Update HISTORY.md and version to 7.8.3

Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags:
Update HISTORY for reverts
2026-11-06 19:12:36 -05:00 · 2022-11-29 06:51:03 -08:00 · 2022-11-28 09:37:05 -08:00 · 2022-11-28 09:34:16 -08:00 · 2022-11-28 09:33:56 -08:00 · 2022-11-27 23:14:56 -08:00
444 changed files with 21767 additions and 8632 deletions
@@ -72,6 +72,14 @@ commands:
          path: /tmp/core_dumps
          when: on_fail

+  upgrade-cmake:
+    steps:
+      - run:
+          name: Upgrade cmake
+          command: |
+            sudo apt remove --purge cmake
+            sudo snap install cmake --classic
+
  install-gflags:
    steps:
      - run:
@@ -93,6 +101,13 @@ commands:
          command: |
            make checkout_folly

+  build-folly:
+    steps:
+      - run:
+          name: Build folly and dependencies
+          command: |
+            make build_folly
+
  build-for-benchmarks:
    steps:
      - pre-steps
@@ -109,10 +124,19 @@ commands:
          environment:
            LD_LIBRARY_PATH: /usr/local/lib
            # How long to run parts of the test(s)
-            DURATION_RO: 450
-            DURATION_RW: 450
+            DURATION_RO: 400
+            DURATION_RW: 700
+            # Keep threads within physical capacity of server (much lower than default)
+            NUM_THREADS: 1
+            MAX_BACKGROUND_JOBS: 3
+            # Don't run a couple of "optional" initial tests
+            CI_TESTS_ONLY: "true"
+            # Reduce configured size of levels to ensure more levels in the leveled compaction LSM tree
+            WRITE_BUFFER_SIZE_MB: 16
+            TARGET_FILE_SIZE_BASE_MB: 16
+            MAX_BYTES_FOR_LEVEL_BASE_MB: 64
            # The benchmark host has 32GB memory
-            # The folllowing values are tailored to work with that
+            # The following values are tailored to work with that
            # Note, tests may not exercise the targeted issues if the memory is increased on new test hosts.


@@ -149,7 +173,7 @@ executors:
      # to run the CI image environment locally:
      #  $ docker run --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -it zjay437/rocksdb:0.5 bash
      # option `--cap-add=SYS_PTRACE --security-opt seccomp=unconfined` is used to enable gdb to attach an existing process
-      - image: zjay437/rocksdb:0.5
+      - image: zjay437/rocksdb:0.6

 jobs:
  build-macos:
@@ -215,7 +239,7 @@ jobs:
      - pre-steps
      - run: ENCRYPTED_ENV=1 ROCKSDB_DISABLE_SNAPPY=1 ROCKSDB_DISABLE_ZLIB=1 ROCKSDB_DISABLE_BZIP=1 ROCKSDB_DISABLE_LZ4=1 ROCKSDB_DISABLE_ZSTD=1 make V=1 J=32 -j32 check
      - run: |
-          ./sst_dump --help | egrep -q 'Supported compression types: kNoCompression$' # Verify no compiled in compression
+          ./sst_dump --help | grep -E -q 'Supported compression types: kNoCompression$' # Verify no compiled in compression
      - post-steps

  build-linux-shared_lib-alt_namespace-status_checked:
@@ -346,7 +370,17 @@ jobs:
    steps:
      - pre-steps
      - setup-folly
-      - run: (mkdir build && cd build && cmake -DUSE_FOLLY=1 -DWITH_GFLAGS=1 .. && make V=1 -j20 && ctest -j20)
+      - build-folly
+      - run: (mkdir build && cd build && cmake -DUSE_FOLLY=1 -DWITH_GFLAGS=1 -DROCKSDB_BUILD_SHARED=0 .. && make V=1 -j20 && ctest -j20)
+      - post-steps
+
+  build-linux-cmake-with-folly-lite-no-test:
+    executor: linux-docker
+    resource_class: 2xlarge
+    steps:
+      - pre-steps
+      - setup-folly
+      - run: (mkdir build && cd build && cmake -DUSE_FOLLY_LITE=1 -DWITH_GFLAGS=1 .. && make V=1 -j20)
      - post-steps

  build-linux-cmake-with-benchmark:
@@ -376,9 +410,19 @@ jobs:
    steps:
      - pre-steps
      - setup-folly
+      - build-folly
      - run: USE_FOLLY=1 CC=gcc-7 CXX=g++-7 V=1 make -j32 check
      - post-steps

+  build-linux-gcc-7-with-folly-lite-no-test:
+    executor: linux-docker
+    resource_class: 2xlarge
+    steps:
+      - pre-steps
+      - setup-folly
+      - run: USE_FOLLY_LITE=1 CC=gcc-7 CXX=g++-7 V=1 make -j32 all
+      - post-steps
+
  build-linux-gcc-8-no_test_run:
    executor: linux-docker
    resource_class: 2xlarge
@@ -387,6 +431,19 @@ jobs:
      - run: CC=gcc-8 CXX=g++-8 V=1 make -j32 all
      - post-steps

+  build-linux-cmake-with-folly-coroutines:
+    executor: linux-docker
+    resource_class: 2xlarge
+    environment:
+      CC: gcc-10
+      CXX: g++-10
+    steps:
+      - pre-steps
+      - setup-folly
+      - build-folly
+      - run: (mkdir build && cd build && cmake -DUSE_COROUTINES=1 -DWITH_GFLAGS=1 -DROCKSDB_BUILD_SHARED=0 .. && make V=1 -j20 && ctest -j20)
+      - post-steps
+
  build-linux-gcc-10-cxx20-no_test_run:
    executor: linux-docker
    resource_class: 2xlarge
@@ -418,6 +475,7 @@ jobs:
    steps:
      - pre-steps
      - setup-folly
+      - build-folly
      - run: CC=clang-13 CXX=clang++-13 USE_CLANG=1 USE_FOLLY=1 COMPILE_WITH_UBSAN=1 COMPILE_WITH_ASAN=1 make -j32 check
      - post-steps

@@ -435,7 +493,7 @@ jobs:
    resource_class: large
    steps:
      - pre-steps
-      - run: ulimit -S -n `ulimit -H -n` && make V=1 -j8 CRASH_TEST_EXT_ARGS=--duration=960 blackbox_crash_test_with_atomic_flush
+      - run: ulimit -S -n `ulimit -H -n` && make V=1 -j8 CRASH_TEST_EXT_ARGS='--duration=960 --max_key=2500000' blackbox_crash_test_with_atomic_flush
      - post-steps

  build-linux-crashtest-tiered-storage-bb:
@@ -584,6 +642,7 @@ jobs:
      - run:
          name: "Test RocksDBJava"
          command: make V=1 J=16 -j16 jtest
+          no_output_timeout: 20m
      - post-steps

  build-macos-java-static:
@@ -608,6 +667,7 @@ jobs:
      - run:
          name: "Build RocksDBJava x86 and ARM Static Libraries"
          command: make V=1 J=16 -j16 rocksdbjavastaticosx
+          no_output_timeout: 20m
      - post-steps

  build-macos-java-static-universal:
@@ -632,6 +692,7 @@ jobs:
      - run:
          name: "Build RocksDBJava Universal Binary Static Library"
          command: make V=1 J=16 -j16 rocksdbjavastaticosx_ub
+          no_output_timeout: 20m
      - post-steps

