
+    ASSERT_TRUE(statuses[1].IsCorruption());
+  }
 }

 TEST_P(DBIOCorruptionTest, CompactionReadCorruptionRetry) {
@@ -793,13 +815,18 @@ TEST_P(DBIOCorruptionTest, CompactionReadCorruptionRetry) {
  ASSERT_OK(Put("key2", "val2"));
  ASSERT_OK(Flush());
  fs->SetCorruptionTrigger(1);
-  ASSERT_OK(dbfull()->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  Status s = dbfull()->CompactRange(CompactRangeOptions(), nullptr, nullptr);
+  if (std::get<2>(GetParam())) {
+    ASSERT_OK(s);

-  std::string val;
-  ReadOptions ro;
-  ro.async_io = std::get<1>(GetParam());
-  ASSERT_OK(dbfull()->Get(ro, "key1", &val));
-  ASSERT_EQ(val, "val1");
+    std::string val;
+    ReadOptions ro;
+    ro.async_io = std::get<1>(GetParam());
+    ASSERT_OK(dbfull()->Get(ro, "key1", &val));
+    ASSERT_EQ(val, "val1");
+  } else {
+    ASSERT_TRUE(s.IsCorruption());
+  }
 }

 TEST_P(DBIOCorruptionTest, FlushReadCorruptionRetry) {
@@ -808,17 +835,44 @@ TEST_P(DBIOCorruptionTest, FlushReadCorruptionRetry) {

  ASSERT_OK(Put("key1", "val1"));
  fs->SetCorruptionTrigger(1);
-  ASSERT_OK(Flush());
+  Status s = Flush();
+  if (std::get<2>(GetParam())) {
+    ASSERT_OK(s);

-  std::string val;
-  ReadOptions ro;
-  ro.async_io = std::get<1>(GetParam());
-  ASSERT_OK(dbfull()->Get(ro, "key1", &val));
-  ASSERT_EQ(val, "val1");
+    std::string val;
+    ReadOptions ro;
+    ro.async_io = std::get<1>(GetParam());
+    ASSERT_OK(dbfull()->Get(ro, "key1", &val));
+    ASSERT_EQ(val, "val1");
+  } else {
+    ASSERT_NOK(s);
+  }
 }

+TEST_P(DBIOCorruptionTest, ManifestCorruptionRetry) {
+  CorruptionFS* fs =
+      static_cast<CorruptionFS*>(env_guard_->GetFileSystem().get());
+
+  ASSERT_OK(Put("key1", "val1"));
+  ASSERT_OK(Flush());
+  SyncPoint::GetInstance()->SetCallBack(
+      "VersionSet::Recover:StartManifestRead",
+      [&](void* /*arg*/) { fs->SetCorruptionTrigger(0); });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  if (std::get<2>(GetParam())) {
+    ASSERT_OK(ReopenDB());
+  } else {
+    ASSERT_EQ(ReopenDB(), Status::Corruption());
+  }
+  SyncPoint::GetInstance()->DisableProcessing();
+}
+
+// The parameters are - 1. Use FS provided buffer, 2. Use async IO ReadOption,
+// 3. Retry with verify_and_reconstruct_read IOOption
 INSTANTIATE_TEST_CASE_P(DBIOCorruptionTest, DBIOCorruptionTest,
-                        testing::Combine(testing::Bool(), testing::Bool()));
+                        testing::Combine(testing::Bool(), testing::Bool(),
+                                         testing::Bool()));
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
@@ -109,6 +109,16 @@ Status DBIter::GetProperty(std::string prop_name, std::string* prop) {
      *prop = "Iterator is not valid.";
    }
    return Status::OK();
+  } else if (prop_name == "rocksdb.iterator.is-value-pinned") {
+    if (valid_) {
+      *prop = (pin_thru_lifetime_ && iter_.Valid() &&
+               iter_.value().data() == value_.data())
+                  ? "1"
+                  : "0";
+    } else {
+      *prop = "Iterator is not valid.";
+    }
+    return Status::OK();
  } else if (prop_name == "rocksdb.iterator.internal-key") {
    *prop = saved_key_.GetUserKey().ToString();
    return Status::OK();
@@ -231,6 +241,7 @@ bool DBIter::SetValueAndColumnsFromEntity(Slice slice) {
  if (!s.ok()) {
    status_ = s;
    valid_ = false;
+    wide_columns_.clear();
    return false;
  }

@@ -133,11 +133,17 @@ TEST_P(DBIteratorTest, IteratorProperty) {
    ASSERT_NOK(iter->GetProperty("non_existing.value", &prop_value));
    ASSERT_OK(iter->GetProperty("rocksdb.iterator.is-key-pinned", &prop_value));
    ASSERT_EQ("0", prop_value);
+    ASSERT_OK(
+        iter->GetProperty("rocksdb.iterator.is-value-pinned", &prop_value));
+    ASSERT_EQ("0", prop_value);
    ASSERT_OK(iter->GetProperty("rocksdb.iterator.internal-key", &prop_value));
    ASSERT_EQ("1", prop_value);
    iter->Next();
    ASSERT_OK(iter->GetProperty("rocksdb.iterator.is-key-pinned", &prop_value));
    ASSERT_EQ("Iterator is not valid.", prop_value);
+    ASSERT_OK(
+        iter->GetProperty("rocksdb.iterator.is-value-pinned", &prop_value));
+    ASSERT_EQ("Iterator is not valid.", prop_value);

    // Get internal key at which the iteration stopped (tombstone in this case).
    ASSERT_OK(iter->GetProperty("rocksdb.iterator.internal-key", &prop_value));
@@ -1680,12 +1686,15 @@ TEST_P(DBIteratorTest, PinnedDataIteratorMultipleFiles) {
  ro.pin_data = true;
  auto iter = NewIterator(ro);

-  std::vector<std::pair<Slice, std::string>> results;
+  std::vector<std::pair<Slice, Slice>> results;
  for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
    std::string prop_value;
    ASSERT_OK(iter->GetProperty("rocksdb.iterator.is-key-pinned", &prop_value));
    ASSERT_EQ("1", prop_value);
-    results.emplace_back(iter->key(), iter->value().ToString());
+    ASSERT_OK(
+        iter->GetProperty("rocksdb.iterator.is-value-pinned", &prop_value));
+    ASSERT_EQ("1", prop_value);
+    results.emplace_back(iter->key(), iter->value());
  }

  ASSERT_EQ(results.size(), true_data.size());
@@ -1739,6 +1748,9 @@ TEST_P(DBIteratorTest, PinnedDataIteratorMergeOperator) {
    std::string prop_value;
    ASSERT_OK(iter->GetProperty("rocksdb.iterator.is-key-pinned", &prop_value));
    ASSERT_EQ("1", prop_value);
+    ASSERT_OK(
+        iter->GetProperty("rocksdb.iterator.is-value-pinned", &prop_value));
+    ASSERT_EQ("0", prop_value);
    results.emplace_back(iter->key(), iter->value().ToString());
  }
  ASSERT_OK(iter->status());
@@ -1792,12 +1804,15 @@ TEST_P(DBIteratorTest, PinnedDataIteratorReadAfterUpdate) {
    }
  }

-  std::vector<std::pair<Slice, std::string>> results;
+  std::vector<std::pair<Slice, Slice>> results;
  for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
    std::string prop_value;
    ASSERT_OK(iter->GetProperty("rocksdb.iterator.is-key-pinned", &prop_value));
    ASSERT_EQ("1", prop_value);
-    results.emplace_back(iter->key(), iter->value().ToString());
+    ASSERT_OK(
+        iter->GetProperty("rocksdb.iterator.is-value-pinned", &prop_value));
+    ASSERT_EQ("1", prop_value);
+    results.emplace_back(iter->key(), iter->value());
  }
  ASSERT_OK(iter->status());

@@ -3555,6 +3570,121 @@ TEST_F(DBIteratorTest, ErrorWhenReadFile) {
  iter->Reset();
 }

+TEST_F(DBIteratorTest, IteratorsConsistentViewImplicitSnapshot) {
+  Options options = GetDefaultOptions();
+  CreateAndReopenWithCF({"cf_1", "cf_2"}, options);
+
+  for (int i = 0; i < 3; ++i) {
+    ASSERT_OK(Put(i, "cf" + std::to_string(i) + "_key",
+                  "cf" + std::to_string(i) + "_val"));
+  }
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+      {{"DBImpl::BGWorkFlush:done",
+        "DBImpl::MultiCFSnapshot::BeforeCheckingSnapshot"}});
+
+  bool flushed = false;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::MultiCFSnapshot::AfterRefSV", [&](void* /*arg*/) {
+        if (!flushed) {
+          for (int i = 0; i < 3; ++i) {
+            ASSERT_OK(Put(i, "cf" + std::to_string(i) + "_key",
+                          "cf" + std::to_string(i) + "_val_new"));
+          }
+          // After SV is obtained for the first CF, flush for the second CF
+          ASSERT_OK(Flush(1));
+          flushed = true;
+        }
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+  ReadOptions read_options;
+  std::vector<Iterator*> iters;
+  ASSERT_OK(db_->NewIterators(read_options, handles_, &iters));
+
+  for (int i = 0; i < 3; ++i) {
+    auto iter = iters[i];
+    ASSERT_OK(iter->status());
+    iter->SeekToFirst();
+    ASSERT_EQ(IterStatus(iter), "cf" + std::to_string(i) + "_key->cf" +
+                                    std::to_string(i) + "_val_new");
+  }
+  for (auto* iter : iters) {
+    delete iter;
+  }
+
+  // Thread-local SVs are no longer obsolete nor in use
+  for (int i = 0; i < 3; ++i) {
+    auto* cfd =
+        static_cast_with_check<ColumnFamilyHandleImpl>(handles_[i])->cfd();
+    ASSERT_NE(cfd->TEST_GetLocalSV()->Get(), SuperVersion::kSVObsolete);
+    ASSERT_NE(cfd->TEST_GetLocalSV()->Get(), SuperVersion::kSVInUse);
+  }
+}
+
+TEST_F(DBIteratorTest, IteratorsConsistentViewExplicitSnapshot) {
+  Options options = GetDefaultOptions();
+  options.atomic_flush = true;
+  CreateAndReopenWithCF({"cf_1", "cf_2"}, options);
+
+  for (int i = 0; i < 3; ++i) {
+    ASSERT_OK(Put(i, "cf" + std::to_string(i) + "_key",
+                  "cf" + std::to_string(i) + "_val"));
+  }
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+      {{"DBImpl::BGWorkFlush:done",
+        "DBImpl::MultiCFSnapshot::BeforeCheckingSnapshot"}});
+
+  bool flushed = false;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::MultiCFSnapshot::AfterRefSV", [&](void* /*arg*/) {
+        if (!flushed) {
+          for (int i = 0; i < 3; ++i) {
+            ASSERT_OK(Put(i, "cf" + std::to_string(i) + "_key",
+                          "cf" + std::to_string(i) + "_val_new"));
+          }
+          // After SV is obtained for the first CF, do the atomic flush()
+          ASSERT_OK(Flush());
+          flushed = true;
+        }
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+  // Explicit snapshot wouldn't force reloading all svs. We should expect old
+  // values
+  const Snapshot* snapshot = db_->GetSnapshot();
+  ReadOptions read_options;
+  read_options.snapshot = snapshot;
+  std::vector<Iterator*> iters;
+  ASSERT_OK(db_->NewIterators(read_options, handles_, &iters));
+
+  for (int i = 0; i < 3; ++i) {
+    auto iter = iters[i];
+    ASSERT_OK(iter->status());
+    iter->SeekToFirst();
+    ASSERT_EQ(IterStatus(iter), "cf" + std::to_string(i) + "_key->cf" +
+                                    std::to_string(i) + "_val");
+  }
+  db_->ReleaseSnapshot(snapshot);
+  for (auto* iter : iters) {
+    delete iter;
+  }
+
+  // Thread-local SV for cf_0 is obsolete (atomic flush happened after the first
+  // SV Ref)
+  auto* cfd0 =
+      static_cast_with_check<ColumnFamilyHandleImpl>(handles_[0])->cfd();
+  ASSERT_EQ(cfd0->TEST_GetLocalSV()->Get(), SuperVersion::kSVObsolete);
+  ASSERT_NE(cfd0->TEST_GetLocalSV()->Get(), SuperVersion::kSVInUse);
+
+  // Rest are not InUse nor Obsolete
+  for (int i = 1; i < 3; ++i) {
+    auto* cfd =
+        static_cast_with_check<ColumnFamilyHandleImpl>(handles_[i])->cfd();
+    ASSERT_NE(cfd->TEST_GetLocalSV()->Get(), SuperVersion::kSVObsolete);
+    ASSERT_NE(cfd->TEST_GetLocalSV()->Get(), SuperVersion::kSVInUse);
+  }
+}
+
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
@@ -732,7 +732,7 @@ TEST_P(DbMemtableKVChecksumTest, GetWithCorruptAfterMemtableInsert) {
      });

  SyncPoint::GetInstance()->SetCallBack(
-      "Memtable::SaveValue:Begin:entry", [&](void* entry) {
+      "Memtable::SaveValue:Found:entry", [&](void* entry) {
        char* buf = *static_cast<char**>(entry);
        buf[corrupt_byte_offset_] += corrupt_byte_addend_;
        ++corrupt_byte_offset_;
@@ -769,7 +769,7 @@ TEST_P(DbMemtableKVChecksumTest,
      });

  SyncPoint::GetInstance()->SetCallBack(
-      "Memtable::SaveValue:Begin:entry", [&](void* entry) {
+      "Memtable::SaveValue:Found:entry", [&](void* entry) {
        char* buf = *static_cast<char**>(entry);
        buf[corrupt_byte_offset_] += corrupt_byte_addend_;
        ++corrupt_byte_offset_;
@@ -313,6 +313,10 @@ TEST_F(DBMemTableTest, InsertWithHint) {
  ASSERT_EQ("foo_v3", Get("foo_k3"));
  ASSERT_EQ("bar_v1", Get("bar_k1"));
  ASSERT_EQ("bar_v2", Get("bar_k2"));
+  ASSERT_OK(db_->DeleteRange(WriteOptions(), "foo_k1", "foo_k4"));
+  ASSERT_EQ(hint_bar, rep->last_hint_in());
+  ASSERT_EQ(hint_bar, rep->last_hint_out());
+  ASSERT_EQ(5, rep->num_insert_with_hint());
  ASSERT_EQ("vvv", Get("NotInPrefixDomain"));
 }

@@ -192,6 +192,49 @@ TEST_F(DBPropertiesTest, GetAggregatedIntPropertyTest) {
  }
 }

+TEST_F(DBPropertiesTest, AggregateBlockCacheProperty) {
+  constexpr size_t kCapacity = 1000;
+  LRUCacheOptions co;
+  co.capacity = kCapacity;
+  co.num_shard_bits = 0;
+  co.metadata_charge_policy = kDontChargeCacheMetadata;
+  auto block_cache = NewLRUCache(co);
+
+  // All columns families share the same block cache.
+  Options options = CurrentOptions();
+  BlockBasedTableOptions table_opt;
+  table_opt.no_block_cache = false;
+  table_opt.block_cache = block_cache;
+  options.table_factory.reset(NewBlockBasedTableFactory(table_opt));
+
+  CreateAndReopenWithCF({"one", "two", "three", "four"}, options);
+
+  // Insert unpinned block to the cache
+  constexpr size_t kSize1 = 100;
+  ASSERT_OK(block_cache->Insert("block1", nullptr /*value*/,
+                                &kNoopCacheItemHelper, kSize1));
+  // Insert pinned block to the cache
+  constexpr size_t kSize2 = 200;
+  Cache::Handle* block2 = nullptr;
+  ASSERT_OK(block_cache->Insert("block2", nullptr /*value*/,
+                                &kNoopCacheItemHelper, kSize2, &block2));
+
+  uint64_t value;
+  ASSERT_TRUE(db_->GetAggregatedIntProperty(DB::Properties::kBlockCacheCapacity,
+                                            &value));
+  ASSERT_EQ(value, kCapacity);
+
+  ASSERT_TRUE(
+      db_->GetAggregatedIntProperty(DB::Properties::kBlockCacheUsage, &value));
+  ASSERT_EQ(value, kSize1 + kSize2);
+
+  ASSERT_TRUE(db_->GetAggregatedIntProperty(
+      DB::Properties::kBlockCachePinnedUsage, &value));
+  ASSERT_EQ(value, kSize2);
+
+  block_cache->Release(block2);
+}
+
 namespace {
 void VerifySimilar(uint64_t a, uint64_t b, double bias) {
  ASSERT_EQ(a == 0U, b == 0U);
@@ -1083,7 +1126,6 @@ TEST_F(DBPropertiesTest, EstimateCompressionRatio) {
  ASSERT_GT(CompressionRatioAtLevel(1), 10.0);
 }

-
 class CountingUserTblPropCollector : public TablePropertiesCollector {
 public:
  const char* Name() const override { return "CountingUserTblPropCollector"; }
@@ -2366,7 +2408,6 @@ TEST_F(DBPropertiesTest, TableMetaIndexKeys) {
  } while (ChangeOptions());
 }

-
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
@@ -3806,6 +3806,25 @@ TEST_F(DBRangeDelTest, RefreshWithSnapshot) {
  iter.reset();
  db_->ReleaseSnapshot(snapshot);
 }
+
+TEST_F(DBRangeDelTest, RowCache) {
+  Options options = CurrentOptions();
+  options.row_cache = NewLRUCache(8 << 10);
+  DestroyAndReopen(options);
+  ASSERT_OK(Put(Key(3), "val"));
+  ASSERT_TRUE(db_->DeleteRange(WriteOptions(), db_->DefaultColumnFamily(),
+                               Key(3), Key(5))
+                  .IsNotSupported());
+  WriteBatch wb;
+  ASSERT_OK(wb.Put(Key(6), "abc"));
+  ASSERT_OK(wb.DeleteRange(Key(1), Key(5)));
+  ASSERT_TRUE(db_->Write(WriteOptions(), &wb).IsNotSupported());
+  ASSERT_EQ(Get(Key(3)), "val");
+  // By default, memtable insertion failure will turn the DB to read-only mode.
+  // The check for delete range should happen before that to fail early
+  // and should not turn db into read-only mdoe.
+  ASSERT_OK(Put(Key(5), "foo"));
+}
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
@@ -21,7 +21,8 @@ namespace ROCKSDB_NAMESPACE {

 class DBSSTTest : public DBTestBase {
 public:
-  DBSSTTest() : DBTestBase("db_sst_test", /*env_do_fsync=*/true) {}
+  DBSSTTest(const std::string& test_name = "db_sst_test")
+      : DBTestBase(test_name, /*env_do_fsync=*/true) {}
 };

 // A class which remembers the name of each flushed file.
@@ -506,6 +507,23 @@ TEST_F(DBSSTTest, DBWithSstFileManagerForBlobFiles) {
  ASSERT_EQ(files_deleted, 0);
  ASSERT_EQ(files_scheduled_to_delete, 0);
  Close();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "SstFileManagerImpl::ScheduleUnaccountedFileDeletion", [&](void* arg) {
+        assert(arg);
+        const std::string* const file_path =
+            static_cast<const std::string*>(arg);
+        if (EndsWith(*file_path, ".blob")) {
+          ++files_scheduled_to_delete;
+        }
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "DeleteScheduler::OnDeleteFile", [&](void* arg) {
+        const std::string* const file_path =
+            static_cast<const std::string*>(arg);
+        if (EndsWith(*file_path, ".blob")) {
+          files_deleted++;
+        }
+      });
  ASSERT_OK(DestroyDB(dbname_, options));
  ASSERT_EQ(files_deleted, blob_files.size());
  ASSERT_EQ(files_scheduled_to_delete, blob_files.size());
@@ -648,6 +666,23 @@ TEST_F(DBSSTTest, DBWithSstFileManagerForBlobFilesWithGC) {
  }

  Close();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "SstFileManagerImpl::ScheduleUnaccountedFileDeletion", [&](void* arg) {
+        assert(arg);
+        const std::string* const file_path =
+            static_cast<const std::string*>(arg);
+        if (EndsWith(*file_path, ".blob")) {
+          ++files_scheduled_to_delete;
+        }
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "DeleteScheduler::OnDeleteFile", [&](void* arg) {
+        const std::string* const file_path =
+            static_cast<const std::string*>(arg);
+        if (EndsWith(*file_path, ".blob")) {
+          files_deleted++;
+        }
+      });
  ASSERT_OK(DestroyDB(dbname_, options));
  sfm->WaitForEmptyTrash();
  ASSERT_EQ(files_deleted, 5);
@@ -886,7 +921,7 @@ TEST_P(DBWALTestWithParam, WALTrashCleanupOnOpen) {
      // before restarting the DB.
      // We have to set this on the 2nd to last file for it to delay deletion
      // on the last file. (Quirk of DeleteScheduler::BackgroundEmptyTrash())
-      options.sst_file_manager->SetDeleteRateBytesPerSecond(1);
+      options.sst_file_manager->SetDeleteRateBytesPerSecond(1024 * 1024);
    }
    ASSERT_OK(Put("Key2", DummyString(1024, v)));
    ASSERT_OK(Put("Key3", DummyString(1024, v)));
@@ -937,15 +972,23 @@ INSTANTIATE_TEST_CASE_P(DBWALTestWithParam, DBWALTestWithParam,
                        ::testing::Values(std::make_tuple("", true),
                                          std::make_tuple("_wal_dir", false)));

-TEST_F(DBSSTTest, OpenDBWithExistingTrashAndObsoleteSstFile) {
+// Test param: max_trash_db_ratio for DeleteScheduler
+class DBObsoleteFileDeletionOnOpenTest
+    : public DBSSTTest,
+      public ::testing::WithParamInterface<double> {
+ public:
+  explicit DBObsoleteFileDeletionOnOpenTest()
+      : DBSSTTest("db_sst_deletion_on_open_test") {}
+};
+
+TEST_P(DBObsoleteFileDeletionOnOpenTest, Basic) {
  Options options = CurrentOptions();
  options.sst_file_manager.reset(
      NewSstFileManager(env_, nullptr, "", 1024 * 1024 /* 1 MB/sec */));
  auto sfm = static_cast<SstFileManagerImpl*>(options.sst_file_manager.get());
-  // Set an extra high trash ratio to prevent immediate/non-rate limited
-  // deletions
  sfm->SetDeleteRateBytesPerSecond(1024 * 1024);
-  sfm->delete_scheduler()->SetMaxTrashDBRatio(1000.0);
+  double max_trash_db_ratio = GetParam();
+  sfm->delete_scheduler()->SetMaxTrashDBRatio(max_trash_db_ratio);

  int bg_delete_file = 0;
  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
@@ -965,10 +1008,19 @@ TEST_F(DBSSTTest, OpenDBWithExistingTrashAndObsoleteSstFile) {
      WriteStringToFile(env_, "abc", dbname_ + "/" + "003.sst.trash", false));
  // Manually add an obsolete sst file. Obsolete SST files are discovered and
  // deleted upon recovery.
-  constexpr uint64_t kSstFileNumber = 100;
-  const std::string kObsoleteSstFile =
-      MakeTableFileName(dbname_, kSstFileNumber);
-  ASSERT_OK(WriteStringToFile(env_, "abc", kObsoleteSstFile, false));
+  uint64_t sst_file_number = 100;
+  const std::string kObsoleteSstFileOne =
+      MakeTableFileName(dbname_, sst_file_number);
+  ASSERT_OK(WriteStringToFile(env_, "abc", kObsoleteSstFileOne, false));
+  // The slow deletion on recovery had a bug before where a file's size is not
+  // first tracked in `total_size_` in SstFileManager before passed to
+  // DeleteScheduler. The first obsolete file is still slow deleted because
+  // 0 (total_trash_size_) > 0 (total_size_) * 1000 (max_trash_db_ratio)
+  // is always false.
+  // Here we explicitly create a second obsolete file to verify this bug's fix
+  const std::string kObsoleteSstFileTwo =
+      MakeTableFileName(dbname_, sst_file_number - 1);
+  ASSERT_OK(WriteStringToFile(env_, "abc", kObsoleteSstFileTwo, false));

  // Reopen the DB and verify that it deletes existing trash files and obsolete
  // SST files with rate limiting.
@@ -977,10 +1029,26 @@ TEST_F(DBSSTTest, OpenDBWithExistingTrashAndObsoleteSstFile) {
  ASSERT_NOK(env_->FileExists(dbname_ + "/" + "001.sst.trash"));
  ASSERT_NOK(env_->FileExists(dbname_ + "/" + "002.sst.trash"));
  ASSERT_NOK(env_->FileExists(dbname_ + "/" + "003.sst.trash"));
-  ASSERT_NOK(env_->FileExists(kObsoleteSstFile));
-  ASSERT_EQ(bg_delete_file, 4);
+  ASSERT_NOK(env_->FileExists(kObsoleteSstFileOne));
+  ASSERT_NOK(env_->FileExists(kObsoleteSstFileTwo));
+  // The files in the DB's directory are all either trash or obsolete sst files.
+  // So the trash/db ratio is 1. A ratio equal to or higher than 1 should
+  // schedule all files' deletion in background. A ratio lower than 1 may
+  // send some files to be deleted immediately.
+  if (max_trash_db_ratio < 1) {
+    ASSERT_LE(bg_delete_file, 5);
+  } else {
+    ASSERT_EQ(bg_delete_file, 5);
+  }
+
+  ASSERT_EQ(sfm->GetTotalSize(), 0);
+  ASSERT_EQ(sfm->delete_scheduler()->GetTotalTrashSize(), 0);
 }

+INSTANTIATE_TEST_CASE_P(DBObsoleteFileDeletionOnOpenTest,
+                        DBObsoleteFileDeletionOnOpenTest,
+                        ::testing::Values(0, 0.5, 1, 1.2));
+
 // Create a DB with 2 db_paths, and generate multiple files in the 2
 // db_paths using CompactRangeOptions, make sure that files that were
 // deleted from first db_path were deleted using DeleteScheduler and
@@ -1868,6 +1936,24 @@ TEST_F(DBSSTTest, DBWithSFMForBlobFilesAtomicFlush) {
  ASSERT_EQ(files_deleted, 1);

  Close();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "SstFileManagerImpl::ScheduleUnaccountedFileDeletion", [&](void* arg) {
+        assert(arg);
+        const std::string* const file_path =
+            static_cast<const std::string*>(arg);
+        if (EndsWith(*file_path, ".blob")) {
+          ++files_scheduled_to_delete;
+        }
+      });
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "DeleteScheduler::OnDeleteFile", [&](void* arg) {
+        const std::string* const file_path =
+            static_cast<const std::string*>(arg);
+        if (EndsWith(*file_path, ".blob")) {
+          files_deleted++;
+        }
+      });
  ASSERT_OK(DestroyDB(dbname_, options));

  ASSERT_EQ(files_scheduled_to_delete, 4);
@@ -26,6 +26,7 @@
 #endif

 #include "cache/lru_cache.h"
+#include "db/attribute_group_iterator_impl.h"
 #include "db/blob/blob_index.h"
 #include "db/blob/blob_log_format.h"
 #include "db/db_impl/db_impl.h"
@@ -3199,11 +3200,18 @@ class ModelDB : public DB {
    return Status::NotSupported("Not supported yet");
  }

-  // UNDER CONSTRUCTION - DO NOT USE
-  std::unique_ptr<Iterator> NewMultiCfIterator(
+  std::unique_ptr<Iterator> NewCoalescingIterator(
      const ReadOptions& /*options*/,
      const std::vector<ColumnFamilyHandle*>& /*column_families*/) override {
-    return nullptr;
+    return std::unique_ptr<Iterator>(
+        NewErrorIterator(Status::NotSupported("Not supported yet")));
+  }
+
+  std::unique_ptr<AttributeGroupIterator> NewAttributeGroupIterator(
+      const ReadOptions& /*options*/,
+      const std::vector<ColumnFamilyHandle*>& /*column_families*/) override {
+    return NewAttributeGroupErrorIterator(
+        Status::NotSupported("Not supported yet"));
  }

  const Snapshot* GetSnapshot() override {
@@ -3836,6 +3844,9 @@ TEST_P(DBTestWithParam, FIFOCompactionTest) {
    } else {
      CompactRangeOptions cro;
      cro.exclusive_manual_compaction = exclusive_manual_compaction_;
+      cro.change_level = true;
+      ASSERT_TRUE(db_->CompactRange(cro, nullptr, nullptr).IsNotSupported());
+      cro.change_level = false;
      ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
    }
    // only 5 files should survive
@@ -4666,24 +4677,27 @@ TEST_F(DBTest, DynamicMemtableOptions) {

 #ifdef ROCKSDB_USING_THREAD_STATUS
 namespace {
-void VerifyOperationCount(Env* env, ThreadStatus::OperationType op_type,
+bool VerifyOperationCount(Env* env, ThreadStatus::OperationType op_type,
                          int expected_count) {
  int op_count = 0;
  std::vector<ThreadStatus> thread_list;
-  ASSERT_OK(env->GetThreadList(&thread_list));
+  EXPECT_OK(env->GetThreadList(&thread_list));
  for (const auto& thread : thread_list) {
    if (thread.operation_type == op_type) {
      op_count++;
    }
  }
  if (op_count != expected_count) {
+    fprintf(stderr, "op_count: %d, expected_count %d\n", op_count,
+            expected_count);
    for (const auto& thread : thread_list) {
-      fprintf(stderr, "thread id: %" PRIu64 ", thread status: %s\n",
+      fprintf(stderr, "thread id: %" PRIu64 ", thread status: %s, cf_name %s\n",
              thread.thread_id,
-              thread.GetOperationName(thread.operation_type).c_str());
+              thread.GetOperationName(thread.operation_type).c_str(),
+              thread.cf_name.c_str());
    }
  }
-  ASSERT_EQ(op_count, expected_count);
+  return op_count == expected_count;
 }
 }  // anonymous namespace

@@ -4783,11 +4797,11 @@ TEST_F(DBTest, ThreadStatusFlush) {
  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();

  CreateAndReopenWithCF({"pikachu"}, options);
-  VerifyOperationCount(env_, ThreadStatus::OP_FLUSH, 0);
+  ASSERT_TRUE(VerifyOperationCount(env_, ThreadStatus::OP_FLUSH, 0));

  ASSERT_OK(Put(1, "foo", "v1"));
  ASSERT_EQ("v1", Get(1, "foo"));
-  VerifyOperationCount(env_, ThreadStatus::OP_FLUSH, 0);
+  ASSERT_TRUE(VerifyOperationCount(env_, ThreadStatus::OP_FLUSH, 0));

  uint64_t num_running_flushes = 0;
  ASSERT_TRUE(db_->GetIntProperty(DB::Properties::kNumRunningFlushes,
@@ -4800,7 +4814,7 @@ TEST_F(DBTest, ThreadStatusFlush) {
  // The first sync point is to make sure there's one flush job
  // running when we perform VerifyOperationCount().
  TEST_SYNC_POINT("DBTest::ThreadStatusFlush:1");
-  VerifyOperationCount(env_, ThreadStatus::OP_FLUSH, 1);
+  ASSERT_TRUE(VerifyOperationCount(env_, ThreadStatus::OP_FLUSH, 1));
  ASSERT_TRUE(db_->GetIntProperty(DB::Properties::kNumRunningFlushes,
                                  &num_running_flushes));
  ASSERT_EQ(num_running_flushes, 1);
@@ -4811,17 +4825,11 @@ TEST_F(DBTest, ThreadStatusFlush) {
 }

 TEST_P(DBTestWithParam, ThreadStatusSingleCompaction) {
-  const int kTestKeySize = 16;
  const int kTestValueSize = 984;
-  const int kEntrySize = kTestKeySize + kTestValueSize;
  const int kEntriesPerBuffer = 100;
  Options options;
  options.create_if_missing = true;
-  options.write_buffer_size = kEntrySize * kEntriesPerBuffer;
  options.compaction_style = kCompactionStyleLevel;
-  options.target_file_size_base = options.write_buffer_size;
-  options.max_bytes_for_level_base = options.target_file_size_base * 2;
-  options.max_bytes_for_level_multiplier = 2;
  options.compression = kNoCompression;
  options = CurrentOptions(options);
  options.env = env_;
@@ -4856,7 +4864,7 @@ TEST_P(DBTestWithParam, ThreadStatusSingleCompaction) {
                                    &num_running_compactions));
    ASSERT_EQ(num_running_compactions, 0);
    TEST_SYNC_POINT("DBTest::ThreadStatusSingleCompaction:0");
-    ASSERT_GE(NumTableFilesAtLevel(0),
+    ASSERT_EQ(NumTableFilesAtLevel(0),
              options.level0_file_num_compaction_trigger);

    // This makes sure at least one compaction is running.
@@ -4864,10 +4872,18 @@ TEST_P(DBTestWithParam, ThreadStatusSingleCompaction) {

    if (options.enable_thread_tracking) {
      // expecting one single L0 to L1 compaction
-      VerifyOperationCount(env_, ThreadStatus::OP_COMPACTION, 1);
+      // This test is flaky and fails here.
+      bool match = VerifyOperationCount(env_, ThreadStatus::OP_COMPACTION, 1);
+      if (!match) {
+        ASSERT_TRUE(db_->GetIntProperty(DB::Properties::kNumRunningCompactions,
+                                        &num_running_compactions));
+        fprintf(stderr, "running compaction: %" PRIu64 " lsm state: %s\n",
+                num_running_compactions, FilesPerLevel().c_str());
+      }
+      ASSERT_TRUE(match);
    } else {
      // If thread tracking is not enabled, compaction count should be 0.
-      VerifyOperationCount(env_, ThreadStatus::OP_COMPACTION, 0);
+      ASSERT_TRUE(VerifyOperationCount(env_, ThreadStatus::OP_COMPACTION, 0));
    }
    ASSERT_TRUE(db_->GetIntProperty(DB::Properties::kNumRunningCompactions,
                                    &num_running_compactions));
@@ -5629,6 +5645,8 @@ TEST_F(DBTest, DynamicUniversalCompactionOptions) {
  ASSERT_EQ(
      dbfull()->GetOptions().compaction_options_universal.allow_trivial_move,
      false);
+  ASSERT_EQ(dbfull()->GetOptions().compaction_options_universal.max_read_amp,
+            -1);

  ASSERT_OK(dbfull()->SetOptions(
      {{"compaction_options_universal", "{size_ratio=7;}"}}));
@@ -5650,9 +5668,11 @@ TEST_F(DBTest, DynamicUniversalCompactionOptions) {
  ASSERT_EQ(
      dbfull()->GetOptions().compaction_options_universal.allow_trivial_move,
      false);
+  ASSERT_EQ(dbfull()->GetOptions().compaction_options_universal.max_read_amp,
+            -1);

-  ASSERT_OK(dbfull()->SetOptions(
-      {{"compaction_options_universal", "{min_merge_width=11;}"}}));
+  ASSERT_OK(dbfull()->SetOptions({{"compaction_options_universal",
+                                   "{min_merge_width=11;max_read_amp=0;}"}}));
  ASSERT_EQ(dbfull()->GetOptions().compaction_options_universal.size_ratio, 7u);
  ASSERT_EQ(dbfull()->GetOptions().compaction_options_universal.min_merge_width,
            11u);
@@ -5671,6 +5691,8 @@ TEST_F(DBTest, DynamicUniversalCompactionOptions) {
  ASSERT_EQ(
      dbfull()->GetOptions().compaction_options_universal.allow_trivial_move,
      false);
+  ASSERT_EQ(dbfull()->GetOptions().compaction_options_universal.max_read_amp,
+            0);
 }

 TEST_F(DBTest, FileCreationRandomFailure) {
@@ -6331,6 +6353,8 @@ TEST_F(DBTest, PromoteL0) {
  Options options = CurrentOptions();
  options.disable_auto_compactions = true;
  options.write_buffer_size = 10 * 1024 * 1024;
+  // Exercise what was a use-after-free (ASAN failure) under ~VersionSet()
+  options.uncache_aggressiveness = 300;
  DestroyAndReopen(options);

  // non overlapping ranges
@@ -4154,7 +4154,7 @@ TEST_F(DBTest2, LiveFilesOmitObsoleteFiles) {
  TEST_SYNC_POINT("DBTest2::LiveFilesOmitObsoleteFiles:FlushTriggered");

  ASSERT_OK(db_->DisableFileDeletions());
-  VectorLogPtr log_files;
+  VectorWalPtr log_files;
  ASSERT_OK(db_->GetSortedWalFiles(log_files));
  TEST_SYNC_POINT("DBTest2::LiveFilesOmitObsoleteFiles:LiveFilesCaptured");
  for (const auto& log_file : log_files) {
@@ -154,6 +154,9 @@ bool DBTestBase::ShouldSkipOptions(int option_config, int skip_mask) {
  if ((skip_mask & kSkipMmapReads) && option_config == kWalDirAndMmapReads) {
    return true;
  }
+  if ((skip_mask & kSkipRowCache) && option_config == kRowCache) {
+    return true;
+  }
  return false;
 }

@@ -766,7 +769,9 @@ Status DBTestBase::TimedPut(const Slice& k, const Slice& v,

 Status DBTestBase::TimedPut(int cf, const Slice& k, const Slice& v,
                            uint64_t write_unix_time, WriteOptions wo) {
-  WriteBatch wb;
+  WriteBatch wb(/*reserved_bytes=*/0, /*max_bytes=*/0,
+                wo.protection_bytes_per_key,
+                /*default_cf_ts_sz=*/0);
  ColumnFamilyHandle* cfh;
  if (cf != 0) {
    cfh = handles_[cf];
@@ -276,16 +276,16 @@ class SpecialEnv : public EnvWrapper {
      SpecialEnv* env_;
      std::unique_ptr<WritableFile> base_;
    };
-    class WalFile : public WritableFile {
+    class SpecialWalFile : public WritableFile {
     public:
-      WalFile(SpecialEnv* env, std::unique_ptr<WritableFile>&& b)
+      SpecialWalFile(SpecialEnv* env, std::unique_ptr<WritableFile>&& b)
          : env_(env), base_(std::move(b)) {
        env_->num_open_wal_file_.fetch_add(1);
      }
-      virtual ~WalFile() { env_->num_open_wal_file_.fetch_add(-1); }
+      virtual ~SpecialWalFile() { env_->num_open_wal_file_.fetch_add(-1); }
      Status Append(const Slice& data) override {
 #if !(defined NDEBUG) || !defined(OS_WIN)
-        TEST_SYNC_POINT("SpecialEnv::WalFile::Append:1");
+        TEST_SYNC_POINT("SpecialEnv::SpecialWalFile::Append:1");
 #endif
        Status s;
        if (env_->log_write_error_.load(std::memory_order_acquire)) {
@@ -299,7 +299,7 @@ class SpecialEnv : public EnvWrapper {
          s = base_->Append(data);
        }
 #if !(defined NDEBUG) || !defined(OS_WIN)
-        TEST_SYNC_POINT("SpecialEnv::WalFile::Append:2");
+        TEST_SYNC_POINT("SpecialEnv::SpecialWalFile::Append:2");
 #endif
        return s;
      }
@@ -419,7 +419,7 @@ class SpecialEnv : public EnvWrapper {
      } else if (strstr(f.c_str(), "MANIFEST") != nullptr) {
        r->reset(new ManifestFile(this, std::move(*r)));
      } else if (strstr(f.c_str(), "log") != nullptr) {
-        r->reset(new WalFile(this, std::move(*r)));
+        r->reset(new SpecialWalFile(this, std::move(*r)));
      } else {
        r->reset(new OtherFile(this, std::move(*r)));
      }
@@ -1071,6 +1071,7 @@ class DBTestBase : public testing::Test {
    kSkipNoSeekToLast = 32,
    kSkipFIFOCompaction = 128,
    kSkipMmapReads = 256,
+    kSkipRowCache = 512,
  };

  const int kRangeDelSkipConfigs =
@@ -1078,7 +1079,9 @@ class DBTestBase : public testing::Test {
      kSkipPlainTable |
      // MmapReads disables the iterator pinning that RangeDelAggregator
      // requires.
-      kSkipMmapReads;
+      kSkipMmapReads |
+      // Not compatible yet.
+      kSkipRowCache;

  // `env_do_fsync` decides whether the special Env would do real
  // fsync for files and directories. Skipping fsync can speed up
@@ -14,6 +14,7 @@
 #include "port/stack_trace.h"
 #include "rocksdb/file_system.h"
 #include "test_util/sync_point.h"
+#include "util/defer.h"
 #include "util/udt_util.h"
 #include "utilities/fault_injection_env.h"
 #include "utilities/fault_injection_fs.h"
@@ -270,8 +271,10 @@ TEST_F(DBWALTest, SyncWALNotWaitWrite) {
  ASSERT_OK(Put("foo3", "bar3"));

  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({
-      {"SpecialEnv::WalFile::Append:1", "DBWALTest::SyncWALNotWaitWrite:1"},
-      {"DBWALTest::SyncWALNotWaitWrite:2", "SpecialEnv::WalFile::Append:2"},
+      {"SpecialEnv::SpecialWalFile::Append:1",
+       "DBWALTest::SyncWALNotWaitWrite:1"},
+      {"DBWALTest::SyncWALNotWaitWrite:2",
+       "SpecialEnv::SpecialWalFile::Append:2"},
  });
  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();

@@ -814,6 +817,7 @@ TEST_F(DBWALTest, WALWithChecksumHandoff) {
    writeOpt.disableWAL = false;
    // Data is persisted in the WAL
    ASSERT_OK(dbfull()->Put(writeOpt, handles_[1], "zoo", "v3"));
+    ASSERT_OK(dbfull()->SyncWAL());
    // The hash does not match, write fails
    fault_fs->SetChecksumHandoffFuncType(ChecksumType::kxxHash);
    writeOpt.disableWAL = false;
@@ -1468,6 +1472,93 @@ TEST_F(DBWALTest, SyncMultipleLogs) {
  ASSERT_OK(dbfull()->SyncWAL());
 }

+TEST_F(DBWALTest, SyncWalPartialFailure) {
+  class MyTestFileSystem : public FileSystemWrapper {
+   public:
+    explicit MyTestFileSystem(std::shared_ptr<FileSystem> base)
+        : FileSystemWrapper(std::move(base)) {}
+
+    static const char* kClassName() { return "MyTestFileSystem"; }
+    const char* Name() const override { return kClassName(); }
+    IOStatus NewWritableFile(const std::string& fname,
+                             const FileOptions& file_opts,
+                             std::unique_ptr<FSWritableFile>* result,
+                             IODebugContext* dbg) override {
+      IOStatus s = target()->NewWritableFile(fname, file_opts, result, dbg);
+      if (s.ok()) {
+        *result =
+            std::make_unique<MyTestWritableFile>(std::move(*result), *this);
+      }
+      return s;
+    }
+
+    AcqRelAtomic<uint32_t> syncs_before_failure_{UINT32_MAX};
+
+   protected:
+    class MyTestWritableFile : public FSWritableFileOwnerWrapper {
+     public:
+      MyTestWritableFile(std::unique_ptr<FSWritableFile>&& file,
+                         MyTestFileSystem& fs)
+          : FSWritableFileOwnerWrapper(std::move(file)), fs_(fs) {}
+
+      IOStatus Sync(const IOOptions& options, IODebugContext* dbg) override {
+        int prev_val = fs_.syncs_before_failure_.FetchSub(1);
+        if (prev_val == 0) {
+          return IOStatus::IOError("fault");
+        } else {
+          return target()->Sync(options, dbg);
+        }
+      }
+
+     protected:
+      MyTestFileSystem& fs_;
+    };
+  };
+
+  Options options = CurrentOptions();
+  options.max_write_buffer_number = 4;
+  options.track_and_verify_wals_in_manifest = true;
+  options.max_bgerror_resume_count = 0;  // manual resume
+
+  auto custom_fs =
+      std::make_shared<MyTestFileSystem>(options.env->GetFileSystem());
+  std::unique_ptr<Env> fault_fs_env(NewCompositeEnv(custom_fs));
+  options.env = fault_fs_env.get();
+  Reopen(options);
+  Defer closer([this]() { Close(); });
+
+  // This is the simplest way to get
+  // * one inactive WAL, synced
+  // * one inactive WAL, not synced, and
+  // * one active WAL, not synced
+  // with a single thread, to exercise as much logic as we reasonably can.
+  ASSERT_OK(static_cast_with_check<DBImpl>(db_)->PauseBackgroundWork());
+  ASSERT_OK(Put("key1", "val1"));
+  ASSERT_OK(static_cast_with_check<DBImpl>(db_)->TEST_SwitchMemtable());
+  ASSERT_OK(db_->SyncWAL());
+  ASSERT_OK(Put("key2", "val2"));
+  ASSERT_OK(static_cast_with_check<DBImpl>(db_)->TEST_SwitchMemtable());
+  ASSERT_OK(Put("key3", "val3"));
+
+  // Allow 1 of the WALs to sync, but another won't
+  custom_fs->syncs_before_failure_.Store(1);
+  ASSERT_NOK(db_->SyncWAL());
+
+  // Stuck in this state. (This could previously cause a segfault.)
+  ASSERT_NOK(db_->SyncWAL());
+
+  // Can't Resume because WAL write failure is considered non-recoverable,
+  // regardless of the IOStatus itself. (Can/should be fixed?)
+  ASSERT_NOK(db_->Resume());
+
+  // Verify no data loss after reopen.
+  // Also Close() could previously crash in this state.
+  Reopen(options);
+  ASSERT_EQ("val1", Get("key1"));
+  ASSERT_EQ("val2", Get("key2"));
+  ASSERT_EQ("val3", Get("key3"));
+}
+
 // Github issue 1339. Prior the fix we read sequence id from the first log to
 // a local variable, then keep increase the variable as we replay logs,
 // ignoring actual sequence id of the records. This is incorrect if some writes
@@ -666,11 +666,25 @@ TEST_P(DBWriteTest, LockWALInEffect) {
  // try the 1st WAL created during open
  ASSERT_OK(Put("key0", "value"));
  ASSERT_NE(options.manual_wal_flush, dbfull()->WALBufferIsEmpty());
+
  ASSERT_OK(db_->LockWAL());
+
  ASSERT_TRUE(dbfull()->WALBufferIsEmpty());
+  uint64_t wal_num = dbfull()->TEST_GetCurrentLogNumber();
+  // Manual flush with wait=false should abruptly fail with TryAgain
+  FlushOptions flush_opts;
+  flush_opts.wait = false;
+  for (bool allow_write_stall : {true, false}) {
+    flush_opts.allow_write_stall = allow_write_stall;
+    ASSERT_TRUE(db_->Flush(flush_opts).IsTryAgain());
+  }
+  ASSERT_EQ(wal_num, dbfull()->TEST_GetCurrentLogNumber());
+
  ASSERT_OK(db_->UnlockWAL());
-  // try the 2nd wal created during SwitchWAL
+
+  // try the 2nd wal created during SwitchWAL (not locked this time)
  ASSERT_OK(dbfull()->TEST_SwitchWAL());
+  ASSERT_NE(wal_num, dbfull()->TEST_GetCurrentLogNumber());
  ASSERT_OK(Put("key1", "value"));
  ASSERT_NE(options.manual_wal_flush, dbfull()->WALBufferIsEmpty());
  ASSERT_OK(db_->LockWAL());
@@ -709,21 +723,57 @@ TEST_P(DBWriteTest, LockWALInEffect) {
 }