  build-examples:
@@ -765,7 +826,10 @@ workflows:
    jobs:
      - build-linux
      - build-linux-cmake-with-folly
+      - build-linux-cmake-with-folly-lite-no-test
      - build-linux-gcc-7-with-folly
+      - build-linux-gcc-7-with-folly-lite-no-test
+      - build-linux-cmake-with-folly-coroutines
      - build-linux-cmake-with-benchmark
      - build-linux-encrypted_env-no_compression
      - build-linux-lite
@@ -1,5 +1,8 @@
 name: Check buck targets and code format
 on: [push, pull_request]
+permissions:
+  contents: read
+
 jobs:
  check:
    name: Check TARGETS file and code format
@@ -56,6 +56,7 @@ io_tracer_parser
 .DS_Store
 .vs
 .vscode
+.clangd

 java/out
 java/target
@@ -74,6 +74,7 @@ option(WITH_WINDOWS_UTF8_FILENAMES "use UTF8 as characterset for opening files,
 if (WITH_WINDOWS_UTF8_FILENAMES)
  add_definitions(-DROCKSDB_WINDOWS_UTF8_FILENAMES)
 endif()
+option(ROCKSDB_BUILD_SHARED "Build shared versions of the RocksDB libraries" ON)

 if ($ENV{CIRCLECI})
  message(STATUS "Build for CircieCI env, a few tests may be disabled")
@@ -584,10 +585,61 @@ endif()

 include_directories(${PROJECT_SOURCE_DIR})
 include_directories(${PROJECT_SOURCE_DIR}/include)
+
+if(USE_COROUTINES)
+  if(USE_FOLLY OR USE_FOLLY_LITE)
+    message(FATAL_ERROR "Please specify exactly one of USE_COROUTINES,"
+    " USE_FOLLY, and USE_FOLLY_LITE")
+  endif()
+  set(CMAKE_CXX_STANDARD 20)
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fcoroutines -Wno-maybe-uninitialized")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-deprecated")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-redundant-move")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-invalid-memory-model")
+  add_compile_definitions(USE_COROUTINES)
+  set(USE_FOLLY 1)
+endif()
+
 if(USE_FOLLY)
-  include_directories(${PROJECT_SOURCE_DIR}/third-party/folly)
-  add_definitions(-DUSE_FOLLY -DFOLLY_NO_CONFIG)
-  list(APPEND THIRDPARTY_LIBS glog)
+  if(USE_FOLLY_LITE)
+    message(FATAL_ERROR "Please specify one of USE_FOLLY or USE_FOLLY_LITE")
+  endif()
+  if(ROCKSDB_BUILD_SHARED)
+    message(FATAL_ERROR "Cannot build RocksDB shared library with folly")
+  endif()
+  set(ROCKSDB_BUILD_SHARED OFF)
+  set(GFLAGS_SHARED FALSE)
+  find_package(folly)
+  # If cmake could not find the folly-config.cmake file, fall back
+  # to looking in third-party/folly for folly and its dependencies
+  if(NOT FOLLY_LIBRARIES)
+    exec_program(python3 ${PROJECT_SOURCE_DIR}/third-party/folly ARGS
+    build/fbcode_builder/getdeps.py show-inst-dir OUTPUT_VARIABLE
+    FOLLY_INST_PATH)
+    exec_program(ls ARGS -d ${FOLLY_INST_PATH}/../boost* OUTPUT_VARIABLE
+    BOOST_INST_PATH)
+    exec_program(ls ARGS -d ${FOLLY_INST_PATH}/../fmt* OUTPUT_VARIABLE
+    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)
+    if(EXISTS ${FMT_INST_PATH}/lib64)
+      set(fmt_DIR ${FMT_INST_PATH}/lib64/cmake/fmt)
+    else()
+      set(fmt_DIR ${FMT_INST_PATH}/lib/cmake/fmt)
+    endif()
+    set(gflags_DIR ${GFLAGS_INST_PATH}/lib/cmake/gflags)
+
+    exec_program(sed ARGS -i 's/gflags_shared//g'
+    ${FOLLY_INST_PATH}/lib/cmake/folly/folly-targets.cmake)
+
+    include(${FOLLY_INST_PATH}/lib/cmake/folly/folly-config.cmake)
+  endif()
+
+  add_compile_definitions(USE_FOLLY FOLLY_NO_CONFIG HAVE_CXX11_ATOMIC)
+  list(APPEND THIRDPARTY_LIBS Folly::folly)
+  set(FOLLY_LIBS Folly::folly)
+  set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -Wl,--copy-dt-needed-entries")
 endif()
 find_package(Threads REQUIRED)

@@ -605,6 +657,7 @@ set(SOURCES
        cache/lru_cache.cc
        cache/sharded_cache.cc
        db/arena_wrapped_db_iter.cc
+        db/blob/blob_contents.cc
        db/blob/blob_fetcher.cc
        db/blob/blob_file_addition.cc
        db/blob/blob_file_builder.cc
@@ -667,7 +720,7 @@ set(SOURCES
        db/merge_helper.cc
        db/merge_operator.cc
        db/output_validator.cc
-        db/periodic_work_scheduler.cc
+        db/periodic_task_scheduler.cc
        db/range_del_aggregator.cc
        db/range_tombstone_fragmenter.cc
        db/repair.cc
@@ -827,6 +880,7 @@ set(SOURCES
        util/slice.cc
        util/file_checksum_helper.cc
        util/status.cc
+        util/stderr_logger.cc
        util/string_util.cc
        util/thread_local.cc
        util/threadpool_imp.cc
@@ -981,7 +1035,7 @@ else()
    env/io_posix.cc)
 endif()

-if(USE_FOLLY)
+if(USE_FOLLY_LITE)
  list(APPEND SOURCES
    third-party/folly/folly/container/detail/F14Table.cpp
    third-party/folly/folly/detail/Futex.cpp
@@ -991,13 +1045,14 @@ if(USE_FOLLY)
    third-party/folly/folly/synchronization/AtomicNotification.cpp
    third-party/folly/folly/synchronization/DistributedMutex.cpp
    third-party/folly/folly/synchronization/ParkingLot.cpp)
+  include_directories(${PROJECT_SOURCE_DIR}/third-party/folly)
+  add_definitions(-DUSE_FOLLY -DFOLLY_NO_CONFIG)
+  list(APPEND THIRDPARTY_LIBS glog)
 endif()

 set(ROCKSDB_STATIC_LIB rocksdb${ARTIFACT_SUFFIX})
 set(ROCKSDB_SHARED_LIB rocksdb-shared${ARTIFACT_SUFFIX})