 TEST_P(DBWriteTest, LockWALConcurrentRecursive) {
+  // This is a micro-stress test of LockWAL and concurrency handling.
+  // It is considered the most convenient way to balance functional
+  // coverage and reproducibility (vs. the two extremes of (a) unit tests
+  // tailored to specific interleavings and (b) db_stress)
  Options options = GetOptions();
  Reopen(options);
-  ASSERT_OK(Put("k1", "val"));
+  ASSERT_OK(Put("k1", "k1_orig"));
  ASSERT_OK(db_->LockWAL());  // 0 -> 1
  auto frozen_seqno = db_->GetLatestSequenceNumber();
-  std::atomic<bool> t1_completed{false};
-  port::Thread t1{[&]() {
-    // Won't finish until WAL unlocked
-    ASSERT_OK(Put("k1", "val2"));
-    t1_completed = true;
+
+  std::string ingest_file = dbname_ + "/external.sst";
+  {
+    SstFileWriter sst_file_writer(EnvOptions(), options);
+    ASSERT_OK(sst_file_writer.Open(ingest_file));
+    ASSERT_OK(sst_file_writer.Put("k2", "k2_val"));
+    ExternalSstFileInfo external_info;
+    ASSERT_OK(sst_file_writer.Finish(&external_info));
+  }
+  AcqRelAtomic<bool> parallel_ingest_completed{false};
+  port::Thread parallel_ingest{[&]() {
+    IngestExternalFileOptions ingest_opts;
+    ingest_opts.move_files = true;  // faster than copy
+    // Shouldn't finish until WAL unlocked
+    ASSERT_OK(db_->IngestExternalFile({ingest_file}, ingest_opts));
+    parallel_ingest_completed.Store(true);
+  }};
+
+  AcqRelAtomic<bool> flush_completed{false};
+  port::Thread parallel_flush{[&]() {
+    FlushOptions flush_opts;
+    // NB: Flush with wait=false case is tested above in LockWALInEffect
+    flush_opts.wait = true;
+    // allow_write_stall = true blocks in fewer cases
+    flush_opts.allow_write_stall = true;
+    // Shouldn't finish until WAL unlocked
+    ASSERT_OK(db_->Flush(flush_opts));
+    flush_completed.Store(true);
+  }};
+
+  AcqRelAtomic<bool> parallel_put_completed{false};
+  port::Thread parallel_put{[&]() {
+    // This can make certain failure scenarios more likely:
+    //   sleep(1);
+    // Shouldn't finish until WAL unlocked
+    ASSERT_OK(Put("k1", "k1_mod"));
+    parallel_put_completed.Store(true);
  }};

  ASSERT_OK(db_->LockWAL());  // 1 -> 2
  // Read-only ops are OK
-  ASSERT_EQ(Get("k1"), "val");
+  ASSERT_EQ(Get("k1"), "k1_orig");
  {
    std::vector<LiveFileStorageInfo> files;
    LiveFilesStorageInfoOptions lf_opts;
@@ -732,29 +782,35 @@ TEST_P(DBWriteTest, LockWALConcurrentRecursive) {
    ASSERT_OK(db_->GetLiveFilesStorageInfo({lf_opts}, &files));
  }

-  port::Thread t2{[&]() {
+  port::Thread parallel_lock_wal{[&]() {
    ASSERT_OK(db_->LockWAL());  // 2 -> 3 or 1 -> 2
  }};

  ASSERT_OK(db_->UnlockWAL());  // 2 -> 1 or 3 -> 2
-  // Give t1 an extra chance to jump in case of bug
+  // Give parallel_put an extra chance to jump in case of bug
  std::this_thread::yield();
-  t2.join();
-  ASSERT_FALSE(t1_completed.load());
+  parallel_lock_wal.join();
+  ASSERT_FALSE(parallel_put_completed.Load());
+  ASSERT_FALSE(parallel_ingest_completed.Load());
+  ASSERT_FALSE(flush_completed.Load());

  // Should now have 2 outstanding LockWAL
-  ASSERT_EQ(Get("k1"), "val");
+  ASSERT_EQ(Get("k1"), "k1_orig");

  ASSERT_OK(db_->UnlockWAL());  // 2 -> 1

-  ASSERT_FALSE(t1_completed.load());
-  ASSERT_EQ(Get("k1"), "val");
+  ASSERT_FALSE(parallel_put_completed.Load());
+  ASSERT_FALSE(parallel_ingest_completed.Load());
+  ASSERT_FALSE(flush_completed.Load());
+
+  ASSERT_EQ(Get("k1"), "k1_orig");
+  ASSERT_EQ(Get("k2"), "NOT_FOUND");
  ASSERT_EQ(frozen_seqno, db_->GetLatestSequenceNumber());

  // Ensure final Unlock is concurrency safe and extra Unlock is safe but
  // non-OK
  std::atomic<int> unlock_ok{0};
-  port::Thread t3{[&]() {
+  port::Thread parallel_stuff{[&]() {
    if (db_->UnlockWAL().ok()) {
      unlock_ok++;
    }
@@ -767,18 +823,23 @@ TEST_P(DBWriteTest, LockWALConcurrentRecursive) {
  if (db_->UnlockWAL().ok()) {
    unlock_ok++;
  }
-  t3.join();
+  parallel_stuff.join();

  // There was one extra unlock, so just one non-ok
  ASSERT_EQ(unlock_ok.load(), 2);

  // Write can proceed
-  t1.join();
-  ASSERT_TRUE(t1_completed.load());
-  ASSERT_EQ(Get("k1"), "val2");
+  parallel_put.join();
+  ASSERT_TRUE(parallel_put_completed.Load());
+  ASSERT_EQ(Get("k1"), "k1_mod");
+  parallel_ingest.join();
+  ASSERT_TRUE(parallel_ingest_completed.Load());
+  ASSERT_EQ(Get("k2"), "k2_val");
+  parallel_flush.join();
+  ASSERT_TRUE(flush_completed.Load());
  // And new writes
-  ASSERT_OK(Put("k2", "val"));
-  ASSERT_EQ(Get("k2"), "val");
+  ASSERT_OK(Put("k3", "val"));
+  ASSERT_EQ(Get("k3"), "val");
 }

 TEST_P(DBWriteTest, ConcurrentlyDisabledWAL) {
@@ -795,7 +856,7 @@ TEST_P(DBWriteTest, ConcurrentlyDisabledWAL) {
  std::thread threads[10];
  for (int t = 0; t < 10; t++) {
    threads[t] = std::thread([t, wal_key_prefix, wal_value, no_wal_key_prefix,
-                              no_wal_value, this] {
+                              no_wal_value, &options, this] {
      for (int i = 0; i < 10; i++) {
        ROCKSDB_NAMESPACE::WriteOptions write_option_disable;
        write_option_disable.disableWAL = true;
@@ -806,7 +867,10 @@ TEST_P(DBWriteTest, ConcurrentlyDisabledWAL) {
        std::string wal_key =
            wal_key_prefix + std::to_string(i) + "_" + std::to_string(i);
        ASSERT_OK(this->Put(wal_key, wal_value, write_option_default));
-        ASSERT_OK(dbfull()->SyncWAL());
+        ASSERT_OK(dbfull()->SyncWAL())
+            << "options.env: " << options.env << ", env_: " << env_
+            << ", env_->is_wal_sync_thread_safe_: "
+            << env_->is_wal_sync_thread_safe_.load();
      }
      return;
    });
@@ -910,6 +974,32 @@ TEST_P(DBWriteTest, RecycleLogTestCFAheadOfWAL) {
            Status::Corruption());
 }

+TEST_P(DBWriteTest, RecycleLogToggleTest) {
+  Options options = GetOptions();
+  options.recycle_log_file_num = 0;
+  options.avoid_flush_during_recovery = true;
+  options.wal_recovery_mode = WALRecoveryMode::kPointInTimeRecovery;
+
+  Destroy(options);
+  Reopen(options);
+  // After opening, a new log gets created, say 1.log
+  ASSERT_OK(Put(Key(1), "val1"));
+
+  options.recycle_log_file_num = 1;
+  Reopen(options);
+  // 1.log is added to alive_log_files_
+  ASSERT_OK(Put(Key(2), "val1"));
+  ASSERT_OK(Flush());
+  // 1.log should be deleted and not recycled, since it
+  // was created by the previous Reopen
+  ASSERT_OK(Put(Key(1), "val2"));
+  ASSERT_OK(Flush());
+
+  options.recycle_log_file_num = 1;
+  Reopen(options);
+  ASSERT_EQ(Get(Key(1)), "val2");
+}
+
 INSTANTIATE_TEST_CASE_P(DBWriteTestInstance, DBWriteTest,
                        testing::Values(DBTestBase::kDefault,
                                        DBTestBase::kConcurrentWALWrites,
@@ -47,6 +47,8 @@ EntryType GetEntryType(ValueType value_type) {
      return kEntryBlobIndex;
    case kTypeWideColumnEntity:
      return kEntryWideColumnEntity;
+    case kTypeValuePreferredSeqno:
+      return kEntryTimedPut;
    default:
      return kEntryOther;
  }
@@ -155,10 +157,23 @@ void ReplaceInternalKeyWithMinTimestamp(std::string* result, const Slice& key,
  result->append(key.data() + key_sz - kNumInternalBytes, kNumInternalBytes);
 }

-std::string ParsedInternalKey::DebugString(bool log_err_key, bool hex) const {
+std::string ParsedInternalKey::DebugString(bool log_err_key, bool hex,
+                                           const Comparator* ucmp) const {
  std::string result = "'";
+  size_t ts_sz_for_debug = ucmp == nullptr ? 0 : ucmp->timestamp_size();
  if (log_err_key) {
-    result += user_key.ToString(hex);
+    if (ts_sz_for_debug == 0) {
+      result += user_key.ToString(hex);
+    } else {
+      assert(user_key.size() >= ts_sz_for_debug);
+      Slice user_key_without_ts = user_key;
+      user_key_without_ts.remove_suffix(ts_sz_for_debug);
+      result += user_key_without_ts.ToString(hex);
+      Slice ts = Slice(user_key.data() + user_key.size() - ts_sz_for_debug,
+                       ts_sz_for_debug);
+      result += "|timestamp:";
+      result += ucmp->TimestampToString(ts);
+    }
  } else {
    result += "<redacted>";
  }
@@ -171,11 +186,11 @@ std::string ParsedInternalKey::DebugString(bool log_err_key, bool hex) const {
  return result;
 }

-std::string InternalKey::DebugString(bool hex) const {
+std::string InternalKey::DebugString(bool hex, const Comparator* ucmp) const {
  std::string result;
  ParsedInternalKey parsed;
  if (ParseInternalKey(rep_, &parsed, false /* log_err_key */).ok()) {
-    result = parsed.DebugString(true /* log_err_key */, hex);  // TODO
+    result = parsed.DebugString(true /* log_err_key */, hex, ucmp);  // TODO
  } else {
    result = "(bad)";
    result.append(EscapeString(rep_));
@@ -148,7 +148,8 @@ struct ParsedInternalKey {
  // u contains timestamp if user timestamp feature is enabled.
  ParsedInternalKey(const Slice& u, const SequenceNumber& seq, ValueType t)
      : user_key(u), sequence(seq), type(t) {}
-  std::string DebugString(bool log_err_key, bool hex) const;
+  std::string DebugString(bool log_err_key, bool hex,
+                          const Comparator* ucmp = nullptr) const;

  void clear() {
    user_key.clear();
@@ -503,7 +504,7 @@ class InternalKey {
    AppendInternalKeyFooter(&rep_, s, t);
  }

-  std::string DebugString(bool hex) const;
+  std::string DebugString(bool hex, const Comparator* ucmp = nullptr) const;
 };

 inline int InternalKeyComparator::Compare(const InternalKey& a,
@@ -223,6 +223,49 @@ TEST_F(DeleteFileTest, PurgeObsoleteFilesTest) {
  CheckFileTypeCounts(dbname_, 0, 1, 1);
 }

+TEST_F(DeleteFileTest, WaitForCompactWithWaitForPurgeOptionTest) {
+  Options options = CurrentOptions();
+  SetOptions(&options);
+  Destroy(options);
+  options.create_if_missing = true;
+  Reopen(options);
+
+  std::string first("0"), last("999999");
+  CompactRangeOptions compact_options;
+  compact_options.change_level = true;
+  compact_options.target_level = 2;
+  Slice first_slice(first), last_slice(last);
+
+  CreateTwoLevels();
+  Iterator* itr = nullptr;
+  ReadOptions read_options;
+  read_options.background_purge_on_iterator_cleanup = true;
+  itr = db_->NewIterator(read_options);
+  ASSERT_OK(itr->status());
+  ASSERT_OK(db_->CompactRange(compact_options, &first_slice, &last_slice));
+  SyncPoint::GetInstance()->LoadDependency(
+      {{"DBImpl::BGWorkPurge:start", "DeleteFileTest::WaitForPurgeTest"},
+       {"DBImpl::WaitForCompact:InsideLoop",
+        "DBImpl::BackgroundCallPurge:beforeMutexLock"}});
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  delete itr;
+
+  TEST_SYNC_POINT("DeleteFileTest::WaitForPurgeTest");
+  // At this point, purge got started, but can't finish due to sync points
+  // not purged yet
+  CheckFileTypeCounts(dbname_, 0, 3, 1);
+
+  // The sync point in WaitForCompact should unblock the purge
+  WaitForCompactOptions wait_for_compact_options;
+  wait_for_compact_options.wait_for_purge = true;
+  Status s = dbfull()->WaitForCompact(wait_for_compact_options);
+  ASSERT_OK(s);
+
+  // Now files should be purged
+  CheckFileTypeCounts(dbname_, 0, 1, 1);
+}
+
 TEST_F(DeleteFileTest, BackgroundPurgeIteratorTest) {
  Options options = CurrentOptions();
  SetOptions(&options);
@@ -600,4 +643,3 @@ int main(int argc, char** argv) {
  RegisterCustomObjects(argc, argv);
  return RUN_ALL_TESTS();
 }
-
@@ -30,6 +30,14 @@ class DBErrorHandlingFSTest : public DBTestBase {
    fault_env_.reset(new CompositeEnvWrapper(env_, fault_fs_));
  }

+  ~DBErrorHandlingFSTest() {
+    // Before destroying fault_env_
+    SyncPoint::GetInstance()->DisableProcessing();
+    SyncPoint::GetInstance()->LoadDependency({});
+    SyncPoint::GetInstance()->ClearAllCallBacks();
+    Close();
+  }
+
  std::string GetManifestNameFromLiveFiles() {
    std::vector<std::string> live_files;
    uint64_t manifest_size;
@@ -417,7 +425,7 @@ TEST_F(DBErrorHandlingFSTest, FlushWALWriteRetryableError) {

  listener->EnableAutoRecovery(false);
  SyncPoint::GetInstance()->SetCallBack(
-      "DBImpl::SyncClosedLogs:Start",
+      "DBImpl::SyncClosedWals:Start",
      [&](void*) { fault_fs_->SetFilesystemActive(false, error_msg); });
  SyncPoint::GetInstance()->EnableProcessing();

@@ -462,7 +470,7 @@ TEST_F(DBErrorHandlingFSTest, FlushWALAtomicWriteRetryableError) {

  listener->EnableAutoRecovery(false);
  SyncPoint::GetInstance()->SetCallBack(
-      "DBImpl::SyncClosedLogs:Start",
+      "DBImpl::SyncClosedWals:Start",
      [&](void*) { fault_fs_->SetFilesystemActive(false, error_msg); });
  SyncPoint::GetInstance()->EnableProcessing();

@@ -1238,6 +1246,7 @@ TEST_F(DBErrorHandlingFSTest, AutoRecoverFlushError) {
                   ERROR_HANDLER_AUTORESUME_RETRY_TOTAL_COUNT));
  ASSERT_EQ(0, options.statistics->getAndResetTickerCount(
                   ERROR_HANDLER_AUTORESUME_SUCCESS_COUNT));
+  ASSERT_OK(dbfull()->SyncWAL());

  Reopen(options);
  ASSERT_EQ("val", Get(Key(0)));
@@ -342,8 +342,7 @@ Status ExternalSstFileIngestionJob::NeedsFlush(bool* flush_needed,
  autovector<UserKeyRange> ranges;
  ranges.reserve(n);
  for (const IngestedFileInfo& file_to_ingest : files_to_ingest_) {
-    ranges.emplace_back(file_to_ingest.smallest_internal_key.user_key(),
-                        file_to_ingest.largest_internal_key.user_key());
+    ranges.emplace_back(file_to_ingest.start_ukey, file_to_ingest.limit_ukey);
  }
  Status status = cfd_->RangesOverlapWithMemtables(
      ranges, super_version, db_options_.allow_data_in_errors, flush_needed);
@@ -930,6 +929,17 @@ Status ExternalSstFileIngestionJob::GetIngestedFileInfo(
    }
  }

+  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();
+  if (ts_sz > 0) {
+    AppendUserKeyWithMaxTimestamp(&file_to_ingest->start_ukey, smallest, ts_sz);
+    AppendUserKeyWithMinTimestamp(&file_to_ingest->limit_ukey, largest, ts_sz);
+  } else {
+    file_to_ingest->start_ukey.assign(smallest.data(), smallest.size());
+    file_to_ingest->limit_ukey.assign(largest.data(), largest.size());
+  }
+
  auto s =
      GetSstInternalUniqueId(file_to_ingest->table_properties.db_id,
                             file_to_ingest->table_properties.db_session_id,
@@ -953,13 +963,15 @@ Status ExternalSstFileIngestionJob::AssignLevelAndSeqnoForIngestedFile(
  *assigned_seqno = 0;
  auto ucmp = cfd_->user_comparator();
  const size_t ts_sz = ucmp->timestamp_size();
-  if (force_global_seqno || files_overlap_) {
+  if (force_global_seqno || files_overlap_ ||
+      compaction_style == kCompactionStyleFIFO) {
    *assigned_seqno = last_seqno + 1;
    // If files overlap, we have to ingest them at level 0.
-    if (files_overlap_) {
+    if (files_overlap_ || compaction_style == kCompactionStyleFIFO) {
      assert(ts_sz == 0);
      file_to_ingest->picked_level = 0;
-      if (ingestion_options_.fail_if_not_bottommost_level) {
+      if (ingestion_options_.fail_if_not_bottommost_level &&
+          cfd_->NumberLevels() > 1) {
        status = Status::TryAgain(
            "Files cannot be ingested to Lmax. Please make sure key range of "
            "Lmax does not overlap with files to ingest.");
@@ -980,9 +992,8 @@ Status ExternalSstFileIngestionJob::AssignLevelAndSeqnoForIngestedFile(
    if (lvl > 0 && lvl < vstorage->base_level()) {
      continue;
    }
-    if (cfd_->RangeOverlapWithCompaction(
-            file_to_ingest->smallest_internal_key.user_key(),
-            file_to_ingest->largest_internal_key.user_key(), lvl)) {
+    if (cfd_->RangeOverlapWithCompaction(file_to_ingest->start_ukey,
+                                         file_to_ingest->limit_ukey, lvl)) {
      // We must use L0 or any level higher than `lvl` to be able to overwrite
      // the compaction output keys that we overlap with in this level, We also
      // need to assign this file a seqno to overwrite the compaction output
@@ -992,9 +1003,8 @@ Status ExternalSstFileIngestionJob::AssignLevelAndSeqnoForIngestedFile(
    } else if (vstorage->NumLevelFiles(lvl) > 0) {
      bool overlap_with_level = false;
      status = sv->current->OverlapWithLevelIterator(
-          ro, env_options_, file_to_ingest->smallest_internal_key.user_key(),
-          file_to_ingest->largest_internal_key.user_key(), lvl,
-          &overlap_with_level);
+          ro, env_options_, file_to_ingest->start_ukey,
+          file_to_ingest->limit_ukey, lvl, &overlap_with_level);
      if (!status.ok()) {
        return status;
      }
@@ -1163,9 +1173,8 @@ bool ExternalSstFileIngestionJob::IngestedFileFitInLevel(
  }

  auto* vstorage = cfd_->current()->storage_info();
-  Slice file_smallest_user_key(
-      file_to_ingest->smallest_internal_key.user_key());
-  Slice file_largest_user_key(file_to_ingest->largest_internal_key.user_key());
+  Slice file_smallest_user_key(file_to_ingest->start_ukey);
+  Slice file_largest_user_key(file_to_ingest->limit_ukey);

  if (vstorage->OverlapInLevel(level, &file_smallest_user_key,
                               &file_largest_user_key)) {
@@ -32,6 +32,17 @@ struct IngestedFileInfo {
  InternalKey smallest_internal_key;
  // Largest internal key in external file
  InternalKey largest_internal_key;
+  // NOTE: use below two fields for all `*Overlap*` types of checks instead of
+  // smallest_internal_key.user_key() and largest_internal_key.user_key().
+  // The smallest / largest user key contained in the file for key range checks.
+  // These could be different from smallest_internal_key.user_key(), and
+  // largest_internal_key.user_key() when user-defined timestamps are enabled,
+  // because the check is about making sure the user key without timestamps part
+  // does not overlap. To achieve that, the smallest user key will be updated
+  // with the maximum timestamp while the largest user key will be updated with
+  // the min timestamp. It's otherwise the same.
+  std::string start_ukey;
+  std::string limit_ukey;
  // Sequence number for keys in external file
  SequenceNumber original_seqno;
  // Offset of the global sequence number field in the file, will
@@ -3082,6 +3082,68 @@ TEST_P(ExternalSSTFileTest,
  delete iter;
 }

+TEST_F(ExternalSSTFileTest, FIFOCompaction) {
+  // FIFO always ingests SST files to L0 and assign latest sequence number.
+  Options options = CurrentOptions();
+  options.num_levels = 1;
+  options.compaction_style = kCompactionStyleFIFO;
+  options.max_open_files = -1;
+  DestroyAndReopen(options);
+  std::map<std::string, std::string> true_data;
+
+  for (int i = 0; i < 100; ++i) {
+    ASSERT_OK(Put(Key(i), Key(i) + "_val"));
+    true_data[Key(i)] = Key(i) + "_val";
+  }
+  ASSERT_OK(Flush());
+  ASSERT_EQ("1", FilesPerLevel());
+  std::vector<std::pair<std::string, std::string>> file_data;
+  for (int i = 0; i <= 20; i++) {
+    file_data.emplace_back(Key(i), Key(i) + "_ingest");
+  }
+  // Overlaps with memtable, will trigger flush
+  ASSERT_OK(GenerateAndAddExternalFile(options, file_data, -1,
+                                       /*allow_global_seqno=*/true, true, false,
+                                       false, false, &true_data));
+  ASSERT_EQ("2", FilesPerLevel());
+
+  file_data.clear();
+  for (int i = 100; i <= 120; i++) {
+    file_data.emplace_back(Key(i), Key(i) + "_ingest");
+  }
+  // global sequence number is always assigned, so this will fail
+  ASSERT_NOK(GenerateAndAddExternalFile(options, file_data, -1,
+                                        /*allow_global_seqno=*/false, true,
+                                        false, false, false, &true_data));
+  ASSERT_OK(GenerateAndAddExternalFile(options, file_data, -1,
+                                       /*allow_global_seqno=*/true, true, false,
+                                       false, false, &true_data));
+
+  // Compact to data to lower level to test multi-level FIFO later
+  options.num_levels = 7;
+  options.compaction_style = kCompactionStyleUniversal;
+  ASSERT_OK(TryReopen(options));
+  CompactRangeOptions cro;
+  cro.bottommost_level_compaction = BottommostLevelCompaction::kForceOptimized;
+  ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
+  ASSERT_EQ("0,0,0,0,0,0,1", FilesPerLevel());
+
+  options.num_levels = 7;
+  options.compaction_style = kCompactionStyleFIFO;
+  ASSERT_OK(TryReopen(options));
+  file_data.clear();
+  for (int i = 200; i <= 220; i++) {
+    file_data.emplace_back(Key(i), Key(i) + "_ingest");
+  }
+  // Files are ingested into L0 for multi-level FIFO
+  ASSERT_OK(GenerateAndAddExternalFile(options, file_data, -1,
+                                       /*allow_global_seqno=*/true, true, false,
+                                       false, false, &true_data));
+
+  ASSERT_EQ("1,0,0,0,0,0,1", FilesPerLevel());
+  VerifyDBFromMap(true_data);
+}
+
 class ExternalSSTFileWithTimestampTest : public ExternalSSTFileTest {
 public:
  ExternalSSTFileWithTimestampTest() = default;
@@ -3296,11 +3358,60 @@ TEST_F(ExternalSSTFileWithTimestampTest, Basic) {

    ASSERT_OK(IngestExternalUDTFile({file4}));

+    for (int k = 200; k < 250; k++) {
+      VerifyValueAndTs(Key(k), EncodeAsUint64(k), Key(k) + "_val",
+                       EncodeAsUint64(k));
+    }
+
    // In UDT mode, any external file that can be successfully ingested also
    // should not overlap with the db. As a result, they can always get the
    // seq 0 assigned.
    ASSERT_EQ(db_->GetLatestSequenceNumber(), seq_num_before_ingestion);

+    // file5.sst (Key(200), ts = 199)
+    // While DB has (Key(200), ts = 200) => user key without timestamp overlaps
+    std::string file5 = sst_files_dir_ + "file5.sst";
+    ASSERT_OK(sst_file_writer.Open(file5));
+    ASSERT_OK(
+        sst_file_writer.Put(Key(200), EncodeAsUint64(199), Key(200) + "_val"));
+
+    ExternalSstFileInfo file5_info;
+    ASSERT_OK(sst_file_writer.Finish(&file5_info));
+    ASSERT_TRUE(IngestExternalUDTFile({file5}).IsInvalidArgument());
+
+    // file6.sst (Key(200), ts = 201)
+    // While DB has (Key(200), ts = 200) => user key without timestamp overlaps
+    std::string file6 = sst_files_dir_ + "file6.sst";
+    ASSERT_OK(sst_file_writer.Open(file6));
+    ASSERT_OK(
+        sst_file_writer.Put(Key(200), EncodeAsUint64(201), Key(0) + "_val"));
+
+    ExternalSstFileInfo file6_info;
+    ASSERT_OK(sst_file_writer.Finish(&file6_info));
+    ASSERT_TRUE(IngestExternalUDTFile({file6}).IsInvalidArgument());
+
+    // Check memtable overlap.
+    ASSERT_OK(dbfull()->Put(WriteOptions(), Key(250), EncodeAsUint64(250),
+                            Key(250) + "_val"));
+
+    std::string file7 = sst_files_dir_ + "file7.sst";
+    ASSERT_OK(sst_file_writer.Open(file7));
+    ASSERT_OK(
+        sst_file_writer.Put(Key(250), EncodeAsUint64(249), Key(250) + "_val2"));
+
+    ExternalSstFileInfo file7_info;
+    ASSERT_OK(sst_file_writer.Finish(&file7_info));
+    ASSERT_TRUE(IngestExternalUDTFile({file7}).IsInvalidArgument());
+
+    std::string file8 = sst_files_dir_ + "file8.sst";
+    ASSERT_OK(sst_file_writer.Open(file8));
+    ASSERT_OK(
+        sst_file_writer.Put(Key(250), EncodeAsUint64(251), Key(250) + "_val3"));
+
+    ExternalSstFileInfo file8_info;
+    ASSERT_OK(sst_file_writer.Finish(&file8_info));
+    ASSERT_TRUE(IngestExternalUDTFile({file8}).IsInvalidArgument());
+
    DestroyAndRecreateExternalSSTFilesDir();
  } while (ChangeOptions(kSkipPlainTable | kSkipFIFOCompaction |
                         kRangeDelSkipConfigs));
@@ -12,6 +12,7 @@
 // file data (or entire files) not protected by a "sync".

 #include "db/db_impl/db_impl.h"
+#include "db/db_test_util.h"
 #include "db/log_format.h"
 #include "db/version_set.h"
 #include "env/mock_env.h"
@@ -545,76 +546,6 @@ TEST_P(FaultInjectionTest, WriteBatchWalTerminationTest) {
  ASSERT_EQ(db_->Get(ro, "boys", &val), Status::NotFound());
 }

-TEST_P(FaultInjectionTest, NoDuplicateTrailingEntries) {
-  auto fault_fs = std::make_shared<FaultInjectionTestFS>(FileSystem::Default());
-  fault_fs->EnableWriteErrorInjection();
-  fault_fs->SetFilesystemDirectWritable(false);
-  const std::string file_name = NormalizePath(dbname_ + "/test_file");
-  std::unique_ptr<log::Writer> log_writer = nullptr;
-  constexpr uint64_t log_number = 0;
-  {
-    std::unique_ptr<FSWritableFile> file;
-    const Status s =
-        fault_fs->NewWritableFile(file_name, FileOptions(), &file, nullptr);
-    ASSERT_OK(s);
-    std::unique_ptr<WritableFileWriter> fwriter(
-        new WritableFileWriter(std::move(file), file_name, FileOptions()));
-    log_writer.reset(new log::Writer(std::move(fwriter), log_number,
-                                     /*recycle_log_files=*/false));
-  }
-
-  fault_fs->SetRandomWriteError(
-      0xdeadbeef, /*one_in=*/1, IOStatus::IOError("Injected IOError"),
-      /*inject_for_all_file_types=*/true, /*types=*/{});
-
-  {
-    VersionEdit edit;
-    edit.SetColumnFamily(0);
-    std::string buf;
-    assert(edit.EncodeTo(&buf));
-    const Status s = log_writer->AddRecord(WriteOptions(), buf);
-    ASSERT_NOK(s);
-  }
-
-  fault_fs->DisableWriteErrorInjection();
-
-  // Closing the log writer will cause WritableFileWriter::Close() and flush
-  // remaining data from its buffer to underlying file.
-  log_writer.reset();
-
-  {
-    std::unique_ptr<FSSequentialFile> file;
-    Status s =
-        fault_fs->NewSequentialFile(file_name, FileOptions(), &file, nullptr);
-    ASSERT_OK(s);
-    std::unique_ptr<SequentialFileReader> freader(
-        new SequentialFileReader(std::move(file), file_name));
-    Status log_read_s;
-    class LogReporter : public log::Reader::Reporter {
-     public:
-      Status* status_;
-      explicit LogReporter(Status* _s) : status_(_s) {}
-      void Corruption(size_t /*bytes*/, const Status& _s) override {
-        if (status_->ok()) {
-          *status_ = _s;
-        }
-      }
-    } reporter(&log_read_s);
-    std::unique_ptr<log::Reader> log_reader(new log::Reader(
-        nullptr, std::move(freader), &reporter, /*checksum=*/true, log_number));
-    Slice record;
-    std::string data;
-    size_t count = 0;
-    while (log_reader->ReadRecord(&record, &data) && log_read_s.ok()) {
-      VersionEdit edit;
-      ASSERT_OK(edit.DecodeFrom(data));
-      ++count;
-    }
-    // Verify that only one version edit exists in the file.
-    ASSERT_EQ(1, count);
-  }
-}
-
 INSTANTIATE_TEST_CASE_P(
    FaultTest, FaultInjectionTest,
    ::testing::Values(std::make_tuple(false, kDefault, kEnd),
@@ -627,6 +558,59 @@ INSTANTIATE_TEST_CASE_P(
                      std::make_tuple(false, kSyncWal, kEnd),
                      std::make_tuple(true, kSyncWal, kEnd)));

+class FaultInjectionDBTest : public DBTestBase {
+ public:
+  FaultInjectionDBTest()
+      : DBTestBase("fault_injection_fs_test", /*env_do_fsync=*/false) {}
+};
+
+TEST(FaultInjectionFSTest, ReadUnsyncedData) {
+  std::shared_ptr<FaultInjectionTestFS> fault_fs =
+      std::make_shared<FaultInjectionTestFS>(FileSystem::Default());
+  ASSERT_TRUE(fault_fs->ReadUnsyncedData());
+
+  uint32_t len = Random::GetTLSInstance()->Uniform(10000) + 1;
+  Random rnd(len);
+
+  // Create partially synced file
+  std::string f = test::PerThreadDBPath("read_unsynced.data");
+  std::string data = rnd.RandomString(len);
+  {
+    std::unique_ptr<FSWritableFile> w;
+    ASSERT_OK(fault_fs->NewWritableFile(f, {}, &w, nullptr));
+    uint32_t synced_len = rnd.Uniform(len + 1);
+    ASSERT_OK(w->Append(Slice(data.data(), synced_len), {}, nullptr));
+    ASSERT_OK(w->Sync({}, nullptr));
+    ASSERT_OK(w->Append(Slice(data.data() + synced_len, len - synced_len), {},
+                        nullptr));
+    // no sync here
+    ASSERT_OK(w->Close({}, nullptr));
+  }
+  // Test file size includes unsynced data
+  {
+    uint64_t file_size;
+    ASSERT_OK(fault_fs->GetFileSize(f, {}, &file_size, nullptr));
+    ASSERT_EQ(len, file_size);
+  }
+  // Test read file contents, with two reads that probably don't
+  // align with the unsynced split.
+  {
+    std::unique_ptr<FSSequentialFile> r;
+    ASSERT_OK(fault_fs->NewSequentialFile(f, {}, &r, nullptr));
+    uint32_t first_read_len = rnd.Uniform(len + 1);
+    Slice sl;
+    std::unique_ptr<char[]> scratch(new char[first_read_len]);
+    ASSERT_OK(r->Read(first_read_len, {}, &sl, scratch.get(), nullptr));
+    ASSERT_EQ(first_read_len, sl.size());
+    ASSERT_EQ(0, sl.compare(Slice(data.data(), first_read_len)));
+    uint32_t second_read_len = len - first_read_len;
+    scratch.reset(new char[second_read_len]);
+    ASSERT_OK(r->Read(second_read_len, {}, &sl, scratch.get(), nullptr));
+    ASSERT_EQ(second_read_len, sl.size());
+    ASSERT_EQ(0,
+              sl.compare(Slice(data.data() + first_read_len, second_read_len)));
+  }
+}
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
@@ -115,6 +115,9 @@ FlushJob::FlushJob(
      db_mutex_(db_mutex),
      shutting_down_(shutting_down),
      existing_snapshots_(std::move(existing_snapshots)),
+      earliest_snapshot_(existing_snapshots_.empty()
+                             ? kMaxSequenceNumber
+                             : existing_snapshots_.at(0)),
      earliest_write_conflict_snapshot_(earliest_write_conflict_snapshot),
      snapshot_checker_(snapshot_checker),
      job_context_(job_context),
@@ -194,6 +197,7 @@ void FlushJob::PickMemTable() {
  // Track effective cutoff user-defined timestamp during flush if
  // user-defined timestamps can be stripped.
  GetEffectiveCutoffUDTForPickedMemTables();
+  GetPrecludeLastLevelMinSeqno();

  ReportFlushInputSize(mems_);

@@ -502,7 +506,7 @@ Status FlushJob::MemPurge() {
    const std::atomic<bool> kManualCompactionCanceledFalse{false};
    CompactionIterator c_iter(
        iter.get(), (cfd_->internal_comparator()).user_comparator(), &merge,
-        kMaxSequenceNumber, &existing_snapshots_,
+        kMaxSequenceNumber, &existing_snapshots_, earliest_snapshot_,
        earliest_write_conflict_snapshot_, job_snapshot_seq, snapshot_checker_,
        env, ShouldReportDetailedTime(env, ioptions->stats),
        true /* internal key corruption is not ok */, range_del_agg.get(),
@@ -968,14 +972,17 @@ Status FlushJob::WriteLevel0Table() {
          cfd_->GetID(), cfd_->GetName(), 0 /* level */,
          false /* is_bottommost */, TableFileCreationReason::kFlush,
          oldest_key_time, current_time, db_id_, db_session_id_,
-          0 /* target_file_size */, meta_.fd.GetNumber());
+          0 /* target_file_size */, meta_.fd.GetNumber(),
+          preclude_last_level_min_seqno_ == kMaxSequenceNumber
+              ? preclude_last_level_min_seqno_
+              : std::min(earliest_snapshot_, preclude_last_level_min_seqno_));
      const SequenceNumber job_snapshot_seq =
          job_context_->GetJobSnapshotSequence();

      s = BuildTable(
          dbname_, versions_, db_options_, tboptions, file_options_,
          cfd_->table_cache(), iter.get(), std::move(range_del_iters), &meta_,
-          &blob_file_additions, existing_snapshots_,
+          &blob_file_additions, existing_snapshots_, earliest_snapshot_,
          earliest_write_conflict_snapshot_, job_snapshot_seq,
          snapshot_checker_, mutable_cf_options_.paranoid_file_checks,
          cfd_->internal_stats(), &io_s, io_tracer_,
@@ -1154,6 +1161,26 @@ void FlushJob::GetEffectiveCutoffUDTForPickedMemTables() {
  }
 }

+void FlushJob::GetPrecludeLastLevelMinSeqno() {
+  if (cfd_->ioptions()->preclude_last_level_data_seconds == 0) {
+    return;
+  }
+  int64_t current_time = 0;
+  Status s = db_options_.clock->GetCurrentTime(&current_time);
+  if (!s.ok()) {
+    ROCKS_LOG_WARN(db_options_.info_log,
+                   "Failed to get current time in Flush: Status: %s",
+                   s.ToString().c_str());
+  } else {
+    SequenceNumber preserve_time_min_seqno;
+    seqno_to_time_mapping_->GetCurrentTieringCutoffSeqnos(
+        static_cast<uint64_t>(current_time),
+        cfd_->ioptions()->preserve_internal_time_seconds,
+        cfd_->ioptions()->preclude_last_level_data_seconds,
+        &preserve_time_min_seqno, &preclude_last_level_min_seqno_);
+  }
+}
+
 Status FlushJob::MaybeIncreaseFullHistoryTsLowToAboveCutoffUDT() {
  db_mutex_->AssertHeld();
  const auto* ucmp = cfd_->user_comparator();
@@ -143,6 +143,13 @@ class FlushJob {
  // `MaybeIncreaseFullHistoryTsLowToAboveCutoffUDT` for details.
  void GetEffectiveCutoffUDTForPickedMemTables();

+  // If this column family enables tiering feature, it will find the current
+  // `preclude_last_level_min_seqno_`, and the smaller one between this and
+  // the `earliset_snapshot_` will later be announced to user property
+  // collectors. It indicates to tiering use cases which data are old enough to
+  // be placed on the last level.
+  void GetPrecludeLastLevelMinSeqno();
+
  Status MaybeIncreaseFullHistoryTsLowToAboveCutoffUDT();

  const std::string& dbname_;
@@ -161,6 +168,7 @@ class FlushJob {
  InstrumentedMutex* db_mutex_;
  std::atomic<bool>* shutting_down_;
  std::vector<SequenceNumber> existing_snapshots_;
+  SequenceNumber earliest_snapshot_;
  SequenceNumber earliest_write_conflict_snapshot_;
  SnapshotChecker* snapshot_checker_;
  JobContext* job_context_;
@@ -221,6 +229,12 @@ class FlushJob {
  // Keeps track of the newest user-defined timestamp for this flush job if
  // `persist_user_defined_timestamps` flag is false.
  std::string cutoff_udt_;
+
+  // The current minimum seqno that compaction jobs will preclude the data from
+  // the last level. Data with seqnos larger than this or larger than
+  // `earliest_snapshot_` will be output to the penultimate level had it gone
+  // through a compaction to the last level.
+  SequenceNumber preclude_last_level_min_seqno_ = kMaxSequenceNumber;
 };

 }  // namespace ROCKSDB_NAMESPACE
@@ -65,11 +65,11 @@ Status ImportColumnFamilyJob::Prepare(uint64_t next_file_number,
        largest = file_to_import.largest_internal_key;
      } else {
        if (cfd_->internal_comparator().Compare(
-                smallest, file_to_import.smallest_internal_key) < 0) {
+                smallest, file_to_import.smallest_internal_key) > 0) {
          smallest = file_to_import.smallest_internal_key;
        }
        if (cfd_->internal_comparator().Compare(
-                largest, file_to_import.largest_internal_key) > 0) {
+                largest, file_to_import.largest_internal_key) < 0) {
          largest = file_to_import.largest_internal_key;
        }
      }
@@ -175,22 +175,29 @@ Status ImportColumnFamilyJob::Run() {
        static_cast<uint64_t>(temp_current_time);
  }

-  // Recover files' epoch number using dummy VersionStorageInfo
-  VersionBuilder dummy_version_builder(
-      cfd_->current()->version_set()->file_options(), cfd_->ioptions(),
-      cfd_->table_cache(), cfd_->current()->storage_info(),
-      cfd_->current()->version_set(),
-      cfd_->GetFileMetadataCacheReservationManager());
-  VersionStorageInfo dummy_vstorage(
-      &cfd_->internal_comparator(), cfd_->user_comparator(),
-      cfd_->NumberLevels(), cfd_->ioptions()->compaction_style,
-      nullptr /* src_vstorage */, cfd_->ioptions()->force_consistency_checks,
-      EpochNumberRequirement::kMightMissing, cfd_->ioptions()->clock,
-      cfd_->GetLatestMutableCFOptions()->bottommost_file_compaction_delay,
-      cfd_->current()->version_set()->offpeak_time_option());
  Status s;
-
+  // When importing multiple CFs, we should not reuse epoch number from ingested
+  // files. Since these epoch numbers were assigned by different CFs, there may
+  // be different files from different CFs with the same epoch number. With a
+  // subsequent intra-L0 compaction we may end up with files with overlapping
+  // key range but the same epoch number. Here we will create a dummy
+  // VersionStorageInfo per CF being imported. Each CF's files will be assigned
+  // increasing epoch numbers to avoid duplicated epoch number. This is done by
+  // only resetting epoch number of the new CF in the first call to
+  // RecoverEpochNumbers() below.
  for (size_t i = 0; s.ok() && i < files_to_import_.size(); ++i) {
+    VersionBuilder dummy_version_builder(
+        cfd_->current()->version_set()->file_options(), cfd_->ioptions(),
+        cfd_->table_cache(), cfd_->current()->storage_info(),
+        cfd_->current()->version_set(),
+        cfd_->GetFileMetadataCacheReservationManager());
+    VersionStorageInfo dummy_vstorage(
+        &cfd_->internal_comparator(), cfd_->user_comparator(),
+        cfd_->NumberLevels(), cfd_->ioptions()->compaction_style,
+        nullptr /* src_vstorage */, cfd_->ioptions()->force_consistency_checks,
+        EpochNumberRequirement::kMightMissing, cfd_->ioptions()->clock,
+        cfd_->GetLatestMutableCFOptions()->bottommost_file_compaction_delay,
+        cfd_->current()->version_set()->offpeak_time_option());
    for (size_t j = 0; s.ok() && j < files_to_import_[i].size(); ++j) {
      const auto& f = files_to_import_[i][j];
      const auto& file_metadata = *metadatas_[i][j];
@@ -218,42 +225,39 @@ Status ImportColumnFamilyJob::Run() {
              f.table_properties.user_defined_timestamps_persisted));
      s = dummy_version_builder.Apply(&dummy_version_edit);
    }
-  }
-
-  if (s.ok()) {
-    s = dummy_version_builder.SaveTo(&dummy_vstorage);
-  }
-  if (s.ok()) {
-    dummy_vstorage.RecoverEpochNumbers(cfd_);
-  }
-
-  // Record changes from this CF import in VersionEdit, including files with
-  // recovered epoch numbers
-  if (s.ok()) {
-    edit_.SetColumnFamily(cfd_->GetID());
+    if (s.ok()) {
+      s = dummy_version_builder.SaveTo(&dummy_vstorage);
+    }
+    if (s.ok()) {
+      // force resetting epoch number for each file
+      dummy_vstorage.RecoverEpochNumbers(cfd_, /*restart_epoch=*/i == 0,
+                                         /*force=*/true);
+      edit_.SetColumnFamily(cfd_->GetID());

+      for (int level = 0; level < dummy_vstorage.num_levels(); level++) {
+        for (FileMetaData* file_meta : dummy_vstorage.LevelFiles(level)) {
+          edit_.AddFile(level, *file_meta);
+          // If incoming sequence number is higher, update local sequence
+          // number.
+          if (file_meta->fd.largest_seqno > versions_->LastSequence()) {
+            versions_->SetLastAllocatedSequence(file_meta->fd.largest_seqno);
+            versions_->SetLastPublishedSequence(file_meta->fd.largest_seqno);
+            versions_->SetLastSequence(file_meta->fd.largest_seqno);
+          }
+        }
+      }
+    }
+    // Release resources occupied by the dummy VersionStorageInfo
    for (int level = 0; level < dummy_vstorage.num_levels(); level++) {
      for (FileMetaData* file_meta : dummy_vstorage.LevelFiles(level)) {
-        edit_.AddFile(level, *file_meta);
-        // If incoming sequence number is higher, update local sequence number.
-        if (file_meta->fd.largest_seqno > versions_->LastSequence()) {
-          versions_->SetLastAllocatedSequence(file_meta->fd.largest_seqno);
-          versions_->SetLastPublishedSequence(file_meta->fd.largest_seqno);
-          versions_->SetLastSequence(file_meta->fd.largest_seqno);
+        file_meta->refs--;
+        if (file_meta->refs <= 0) {
+          delete file_meta;
        }
      }
    }
  }

-  // Release resources occupied by the dummy VersionStorageInfo
-  for (int level = 0; level < dummy_vstorage.num_levels(); level++) {
-    for (FileMetaData* file_meta : dummy_vstorage.LevelFiles(level)) {
-      file_meta->refs--;
-      if (file_meta->refs <= 0) {
-        delete file_meta;
-      }
-    }
-  }
  return s;
 }