-option(ROCKSDB_BUILD_SHARED "Build shared versions of the RocksDB libraries" ON)
-

 if(WIN32)
  set(SYSTEM_LIBS ${SYSTEM_LIBS} shlwapi.lib rpcrt4.lib)
@@ -1008,31 +1063,36 @@ endif()
 set(ROCKSDB_PLUGIN_EXTERNS "")
 set(ROCKSDB_PLUGIN_BUILTINS "")
 message(STATUS "ROCKSDB PLUGINS TO BUILD ${ROCKSDB_PLUGINS}")
-list(APPEND PLUGINS ${ROCKSDB_PLUGINS})
 foreach(PLUGIN IN LISTS PLUGINS)
  set(PLUGIN_ROOT "${CMAKE_CURRENT_SOURCE_DIR}/plugin/${PLUGIN}/")
-  message("including rocksb plugin ${PLUGIN_ROOT}")
+  message(STATUS "PLUGIN ${PLUGIN} including rocksb plugin ${PLUGIN_ROOT}")
  set(PLUGINMKFILE "${PLUGIN_ROOT}${PLUGIN}.mk")
  if (NOT EXISTS ${PLUGINMKFILE})
-    message(FATAL_ERROR "Missing plugin makefile: ${PLUGINMKFILE}")
+    message(FATAL_ERROR "PLUGIN ${PLUGIN} Missing plugin makefile: ${PLUGINMKFILE}")
  endif()
  file(READ ${PLUGINMKFILE} PLUGINMK)
+
  string(REGEX MATCH "SOURCES = ([^\n]*)" FOO ${PLUGINMK})
  set(MK_SOURCES ${CMAKE_MATCH_1})
  separate_arguments(MK_SOURCES)
  foreach(MK_FILE IN LISTS MK_SOURCES)
    list(APPEND SOURCES "${PLUGIN_ROOT}${MK_FILE}")
+    message(STATUS "PLUGIN ${PLUGIN} Appending ${PLUGIN_ROOT}${MK_FILE} to SOURCES")
  endforeach()
+
  string(REGEX MATCH "_FUNC = ([^\n]*)" FOO ${PLUGINMK})
  if (NOT ${CMAKE_MATCH_1} STREQUAL "")
    string(APPEND ROCKSDB_PLUGIN_BUILTINS "{\"${PLUGIN}\", " ${CMAKE_MATCH_1} "},")
    string(APPEND ROCKSDB_PLUGIN_EXTERNS "int " ${CMAKE_MATCH_1} "(ROCKSDB_NAMESPACE::ObjectLibrary&, const std::string&); ")
  endif()
+
  string(REGEX MATCH "_LIBS = ([^\n]*)" FOO ${PLUGINMK})
-  if (NOT ${CMAKE_MATCH_1} STREQUAL "")
-    list(APPEND THIRDPARTY_LIBS "${CMAKE_MATCH_1}")
-  endif()
-  message("THIRDPARTY_LIBS=${THIRDPARTY_LIBS}")
+  separate_arguments(CMAKE_MATCH_1)
+  foreach(MK_LIB IN LISTS CMAKE_MATCH_1)
+    list(APPEND THIRDPARTY_LIBS "${MK_LIB}")
+  endforeach()
+  message(STATUS "PLUGIN ${PLUGIN} THIRDPARTY_LIBS=${THIRDPARTY_LIBS}")
+
  #TODO: We need to set any compile/link-time flags and add any link libraries
 endforeach()

@@ -1297,7 +1357,7 @@ if(WITH_TESTS)
        db/merge_test.cc
        db/options_file_test.cc
        db/perf_context_test.cc
-        db/periodic_work_scheduler_test.cc
+        db/periodic_task_scheduler_test.cc
        db/plain_table_db_test.cc
        db/seqno_time_test.cc
        db/prefix_test.cc
@@ -1424,7 +1484,7 @@ if(WITH_TESTS)
  add_custom_target(check COMMAND ${CMAKE_CTEST_COMMAND})
  set(TESTUTILLIB testutillib${ARTIFACT_SUFFIX})
  add_library(${TESTUTILLIB} STATIC ${TESTUTIL_SOURCE})
-  target_link_libraries(${TESTUTILLIB} ${ROCKSDB_LIB})
+  target_link_libraries(${TESTUTILLIB} ${ROCKSDB_LIB} ${FOLLY_LIBS})
  if(MSVC)
    set_target_properties(${TESTUTILLIB} PROPERTIES COMPILE_FLAGS "/Fd${CMAKE_CFG_INTDIR}/testutillib${ARTIFACT_SUFFIX}.pdb")
  endif()
@@ -1482,32 +1542,32 @@ if(WITH_BENCHMARK_TOOLS)
    cache/cache_bench.cc
    cache/cache_bench_tool.cc)
  target_link_libraries(cache_bench${ARTIFACT_SUFFIX}
-    ${ROCKSDB_LIB} ${GFLAGS_LIB})
+    ${ROCKSDB_LIB} ${GFLAGS_LIB} ${FOLLY_LIBS})

  add_executable(memtablerep_bench${ARTIFACT_SUFFIX}
    memtable/memtablerep_bench.cc)
  target_link_libraries(memtablerep_bench${ARTIFACT_SUFFIX}
-    ${ROCKSDB_LIB} ${GFLAGS_LIB})
+    ${ROCKSDB_LIB} ${GFLAGS_LIB} ${FOLLY_LIBS})

  add_executable(range_del_aggregator_bench${ARTIFACT_SUFFIX}
    db/range_del_aggregator_bench.cc)
  target_link_libraries(range_del_aggregator_bench${ARTIFACT_SUFFIX}
-    ${ROCKSDB_LIB} ${GFLAGS_LIB})
+    ${ROCKSDB_LIB} ${GFLAGS_LIB} ${FOLLY_LIBS})

  add_executable(table_reader_bench${ARTIFACT_SUFFIX}
    table/table_reader_bench.cc)
  target_link_libraries(table_reader_bench${ARTIFACT_SUFFIX}
-    ${ROCKSDB_LIB} testharness ${GFLAGS_LIB})
+    ${ROCKSDB_LIB} testharness ${GFLAGS_LIB} ${FOLLY_LIBS})

  add_executable(filter_bench${ARTIFACT_SUFFIX}
    util/filter_bench.cc)
  target_link_libraries(filter_bench${ARTIFACT_SUFFIX}
-    ${ROCKSDB_LIB} ${GFLAGS_LIB})
+    ${ROCKSDB_LIB} ${GFLAGS_LIB} ${FOLLY_LIBS})

  add_executable(hash_table_bench${ARTIFACT_SUFFIX}
    utilities/persistent_cache/hash_table_bench.cc)
  target_link_libraries(hash_table_bench${ARTIFACT_SUFFIX}
-    ${ROCKSDB_LIB} ${GFLAGS_LIB})
+    ${ROCKSDB_LIB} ${GFLAGS_LIB} ${FOLLY_LIBS})
 endif()

 option(WITH_TRACE_TOOLS "build with trace tools" ON)
@@ -1515,12 +1575,12 @@ if(WITH_TRACE_TOOLS)
  add_executable(block_cache_trace_analyzer_tool${ARTIFACT_SUFFIX}
    tools/block_cache_analyzer/block_cache_trace_analyzer_tool.cc)
  target_link_libraries(block_cache_trace_analyzer_tool${ARTIFACT_SUFFIX}
-    ${ROCKSDB_LIB} ${GFLAGS_LIB})
+    ${ROCKSDB_LIB} ${GFLAGS_LIB} ${FOLLY_LIBS})

  add_executable(trace_analyzer${ARTIFACT_SUFFIX}
    tools/trace_analyzer.cc)
  target_link_libraries(trace_analyzer${ARTIFACT_SUFFIX}
-    ${ROCKSDB_LIB} ${GFLAGS_LIB})
+    ${ROCKSDB_LIB} ${GFLAGS_LIB} ${FOLLY_LIBS})

 endif()

@@ -1,4 +1,101 @@
 # Rocksdb Change Log
+## 7.8.3 (11/29/2022)
+* Revert an internal change in 7.8.0 associated with some memory usage churn.
+
+## 7.8.2 (11/27/2022)
+### Behavior changes
+* Make best-efforts recovery verify SST unique ID before Version construction (#10962)
+* Fix failed memtable flush retry bug that could cause wrongly ordered updates, which would surface to writers as `Status::Corruption` in case of `force_consistency_checks=true` (default). It affects use cases that enable both parallel flush (`max_background_flushes > 1` or `max_background_jobs >= 8`) and non-default memtable count (`max_write_buffer_number > 2`).
+* Tiered Storage: fixed excessive keys written to penultimate level in non-debug builds.
+
+### Bug Fixes
+* Fixed a regression in scan for async_io. During seek, valid buffers were getting cleared causing a regression.
+* Fixed a performance regression in iterator where range tombstones after `iterate_upper_bound` is processed.
+
+## 7.8.1 (11/2/2022)
+### Bug Fixes
+* Fix memory corruption error in scans if async_io is enabled. Memory corruption happened if there is IOError while reading the data leading to empty buffer and other buffer already in progress of async read goes again for reading.
+
+## 7.8.0 (10/22/2022)
+### New Features
+* `DeleteRange()` now supports user-defined timestamp.
+* Provide support for async_io with tailing iterators when ReadOptions.tailing is enabled during scans.
+* Tiered Storage: allow data moving up from the last level to the penultimate level if the input level is penultimate level or above.
+* Added `DB::Properties::kFastBlockCacheEntryStats`, which is similar to `DB::Properties::kBlockCacheEntryStats`, except returns cached (stale) values in more cases to reduce overhead.
+* FIFO compaction now supports migrating from a multi-level DB via DB::Open(). During the migration phase, FIFO compaction picker will:
+* picks the sst file with the smallest starting key in the bottom-most non-empty level.
+* Note that during the migration phase, the file purge order will only be an approximation of "FIFO" as files in lower-level might sometime contain newer keys than files in upper-level.
+* Added an option `ignore_max_compaction_bytes_for_input` to ignore max_compaction_bytes limit when adding files to be compacted from input level. This should help reduce write amplification. The option is enabled by default.
+* Tiered Storage: allow data moving up from the last level even if it's a last level only compaction, as long as the penultimate level is empty.
+* Add a new option IOOptions.do_not_recurse that can be used by underlying file systems to skip recursing through sub directories and list only files in GetChildren API.
+* Add option `preserve_internal_time_seconds` to preserve the time information for the latest data. Which can be used to determine the age of data when `preclude_last_level_data_seconds` is enabled. The time information is attached with SST in table property `rocksdb.seqno.time.map` which can be parsed by tool ldb or sst_dump.
+
+### Bug Fixes
+* Fix a bug in io_uring_prep_cancel in AbortIO API for posix which expects sqe->addr to match with read request submitted and wrong paramter was being passed.
+* Fixed a regression in iterator performance when the entire DB is a single memtable introduced in #10449. The fix is in #10705 and #10716.
+* Fixed an optimistic transaction validation bug caused by DBImpl::GetLatestSequenceForKey() returning non-latest seq for merge (#10724).
+* Fixed a bug in iterator refresh which could segfault for DeleteRange users (#10739).
+* Fixed a bug causing manual flush with `flush_opts.wait=false` to stall when database has stopped all writes (#10001).
+* Fixed a bug in iterator refresh that was not freeing up SuperVersion, which could cause excessive resource pinniung (#10770).
+* Fixed a bug where RocksDB could be doing compaction endlessly when allow_ingest_behind is true and the bottommost level is not filled (#10767).
+* Fixed a memory safety bug in experimental HyperClockCache (#10768)
+* Fixed some cases where `ldb update_manifest` and `ldb unsafe_remove_sst_file` are not usable because they were requiring the DB files to match the existing manifest state (before updating the manifest to match a desired state).
+
+### Performance Improvements
+* Try to align the compaction output file boundaries to the next level ones, which can reduce more than 10% compaction load for the default level compaction. The feature is enabled by default, to disable, set `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size` to false. As a side effect, it can create SSTs larger than the target_file_size (capped at 2x target_file_size) or smaller files.
+* Improve RoundRobin TTL compaction, which is going to be the same as normal RoundRobin compaction to move the compaction cursor.
+* Fix a small CPU regression caused by a change that UserComparatorWrapper was made Customizable, because Customizable itself has small CPU overhead for initialization.
+* Fixed an iterator performance regression for delete range users when scanning through a consecutive sequence of range tombstones (#10877).
+
+### Behavior Changes
+* Sanitize min_write_buffer_number_to_merge to 1 if atomic flush is enabled to prevent unexpected data loss when WAL is disabled in a multi-column-family setting (#10773).
+
+### Public API changes
+* Make kXXH3 checksum the new default, because it is faster on common hardware, especially with kCRC32c affected by a performance bug in some versions of clang (https://github.com/facebook/rocksdb/issues/9891). DBs written with this new setting can be read by RocksDB 6.27 and newer.
+* Refactor the classes, APIs and data structures for block cache tracing to allow a user provided trace writer to be used. Introduced an abstract BlockCacheTraceWriter class that takes a structured BlockCacheTraceRecord. The BlockCacheTraceWriter implementation can then format and log the record in whatever way it sees fit. The default BlockCacheTraceWriterImpl does file tracing using a user provided TraceWriter. More details in rocksdb/includb/block_cache_trace_writer.h.
+
+## 7.7.0 (09/18/2022)
+### Bug Fixes
+* Fixed a hang when an operation such as `GetLiveFiles` or `CreateNewBackup` is asked to trigger and wait for memtable flush on a read-only DB. Such indirect requests for memtable flush are now ignored on a read-only DB.
+* Fixed bug where `FlushWAL(true /* sync */)` (used by `GetLiveFilesStorageInfo()`, which is used by checkpoint and backup) could cause parallel writes at the tail of a WAL file to never be synced.
+* Fix periodic_task unable to re-register the same task type, which may cause `SetOptions()` fail to update periodical_task time like: `stats_dump_period_sec`, `stats_persist_period_sec`.
+* Fixed a bug in the rocksdb.prefetched.bytes.discarded stat. It was counting the prefetch buffer size, rather than the actual number of bytes discarded from the buffer.
+* Fix bug where the directory containing CURRENT can left unsynced after CURRENT is updated to point to the latest MANIFEST, which leads to risk of unsync data loss of CURRENT.
+* Update rocksdb.multiget.io.batch.size stat in non-async MultiGet as well.
+* Fix a bug in key range overlap checking with concurrent compactions when user-defined timestamp is enabled. User-defined timestamps should be EXCLUDED when checking if two ranges overlap.
+* Fixed a bug where the blob cache prepopulating logic did not consider the secondary cache (see #10603).
+* Fixed the rocksdb.num.sst.read.per.level, rocksdb.num.index.and.filter.blocks.read.per.level and rocksdb.num.level.read.per.multiget stats in the MultiGet coroutines
+
+### Public API changes
+* Add `rocksdb_column_family_handle_get_id`, `rocksdb_column_family_handle_get_name` to get name, id of column family in C API
+* Add a new stat rocksdb.async.prefetch.abort.micros to measure time spent waiting for async prefetch reads to abort
+
+### Java API Changes
+* Add CompactionPriority.RoundRobin.
+* Revert to using the default metadata charge policy when creating an LRU cache via the Java API.
+
+### Behavior Change
+* DBOptions::verify_sst_unique_id_in_manifest is now an on-by-default feature that verifies SST file identity whenever they are opened by a DB, rather than only at DB::Open time.
+* Right now, when the option migration tool (OptionChangeMigration()) migrates to FIFO compaction, it compacts all the data into one single SST file and move to L0. This might create a problem for some users: the giant file may be soon deleted to satisfy max_table_files_size, and might cayse the DB to be almost empty. We change the behavior so that the files are cut to be smaller, but these files might not follow the data insertion order. With the change, after the migration, migrated data might not be dropped by insertion order by FIFO compaction.
+* When a block is firstly found from `CompressedSecondaryCache`, we just insert a dummy block into the primary cache and don’t erase the block from `CompressedSecondaryCache`. A standalone handle is returned to the caller. Only if the block is found again from `CompressedSecondaryCache` before the dummy block is evicted, we erase the block from `CompressedSecondaryCache` and insert it into the primary cache.
+* When a block is firstly evicted from the primary cache to `CompressedSecondaryCache`, we just insert a dummy block in `CompressedSecondaryCache`. Only if it is evicted again before the dummy block is evicted from the cache, it is treated as a hot block and is inserted into `CompressedSecondaryCache`.
+* Improved the estimation of memory used by cached blobs by taking into account the size of the object owning the blob value and also the allocator overhead if `malloc_usable_size` is available (see #10583).
+* Blob values now have their own category in the cache occupancy statistics, as opposed to being lumped into the "Misc" bucket (see #10601).
+* Change the optimize_multiget_for_io experimental ReadOptions flag to default on.
+
+### New Features
+*  RocksDB does internal auto prefetching if it notices 2 sequential reads if readahead_size is not specified. New option `num_file_reads_for_auto_readahead` is added in BlockBasedTableOptions which indicates after how many sequential reads internal auto prefetching should be start (default is 2).
+* Added new perf context counters `block_cache_standalone_handle_count`, `block_cache_real_handle_count`,`compressed_sec_cache_insert_real_count`, `compressed_sec_cache_insert_dummy_count`, `compressed_sec_cache_uncompressed_bytes`, and `compressed_sec_cache_compressed_bytes`.
+* Memory for blobs which are to be inserted into the blob cache is now allocated using the cache's allocator (see #10628 and #10647).
+* HyperClockCache is an experimental, lock-free Cache alternative for block cache that offers much improved CPU efficiency under high parallel load or high contention, with some caveats. As much as 4.5x higher ops/sec vs. LRUCache has been seen in db_bench under high parallel load.
+* `CompressedSecondaryCacheOptions::enable_custom_split_merge` is added for enabling the custom split and merge feature, which split the compressed value into chunks so that they may better fit jemalloc bins.
+
+### Performance Improvements
+* Iterator performance is improved for `DeleteRange()` users. Internally, iterator will skip to the end of a range tombstone when possible, instead of looping through each key and check individually if a key is range deleted.
+* Eliminated some allocations and copies in the blob read path. Also, `PinnableSlice` now only points to the blob value and pins the backing resource (cache entry or buffer) in all cases, instead of containing a copy of the blob value. See #10625 and #10647.
+* In case of scans with async_io enabled, few optimizations have been added to issue more asynchronous requests in parallel in order to avoid synchronous prefetching.
+* `DeleteRange()` users should see improvement in get/iterator performance from mutable memtable (see #10547).
+
 ## 7.6.0 (08/19/2022)
 ### New Features
 * Added `prepopulate_blob_cache` to ColumnFamilyOptions. If enabled, prepopulate warm/hot blobs which are already in memory into blob cache at the time of flush. On a flush, the blob that is in memory (in memtables) get flushed to the device. If using Direct IO, additional IO is incurred to read this blob back into memory again, which is avoided by enabling this option. This further helps if the workload exhibits high temporal locality, where most of the reads go to recently written data. This also helps in case of the remote file system since it involves network traffic and higher latencies.
@@ -6,7 +6,9 @@ This is the list of all known third-party language bindings for RocksDB. If some
    * http://pyrocksdb.readthedocs.org/en/latest/ (unmaintained)
 * Perl - https://metacpan.org/pod/RocksDB
 * Node.js - https://npmjs.org/package/rocksdb
-* Go - https://github.com/tecbot/gorocksdb
+* Go 
+  * https://github.com/linxGnu/grocksdb
+  * https://github.com/tecbot/gorocksdb (unmaintained)
 * Ruby - http://rubygems.org/gems/rocksdb-ruby
 * Haskell - https://hackage.haskell.org/package/rocksdb-haskell
 * PHP - https://github.com/Photonios/rocksdb-php
@@ -139,9 +139,14 @@ endif
 GIT_COMMAND ?= git
 ifeq ($(USE_COROUTINES), 1)
 	USE_FOLLY = 1
-	OPT += -DUSE_COROUTINES
+	# glog/logging.h requires HAVE_CXX11_ATOMIC
+	OPT += -DUSE_COROUTINES -DHAVE_CXX11_ATOMIC
 	ROCKSDB_CXX_STANDARD = c++2a
 	USE_RTTI = 1
+ifneq ($(USE_CLANG), 1)
+	ROCKSDB_CXX_STANDARD = c++20
+	PLATFORM_CXXFLAGS += -fcoroutines
+endif
 endif