@@ -881,6 +881,173 @@ TEST_F(ImportColumnFamilyTest, ImportMultiColumnFamilyWithOverlap) {
  delete db_copy;
  ASSERT_OK(DestroyDir(env_, dbname_ + "/db_copy"));
 }
+
+TEST_F(ImportColumnFamilyTest, ImportMultiColumnFamilySeveralFilesWithOverlap) {
+  Options options = CurrentOptions();
+  CreateAndReopenWithCF({"koko"}, options);
+
+  SstFileWriter sfw_cf1(EnvOptions(), options, handles_[1]);
+  const std::string file1_sst_name = "file1.sst";
+  const std::string file1_sst = sst_files_dir_ + file1_sst_name;
+  ASSERT_OK(sfw_cf1.Open(file1_sst));
+  ASSERT_OK(sfw_cf1.Put("K1", "V1"));
+  ASSERT_OK(sfw_cf1.Put("K2", "V2"));
+  ASSERT_OK(sfw_cf1.Finish());
+
+  SstFileWriter sfw_cf2(EnvOptions(), options, handles_[1]);
+  const std::string file2_sst_name = "file2.sst";
+  const std::string file2_sst = sst_files_dir_ + file2_sst_name;
+  ASSERT_OK(sfw_cf2.Open(file2_sst));
+  ASSERT_OK(sfw_cf2.Put("K2", "V2"));
+  ASSERT_OK(sfw_cf2.Put("K3", "V3"));
+  ASSERT_OK(sfw_cf2.Finish());
+
+  ColumnFamilyHandle* second_cfh = nullptr;
+  ASSERT_OK(db_->CreateColumnFamily(options, "toto", &second_cfh));
+
+  SstFileWriter sfw_cf3(EnvOptions(), options, second_cfh);
+  const std::string file3_sst_name = "file3.sst";
+  const std::string file3_sst = sst_files_dir_ + file3_sst_name;
+  ASSERT_OK(sfw_cf3.Open(file3_sst));
+  ASSERT_OK(sfw_cf3.Put("K3", "V3"));
+  ASSERT_OK(sfw_cf3.Put("K4", "V4"));
+  ASSERT_OK(sfw_cf3.Finish());
+
+  SstFileWriter sfw_cf4(EnvOptions(), options, second_cfh);
+  const std::string file4_sst_name = "file4.sst";
+  const std::string file4_sst = sst_files_dir_ + file4_sst_name;
+  ASSERT_OK(sfw_cf4.Open(file4_sst));
+  ASSERT_OK(sfw_cf4.Put("K4", "V4"));
+  ASSERT_OK(sfw_cf4.Put("K5", "V5"));
+  ASSERT_OK(sfw_cf4.Finish());
+
+  ExportImportFilesMetaData metadata1, metadata2;
+  metadata1.files.push_back(
+      LiveFileMetaDataInit(file1_sst_name, sst_files_dir_, 1, 1, 2));
+  metadata1.files.push_back(
+      LiveFileMetaDataInit(file2_sst_name, sst_files_dir_, 1, 3, 4));
+  metadata1.db_comparator_name = options.comparator->Name();
+  metadata2.files.push_back(
+      LiveFileMetaDataInit(file3_sst_name, sst_files_dir_, 1, 1, 2));
+  metadata2.files.push_back(
+      LiveFileMetaDataInit(file4_sst_name, sst_files_dir_, 1, 3, 4));
+  metadata2.db_comparator_name = options.comparator->Name();
+
+  std::vector<const ExportImportFilesMetaData*> metadatas{&metadata1,
+                                                          &metadata2};
+
+  ASSERT_EQ(db_->CreateColumnFamilyWithImport(ColumnFamilyOptions(), "yoyo",
+                                              ImportColumnFamilyOptions(),
+                                              metadatas, &import_cfh_),
+            Status::InvalidArgument("CFs have overlapping ranges"));
+  ASSERT_EQ(import_cfh_, nullptr);
+
+  ASSERT_OK(db_->DropColumnFamily(second_cfh));
+  ASSERT_OK(db_->DestroyColumnFamilyHandle(second_cfh));
+}
+
+TEST_F(ImportColumnFamilyTest, AssignEpochNumberToMultipleCF) {
+  // Test ingesting CFs where L0 files could have the same epoch number.
+  Options options = CurrentOptions();
+  options.level_compaction_dynamic_level_bytes = true;
+  options.max_background_jobs = 8;
+  env_->SetBackgroundThreads(2, Env::LOW);
+  env_->SetBackgroundThreads(0, Env::BOTTOM);
+  CreateAndReopenWithCF({"CF1", "CF2"}, options);
+
+  // CF1:
+  // L6: [0, 99], [100, 199]
+  // CF2:
+  // L6: [1000, 1099], [1100, 1199]
+  for (int i = 100; i < 200; ++i) {
+    ASSERT_OK(Put(1, Key(i), Key(i) + "_val"));
+    ASSERT_OK(Put(2, Key(1000 + i), Key(1000 + i) + "_val"));
+  }
+  ASSERT_OK(Flush(1));
+  ASSERT_OK(Flush(2));
+  for (int i = 0; i < 100; ++i) {
+    ASSERT_OK(Put(1, Key(i), Key(i) + "_val"));
+    ASSERT_OK(Put(2, Key(1000 + i), Key(1000 + i) + "_val"));
+  }
+  ASSERT_OK(Flush(1));
+  ASSERT_OK(Flush(2));
+  MoveFilesToLevel(6, 1);
+  MoveFilesToLevel(6, 2);
+
+  // CF1:
+  // level 0 epoch: 5 file num 30 smallest key000010 - key000019
+  // level 0 epoch: 4 file num 27 smallest key000000 - key000009
+  // level 0 epoch: 3 file num 23 smallest key000100 - key000199
+  // level 6 epoch: 2 file num 20 smallest key000000 - key000099
+  // level 6 epoch: 1 file num 17 smallest key000100 - key000199
+  // CF2:
+  // level 0 epoch: 5 file num 31 smallest key001010 - key001019
+  // level 0 epoch: 4 file num 28 smallest key001000 - key001009
+  // level 0 epoch: 3 file num 25 smallest key001020 - key001029
+  // level 6 epoch: 2 file num 21 smallest key001000 - key001099
+  // level 6 epoch: 1 file num 18 smallest key001100 - key001199
+  for (int i = 100; i < 200; ++i) {
+    ASSERT_OK(Put(1, Key(i), Key(i) + "_val"));
+  }
+  ASSERT_OK(Flush(1));
+  for (int i = 20; i < 30; ++i) {
+    ASSERT_OK(Put(2, Key(i + 1000), Key(i + 1000) + "_val"));
+  }
+  ASSERT_OK(Flush(2));
+
+  for (int i = 0; i < 20; ++i) {
+    ASSERT_OK(Put(1, Key(i), Key(i) + "_val"));
+    ASSERT_OK(Put(2, Key(i + 1000), Key(i + 1000) + "_val"));
+    if (i % 10 == 9) {
+      ASSERT_OK(Flush(1));
+      ASSERT_OK(Flush(2));
+    }
+  }
+  ASSERT_OK(Flush(1));
+  ASSERT_OK(Flush(2));
+
+  // Create a CF by importing these two CF1 and CF2.
+  // Then two compactions will be triggerred, one to compact from L0
+  // to L6 (files #23 and #17), and another to do intra-L0 compaction
+  // for the rest of the L0 files. Before a bug fix, we used to
+  // directly use the epoch numbers from the ingested files in the new CF.
+  // This means different files from different CFs can have the same epoch
+  // number. If the intra-L0 compaction finishes first, it can cause a
+  // corruption where two L0 files can have the same epoch number but
+  // with overlapping key range.
+  Checkpoint* checkpoint1;
+  ASSERT_OK(Checkpoint::Create(db_, &checkpoint1));
+  ASSERT_OK(checkpoint1->ExportColumnFamily(handles_[1], export_files_dir_,
+                                            &metadata_ptr_));
+  ASSERT_OK(checkpoint1->ExportColumnFamily(handles_[2], export_files_dir2_,
+                                            &metadata_ptr2_));
+  ASSERT_NE(metadata_ptr_, nullptr);
+  ASSERT_NE(metadata_ptr2_, nullptr);
+
+  std::atomic_int compaction_counter = 0;
+  SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::BackgroundCompaction:NonTrivial:BeforeRun",
+      [&compaction_counter](void*) {
+        compaction_counter++;
+        if (compaction_counter == 1) {
+          // Wait for the next compaction to finish
+          TEST_SYNC_POINT("WaitForSecondCompaction");
+        }
+      });
+  SyncPoint::GetInstance()->LoadDependency(
+      {{"DBImpl::BackgroundCompaction:AfterCompaction",
+        "WaitForSecondCompaction"}});
+  SyncPoint::GetInstance()->EnableProcessing();
+  ImportColumnFamilyOptions import_options;
+  import_options.move_files = false;
+  std::vector<const ExportImportFilesMetaData*> metadatas = {metadata_ptr_,
+                                                             metadata_ptr2_};
+  ASSERT_OK(db_->CreateColumnFamilyWithImport(options, "CF3", import_options,
+                                              metadatas, &import_cfh_));
+  WaitForCompactOptions o;
+  ASSERT_OK(db_->WaitForCompact(o));
+  delete checkpoint1;
+}
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
@@ -16,6 +16,7 @@
 #include <limits>
 #include <sstream>
 #include <string>
+#include <unordered_map>
 #include <utility>
 #include <vector>

@@ -33,7 +34,6 @@

 namespace ROCKSDB_NAMESPACE {

-
 const std::map<LevelStatType, LevelStat> InternalStats::compaction_level_stats =
    {
        {LevelStatType::NUM_FILES, LevelStat{"NumFiles", "Files"}},
@@ -2135,5 +2135,64 @@ void InternalStats::DumpCFFileHistogram(std::string* value) {
  value->append(oss.str());
 }

+namespace {
+
+class SumPropertyAggregator : public IntPropertyAggregator {
+ public:
+  SumPropertyAggregator() : aggregated_value_(0) {}
+  virtual ~SumPropertyAggregator() override = default;
+
+  void Add(ColumnFamilyData* cfd, uint64_t value) override {
+    (void)cfd;
+    aggregated_value_ += value;
+  }
+
+  uint64_t Aggregate() const override { return aggregated_value_; }
+
+ private:
+  uint64_t aggregated_value_;
+};
+
+// A block cache may be shared by multiple column families.
+// BlockCachePropertyAggregator ensures that the same cache is only added once.
+class BlockCachePropertyAggregator : public IntPropertyAggregator {
+ public:
+  BlockCachePropertyAggregator() = default;
+  virtual ~BlockCachePropertyAggregator() override = default;
+
+  void Add(ColumnFamilyData* cfd, uint64_t value) override {
+    auto* table_factory = cfd->ioptions()->table_factory.get();
+    assert(table_factory != nullptr);
+    Cache* cache =
+        table_factory->GetOptions<Cache>(TableFactory::kBlockCacheOpts());
+    if (cache != nullptr) {
+      block_cache_properties_.emplace(cache, value);
+    }
+  }
+
+  uint64_t Aggregate() const override {
+    uint64_t sum = 0;
+    for (const auto& p : block_cache_properties_) {
+      sum += p.second;
+    }
+    return sum;
+  }
+
+ private:
+  std::unordered_map<Cache*, uint64_t> block_cache_properties_;
+};
+
+}  // anonymous namespace
+
+std::unique_ptr<IntPropertyAggregator> CreateIntPropertyAggregator(
+    const Slice& property) {
+  if (property == DB::Properties::kBlockCacheCapacity ||
+      property == DB::Properties::kBlockCacheUsage ||
+      property == DB::Properties::kBlockCachePinnedUsage) {
+    return std::make_unique<BlockCachePropertyAggregator>();
+  } else {
+    return std::make_unique<SumPropertyAggregator>();
+  }
+}

 }  // namespace ROCKSDB_NAMESPACE
@@ -432,7 +432,7 @@ class InternalStats {
    explicit CompactionStatsFull() : stats(), penultimate_level_stats() {}

    explicit CompactionStatsFull(CompactionReason reason, int c)
-        : stats(reason, c), penultimate_level_stats(reason, c){}
+        : stats(reason, c), penultimate_level_stats(reason, c) {}

    uint64_t TotalBytesWritten() const {
      uint64_t bytes_written = stats.bytes_written + stats.bytes_written_blob;
@@ -873,5 +873,24 @@ class InternalStats {
  uint64_t started_at_;
 };

+// IntPropertyAggregator aggregates an integer property across all column
+// families.
+class IntPropertyAggregator {
+ public:
+  IntPropertyAggregator() {}
+  virtual ~IntPropertyAggregator() {}
+
+  IntPropertyAggregator(const IntPropertyAggregator&) = delete;
+  void operator=(const IntPropertyAggregator&) = delete;
+
+  // Add a column family's property value to the aggregator.
+  virtual void Add(ColumnFamilyData* cfd, uint64_t value) = 0;
+
+  // Get the aggregated value.
+  virtual uint64_t Aggregate() const = 0;
+};
+
+std::unique_ptr<IntPropertyAggregator> CreateIntPropertyAggregator(
+    const Slice& property);

 }  // namespace ROCKSDB_NAMESPACE
@@ -232,7 +232,9 @@ bool Reader::ReadRecord(Slice* record, std::string* scratch,
            // produce a hole in the recovered data. Report an error here, which
            // higher layers can choose to ignore when it's provable there is no
            // hole.
-            ReportCorruption(scratch->size(), "error reading trailing data");
+            ReportCorruption(
+                scratch->size(),
+                "error reading trailing data due to encountering EOF");
          }
          // This can be caused by the writer dying immediately after
          //  writing a physical record but before completing the next; don't
@@ -252,7 +254,9 @@ bool Reader::ReadRecord(Slice* record, std::string* scratch,
              // produce a hole in the recovered data. Report an error here,
              // which higher layers can choose to ignore when it's provable
              // there is no hole.
-              ReportCorruption(scratch->size(), "error reading trailing data");
+              ReportCorruption(
+                  scratch->size(),
+                  "error reading trailing data due to encountering old record");
            }
            // This can be caused by the writer dying immediately after
            //  writing a physical record but before completing the next; don't
@@ -483,9 +487,11 @@ unsigned int Reader::ReadPhysicalRecord(Slice* result, size_t* drop_size,
         type == kRecyclableUserDefinedTimestampSizeType);
    if (is_recyclable_type) {
      header_size = kRecyclableHeaderSize;
-      if (end_of_buffer_offset_ - buffer_.size() == 0) {
-        recycled_ = true;
+      if (first_record_read_ && !recycled_) {
+        // A recycled log should have started with a recycled record
+        return kBadRecord;
      }
+      recycled_ = true;
      // We need enough for the larger header
      if (buffer_.size() < static_cast<size_t>(kRecyclableHeaderSize)) {
        int r = kEof;
@@ -863,9 +869,12 @@ bool FragmentBufferedReader::TryReadFragment(
  int header_size = kHeaderSize;
  if ((type >= kRecyclableFullType && type <= kRecyclableLastType) ||
      type == kRecyclableUserDefinedTimestampSizeType) {
-    if (end_of_buffer_offset_ - buffer_.size() == 0) {
-      recycled_ = true;
+    if (first_record_read_ && !recycled_) {
+      // A recycled log should have started with a recycled record
+      *fragment_type_or_err = kBadRecord;
+      return true;
    }
+    recycled_ = true;
    header_size = kRecyclableHeaderSize;
    while (buffer_.size() < static_cast<size_t>(kRecyclableHeaderSize)) {
      size_t old_size = buffer_.size();
@@ -774,6 +774,32 @@ TEST_P(LogTest, RecycleWithTimestampSize) {
  ASSERT_EQ("EOF", Read());
 }

+// Validates that `MaybeAddUserDefinedTimestampSizeRecord`` adds padding to the
+// tail of a block and switches to a new block, if there's not enough space for
+// the record.
+TEST_P(LogTest, TimestampSizeRecordPadding) {
+  bool recyclable_log = (std::get<0>(GetParam()) != 0);
+  const size_t header_size =
+      recyclable_log ? kRecyclableHeaderSize : kHeaderSize;
+  const size_t data_len = kBlockSize - 2 * header_size;
+
+  const auto first_str = BigString("foo", data_len);
+  Write(first_str);
+
+  UnorderedMap<uint32_t, size_t> ts_sz = {
+      {2, sizeof(uint64_t)},
+  };
+  ASSERT_OK(
+      writer_->MaybeAddUserDefinedTimestampSizeRecord(WriteOptions(), ts_sz));
+  ASSERT_LT(writer_->TEST_block_offset(), kBlockSize);
+
+  const auto second_str = BigString("bar", 1000);
+  Write(second_str);
+
+  ASSERT_EQ(first_str, Read());
+  CheckRecordAndTimestampSize(second_str, ts_sz);
+}
+
 // Do NOT enable compression for this instantiation.
 INSTANTIATE_TEST_CASE_P(
    Log, LogTest,
@@ -1131,6 +1157,42 @@ TEST_P(CompressionLogTest, AlignedFragmentation) {
  ASSERT_EQ("EOF", Read());
 }

+TEST_P(CompressionLogTest, ChecksumMismatch) {
+  const CompressionType kCompressionType = std::get<2>(GetParam());
+  const bool kCompressionEnabled = kCompressionType != kNoCompression;
+  const bool kRecyclableLog = (std::get<0>(GetParam()) != 0);
+  if (!StreamingCompressionTypeSupported(kCompressionType)) {
+    ROCKSDB_GTEST_SKIP("Test requires support for compression type");
+    return;
+  }
+  ASSERT_OK(SetupTestEnv());
+
+  Write("foooooo");
+  int header_len;
+  if (kRecyclableLog) {
+    header_len = kRecyclableHeaderSize;
+  } else {
+    header_len = kHeaderSize;
+  }
+  int compression_record_len;
+  if (kCompressionEnabled) {
+    compression_record_len = header_len + 4;
+  } else {
+    compression_record_len = 0;
+  }
+  IncrementByte(compression_record_len + header_len /* offset */,
+                14 /* delta */);
+
+  ASSERT_EQ("EOF", Read());
+  if (!kRecyclableLog) {
+    ASSERT_GT(DroppedBytes(), 0U);
+    ASSERT_EQ("OK", MatchError("checksum mismatch"));
+  } else {
+    ASSERT_EQ(0U, DroppedBytes());
+    ASSERT_EQ("", ReportMessage());
+  }
+}
+
 INSTANTIATE_TEST_CASE_P(
    Compression, CompressionLogTest,
    ::testing::Combine(::testing::Values(0, 1), ::testing::Bool(),
@@ -27,6 +27,8 @@ Writer::Writer(std::unique_ptr<WritableFileWriter>&& dest, uint64_t log_number,
      block_offset_(0),
      log_number_(log_number),
      recycle_log_files_(recycle_log_files),
+      // Header size varies depending on whether we are recycling or not.
+      header_size_(recycle_log_files ? kRecyclableHeaderSize : kHeaderSize),
      manual_flush_(manual_flush),
      compression_type_(compression_type),
      compress_(nullptr) {
@@ -51,6 +53,11 @@ Writer::~Writer() {

 IOStatus Writer::WriteBuffer(const WriteOptions& write_options) {
  if (dest_->seen_error()) {
+#ifndef NDEBUG
+    if (dest_->seen_injected_error()) {
+      return IOStatus::IOError("Seen injected error. Skip writing buffer.");
+    }
+#endif  // NDEBUG
    return IOStatus::IOError("Seen error. Skip writing buffer.");
  }
  IOOptions opts;
@@ -72,18 +79,28 @@ IOStatus Writer::Close(const WriteOptions& write_options) {
  return s;
 }

+bool Writer::PublishIfClosed() {
+  if (dest_->IsClosed()) {
+    dest_.reset();
+    return true;
+  } else {
+    return false;
+  }
+}
+
 IOStatus Writer::AddRecord(const WriteOptions& write_options,
                           const Slice& slice) {
  if (dest_->seen_error()) {
+#ifndef NDEBUG
+    if (dest_->seen_injected_error()) {
+      return IOStatus::IOError("Seen injected error. Skip writing buffer.");
+    }
+#endif  // NDEBUG
    return IOStatus::IOError("Seen error. Skip writing buffer.");
  }
  const char* ptr = slice.data();
  size_t left = slice.size();

-  // Header size varies depending on whether we are recycling or not.
-  const int header_size =
-      recycle_log_files_ ? kRecyclableHeaderSize : kHeaderSize;
-
  // Fragment the record if necessary and emit it.  Note that if slice
  // is empty, we still want to iterate once to emit a single
  // zero-length record
@@ -102,12 +119,12 @@ IOStatus Writer::AddRecord(const WriteOptions& write_options,
    do {
      const int64_t leftover = kBlockSize - block_offset_;
      assert(leftover >= 0);
-      if (leftover < header_size) {
+      if (leftover < header_size_) {
        // Switch to a new block
        if (leftover > 0) {
          // Fill the trailer (literal below relies on kHeaderSize and
          // kRecyclableHeaderSize being <= 11)
-          assert(header_size <= 11);
+          assert(header_size_ <= 11);
          s = dest_->Append(opts,
                            Slice("\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00",
                                  static_cast<size_t>(leftover)),
@@ -120,9 +137,9 @@ IOStatus Writer::AddRecord(const WriteOptions& write_options,
      }

      // Invariant: we never leave < header_size bytes in a block.
-      assert(static_cast<int64_t>(kBlockSize - block_offset_) >= header_size);
+      assert(static_cast<int64_t>(kBlockSize - block_offset_) >= header_size_);

-      const size_t avail = kBlockSize - block_offset_ - header_size;
+      const size_t avail = kBlockSize - block_offset_ - header_size_;

      // Compress the record if compression is enabled.
      // Compress() is called at least once (compress_start=true) and after the
@@ -186,6 +203,11 @@ IOStatus Writer::AddCompressionTypeRecord(const WriteOptions& write_options) {
  }

  if (dest_->seen_error()) {
+#ifndef NDEBUG
+    if (dest_->seen_injected_error()) {
+      return IOStatus::IOError("Seen injected error. Skip writing buffer.");
+    }
+#endif  // NDEBUG
    return IOStatus::IOError("Seen error. Skip writing buffer.");
  }

@@ -203,8 +225,7 @@ IOStatus Writer::AddCompressionTypeRecord(const WriteOptions& write_options) {
      }
    }
    // Initialize fields required for compression
-    const size_t max_output_buffer_len =
-        kBlockSize - (recycle_log_files_ ? kRecyclableHeaderSize : kHeaderSize);
+    const size_t max_output_buffer_len = kBlockSize - header_size_;
    CompressionOptions opts;
    constexpr uint32_t compression_format_version = 2;
    compress_ = StreamingCompress::Create(compression_type_, opts,
@@ -244,6 +265,25 @@ IOStatus Writer::MaybeAddUserDefinedTimestampSizeRecord(
  record.EncodeTo(&encoded);
  RecordType type = recycle_log_files_ ? kRecyclableUserDefinedTimestampSizeType
                                       : kUserDefinedTimestampSizeType;
+
+  // If there's not enough space for this record, switch to a new block.
+  const int64_t leftover = kBlockSize - block_offset_;
+  if (leftover < header_size_ + (int)encoded.size()) {
+    IOOptions opts;
+    IOStatus s = WritableFileWriter::PrepareIOOptions(write_options, opts);
+    if (!s.ok()) {
+      return s;
+    }
+
+    std::vector<char> trailer(leftover, '\x00');
+    s = dest_->Append(opts, Slice(trailer.data(), trailer.size()));
+    if (!s.ok()) {
+      return s;
+    }
+
+    block_offset_ = 0;
+  }
+
  return EmitPhysicalRecord(write_options, type, encoded.data(),
                            encoded.size());
 }
@@ -107,13 +107,21 @@ class Writer {

  IOStatus Close(const WriteOptions& write_options);

+  // If closing the writer through file(), call this afterwards to modify
+  // this object's state to reflect that. Returns true if the destination file
+  // has been closed. If it hasn't been closed, returns false with no change.
+  bool PublishIfClosed();
+
  bool BufferIsEmpty();

+  size_t TEST_block_offset() const { return block_offset_; }
+
 private:
  std::unique_ptr<WritableFileWriter> dest_;
  size_t block_offset_;  // Current offset in block
  uint64_t log_number_;
  bool recycle_log_files_;
+  int header_size_;

  // crc32c values for all supported record types.  These are
  // pre-computed to reduce the overhead of computing the crc of the
@@ -13,6 +13,7 @@
 #include <array>
 #include <limits>
 #include <memory>
+#include <optional>

 #include "db/dbformat.h"
 #include "db/kv_checksum.h"
@@ -617,8 +618,9 @@ FragmentedRangeTombstoneIterator* MemTable::NewRangeTombstoneIteratorInternal(
 }

 void MemTable::ConstructFragmentedRangeTombstones() {
-  assert(!IsFragmentedRangeTombstonesConstructed(false));
-  // There should be no concurrent Construction
+  // There should be no concurrent Construction.
+  // We could also check fragmented_range_tombstone_list_ to avoid repeate
+  // constructions. We just construct them here again to be safe.
  if (!is_range_del_table_empty_.load(std::memory_order_relaxed)) {
    // TODO: plumb Env::IOActivity, Env::IOPriority
    auto* unfragmented_iter = new MemTableIterator(
@@ -765,8 +767,9 @@ Status MemTable::Add(SequenceNumber s, ValueType type,
  Slice key_without_ts = StripTimestampFromUserKey(key, ts_sz_);

  if (!allow_concurrent) {
-    // Extract prefix for insert with hint.
-    if (insert_with_hint_prefix_extractor_ != nullptr &&
+    // Extract prefix for insert with hint. Hints are for point key table
+    // (`table_`) only, not `range_del_table_`.
+    if (table == table_ && insert_with_hint_prefix_extractor_ != nullptr &&
        insert_with_hint_prefix_extractor_->InDomain(key_slice)) {
      Slice prefix = insert_with_hint_prefix_extractor_->Transform(key_slice);
      bool res = table->InsertKeyWithHint(handle, &insert_hints_[prefix]);
@@ -927,21 +930,10 @@ struct Saver {
 }  // anonymous namespace

 static bool SaveValue(void* arg, const char* entry) {
-  TEST_SYNC_POINT_CALLBACK("Memtable::SaveValue:Begin:entry", &entry);
  Saver* s = static_cast<Saver*>(arg);
  assert(s != nullptr);
  assert(!s->value || !s->columns);

-  if (s->protection_bytes_per_key > 0) {
-    *(s->status) = MemTable::VerifyEntryChecksum(
-        entry, s->protection_bytes_per_key, s->allow_data_in_errors);
-    if (!s->status->ok()) {
-      ROCKS_LOG_ERROR(s->logger, "In SaveValue: %s", s->status->getState());
-      // Memtable entry corrupted
-      return false;
-    }
-  }
-
  MergeContext* merge_context = s->merge_context;
  SequenceNumber max_covering_tombstone_seq = s->max_covering_tombstone_seq;
  const MergeOperator* merge_operator = s->merge_operator;
@@ -964,6 +956,22 @@ static bool SaveValue(void* arg, const char* entry) {
  if (user_comparator->EqualWithoutTimestamp(user_key_slice,
                                             s->key->user_key())) {
    // Correct user key
+    TEST_SYNC_POINT_CALLBACK("Memtable::SaveValue:Found:entry", &entry);
+    std::optional<ReadLock> read_lock;
+    if (s->inplace_update_support) {
+      read_lock.emplace(s->mem->GetLock(s->key->user_key()));
+    }
+
+    if (s->protection_bytes_per_key > 0) {
+      *(s->status) = MemTable::VerifyEntryChecksum(
+          entry, s->protection_bytes_per_key, s->allow_data_in_errors);
+      if (!s->status->ok()) {
+        ROCKS_LOG_ERROR(s->logger, "In SaveValue: %s", s->status->getState());
+        // Memtable entry corrupted
+        return false;
+      }
+    }
+
    const uint64_t tag = DecodeFixed64(key_ptr + key_length - 8);
    ValueType type;
    SequenceNumber seq;
@@ -1034,10 +1042,6 @@ static bool SaveValue(void* arg, const char* entry) {
          return false;
        }

-        if (s->inplace_update_support) {
-          s->mem->GetLock(s->key->user_key())->ReadLock();
-        }
-
        Slice v = GetLengthPrefixedSlice(key_ptr + key_length);

        *(s->status) = Status::OK();
@@ -1048,10 +1052,6 @@ static bool SaveValue(void* arg, const char* entry) {
          s->columns->SetPlainValue(v);
        }

-        if (s->inplace_update_support) {
-          s->mem->GetLock(s->key->user_key())->ReadUnlock();
-        }
-
        *(s->found_final_value) = true;
        *(s->is_blob_index) = true;

@@ -1059,10 +1059,6 @@ static bool SaveValue(void* arg, const char* entry) {
      }
      case kTypeValue:
      case kTypeValuePreferredSeqno: {
-        if (s->inplace_update_support) {
-          s->mem->GetLock(s->key->user_key())->ReadLock();
-        }
-
        Slice v = GetLengthPrefixedSlice(key_ptr + key_length);

        if (type == kTypeValuePreferredSeqno) {
@@ -1099,10 +1095,6 @@ static bool SaveValue(void* arg, const char* entry) {
          s->columns->SetPlainValue(v);
        }

-        if (s->inplace_update_support) {
-          s->mem->GetLock(s->key->user_key())->ReadUnlock();
-        }
-
        *(s->found_final_value) = true;

        if (s->is_blob_index != nullptr) {
@@ -1112,10 +1104,6 @@ static bool SaveValue(void* arg, const char* entry) {
        return false;
      }
      case kTypeWideColumnEntity: {
-        if (s->inplace_update_support) {
-          s->mem->GetLock(s->key->user_key())->ReadLock();
-        }
-
        Slice v = GetLengthPrefixedSlice(key_ptr + key_length);

        *(s->status) = Status::OK();
@@ -1157,10 +1145,6 @@ static bool SaveValue(void* arg, const char* entry) {
          *(s->status) = s->columns->SetWideColumnValue(v);
        }

-        if (s->inplace_update_support) {
-          s->mem->GetLock(s->key->user_key())->ReadUnlock();
-        }
-
        *(s->found_final_value) = true;

        if (s->is_blob_index != nullptr) {
@@ -1498,9 +1482,9 @@ Status MemTable::Update(SequenceNumber seq, ValueType value_type,

        // Update value, if new value size  <= previous value size
        if (new_size <= prev_size) {
+          WriteLock wl(GetLock(lkey.user_key()));
          char* p =
              EncodeVarint32(const_cast<char*>(key_ptr) + key_length, new_size);
-          WriteLock wl(GetLock(lkey.user_key()));
          memcpy(p, value.data(), value.size());
          assert((unsigned)((p + value.size()) - entry) ==
                 (unsigned)(VarintLength(key_length) + key_length +
@@ -1621,7 +1605,8 @@ Status MemTable::UpdateCallback(SequenceNumber seq, const Slice& key,
  return Status::NotFound();
 }

-size_t MemTable::CountSuccessiveMergeEntries(const LookupKey& key) {
+size_t MemTable::CountSuccessiveMergeEntries(const LookupKey& key,
+                                             size_t limit) {
  Slice memkey = key.memtable_key();

  // A total ordered iterator is costly for some memtablerep (prefix aware
@@ -1633,7 +1618,7 @@ size_t MemTable::CountSuccessiveMergeEntries(const LookupKey& key) {

  size_t num_successive_merges = 0;

-  for (; iter->Valid(); iter->Next()) {
+  for (; iter->Valid() && num_successive_merges < limit; iter->Next()) {
    const char* entry = iter->key();
    uint32_t key_length = 0;
    const char* iter_key_ptr = GetVarint32Ptr(entry, entry + 5, &key_length);
@@ -326,9 +326,10 @@ class MemTable {
                        const ProtectionInfoKVOS64* kv_prot_info);

  // Returns the number of successive merge entries starting from the newest
-  // entry for the key up to the last non-merge entry or last entry for the
-  // key in the memtable.
-  size_t CountSuccessiveMergeEntries(const LookupKey& key);
+  // entry for the key. The count ends when the oldest entry in the memtable
+  // with which the newest entry would be merged is reached, or the count
+  // reaches `limit`.
+  size_t CountSuccessiveMergeEntries(const LookupKey& key, size_t limit);

  // Update counters and flush status after inserting a whole write batch
  // Used in concurrent memtable inserts.
@@ -533,21 +534,21 @@ class MemTable {
  // Returns a heuristic flush decision
  bool ShouldFlushNow();

+  // Updates `fragmented_range_tombstone_list_` that will be used to serve reads
+  // when this memtable becomes an immutable memtable (in some
+  // MemtableListVersion::memlist_). Should be called when this memtable is
+  // about to become immutable. May be called multiple times since
+  // SwitchMemtable() may fail.
  void ConstructFragmentedRangeTombstones();

  // Returns whether a fragmented range tombstone list is already constructed
  // for this memtable. It should be constructed right before a memtable is
  // added to an immutable memtable list. Note that if a memtable does not have
-  // any range tombstone, then no range tombstone list will ever be constructed.
-  // @param allow_empty Specifies whether a memtable with no range tombstone is
-  // considered to have its fragmented range tombstone list constructed.
-  bool IsFragmentedRangeTombstonesConstructed(bool allow_empty = true) const {
-    if (allow_empty) {
-      return fragmented_range_tombstone_list_.get() != nullptr ||
-             is_range_del_table_empty_;
-    } else {
-      return fragmented_range_tombstone_list_.get() != nullptr;
-    }
+  // any range tombstone, then no range tombstone list will ever be constructed
+  // and true is returned in that case.
+  bool IsFragmentedRangeTombstonesConstructed() const {
+    return fragmented_range_tombstone_list_.get() != nullptr ||
+           is_range_del_table_empty_;
  }

  // Get the newest user-defined timestamp contained in this MemTable. Check
@@ -1,102 +0,0 @@
-//  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/multi_cf_iterator.h"
-
-#include <cassert>
-
-namespace ROCKSDB_NAMESPACE {
-
-template <typename BinaryHeap, typename ChildSeekFuncType>
-void MultiCfIterator::SeekCommon(BinaryHeap& heap,
-                                 ChildSeekFuncType child_seek_func) {
-  heap.clear();
-  int i = 0;
-  for (auto& cfh_iter_pair : cfh_iter_pairs_) {
-    auto& cfh = cfh_iter_pair.first;
-    auto& iter = cfh_iter_pair.second;
-    child_seek_func(iter.get());
-    if (iter->Valid()) {
-      assert(iter->status().ok());
-      heap.push(MultiCfIteratorInfo{iter.get(), cfh, i});
-    } else {
-      considerStatus(iter->status());
-    }
-    ++i;
-  }
-}
-
-template <typename BinaryHeap, typename AdvanceFuncType>
-void MultiCfIterator::AdvanceIterator(BinaryHeap& heap,
-                                      AdvanceFuncType advance_func) {
-  // 1. Keep the top iterator (by popping it from the heap)
-  // 2. Make sure all others have iterated past the top iterator key slice
-  // 3. Advance the top iterator, and add it back to the heap if valid
-  auto top = heap.top();
-  heap.pop();
-  if (!heap.empty()) {
-    auto* current = heap.top().iterator;
-    while (current->Valid() &&
-           comparator_->Compare(top.iterator->key(), current->key()) == 0) {
-      assert(current->status().ok());
-      advance_func(current);
-      if (current->Valid()) {
-        heap.replace_top(heap.top());
-      } else {
-        considerStatus(current->status());
-        heap.pop();
-      }
-      if (!heap.empty()) {
-        current = heap.top().iterator;
-      }
-    }
-  }
-  advance_func(top.iterator);
-  if (top.iterator->Valid()) {
-    assert(top.iterator->status().ok());
-    heap.push(top);
-  } else {
-    considerStatus(top.iterator->status());
-  }
-}
-
-void MultiCfIterator::SeekToFirst() {
-  auto& min_heap = GetHeap<MultiCfMinHeap>([this]() { InitMinHeap(); });
-  SeekCommon(min_heap, [](Iterator* iter) { iter->SeekToFirst(); });
-}
-void MultiCfIterator::Seek(const Slice& target) {
-  auto& min_heap = GetHeap<MultiCfMinHeap>([this]() { InitMinHeap(); });
-  SeekCommon(min_heap, [&target](Iterator* iter) { iter->Seek(target); });
-}
-void MultiCfIterator::SeekToLast() {
-  auto& max_heap = GetHeap<MultiCfMaxHeap>([this]() { InitMaxHeap(); });
-  SeekCommon(max_heap, [](Iterator* iter) { iter->SeekToLast(); });
-}
-void MultiCfIterator::SeekForPrev(const Slice& target) {
-  auto& max_heap = GetHeap<MultiCfMaxHeap>([this]() { InitMaxHeap(); });
-  SeekCommon(max_heap,
-             [&target](Iterator* iter) { iter->SeekForPrev(target); });
-}
-
-void MultiCfIterator::Next() {
-  assert(Valid());
-  auto& min_heap = GetHeap<MultiCfMinHeap>([this]() {
-    Slice target = key();
-    InitMinHeap();
-    Seek(target);
-  });
-  AdvanceIterator(min_heap, [](Iterator* iter) { iter->Next(); });
-}
-void MultiCfIterator::Prev() {
-  assert(Valid());
-  auto& max_heap = GetHeap<MultiCfMaxHeap>([this]() {
-    Slice target = key();
-    InitMaxHeap();
-    SeekForPrev(target);
-  });
-  AdvanceIterator(max_heap, [](Iterator* iter) { iter->Prev(); });
-}
-
-}  // namespace ROCKSDB_NAMESPACE
@@ -1,159 +0,0 @@
-//  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).
-
-#pragma once
-
-#include <variant>
-
-#include "rocksdb/comparator.h"
-#include "rocksdb/iterator.h"
-#include "rocksdb/options.h"
-#include "util/heap.h"
-#include "util/overload.h"
-
-namespace ROCKSDB_NAMESPACE {
-
-// UNDER CONSTRUCTION - DO NOT USE
-// A cross-column-family iterator from a consistent database state.
-// When the same key exists in more than one column families, the iterator
-// selects the value from the first column family containing the key, in the
-// order provided in the `column_families` parameter.
-class MultiCfIterator : public Iterator {
- public:
-  MultiCfIterator(const Comparator* comparator,
-                  const std::vector<ColumnFamilyHandle*>& column_families,
-                  const std::vector<Iterator*>& child_iterators)
-      : comparator_(comparator),
-        heap_(MultiCfMinHeap(
-            MultiCfHeapItemComparator<std::greater<int>>(comparator_))) {
-    assert(column_families.size() > 0 &&
-           column_families.size() == child_iterators.size());
-    cfh_iter_pairs_.reserve(column_families.size());
-    for (size_t i = 0; i < column_families.size(); ++i) {
-      cfh_iter_pairs_.emplace_back(
-          column_families[i], std::unique_ptr<Iterator>(child_iterators[i]));
-    }
-  }
-  ~MultiCfIterator() override { status_.PermitUncheckedError(); }
-
-  // No copy allowed
-  MultiCfIterator(const MultiCfIterator&) = delete;
-  MultiCfIterator& operator=(const MultiCfIterator&) = delete;
-
- private:
-  std::vector<std::pair<ColumnFamilyHandle*, std::unique_ptr<Iterator>>>
-      cfh_iter_pairs_;
-  ReadOptions read_options_;
-  Status status_;
-
-  AttributeGroups attribute_groups_;
-
-  struct MultiCfIteratorInfo {
-    Iterator* iterator;
-    ColumnFamilyHandle* cfh;
-    int order;
-  };
-
-  template <typename CompareOp>
-  class MultiCfHeapItemComparator {
-   public:
-    explicit MultiCfHeapItemComparator(const Comparator* comparator)
-        : comparator_(comparator) {}
-    bool operator()(const MultiCfIteratorInfo& a,
-                    const MultiCfIteratorInfo& b) const {
-      assert(a.iterator);
-      assert(b.iterator);
-      assert(a.iterator->Valid());
-      assert(b.iterator->Valid());
-      int c = comparator_->Compare(a.iterator->key(), b.iterator->key());
-      assert(c != 0 || a.order != b.order);
-      return c == 0 ? a.order - b.order > 0 : CompareOp()(c, 0);
-    }
-
-   private:
-    const Comparator* comparator_;
-  };
-  const Comparator* comparator_;
-  using MultiCfMinHeap =
-      BinaryHeap<MultiCfIteratorInfo,
-                 MultiCfHeapItemComparator<std::greater<int>>>;
-  using MultiCfMaxHeap = BinaryHeap<MultiCfIteratorInfo,
-                                    MultiCfHeapItemComparator<std::less<int>>>;
-
-  using MultiCfIterHeap = std::variant<MultiCfMinHeap, MultiCfMaxHeap>;
-
-  MultiCfIterHeap heap_;
-
-  // TODO: Lower and Upper bounds
-
-  Iterator* current() const {
-    if (std::holds_alternative<MultiCfMaxHeap>(heap_)) {
-      auto& max_heap = std::get<MultiCfMaxHeap>(heap_);
-      return max_heap.top().iterator;
-    }
-    auto& min_heap = std::get<MultiCfMinHeap>(heap_);
-    return min_heap.top().iterator;
-  }
-
-  Slice key() const override {
-    assert(Valid());
-    return current()->key();
-  }
-  Slice value() const override {
-    assert(Valid());
-    return current()->value();
-  }
-  const WideColumns& columns() const override {
-    assert(Valid());
-    return current()->columns();
-  }
-
-  bool Valid() const override {
-    if (std::holds_alternative<MultiCfMaxHeap>(heap_)) {
-      auto& max_heap = std::get<MultiCfMaxHeap>(heap_);
-      return !max_heap.empty() && status_.ok();
-    }
-    auto& min_heap = std::get<MultiCfMinHeap>(heap_);
-    return !min_heap.empty() && status_.ok();
-  }
-
-  Status status() const override { return status_; }
-  void considerStatus(Status s) {
-    if (!s.ok() && status_.ok()) {
-      status_ = std::move(s);
-    }
-  }
-
-  template <typename HeapType, typename InitFunc>
-  HeapType& GetHeap(InitFunc initFunc) {
-    if (!std::holds_alternative<HeapType>(heap_)) {
-      initFunc();
-    }
-    return std::get<HeapType>(heap_);
-  }
-
-  void InitMinHeap() {
-    heap_.emplace<MultiCfMinHeap>(
-        MultiCfHeapItemComparator<std::greater<int>>(comparator_));
-  }
-  void InitMaxHeap() {
-    heap_.emplace<MultiCfMaxHeap>(
-        MultiCfHeapItemComparator<std::less<int>>(comparator_));
-  }
-
-  template <typename BinaryHeap, typename ChildSeekFuncType>
-  void SeekCommon(BinaryHeap& heap, ChildSeekFuncType child_seek_func);
-  template <typename BinaryHeap, typename AdvanceFuncType>
-  void AdvanceIterator(BinaryHeap& heap, AdvanceFuncType advance_func);
-
-  void SeekToFirst() override;
-  void SeekToLast() override;
-  void Seek(const Slice& /*target*/) override;
-  void SeekForPrev(const Slice& /*target*/) override;
-  void Next() override;
-  void Prev() override;
-};
-
-}  // namespace ROCKSDB_NAMESPACE
@@ -0,0 +1,296 @@
+//  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).
+
+#pragma once
+
+#include <functional>
+#include <variant>
+
+#include "rocksdb/comparator.h"
+#include "rocksdb/iterator.h"
+#include "rocksdb/options.h"
+#include "util/heap.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+struct MultiCfIteratorInfo {
+  ColumnFamilyHandle* cfh;
+  Iterator* iterator;
+  int order;
+};
+
+class MultiCfIteratorImpl {
+ public:
+  MultiCfIteratorImpl(
+      const Comparator* comparator,
+      const std::vector<ColumnFamilyHandle*>& column_families,
+      const std::vector<Iterator*>& child_iterators,
+      std::function<void()> reset_func,
+      std::function<void(const autovector<MultiCfIteratorInfo>&)> populate_func)
+      : comparator_(comparator),
+        heap_(MultiCfMinHeap(
+            MultiCfHeapItemComparator<std::greater<int>>(comparator_))),
+        reset_func_(std::move(reset_func)),
+        populate_func_(std::move(populate_func)) {
+    assert(column_families.size() > 0 &&
+           column_families.size() == child_iterators.size());
+    cfh_iter_pairs_.reserve(column_families.size());
+    for (size_t i = 0; i < column_families.size(); ++i) {
+      cfh_iter_pairs_.emplace_back(
+          column_families[i], std::unique_ptr<Iterator>(child_iterators[i]));
+    }
+  }
+  ~MultiCfIteratorImpl() { status_.PermitUncheckedError(); }
+
+  // No copy allowed
+  MultiCfIteratorImpl(const MultiCfIteratorImpl&) = delete;
+  MultiCfIteratorImpl& operator=(const MultiCfIteratorImpl&) = delete;
+
+  Slice key() const {
+    assert(Valid());
+    return current()->key();
+  }
+
+  bool Valid() const {
+    if (std::holds_alternative<MultiCfMaxHeap>(heap_)) {
+      auto& max_heap = std::get<MultiCfMaxHeap>(heap_);
+      return !max_heap.empty() && status_.ok();
+    }
+    auto& min_heap = std::get<MultiCfMinHeap>(heap_);
+    return !min_heap.empty() && status_.ok();
+  }
+
+  Status status() const { return status_; }
+
+  void SeekToFirst() {
+    auto& min_heap = GetHeap<MultiCfMinHeap>([this]() { InitMinHeap(); });
+    SeekCommon(min_heap, [](Iterator* iter) { iter->SeekToFirst(); });
+  }
+  void Seek(const Slice& target) {
+    auto& min_heap = GetHeap<MultiCfMinHeap>([this]() { InitMinHeap(); });
+    SeekCommon(min_heap, [&target](Iterator* iter) { iter->Seek(target); });
+  }
+  void SeekToLast() {
+    auto& max_heap = GetHeap<MultiCfMaxHeap>([this]() { InitMaxHeap(); });
+    SeekCommon(max_heap, [](Iterator* iter) { iter->SeekToLast(); });
+  }
+  void SeekForPrev(const Slice& target) {
+    auto& max_heap = GetHeap<MultiCfMaxHeap>([this]() { InitMaxHeap(); });
+    SeekCommon(max_heap,
+               [&target](Iterator* iter) { iter->SeekForPrev(target); });
+  }
+
+  void Next() {
+    assert(Valid());
+    auto& min_heap = GetHeap<MultiCfMinHeap>([this]() {
+      std::string target(key().data(), key().size());
+      InitMinHeap();
+      Seek(target);
+    });
+    AdvanceIterator(min_heap, [](Iterator* iter) { iter->Next(); });
+  }
+  void Prev() {
+    assert(Valid());
+    auto& max_heap = GetHeap<MultiCfMaxHeap>([this]() {
+      std::string target(key().data(), key().size());
+      InitMaxHeap();
+      SeekForPrev(target);
+    });
+    AdvanceIterator(max_heap, [](Iterator* iter) { iter->Prev(); });
+  }
+
+ private:
+  std::vector<std::pair<ColumnFamilyHandle*, std::unique_ptr<Iterator>>>
+      cfh_iter_pairs_;
+  Status status_;
+
+  template <typename CompareOp>
+  class MultiCfHeapItemComparator {
+   public:
+    explicit MultiCfHeapItemComparator(const Comparator* comparator)
+        : comparator_(comparator) {}
+    bool operator()(const MultiCfIteratorInfo& a,
+                    const MultiCfIteratorInfo& b) const {
+      assert(a.iterator);
+      assert(b.iterator);
+      assert(a.iterator->Valid());
+      assert(b.iterator->Valid());
+      int c = comparator_->Compare(a.iterator->key(), b.iterator->key());
+      assert(c != 0 || a.order != b.order);
+      return c == 0 ? a.order - b.order > 0 : CompareOp()(c, 0);
+    }
+
+   private:
+    const Comparator* comparator_;
+  };
+  const Comparator* comparator_;
+  using MultiCfMinHeap =
+      BinaryHeap<MultiCfIteratorInfo,
+                 MultiCfHeapItemComparator<std::greater<int>>>;
+  using MultiCfMaxHeap = BinaryHeap<MultiCfIteratorInfo,
+                                    MultiCfHeapItemComparator<std::less<int>>>;
+
+  using MultiCfIterHeap = std::variant<MultiCfMinHeap, MultiCfMaxHeap>;
+
+  MultiCfIterHeap heap_;
+
+  std::function<void()> reset_func_;
+  std::function<void(autovector<MultiCfIteratorInfo>)> populate_func_;
+
+  Iterator* current() const {
+    if (std::holds_alternative<MultiCfMaxHeap>(heap_)) {
+      auto& max_heap = std::get<MultiCfMaxHeap>(heap_);
+      return max_heap.top().iterator;
+    }
+    auto& min_heap = std::get<MultiCfMinHeap>(heap_);
+    return min_heap.top().iterator;
+  }
+
+  void considerStatus(Status s) {
+    if (!s.ok() && status_.ok()) {
+      status_ = std::move(s);
+    }
+  }
+
+  template <typename HeapType, typename InitFunc>
+  HeapType& GetHeap(InitFunc initFunc) {
+    if (!std::holds_alternative<HeapType>(heap_)) {
+      initFunc();
+    }
+    return std::get<HeapType>(heap_);
+  }
+
+  void InitMinHeap() {
+    heap_.emplace<MultiCfMinHeap>(
+        MultiCfHeapItemComparator<std::greater<int>>(comparator_));
+  }
+  void InitMaxHeap() {
+    heap_.emplace<MultiCfMaxHeap>(
+        MultiCfHeapItemComparator<std::less<int>>(comparator_));
+  }
+
+  template <typename BinaryHeap, typename ChildSeekFuncType>
+  void SeekCommon(BinaryHeap& heap, ChildSeekFuncType child_seek_func) {
+    reset_func_();
+    heap.clear();
+    int i = 0;
+    for (auto& [cfh, iter] : cfh_iter_pairs_) {
+      child_seek_func(iter.get());
+      if (iter->Valid()) {
+        assert(iter->status().ok());
+        heap.push(MultiCfIteratorInfo{cfh, iter.get(), i});
+      } else {
+        considerStatus(iter->status());
+        if (!status_.ok()) {
+          // Non-OK status from the iterator. Bail out early
+          heap.clear();
+          break;
+        }
+      }
+      ++i;
+    }
+    if (!heap.empty()) {
+      PopulateIterator(heap);
+    }
+  }
+
+  template <typename BinaryHeap, typename AdvanceFuncType>
+  void AdvanceIterator(BinaryHeap& heap, AdvanceFuncType advance_func) {
+    reset_func_();
+    // It is possible for one or more child iters are at invalid keys due to
+    // manual prefix iteration. For such cases, we consider the result of the
+    // multi-cf-iter is also undefined.
+    // https://github.com/facebook/rocksdb/wiki/Prefix-Seek#manual-prefix-iterating
+    // for details about manual prefix iteration
+    if (heap.empty()) {
+      return;
+    }
+
+    // 1. Keep the top iterator (by popping it from the heap)
+    // 2. Make sure all others have iterated past the top iterator key slice
+    // 3. Advance the top iterator, and add it back to the heap if valid
+    auto top = heap.top();
+    heap.pop();
+    if (!heap.empty()) {
+      auto current = heap.top();
+      assert(current.iterator);
+      while (current.iterator->Valid() &&
+             comparator_->Compare(top.iterator->key(),
+                                  current.iterator->key()) == 0) {
+        assert(current.iterator->status().ok());
+        advance_func(current.iterator);
+        if (current.iterator->Valid()) {
+          heap.replace_top(heap.top());
+        } else {
+          considerStatus(current.iterator->status());
+          if (!status_.ok()) {
+            heap.clear();
+            return;
+          } else {
+            heap.pop();
+          }
+        }
+        if (!heap.empty()) {
+          current = heap.top();
+        }
+      }
+    }
+    advance_func(top.iterator);
+    if (top.iterator->Valid()) {
+      assert(top.iterator->status().ok());
+      heap.push(top);
+    } else {
+      considerStatus(top.iterator->status());
+      if (!status_.ok()) {
+        heap.clear();
+        return;
+      }
+    }
+
+    if (!heap.empty()) {
+      PopulateIterator(heap);
+    }
+  }
+
+  template <typename BinaryHeap>
+  void PopulateIterator(BinaryHeap& heap) {
+    // 1. Keep the top iterator (by popping it from the heap) and add it to list
+    //    to populate
+    // 2. For all non-top iterators having the same key as top iter popped
+    //    from the previous step, add them to the same list and pop it
+    //    temporarily from the heap
+    // 3. Once no other iters have the same key as the top iter from step 1,
+    //    populate the value/columns and attribute_groups from the list
+    //    collected in step 1 and 2 and add all the iters back to the heap
+    assert(!heap.empty());
+    auto top = heap.top();
+    heap.pop();
+    autovector<MultiCfIteratorInfo> to_populate;
+    to_populate.push_back(top);
+    if (!heap.empty()) {
+      auto current = heap.top();
+      assert(current.iterator);
+      while (current.iterator->Valid() &&
+             comparator_->Compare(top.iterator->key(),
+                                  current.iterator->key()) == 0) {
+        assert(current.iterator->status().ok());
+        to_populate.push_back(current);
+        heap.pop();
+        if (!heap.empty()) {
+          current = heap.top();
+        } else {
+          break;
+        }
+      }
+    }
+    // Add the items back to the heap
+    for (auto& item : to_populate) {
+      heap.push(item);
+    }
+    populate_func_(to_populate);
+  }
+};
+
+}  // namespace ROCKSDB_NAMESPACE
@@ -3,43 +3,37 @@
 //  COPYING file in the root directory) and Apache 2.0 License
 //  (found in the LICENSE.Apache file in the root directory).