 # if we're compiling for release, compile without debug code (-DNDEBUG)
@@ -185,6 +190,16 @@ ifeq ($(USE_LTO), 1)
 	LDFLAGS += -flto -fuse-linker-plugin
 endif

+# `COERCE_CONTEXT_SWITCH=1` will inject spurious wakeup and
+# random length of sleep or context switch at critical
+# points (e.g, before acquring db mutex) in RocksDB.
+# In this way, it coerces as many excution orders as possible in the hope of
+# exposing the problematic excution order
+COERCE_CONTEXT_SWITCH ?= 0
+ifeq ($(COERCE_CONTEXT_SWITCH), 1)
+OPT += -DCOERCE_CONTEXT_SWITCH
+endif
+
 #-----------------------------------------------
 include src.mk

@@ -235,6 +250,7 @@ dummy := $(shell (export ROCKSDB_ROOT="$(CURDIR)"; \
                  export USE_CLANG="$(USE_CLANG)"; \
                  export LIB_MODE="$(LIB_MODE)"; \
 		  export ROCKSDB_CXX_STANDARD="$(ROCKSDB_CXX_STANDARD)"; \
+		  export USE_FOLLY="$(USE_FOLLY)"; \
                  "$(CURDIR)/build_tools/build_detect_platform" "$(CURDIR)/make_config.mk"))
 # this file is generated by the previous line to set build flags and sources
 include make_config.mk
@@ -415,6 +431,10 @@ ifndef DISABLE_JEMALLOC
 			PLATFORM_CXXFLAGS += -DUSE_JEMALLOC
 			PLATFORM_CCFLAGS  += -DUSE_JEMALLOC
 		endif
+		ifeq ($(USE_FOLLY_LITE),1)
+			PLATFORM_CXXFLAGS += -DUSE_JEMALLOC
+			PLATFORM_CCFLAGS  += -DUSE_JEMALLOC
+		endif
 	endif
 	ifdef WITH_JEMALLOC_FLAG
 		PLATFORM_LDFLAGS += -ljemalloc
@@ -448,10 +468,57 @@ endif

 # This provides a Makefile simulation of a Meta-internal folly integration.
 # It is not validated for general use.
+#
+# USE_FOLLY links the build targets with libfolly.a. The latter could be
+# built using 'make build_folly', or built externally and specified in
+# the CXXFLAGS and EXTRA_LDFLAGS env variables. The build_detect_platform
+# script tries to detect if an external folly dependency has been specified.
+# If not, it exports FOLLY_PATH to the path of the installed Folly and
+# dependency libraries.
+#
+# USE_FOLLY_LITE cherry picks source files from Folly to include in the
+# RocksDB library. Its faster and has fewer dependencies on 3rd party
+# libraries, but with limited functionality. For example, coroutine
+# functionality is not available.
 ifeq ($(USE_FOLLY),1)
-	ifeq (,$(FOLLY_DIR))
-		FOLLY_DIR = ./third-party/folly
+ifeq ($(USE_FOLLY_LITE),1)
+$(error Please specify only one of USE_FOLLY and USE_FOLLY_LITE)
+endif
+ifneq ($(strip $(FOLLY_PATH)),)
+	BOOST_PATH = $(shell (ls -d $(FOLLY_PATH)/../boost*))
+	DBL_CONV_PATH = $(shell (ls -d $(FOLLY_PATH)/../double-conversion*))
+	GFLAGS_PATH = $(shell (ls -d $(FOLLY_PATH)/../gflags*))
+	GLOG_PATH = $(shell (ls -d $(FOLLY_PATH)/../glog*))
+	LIBEVENT_PATH = $(shell (ls -d $(FOLLY_PATH)/../libevent*))
+	XZ_PATH = $(shell (ls -d $(FOLLY_PATH)/../xz*))
+	LIBSODIUM_PATH = $(shell (ls -d $(FOLLY_PATH)/../libsodium*))
+	FMT_PATH = $(shell (ls -d $(FOLLY_PATH)/../fmt*))
+
+	# For some reason, glog and fmt libraries are under either lib or lib64
+	GLOG_LIB_PATH = $(shell (ls -d $(GLOG_PATH)/lib*))
+	FMT_LIB_PATH = $(shell (ls -d $(FMT_PATH)/lib*))
+
+	# AIX: pre-defined system headers are surrounded by an extern "C" block
+	ifeq ($(PLATFORM), OS_AIX)
+		PLATFORM_CCFLAGS += -I$(BOOST_PATH)/include -I$(DBL_CONV_PATH)/include -I$(GLOG_PATH)/include -I$(LIBEVENT_PATH)/include -I$(XZ_PATH)/include -I$(LIBSODIUM_PATH)/include -I$(FOLLY_PATH)/include -I$(FMT_PATH)/include
+		PLATFORM_CXXFLAGS += -I$(BOOST_PATH)/include -I$(DBL_CONV_PATH)/include -I$(GLOG_PATH)/include -I$(LIBEVENT_PATH)/include -I$(XZ_PATH)/include -I$(LIBSODIUM_PATH)/include -I$(FOLLY_PATH)/include -I$(FMT_PATH)/include
+	else
+		PLATFORM_CCFLAGS += -isystem $(BOOST_PATH)/include -isystem $(DBL_CONV_PATH)/include -isystem $(GLOG_PATH)/include -isystem $(LIBEVENT_PATH)/include -isystem $(XZ_PATH)/include -isystem $(LIBSODIUM_PATH)/include -isystem $(FOLLY_PATH)/include -isystem $(FMT_PATH)/include
+		PLATFORM_CXXFLAGS += -isystem $(BOOST_PATH)/include -isystem $(DBL_CONV_PATH)/include -isystem $(GLOG_PATH)/include -isystem $(LIBEVENT_PATH)/include -isystem $(XZ_PATH)/include -isystem $(LIBSODIUM_PATH)/include -isystem $(FOLLY_PATH)/include -isystem $(FMT_PATH)/include
 	endif
+
+	# Add -ldl at the end as gcc resolves a symbol in a library by searching only in libraries specified later
+	# in the command line
+	PLATFORM_LDFLAGS += $(FOLLY_PATH)/lib/libfolly.a $(BOOST_PATH)/lib/libboost_context.a $(BOOST_PATH)/lib/libboost_filesystem.a $(BOOST_PATH)/lib/libboost_atomic.a $(BOOST_PATH)/lib/libboost_program_options.a $(BOOST_PATH)/lib/libboost_regex.a $(BOOST_PATH)/lib/libboost_system.a $(BOOST_PATH)/lib/libboost_thread.a $(DBL_CONV_PATH)/lib/libdouble-conversion.a $(FMT_LIB_PATH)/libfmt.a $(GLOG_LIB_PATH)/libglog.so $(GFLAGS_PATH)/lib/libgflags.so.2.2 $(LIBEVENT_PATH)/lib/libevent-2.1.so -ldl
+	PLATFORM_LDFLAGS += -Wl,-rpath=$(GFLAGS_PATH)/lib -Wl,-rpath=$(GLOG_LIB_PATH) -Wl,-rpath=$(LIBEVENT_PATH)/lib -Wl,-rpath=$(LIBSODIUM_PATH)/lib -Wl,-rpath=$(LIBEVENT_PATH)/lib
+endif
+	PLATFORM_CCFLAGS += -DUSE_FOLLY -DFOLLY_NO_CONFIG
+	PLATFORM_CXXFLAGS += -DUSE_FOLLY -DFOLLY_NO_CONFIG
+endif
+
+ifeq ($(USE_FOLLY_LITE),1)
+	# Path to the Folly source code and include files
+	FOLLY_DIR = ./third-party/folly
 	# AIX: pre-defined system headers are surrounded by an extern "C" block
 	ifeq ($(PLATFORM), OS_AIX)
 		PLATFORM_CCFLAGS += -I$(FOLLY_DIR)
@@ -550,7 +617,7 @@ LIB_OBJECTS += $(patsubst %.c, $(OBJ_DIR)/%.o, $(LIB_SOURCES_C))
 LIB_OBJECTS += $(patsubst %.S, $(OBJ_DIR)/%.o, $(LIB_SOURCES_ASM))
 endif

-ifeq ($(USE_FOLLY),1)
+ifeq ($(USE_FOLLY_LITE),1)
  LIB_OBJECTS += $(patsubst %.cpp, $(OBJ_DIR)/%.o, $(FOLLY_SOURCES))
 endif

@@ -591,7 +658,7 @@ 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' | egrep -v 'port/|plugin/|lua/|range_tree/')
+	DEV_HEADERS := $(shell $(FIND) $(DEV_HEADER_DIRS) -type f -name '*.h' | grep -E -v 'port/|plugin/|lua/|range_tree/')
 else
 	DEV_HEADERS :=
 endif
@@ -728,9 +795,9 @@ TOOLS_LIBRARY=$(STATIC_TOOLS_LIBRARY)
 endif
 STRESS_LIBRARY=$(STATIC_STRESS_LIBRARY)