-#include <memory>
-
 #include "db/db_test_util.h"
+#include "rocksdb/attribute_groups.h"

 namespace ROCKSDB_NAMESPACE {

-class MultiCfIteratorTest : public DBTestBase {
+class CoalescingIteratorTest : public DBTestBase {
 public:
-  MultiCfIteratorTest()
-      : DBTestBase("multi_cf_iterator_test", /*env_do_fsync=*/true) {}
+  CoalescingIteratorTest()
+      : DBTestBase("coalescing_iterator_test", /*env_do_fsync=*/true) {}

-  // Verify Iteration of MultiCfIterator
+  // Verify Iteration of CoalescingIterator
  // by SeekToFirst() + Next() and SeekToLast() + Prev()
-  void verifyMultiCfIterator(
-      const std::vector<ColumnFamilyHandle*>& cfhs,
-      const std::vector<Slice>& expected_keys,
-      const std::optional<std::vector<Slice>>& expected_values = std::nullopt,
-      const std::optional<std::vector<WideColumns>>& expected_wide_columns =
-          std::nullopt,
-      const std::optional<std::vector<AttributeGroups>>&
-          expected_attribute_groups = std::nullopt) {
+  void verifyCoalescingIterator(const std::vector<ColumnFamilyHandle*>& cfhs,
+                                const std::vector<Slice>& expected_keys,
+                                const std::vector<Slice>& expected_values,
+                                const std::optional<std::vector<WideColumns>>&
+                                    expected_wide_columns = std::nullopt,
+                                const Slice* lower_bound = nullptr,
+                                const Slice* upper_bound = nullptr) {
    int i = 0;
+    ReadOptions read_options;
+    read_options.iterate_lower_bound = lower_bound;
+    read_options.iterate_upper_bound = upper_bound;
    std::unique_ptr<Iterator> iter =
-        db_->NewMultiCfIterator(ReadOptions(), cfhs);
+        db_->NewCoalescingIterator(read_options, cfhs);
    for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
      ASSERT_EQ(expected_keys[i], iter->key());
-      if (expected_values.has_value()) {
-        ASSERT_EQ(expected_values.value()[i], iter->value());
-      }
+      ASSERT_EQ(expected_values[i], iter->value());
      if (expected_wide_columns.has_value()) {
        ASSERT_EQ(expected_wide_columns.value()[i], iter->columns());
      }
-      if (expected_attribute_groups.has_value()) {
-        // TODO - Add this back when attribute_groups() API is added
-        // ASSERT_EQ(expected_attribute_groups.value()[i],
-        //           iter->attribute_groups());
-      }
      ++i;
    }
    ASSERT_EQ(expected_keys.size(), i);
@@ -48,17 +42,10 @@ class MultiCfIteratorTest : public DBTestBase {
    int rev_i = i - 1;
    for (iter->SeekToLast(); iter->Valid(); iter->Prev()) {
      ASSERT_EQ(expected_keys[rev_i], iter->key());
-      if (expected_values.has_value()) {
-        ASSERT_EQ(expected_values.value()[rev_i], iter->value());
-      }
+      ASSERT_EQ(expected_values[rev_i], iter->value());
      if (expected_wide_columns.has_value()) {
        ASSERT_EQ(expected_wide_columns.value()[rev_i], iter->columns());
      }
-      if (expected_attribute_groups.has_value()) {
-        // TODO - Add this back when attribute_groups() API is added
-        // ASSERT_EQ(expected_attribute_groups.value()[rev_i],
-        //           iter->attribute_groups());
-      }
      rev_i--;
    }
    ASSERT_OK(iter->status());
@@ -78,20 +65,20 @@ class MultiCfIteratorTest : public DBTestBase {
  }
 };

-TEST_F(MultiCfIteratorTest, InvalidArguments) {
+TEST_F(CoalescingIteratorTest, InvalidArguments) {
  Options options = GetDefaultOptions();
  {
    CreateAndReopenWithCF({"cf_1", "cf_2", "cf_3"}, options);

    // Invalid - No CF is provided
    std::unique_ptr<Iterator> iter_with_no_cf =
-        db_->NewMultiCfIterator(ReadOptions(), {});
+        db_->NewCoalescingIterator(ReadOptions(), {});
    ASSERT_NOK(iter_with_no_cf->status());
    ASSERT_TRUE(iter_with_no_cf->status().IsInvalidArgument());
  }
 }

-TEST_F(MultiCfIteratorTest, SimpleValues) {
+TEST_F(CoalescingIteratorTest, SimpleValues) {
  Options options = GetDefaultOptions();
  {
    // Case 1: Unique key per CF
@@ -109,19 +96,21 @@ TEST_F(MultiCfIteratorTest, SimpleValues) {
    // Test for iteration over CF default->1->2->3
    std::vector<ColumnFamilyHandle*> cfhs_order_0_1_2_3 = {
        handles_[0], handles_[1], handles_[2], handles_[3]};
-    verifyMultiCfIterator(cfhs_order_0_1_2_3, expected_keys, expected_values);
+    verifyCoalescingIterator(cfhs_order_0_1_2_3, expected_keys,
+                             expected_values);

    // Test for iteration over CF 3->1->default_cf->2
    std::vector<ColumnFamilyHandle*> cfhs_order_3_1_0_2 = {
        handles_[3], handles_[1], handles_[0], handles_[2]};
    // Iteration order and the return values should be the same since keys are
    // unique per CF
-    verifyMultiCfIterator(cfhs_order_3_1_0_2, expected_keys, expected_values);
+    verifyCoalescingIterator(cfhs_order_3_1_0_2, expected_keys,
+                             expected_values);

    // Verify Seek()
    {
      std::unique_ptr<Iterator> iter =
-          db_->NewMultiCfIterator(ReadOptions(), cfhs_order_0_1_2_3);
+          db_->NewCoalescingIterator(ReadOptions(), cfhs_order_0_1_2_3);
      iter->Seek("");
      ASSERT_EQ(IterStatus(iter.get()), "key_1->key_1_cf_0_val");
      iter->Seek("key_1");
@@ -136,7 +125,7 @@ TEST_F(MultiCfIteratorTest, SimpleValues) {
    // Verify SeekForPrev()
    {
      std::unique_ptr<Iterator> iter =
-          db_->NewMultiCfIterator(ReadOptions(), cfhs_order_0_1_2_3);
+          db_->NewCoalescingIterator(ReadOptions(), cfhs_order_0_1_2_3);
      iter->SeekForPrev("");
      ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
      iter->SeekForPrev("key_1");
@@ -172,56 +161,358 @@ TEST_F(MultiCfIteratorTest, SimpleValues) {
    // Test for iteration over CFs default->1->2->3
    std::vector<ColumnFamilyHandle*> cfhs_order_0_1_2_3 = {
        handles_[0], handles_[1], handles_[2], handles_[3]};
-    std::vector<Slice> expected_values = {"key_1_cf_0_val", "key_2_cf_1_val",
-                                          "key_3_cf_0_val"};
-    verifyMultiCfIterator(cfhs_order_0_1_2_3, expected_keys, expected_values);
+    std::vector<Slice> expected_values = {"key_1_cf_3_val", "key_2_cf_2_val",
+                                          "key_3_cf_3_val"};
+    verifyCoalescingIterator(cfhs_order_0_1_2_3, expected_keys,
+                             expected_values);

    // Test for iteration over CFs 3->2->default_cf->1
    std::vector<ColumnFamilyHandle*> cfhs_order_3_2_0_1 = {
        handles_[3], handles_[2], handles_[0], handles_[1]};
-    expected_values = {"key_1_cf_3_val", "key_2_cf_2_val", "key_3_cf_3_val"};
-    verifyMultiCfIterator(cfhs_order_3_2_0_1, expected_keys, expected_values);
+    expected_values = {"key_1_cf_0_val", "key_2_cf_1_val", "key_3_cf_1_val"};
+    verifyCoalescingIterator(cfhs_order_3_2_0_1, expected_keys,
+                             expected_values);

    // Verify Seek()
    {
      std::unique_ptr<Iterator> iter =
-          db_->NewMultiCfIterator(ReadOptions(), cfhs_order_3_2_0_1);
+          db_->NewCoalescingIterator(ReadOptions(), cfhs_order_3_2_0_1);
      iter->Seek("");
-      ASSERT_EQ(IterStatus(iter.get()), "key_1->key_1_cf_3_val");
+      ASSERT_EQ(IterStatus(iter.get()), "key_1->key_1_cf_0_val");
      iter->Seek("key_1");
-      ASSERT_EQ(IterStatus(iter.get()), "key_1->key_1_cf_3_val");
+      ASSERT_EQ(IterStatus(iter.get()), "key_1->key_1_cf_0_val");
      iter->Seek("key_2");
-      ASSERT_EQ(IterStatus(iter.get()), "key_2->key_2_cf_2_val");
+      ASSERT_EQ(IterStatus(iter.get()), "key_2->key_2_cf_1_val");
      iter->Next();
-      ASSERT_EQ(IterStatus(iter.get()), "key_3->key_3_cf_3_val");
+      ASSERT_EQ(IterStatus(iter.get()), "key_3->key_3_cf_1_val");
      iter->Seek("key_x");
      ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
    }
    // Verify SeekForPrev()
    {
      std::unique_ptr<Iterator> iter =
-          db_->NewMultiCfIterator(ReadOptions(), cfhs_order_3_2_0_1);
+          db_->NewCoalescingIterator(ReadOptions(), cfhs_order_3_2_0_1);
      iter->SeekForPrev("");
      ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
      iter->SeekForPrev("key_1");
-      ASSERT_EQ(IterStatus(iter.get()), "key_1->key_1_cf_3_val");
+      ASSERT_EQ(IterStatus(iter.get()), "key_1->key_1_cf_0_val");
      iter->Next();
-      ASSERT_EQ(IterStatus(iter.get()), "key_2->key_2_cf_2_val");
+      ASSERT_EQ(IterStatus(iter.get()), "key_2->key_2_cf_1_val");
      iter->SeekForPrev("key_x");
-      ASSERT_EQ(IterStatus(iter.get()), "key_3->key_3_cf_3_val");
+      ASSERT_EQ(IterStatus(iter.get()), "key_3->key_3_cf_1_val");
      iter->Next();
      ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
    }
  }
 }

-TEST_F(MultiCfIteratorTest, EmptyCfs) {
+TEST_F(CoalescingIteratorTest, LowerAndUpperBounds) {
+  Options options = GetDefaultOptions();
+  {
+    // Case 1: Unique key per CF
+    CreateAndReopenWithCF({"cf_1", "cf_2", "cf_3"}, options);
+
+    ASSERT_OK(Put(0, "key_1", "key_1_cf_0_val"));
+    ASSERT_OK(Put(1, "key_2", "key_2_cf_1_val"));
+    ASSERT_OK(Put(2, "key_3", "key_3_cf_2_val"));
+    ASSERT_OK(Put(3, "key_4", "key_4_cf_3_val"));
+
+    std::vector<ColumnFamilyHandle*> cfhs_order_0_1_2_3 = {
+        handles_[0], handles_[1], handles_[2], handles_[3]};
+
+    // with lower_bound
+    {
+      // lower_bound is inclusive
+      Slice lb = Slice("key_2");
+      std::vector<Slice> expected_keys = {"key_2", "key_3", "key_4"};
+      std::vector<Slice> expected_values = {"key_2_cf_1_val", "key_3_cf_2_val",
+                                            "key_4_cf_3_val"};
+      verifyCoalescingIterator(cfhs_order_0_1_2_3, expected_keys,
+                               expected_values, std::nullopt, &lb);
+    }
+    // with upper_bound
+    {
+      // upper_bound is exclusive
+      Slice ub = Slice("key_3");
+      std::vector<Slice> expected_keys = {"key_1", "key_2"};
+      std::vector<Slice> expected_values = {"key_1_cf_0_val", "key_2_cf_1_val"};
+      verifyCoalescingIterator(cfhs_order_0_1_2_3, expected_keys,
+                               expected_values, std::nullopt, nullptr, &ub);
+    }
+    // with lower and upper bound
+    {
+      Slice lb = Slice("key_2");
+      Slice ub = Slice("key_4");
+      std::vector<Slice> expected_keys = {"key_2", "key_3"};
+      std::vector<Slice> expected_values = {"key_2_cf_1_val", "key_3_cf_2_val"};
+      verifyCoalescingIterator(cfhs_order_0_1_2_3, expected_keys,
+                               expected_values, std::nullopt, &lb, &ub);
+    }
+
+    {
+      Slice lb = Slice("key_2");
+      Slice ub = Slice("key_4");
+      ReadOptions read_options;
+      read_options.iterate_lower_bound = &lb;
+      read_options.iterate_upper_bound = &ub;
+      // Verify Seek() with bounds
+      {
+        std::unique_ptr<Iterator> iter =
+            db_->NewCoalescingIterator(read_options, cfhs_order_0_1_2_3);
+        iter->Seek("");
+        ASSERT_EQ(IterStatus(iter.get()), "key_2->key_2_cf_1_val");
+        iter->Next();
+        ASSERT_EQ(IterStatus(iter.get()), "key_3->key_3_cf_2_val");
+        iter->Seek("key_x");
+        ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
+      }
+      // Verify SeekForPrev() with bounds
+      {
+        std::unique_ptr<Iterator> iter =
+            db_->NewCoalescingIterator(read_options, cfhs_order_0_1_2_3);
+        iter->SeekForPrev("");
+        ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
+        iter->SeekForPrev("key_1");
+        ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
+        iter->SeekForPrev("key_2");
+        ASSERT_EQ(IterStatus(iter.get()), "key_2->key_2_cf_1_val");
+        iter->Next();
+        ASSERT_EQ(IterStatus(iter.get()), "key_3->key_3_cf_2_val");
+        iter->SeekForPrev("key_x");
+        ASSERT_EQ(IterStatus(iter.get()), "key_3->key_3_cf_2_val");
+        iter->Prev();
+        ASSERT_EQ(IterStatus(iter.get()), "key_2->key_2_cf_1_val");
+        iter->Next();
+        ASSERT_EQ(IterStatus(iter.get()), "key_3->key_3_cf_2_val");
+        iter->Next();
+        ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
+      }
+    }
+  }
+  {
+    // Case 2: Same key in multiple CFs
+    options = CurrentOptions(options);
+    DestroyAndReopen(options);
+    CreateAndReopenWithCF({"cf_1", "cf_2", "cf_3"}, options);
+
+    ASSERT_OK(Put(0, "key_1", "key_1_cf_0_val"));
+    ASSERT_OK(Put(3, "key_1", "key_1_cf_3_val"));
+    ASSERT_OK(Put(1, "key_2", "key_2_cf_1_val"));
+    ASSERT_OK(Put(2, "key_2", "key_2_cf_2_val"));
+    ASSERT_OK(Put(0, "key_3", "key_3_cf_0_val"));
+    ASSERT_OK(Put(1, "key_3", "key_3_cf_1_val"));
+    ASSERT_OK(Put(3, "key_3", "key_3_cf_3_val"));
+
+    // Test for iteration over CFs default->1->2->3
+    std::vector<ColumnFamilyHandle*> cfhs_order_0_1_2_3 = {
+        handles_[0], handles_[1], handles_[2], handles_[3]};
+    // with lower_bound
+    {
+      // lower_bound is inclusive
+      Slice lb = Slice("key_2");
+      std::vector<Slice> expected_keys = {"key_2", "key_3"};
+      std::vector<Slice> expected_values = {"key_2_cf_2_val", "key_3_cf_3_val"};
+      verifyCoalescingIterator(cfhs_order_0_1_2_3, expected_keys,
+                               expected_values, std::nullopt, &lb);
+    }
+    // with upper_bound
+    {
+      // upper_bound is exclusive
+      Slice ub = Slice("key_3");
+      std::vector<Slice> expected_keys = {"key_1", "key_2"};
+      std::vector<Slice> expected_values = {"key_1_cf_3_val", "key_2_cf_2_val"};
+      verifyCoalescingIterator(cfhs_order_0_1_2_3, expected_keys,
+                               expected_values, std::nullopt, nullptr, &ub);
+    }
+    // with lower and upper bound
+    {
+      Slice lb = Slice("key_2");
+      Slice ub = Slice("key_3");
+      std::vector<Slice> expected_keys = {"key_2"};
+      std::vector<Slice> expected_values = {"key_2_cf_2_val"};
+      verifyCoalescingIterator(cfhs_order_0_1_2_3, expected_keys,
+                               expected_values, std::nullopt, &lb, &ub);
+    }
+
+    // Test for iteration over CFs 3->2->default_cf->1
+    std::vector<ColumnFamilyHandle*> cfhs_order_3_2_0_1 = {
+        handles_[3], handles_[2], handles_[0], handles_[1]};
+    {
+      // lower_bound is inclusive
+      Slice lb = Slice("key_2");
+      std::vector<Slice> expected_keys = {"key_2", "key_3"};
+      std::vector<Slice> expected_values = {"key_2_cf_1_val", "key_3_cf_1_val"};
+      verifyCoalescingIterator(cfhs_order_3_2_0_1, expected_keys,
+                               expected_values, std::nullopt, &lb);
+    }
+    // with upper_bound
+    {
+      // upper_bound is exclusive
+      Slice ub = Slice("key_3");
+      std::vector<Slice> expected_keys = {"key_1", "key_2"};
+      std::vector<Slice> expected_values = {"key_1_cf_0_val", "key_2_cf_1_val"};
+      verifyCoalescingIterator(cfhs_order_3_2_0_1, expected_keys,
+                               expected_values, std::nullopt, nullptr, &ub);
+    }
+    // with lower and upper bound
+    {
+      Slice lb = Slice("key_2");
+      Slice ub = Slice("key_3");
+      std::vector<Slice> expected_keys = {"key_2"};
+      std::vector<Slice> expected_values = {"key_2_cf_1_val"};
+      verifyCoalescingIterator(cfhs_order_3_2_0_1, expected_keys,
+                               expected_values, std::nullopt, &lb, &ub);
+    }
+    {
+      Slice lb = Slice("key_2");
+      Slice ub = Slice("key_3");
+      ReadOptions read_options;
+      read_options.iterate_lower_bound = &lb;
+      read_options.iterate_upper_bound = &ub;
+      // Verify Seek() with bounds
+      {
+        std::unique_ptr<Iterator> iter =
+            db_->NewCoalescingIterator(read_options, cfhs_order_3_2_0_1);
+        iter->Seek("");
+        ASSERT_EQ(IterStatus(iter.get()), "key_2->key_2_cf_1_val");
+        iter->Next();
+        ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
+        iter->Seek("key_x");
+        ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
+      }
+      // Verify SeekForPrev() with bounds
+      {
+        std::unique_ptr<Iterator> iter =
+            db_->NewCoalescingIterator(read_options, cfhs_order_3_2_0_1);
+        iter->SeekForPrev("");
+        ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
+        iter->SeekForPrev("key_1");
+        ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
+        iter->SeekForPrev("key_2");
+        ASSERT_EQ(IterStatus(iter.get()), "key_2->key_2_cf_1_val");
+        iter->SeekForPrev("key_x");
+        ASSERT_EQ(IterStatus(iter.get()), "key_2->key_2_cf_1_val");
+        iter->Next();
+        ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
+      }
+    }
+  }
+}
+
+TEST_F(CoalescingIteratorTest, ConsistentViewExplicitSnapshot) {
+  Options options = GetDefaultOptions();
+  options.atomic_flush = true;
+  CreateAndReopenWithCF({"cf_1", "cf_2", "cf_3"}, options);
+
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_OK(Put(i, "cf" + std::to_string(i) + "_key",
+                  "cf" + std::to_string(i) + "_val"));
+  }
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+      {{"DBImpl::BGWorkFlush:done",
+        "DBImpl::MultiCFSnapshot::BeforeCheckingSnapshot"}});
+
+  bool flushed = false;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::MultiCFSnapshot::AfterRefSV", [&](void* /*arg*/) {
+        if (!flushed) {
+          for (int i = 0; i < 4; ++i) {
+            ASSERT_OK(Put(i, "cf" + std::to_string(i) + "_key",
+                          "cf" + std::to_string(i) + "_val_new"));
+          }
+          ASSERT_OK(Flush());
+          flushed = true;
+        }
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  std::vector<ColumnFamilyHandle*> cfhs_order_0_1_2_3 = {
+      handles_[0], handles_[1], handles_[2], handles_[3]};
+  ReadOptions read_options;
+  const Snapshot* snapshot = db_->GetSnapshot();
+  read_options.snapshot = snapshot;
+  // Verify Seek()
+  {
+    std::unique_ptr<Iterator> iter =
+        db_->NewCoalescingIterator(read_options, cfhs_order_0_1_2_3);
+    iter->Seek("");
+    ASSERT_EQ(IterStatus(iter.get()), "cf0_key->cf0_val");
+    iter->Next();
+    ASSERT_EQ(IterStatus(iter.get()), "cf1_key->cf1_val");
+  }
+  // Verify SeekForPrev()
+  {
+    std::unique_ptr<Iterator> iter =
+        db_->NewCoalescingIterator(read_options, cfhs_order_0_1_2_3);
+    iter->SeekForPrev("");
+    ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
+    iter->SeekForPrev("cf2_key");
+    ASSERT_EQ(IterStatus(iter.get()), "cf2_key->cf2_val");
+    iter->Prev();
+    ASSERT_EQ(IterStatus(iter.get()), "cf1_key->cf1_val");
+  }
+  db_->ReleaseSnapshot(snapshot);
+}
+
+TEST_F(CoalescingIteratorTest, ConsistentViewImplicitSnapshot) {
+  Options options = GetDefaultOptions();
+  CreateAndReopenWithCF({"cf_1", "cf_2", "cf_3"}, options);
+
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_OK(Put(i, "cf" + std::to_string(i) + "_key",
+                  "cf" + std::to_string(i) + "_val"));
+  }
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+      {{"DBImpl::BGWorkFlush:done",
+        "DBImpl::MultiCFSnapshot::BeforeCheckingSnapshot"}});
+
+  bool flushed = false;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::MultiCFSnapshot::AfterRefSV", [&](void* /*arg*/) {
+        if (!flushed) {
+          for (int i = 0; i < 4; ++i) {
+            ASSERT_OK(Put(i, "cf" + std::to_string(i) + "_key",
+                          "cf" + std::to_string(i) + "_val_new"));
+          }
+          ASSERT_OK(Flush(1));
+          flushed = true;
+        }
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  std::vector<ColumnFamilyHandle*> cfhs_order_0_1_2_3 = {
+      handles_[0], handles_[1], handles_[2], handles_[3]};
+  // Verify Seek()
+  {
+    std::unique_ptr<Iterator> iter =
+        db_->NewCoalescingIterator(ReadOptions(), cfhs_order_0_1_2_3);
+    iter->Seek("cf2_key");
+    ASSERT_EQ(IterStatus(iter.get()), "cf2_key->cf2_val_new");
+    iter->Next();
+    ASSERT_EQ(IterStatus(iter.get()), "cf3_key->cf3_val_new");
+  }
+  // Verify SeekForPrev()
+  {
+    std::unique_ptr<Iterator> iter =
+        db_->NewCoalescingIterator(ReadOptions(), cfhs_order_0_1_2_3);
+    iter->SeekForPrev("");
+    ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
+    iter->SeekForPrev("cf1_key");
+    ASSERT_EQ(IterStatus(iter.get()), "cf1_key->cf1_val_new");
+    iter->Prev();
+    ASSERT_EQ(IterStatus(iter.get()), "cf0_key->cf0_val_new");
+  }
+}
+
+TEST_F(CoalescingIteratorTest, EmptyCfs) {
  Options options = GetDefaultOptions();
  {
    // Case 1: No keys in any of the CFs
    CreateAndReopenWithCF({"cf_1", "cf_2", "cf_3"}, options);
    std::unique_ptr<Iterator> iter =
-        db_->NewMultiCfIterator(ReadOptions(), handles_);
+        db_->NewCoalescingIterator(ReadOptions(), handles_);
    iter->SeekToFirst();
    ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
    iter->SeekToLast();
@@ -237,7 +528,7 @@ TEST_F(MultiCfIteratorTest, EmptyCfs) {
    // Case 2: A single key exists in only one of the CF. Rest CFs are empty.
    ASSERT_OK(Put(1, "key_1", "key_1_cf_1_val"));
    std::unique_ptr<Iterator> iter =
-        db_->NewMultiCfIterator(ReadOptions(), handles_);
+        db_->NewCoalescingIterator(ReadOptions(), handles_);
    iter->SeekToFirst();
    ASSERT_EQ(IterStatus(iter.get()), "key_1->key_1_cf_1_val");
    iter->Next();
@@ -251,21 +542,21 @@ TEST_F(MultiCfIteratorTest, EmptyCfs) {
    // Case 3: same key exists in all of the CFs except one (cf_2)
    ASSERT_OK(Put(0, "key_1", "key_1_cf_0_val"));
    ASSERT_OK(Put(3, "key_1", "key_1_cf_3_val"));
-    // handles_ are in the order of 0->1->2->3. We should expect value from cf_0
+    // handles_ are in the order of 0->1->2->3
    std::unique_ptr<Iterator> iter =
-        db_->NewMultiCfIterator(ReadOptions(), handles_);
+        db_->NewCoalescingIterator(ReadOptions(), handles_);
    iter->SeekToFirst();
-    ASSERT_EQ(IterStatus(iter.get()), "key_1->key_1_cf_0_val");
+    ASSERT_EQ(IterStatus(iter.get()), "key_1->key_1_cf_3_val");
    iter->Next();
    ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
    iter->SeekToLast();
-    ASSERT_EQ(IterStatus(iter.get()), "key_1->key_1_cf_0_val");
+    ASSERT_EQ(IterStatus(iter.get()), "key_1->key_1_cf_3_val");
    iter->Prev();
    ASSERT_EQ(IterStatus(iter.get()), "(invalid)");
  }
 }