-ROCKSDB_MAJOR = $(shell egrep "ROCKSDB_MAJOR.[0-9]" include/rocksdb/version.h | cut -d ' ' -f 3)
-ROCKSDB_MINOR = $(shell egrep "ROCKSDB_MINOR.[0-9]" include/rocksdb/version.h | cut -d ' ' -f 3)
-ROCKSDB_PATCH = $(shell egrep "ROCKSDB_PATCH.[0-9]" include/rocksdb/version.h | cut -d ' ' -f 3)
+ROCKSDB_MAJOR = $(shell grep -E "ROCKSDB_MAJOR.[0-9]" include/rocksdb/version.h | cut -d ' ' -f 3)
+ROCKSDB_MINOR = $(shell grep -E "ROCKSDB_MINOR.[0-9]" include/rocksdb/version.h | cut -d ' ' -f 3)
+ROCKSDB_PATCH = $(shell grep -E "ROCKSDB_PATCH.[0-9]" include/rocksdb/version.h | cut -d ' ' -f 3)

 # If NO_UPDATE_BUILD_VERSION is set we don't update util/build_version.cc, but
 # the file needs to already exist or else the build will fail
@@ -1765,7 +1832,7 @@ memtable_list_test: $(OBJ_DIR)/db/memtable_list_test.o $(TEST_LIBRARY) $(LIBRARY
 write_callback_test: $(OBJ_DIR)/db/write_callback_test.o $(TEST_LIBRARY) $(LIBRARY)
 	$(AM_LINK)

-heap_test: $(OBJ_DIR)/util/heap_test.o $(GTEST)
+heap_test: $(OBJ_DIR)/util/heap_test.o $(TEST_LIBRARY) $(LIBRARY)
 	$(AM_LINK)

 point_lock_manager_test: utilities/transactions/lock/point/point_lock_manager_test.o $(TEST_LIBRARY) $(LIBRARY)
@@ -1882,7 +1949,7 @@ blob_garbage_meter_test: $(OBJ_DIR)/db/blob/blob_garbage_meter_test.o $(TEST_LIB
 timer_test: $(OBJ_DIR)/util/timer_test.o $(TEST_LIBRARY) $(LIBRARY)
 	$(AM_LINK)

-periodic_work_scheduler_test: $(OBJ_DIR)/db/periodic_work_scheduler_test.o $(TEST_LIBRARY) $(LIBRARY)
+periodic_task_scheduler_test: $(OBJ_DIR)/db/periodic_task_scheduler_test.o $(TEST_LIBRARY) $(LIBRARY)
 	$(AM_LINK)

 testutil_test: $(OBJ_DIR)/test_util/testutil_test.o $(TEST_LIBRARY) $(LIBRARY)
@@ -2371,13 +2438,26 @@ checkout_folly:
 	@# 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 not needed if using FBCODE compiler config
+	@# NOTE: this hack is only needed if building using USE_FOLLY_LITE
 	perl -pi -e 's/^(#include <boost)/\/\/$$1/' third-party/folly/folly/functional/Invoke.h
 	@# NOTE: this hack is required for clang in some cases
 	perl -pi -e 's/int rv = syscall/int rv = (int)syscall/' third-party/folly/folly/detail/Futex.cpp
 	@# NOTE: this hack is required for gcc in some cases
 	perl -pi -e 's/(__has_include.<experimental.memory_resource>.)/__cpp_rtti && $$1/' third-party/folly/folly/memory/MemoryResource.h

+build_folly:
+	FOLLY_INST_PATH=`cd third-party/folly; $(PYTHON) build/fbcode_builder/getdeps.py show-inst-dir`; \
+	if [ "$$FOLLY_INST_PATH" ]; then \
+		rm -rf $${FOLLY_INST_PATH}/../../*; \
+	else \
+		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 && MAYBE_AVX2=`echo $(CXXFLAGS) | grep -o -- -DHAVE_AVX2 | sed 's/-DHAVE_AVX2/-mavx2/g' || true` && \
+		CXXFLAGS=" $$MAYBE_AVX2 -DHAVE_CXX11_ATOMIC " $(PYTHON) build/fbcode_builder/getdeps.py build --no-tests
+
 # ---------------------------------------------------------------------------
 #   Build size testing
 # ---------------------------------------------------------------------------
@@ -2462,7 +2542,7 @@ endif
 ifneq ($(SKIP_DEPENDS), 1)
 DEPFILES = $(patsubst %.cc, $(OBJ_DIR)/%.cc.d, $(ALL_SOURCES))
 DEPFILES+ = $(patsubst %.c, $(OBJ_DIR)/%.c.d, $(LIB_SOURCES_C) $(TEST_MAIN_SOURCES_C))
-ifeq ($(USE_FOLLY),1)
+ifeq ($(USE_FOLLY_LITE),1)
  DEPFILES +=$(patsubst %.cpp, $(OBJ_DIR)/%.cpp.d, $(FOLLY_SOURCES))
 endif
 endif
@@ -20,6 +20,7 @@ cpp_library_wrapper(name="rocksdb_lib", srcs=[
        "cache/lru_cache.cc",
        "cache/sharded_cache.cc",
        "db/arena_wrapped_db_iter.cc",
+        "db/blob/blob_contents.cc",
        "db/blob/blob_fetcher.cc",
        "db/blob/blob_file_addition.cc",
        "db/blob/blob_file_builder.cc",
@@ -82,7 +83,7 @@ cpp_library_wrapper(name="rocksdb_lib", srcs=[
        "db/merge_helper.cc",
        "db/merge_operator.cc",
        "db/output_validator.cc",
-        "db/periodic_work_scheduler.cc",
+        "db/periodic_task_scheduler.cc",
        "db/range_del_aggregator.cc",
        "db/range_tombstone_fragmenter.cc",
        "db/repair.cc",
@@ -252,6 +253,7 @@ cpp_library_wrapper(name="rocksdb_lib", srcs=[
        "util/ribbon_config.cc",
        "util/slice.cc",
        "util/status.cc",
+        "util/stderr_logger.cc",
        "util/string_util.cc",
        "util/thread_local.cc",
        "util/threadpool_imp.cc",
@@ -357,6 +359,7 @@ cpp_library_wrapper(name="rocksdb_whole_archive_lib", srcs=[
        "cache/lru_cache.cc",
        "cache/sharded_cache.cc",
        "db/arena_wrapped_db_iter.cc",
+        "db/blob/blob_contents.cc",
        "db/blob/blob_fetcher.cc",
        "db/blob/blob_file_addition.cc",
        "db/blob/blob_file_builder.cc",
@@ -419,7 +422,7 @@ cpp_library_wrapper(name="rocksdb_whole_archive_lib", srcs=[
        "db/merge_helper.cc",
        "db/merge_operator.cc",
        "db/output_validator.cc",
-        "db/periodic_work_scheduler.cc",
+        "db/periodic_task_scheduler.cc",
        "db/range_del_aggregator.cc",
        "db/range_tombstone_fragmenter.cc",
        "db/repair.cc",
@@ -589,6 +592,7 @@ cpp_library_wrapper(name="rocksdb_whole_archive_lib", srcs=[
        "util/ribbon_config.cc",
        "util/slice.cc",
        "util/status.cc",
+        "util/stderr_logger.cc",
        "util/string_util.cc",
        "util/thread_local.cc",
        "util/threadpool_imp.cc",
@@ -5598,8 +5602,8 @@ cpp_unittest_wrapper(name="perf_context_test",
            extra_compiler_flags=[])