-TEST_F(MultiCfIteratorTest, WideColumns) {
+TEST_F(CoalescingIteratorTest, WideColumns) {
  // Set up the DB and Column Families
  Options options = GetDefaultOptions();
  CreateAndReopenWithCF({"cf_1", "cf_2", "cf_3"}, options);
@@ -274,30 +565,65 @@ TEST_F(MultiCfIteratorTest, WideColumns) {
  WideColumns key_1_columns_in_cf_2{
      {kDefaultWideColumnName, "cf_2_col_val_0_key_1"},
      {"cf_2_col_name_1", "cf_2_col_val_1_key_1"},
-      {"cf_2_col_name_2", "cf_2_col_val_2_key_1"}};
+      {"cf_2_col_name_2", "cf_2_col_val_2_key_1"},
+      {"cf_overlap_col_name", "cf_2_overlap_value_key_1"}};
  WideColumns key_1_columns_in_cf_3{
      {"cf_3_col_name_1", "cf_3_col_val_1_key_1"},
      {"cf_3_col_name_2", "cf_3_col_val_2_key_1"},
-      {"cf_3_col_name_3", "cf_3_col_val_3_key_1"}};
+      {"cf_3_col_name_3", "cf_3_col_val_3_key_1"},
+      {"cf_overlap_col_name", "cf_3_overlap_value_key_1"}};
+  WideColumns key_1_expected_columns_cfh_order_2_3{
+      {kDefaultWideColumnName, "cf_2_col_val_0_key_1"},
+      {"cf_2_col_name_1", "cf_2_col_val_1_key_1"},
+      {"cf_2_col_name_2", "cf_2_col_val_2_key_1"},
+      {"cf_3_col_name_1", "cf_3_col_val_1_key_1"},
+      {"cf_3_col_name_2", "cf_3_col_val_2_key_1"},
+      {"cf_3_col_name_3", "cf_3_col_val_3_key_1"},
+      {"cf_overlap_col_name", "cf_3_overlap_value_key_1"}};
+  WideColumns key_1_expected_columns_cfh_order_3_2{
+      {kDefaultWideColumnName, "cf_2_col_val_0_key_1"},
+      {"cf_2_col_name_1", "cf_2_col_val_1_key_1"},
+      {"cf_2_col_name_2", "cf_2_col_val_2_key_1"},
+      {"cf_3_col_name_1", "cf_3_col_val_1_key_1"},
+      {"cf_3_col_name_2", "cf_3_col_val_2_key_1"},
+      {"cf_3_col_name_3", "cf_3_col_val_3_key_1"},
+      {"cf_overlap_col_name", "cf_2_overlap_value_key_1"}};

  constexpr char key_2[] = "key_2";
  WideColumns key_2_columns_in_cf_1{
-      {"cf_1_col_name_1", "cf_1_col_val_1_key_2"}};
+      {"cf_overlap_col_name", "cf_1_overlap_value_key_2"}};
  WideColumns key_2_columns_in_cf_2{
      {"cf_2_col_name_1", "cf_2_col_val_1_key_2"},
-      {"cf_2_col_name_2", "cf_2_col_val_2_key_2"}};
+      {"cf_2_col_name_2", "cf_2_col_val_2_key_2"},
+      {"cf_overlap_col_name", "cf_2_overlap_value_key_2"}};
+  WideColumns key_2_expected_columns_cfh_order_1_2{
+      {"cf_2_col_name_1", "cf_2_col_val_1_key_2"},
+      {"cf_2_col_name_2", "cf_2_col_val_2_key_2"},
+      {"cf_overlap_col_name", "cf_2_overlap_value_key_2"}};
+  WideColumns key_2_expected_columns_cfh_order_2_1{
+      {"cf_2_col_name_1", "cf_2_col_val_1_key_2"},
+      {"cf_2_col_name_2", "cf_2_col_val_2_key_2"},
+      {"cf_overlap_col_name", "cf_1_overlap_value_key_2"}};

  constexpr char key_3[] = "key_3";
  WideColumns key_3_columns_in_cf_1{
      {"cf_1_col_name_1", "cf_1_col_val_1_key_3"}};
  WideColumns key_3_columns_in_cf_3{
      {"cf_3_col_name_1", "cf_3_col_val_1_key_3"}};
+  WideColumns key_3_expected_columns{
+      {"cf_1_col_name_1", "cf_1_col_val_1_key_3"},
+      {"cf_3_col_name_1", "cf_3_col_val_1_key_3"},
+  };

  constexpr char key_4[] = "key_4";
  WideColumns key_4_columns_in_cf_0{
      {"cf_0_col_name_1", "cf_0_col_val_1_key_4"}};
  WideColumns key_4_columns_in_cf_2{
      {"cf_2_col_name_1", "cf_2_col_val_1_key_4"}};
+  WideColumns key_4_expected_columns{
+      {"cf_0_col_name_1", "cf_0_col_val_1_key_4"},
+      {"cf_2_col_name_1", "cf_2_col_val_1_key_4"},
+  };

  // Use AttributeGroup PutEntity API to insert them together
  AttributeGroups key_1_attribute_groups{
@@ -318,31 +644,47 @@ TEST_F(MultiCfIteratorTest, WideColumns) {
  ASSERT_OK(db_->PutEntity(WriteOptions(), key_3, key_3_attribute_groups));
  ASSERT_OK(db_->PutEntity(WriteOptions(), key_4, key_4_attribute_groups));

-  // Test for iteration over CF default->1->2->3
-  std::vector<ColumnFamilyHandle*> cfhs_order_0_1_2_3 = {
-      handles_[0], handles_[1], handles_[2], handles_[3]};
+  // Keys should be returned in order regardless of cfh order
  std::vector<Slice> expected_keys = {key_1, key_2, key_3, key_4};
-  // Pick what DBIter would return for value() in the first CF that key exists
+
  // Since value for kDefaultWideColumnName only exists for key_1, rest will
-  // return empty value
+  // return empty value after coalesced
  std::vector<Slice> expected_values = {"cf_2_col_val_0_key_1", "", "", ""};

-  // Pick columns from the first CF that the key exists and value is stored as
-  // wide column
-  std::vector<WideColumns> expected_wide_columns = {
-      {{kDefaultWideColumnName, "cf_2_col_val_0_key_1"},
-       {"cf_2_col_name_1", "cf_2_col_val_1_key_1"},
-       {"cf_2_col_name_2", "cf_2_col_val_2_key_1"}},
-      {{"cf_1_col_name_1", "cf_1_col_val_1_key_2"}},
-      {{"cf_1_col_name_1", "cf_1_col_val_1_key_3"}},
-      {{"cf_0_col_name_1", "cf_0_col_val_1_key_4"}}};
-  verifyMultiCfIterator(cfhs_order_0_1_2_3, expected_keys, expected_values,
-                        expected_wide_columns);
+  // Test for iteration over CF default->1->2->3
+  {
+    std::vector<ColumnFamilyHandle*> cfhs_order_0_1_2_3 = {
+        handles_[0], handles_[1], handles_[2], handles_[3]};
+
+    // Coalesced columns
+    std::vector<WideColumns> expected_wide_columns_0_1_2_3 = {
+        key_1_expected_columns_cfh_order_2_3,
+        key_2_expected_columns_cfh_order_1_2, key_3_expected_columns,
+        key_4_expected_columns};
+
+    verifyCoalescingIterator(cfhs_order_0_1_2_3, expected_keys, expected_values,
+                             expected_wide_columns_0_1_2_3);
+  }
+
+  // Test for iteration over CF 3->2->default->1
+  {
+    std::vector<ColumnFamilyHandle*> cfhs_order_3_2_0_1 = {
+        handles_[3], handles_[2], handles_[0], handles_[1]};
+
+    // Coalesced columns
+    std::vector<WideColumns> expected_wide_columns_3_2_0_1 = {
+        key_1_expected_columns_cfh_order_3_2,
+        key_2_expected_columns_cfh_order_2_1, key_3_expected_columns,
+        key_4_expected_columns};
+
+    verifyCoalescingIterator(cfhs_order_3_2_0_1, expected_keys, expected_values,
+                             expected_wide_columns_3_2_0_1);
+  }
 }

-TEST_F(MultiCfIteratorTest, DifferentComparatorsInMultiCFs) {
+TEST_F(CoalescingIteratorTest, DifferentComparatorsInMultiCFs) {
  // This test creates two column families with two different comparators.
-  // Attempting to create the MultiCFIterator should fail.
+  // Attempting to create the CoalescingIterator should fail.
  Options options = GetDefaultOptions();
  options.create_if_missing = true;
  DestroyAndReopen(options);
@@ -362,15 +704,15 @@ TEST_F(MultiCfIteratorTest, DifferentComparatorsInMultiCFs) {
  verifyExpectedKeys(handles_[1], {"key_3", "key_2", "key_1"});

  std::unique_ptr<Iterator> iter =
-      db_->NewMultiCfIterator(ReadOptions(), handles_);
+      db_->NewCoalescingIterator(ReadOptions(), handles_);
  ASSERT_NOK(iter->status());
  ASSERT_TRUE(iter->status().IsInvalidArgument());
 }

-TEST_F(MultiCfIteratorTest, CustomComparatorsInMultiCFs) {
+TEST_F(CoalescingIteratorTest, CustomComparatorsInMultiCFs) {
  // This test creates two column families with the same custom test
  // comparators (but instantiated independently). Attempting to create the
-  // MultiCFIterator should not fail.
+  // CoalescingIterator should not fail.
  Options options = GetDefaultOptions();
  options.create_if_missing = true;
  DestroyAndReopen(options);
@@ -410,12 +752,12 @@ TEST_F(MultiCfIteratorTest, CustomComparatorsInMultiCFs) {
  std::vector<Slice> expected_keys = {
      "key_001_003", "key_001_002", "key_001_001", "key_002_003", "key_002_002",
      "key_002_001", "key_003_006", "key_003_005", "key_003_004"};
-  std::vector<Slice> expected_values = {"value_0_1", "value_0_2", "value_0_3",
+  std::vector<Slice> expected_values = {"value_1_1", "value_1_2", "value_1_3",
                                        "value_0_4", "value_0_5", "value_0_6",
                                        "value_1_4", "value_1_5", "value_1_6"};
  int i = 0;
  std::unique_ptr<Iterator> iter =
-      db_->NewMultiCfIterator(ReadOptions(), handles_);
+      db_->NewCoalescingIterator(ReadOptions(), handles_);
  for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
    ASSERT_EQ(expected_keys[i], iter->key());
    ASSERT_EQ(expected_values[i], iter->value());
@@ -424,7 +766,59 @@ TEST_F(MultiCfIteratorTest, CustomComparatorsInMultiCFs) {
  ASSERT_OK(iter->status());
 }

-TEST_F(MultiCfIteratorTest, DISABLED_IterateAttributeGroups) {
+class AttributeGroupIteratorTest : public DBTestBase {
+ public:
+  AttributeGroupIteratorTest()
+      : DBTestBase("attribute_group_iterator_test", /*env_do_fsync=*/true) {}
+
+  void verifyAttributeGroupIterator(
+      const std::vector<ColumnFamilyHandle*>& cfhs,
+      const std::vector<Slice>& expected_keys,
+      const std::vector<AttributeGroups>& expected_attribute_groups,
+      const Slice* lower_bound = nullptr, const Slice* upper_bound = nullptr) {
+    int i = 0;
+    ReadOptions read_options;
+    read_options.iterate_lower_bound = lower_bound;
+    read_options.iterate_upper_bound = upper_bound;
+    std::unique_ptr<AttributeGroupIterator> iter =
+        db_->NewAttributeGroupIterator(read_options, cfhs);
+    for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
+      ASSERT_EQ(expected_keys[i], iter->key());
+      auto iterator_attribute_groups = iter->attribute_groups();
+      ASSERT_EQ(expected_attribute_groups[i].size(),
+                iterator_attribute_groups.size());
+      for (size_t cfh_i = 0; cfh_i < iterator_attribute_groups.size();
+           cfh_i++) {
+        ASSERT_EQ(expected_attribute_groups[i][cfh_i].column_family(),
+                  iterator_attribute_groups[cfh_i].column_family());
+        ASSERT_EQ(expected_attribute_groups[i][cfh_i].columns(),
+                  iterator_attribute_groups[cfh_i].columns());
+      }
+      ++i;
+    }
+    ASSERT_EQ(expected_keys.size(), i);
+    ASSERT_OK(iter->status());
+
+    int rev_i = i - 1;
+    for (iter->SeekToLast(); iter->Valid(); iter->Prev()) {
+      ASSERT_EQ(expected_keys[rev_i], iter->key());
+      auto iterator_attribute_groups = iter->attribute_groups();
+      ASSERT_EQ(expected_attribute_groups[rev_i].size(),
+                iterator_attribute_groups.size());
+      for (size_t cfh_i = 0; cfh_i < iterator_attribute_groups.size();
+           cfh_i++) {
+        ASSERT_EQ(expected_attribute_groups[rev_i][cfh_i].column_family(),
+                  iterator_attribute_groups[cfh_i].column_family());
+        ASSERT_EQ(expected_attribute_groups[rev_i][cfh_i].columns(),
+                  iterator_attribute_groups[cfh_i].columns());
+      }
+      rev_i--;
+    }
+    ASSERT_OK(iter->status());
+  }
+};
+
+TEST_F(AttributeGroupIteratorTest, IterateAttributeGroups) {
  // Set up the DB and Column Families
  Options options = GetDefaultOptions();
  CreateAndReopenWithCF({"cf_1", "cf_2", "cf_3"}, options);
@@ -479,13 +873,40 @@ TEST_F(MultiCfIteratorTest, DISABLED_IterateAttributeGroups) {
  // Test for iteration over CF default->1->2->3
  std::vector<ColumnFamilyHandle*> cfhs_order_0_1_2_3 = {
      handles_[0], handles_[1], handles_[2], handles_[3]};
-  std::vector<Slice> expected_keys = {key_1, key_2, key_3, key_4};
-  std::vector<AttributeGroups> expected_attribute_groups = {
-      key_1_attribute_groups, key_2_attribute_groups, key_3_attribute_groups,
-      key_4_attribute_groups};
-  verifyMultiCfIterator(
-      cfhs_order_0_1_2_3, expected_keys, std::nullopt /* expected_values */,
-      std::nullopt /* expected_wide_columns */, expected_attribute_groups);
+  {
+    std::vector<Slice> expected_keys = {key_1, key_2, key_3, key_4};
+    std::vector<AttributeGroups> expected_attribute_groups = {
+        key_1_attribute_groups, key_2_attribute_groups, key_3_attribute_groups,
+        key_4_attribute_groups};
+    verifyAttributeGroupIterator(cfhs_order_0_1_2_3, expected_keys,
+                                 expected_attribute_groups);
+  }
+  Slice lb = Slice("key_2");
+  Slice ub = Slice("key_4");
+  // Test for lower bound only
+  {
+    std::vector<Slice> expected_keys = {key_2, key_3, key_4};
+    std::vector<AttributeGroups> expected_attribute_groups = {
+        key_2_attribute_groups, key_3_attribute_groups, key_4_attribute_groups};
+    verifyAttributeGroupIterator(cfhs_order_0_1_2_3, expected_keys,
+                                 expected_attribute_groups, &lb);
+  }
+  // Test for upper bound only
+  {
+    std::vector<Slice> expected_keys = {key_1, key_2, key_3};
+    std::vector<AttributeGroups> expected_attribute_groups = {
+        key_1_attribute_groups, key_2_attribute_groups, key_3_attribute_groups};
+    verifyAttributeGroupIterator(cfhs_order_0_1_2_3, expected_keys,
+                                 expected_attribute_groups, nullptr, &ub);
+  }
+  // Test for lower and upper bound
+  {
+    std::vector<Slice> expected_keys = {key_2, key_3};
+    std::vector<AttributeGroups> expected_attribute_groups = {
+        key_2_attribute_groups, key_3_attribute_groups};
+    verifyAttributeGroupIterator(cfhs_order_0_1_2_3, expected_keys,
+                                 expected_attribute_groups, &lb, &ub);
+  }
 }

 }  // namespace ROCKSDB_NAMESPACE
@@ -17,16 +17,6 @@
 #include "table/internal_iterator.h"

 namespace ROCKSDB_NAMESPACE {
-struct FragmentedRangeTombstoneList;
-
-struct FragmentedRangeTombstoneListCache {
-  // ensure only the first reader needs to initialize l
-  std::mutex reader_mutex;
-  std::unique_ptr<FragmentedRangeTombstoneList> tombstones = nullptr;
-  // readers will first check this bool to avoid
-  std::atomic<bool> initialized = false;
-};
-
 struct FragmentedRangeTombstoneList {
 public:
  // A compact representation of a "stack" of range tombstone fragments, which
@@ -124,6 +114,14 @@ struct FragmentedRangeTombstoneList {
  uint64_t total_tombstone_payload_bytes_;
 };

+struct FragmentedRangeTombstoneListCache {
+  // ensure only the first reader needs to initialize l
+  std::mutex reader_mutex;
+  std::unique_ptr<FragmentedRangeTombstoneList> tombstones = nullptr;
+  // readers will first check this bool to avoid
+  std::atomic<bool> initialized = false;
+};
+
 // FragmentedRangeTombstoneIterator converts an InternalIterator of a range-del
 // meta block into an iterator over non-overlapping tombstone fragments. The
 // tombstone fragmentation process should be more efficient than the range