-cpp_unittest_wrapper(name="periodic_work_scheduler_test",
-            srcs=["db/periodic_work_scheduler_test.cc"],
+cpp_unittest_wrapper(name="periodic_task_scheduler_test",
+            srcs=["db/periodic_task_scheduler_test.cc"],
            deps=[":rocksdb_test_lib"],
            extra_compiler_flags=[])

@@ -1,19 +1,18 @@
 #!/usr/bin/env python3
 # Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-from __future__ import unicode_literals
+from __future__ import absolute_import, division, print_function, unicode_literals
+
 try:
    from builtins import str
 except ImportError:
    from __builtin__ import str
-from targets_builder import TARGETSBuilder
+import fnmatch
 import json
 import os
-import fnmatch
 import sys

+from targets_builder import TARGETSBuilder
+
 from util import ColorString

 # This script generates TARGETS file for Buck.
@@ -44,13 +43,13 @@ def parse_src_mk(repo_path):
    src_files = {}
    for line in open(src_mk):
        line = line.strip()
-        if len(line) == 0 or line[0] == '#':
+        if len(line) == 0 or line[0] == "#":
            continue
-        if '=' in line:
-            current_src = line.split('=')[0].strip()
+        if "=" in line:
+            current_src = line.split("=")[0].strip()
            src_files[current_src] = []
-        elif '.c' in line:
-            src_path = line.split('\\')[0].strip()
+        elif ".c" in line:
+            src_path = line.split("\\")[0].strip()
            src_files[current_src].append(src_path)
    return src_files

@@ -58,14 +57,16 @@ def parse_src_mk(repo_path):
 # get all .cc / .c files
 def get_cc_files(repo_path):
    cc_files = []
-    for root, dirnames, filenames in os.walk(repo_path):  # noqa: B007 T25377293 Grandfathered in
-        root = root[(len(repo_path) + 1):]
+    for root, _dirnames, filenames in os.walk(
+        repo_path
+    ):  # noqa: B007 T25377293 Grandfathered in
+        root = root[(len(repo_path) + 1) :]
        if "java" in root:
            # Skip java
            continue
-        for filename in fnmatch.filter(filenames, '*.cc'):
+        for filename in fnmatch.filter(filenames, "*.cc"):
            cc_files.append(os.path.join(root, filename))
-        for filename in fnmatch.filter(filenames, '*.c'):
+        for filename in fnmatch.filter(filenames, "*.c"):
            cc_files.append(os.path.join(root, filename))
    return cc_files

@@ -93,14 +94,10 @@ def get_non_parallel_tests(repo_path):

    return s

+
 # Parse extra dependencies passed by user from command line
 def get_dependencies():
-    deps_map = {
-        '': {
-            'extra_deps': [],
-            'extra_compiler_flags': []
-        }
-    }
+    deps_map = {"": {"extra_deps": [], "extra_compiler_flags": []}}
    if len(sys.argv) < 2:
        return deps_map

@@ -111,6 +108,7 @@ def get_dependencies():
                v = encode_dict(v)
            rv[k] = v
        return rv
+
    extra_deps = json.loads(sys.argv[1], object_hook=encode_dict)
    for target_alias, deps in extra_deps.items():
        deps_map[target_alias] = deps
@@ -143,73 +141,73 @@ def generate_targets(repo_path, deps_map):
        "rocksdb_lib",
        src_mk["LIB_SOURCES"] +
        # always add range_tree, it's only excluded on ppc64, which we don't use internally
-        src_mk["RANGE_TREE_SOURCES"] +
-        src_mk["TOOL_LIB_SOURCES"],
+        src_mk["RANGE_TREE_SOURCES"] + src_mk["TOOL_LIB_SOURCES"],
        deps=[
            "//folly/container:f14_hash",
            "//folly/experimental/coro:blocking_wait",
            "//folly/experimental/coro:collect",
            "//folly/experimental/coro:coroutine",
            "//folly/experimental/coro:task",
-            "//folly/synchronization:distributed_mutex"])
+            "//folly/synchronization:distributed_mutex",
+        ],
+    )
    # rocksdb_whole_archive_lib
    TARGETS.add_library(
        "rocksdb_whole_archive_lib",
        src_mk["LIB_SOURCES"] +
        # always add range_tree, it's only excluded on ppc64, which we don't use internally
-        src_mk["RANGE_TREE_SOURCES"] +
-        src_mk["TOOL_LIB_SOURCES"],
+        src_mk["RANGE_TREE_SOURCES"] + src_mk["TOOL_LIB_SOURCES"],
        deps=[
            "//folly/container:f14_hash",
            "//folly/experimental/coro:blocking_wait",
            "//folly/experimental/coro:collect",
            "//folly/experimental/coro:coroutine",
            "//folly/experimental/coro:task",
-            "//folly/synchronization:distributed_mutex"],
+            "//folly/synchronization:distributed_mutex",
+        ],
        headers=None,
        extra_external_deps="",
-        link_whole=True)
+        link_whole=True,
+    )
    # rocksdb_test_lib
    TARGETS.add_library(
        "rocksdb_test_lib",
-        src_mk.get("MOCK_LIB_SOURCES", []) +
-        src_mk.get("TEST_LIB_SOURCES", []) +
-        src_mk.get("EXP_LIB_SOURCES", []) +
-        src_mk.get("ANALYZER_LIB_SOURCES", []),
+        src_mk.get("MOCK_LIB_SOURCES", [])
+        + src_mk.get("TEST_LIB_SOURCES", [])
+        + src_mk.get("EXP_LIB_SOURCES", [])
+        + src_mk.get("ANALYZER_LIB_SOURCES", []),
        [":rocksdb_lib"],
-        extra_test_libs=True
-        )
+        extra_test_libs=True,
+    )
    # rocksdb_tools_lib
    TARGETS.add_library(
        "rocksdb_tools_lib",
-        src_mk.get("BENCH_LIB_SOURCES", []) +
-        src_mk.get("ANALYZER_LIB_SOURCES", []) +
-        ["test_util/testutil.cc"],
-        [":rocksdb_lib"])
+        src_mk.get("BENCH_LIB_SOURCES", [])
+        + src_mk.get("ANALYZER_LIB_SOURCES", [])
+        + ["test_util/testutil.cc"],
+        [":rocksdb_lib"],
+    )
    # rocksdb_cache_bench_tools_lib
    TARGETS.add_library(
        "rocksdb_cache_bench_tools_lib",
        src_mk.get("CACHE_BENCH_LIB_SOURCES", []),
-        [":rocksdb_lib"])
+        [":rocksdb_lib"],
+    )
    # rocksdb_stress_lib
    TARGETS.add_rocksdb_library(
        "rocksdb_stress_lib",
        src_mk.get("ANALYZER_LIB_SOURCES", [])
-        + src_mk.get('STRESS_LIB_SOURCES', [])
-        + ["test_util/testutil.cc"])
+        + src_mk.get("STRESS_LIB_SOURCES", [])
+        + ["test_util/testutil.cc"],
+    )
    # db_stress binary
-    TARGETS.add_binary("db_stress",
-                       ["db_stress_tool/db_stress.cc"],
-                       [":rocksdb_stress_lib"])
+    TARGETS.add_binary(
+        "db_stress", ["db_stress_tool/db_stress.cc"], [":rocksdb_stress_lib"]
+    )
    # bench binaries
    for src in src_mk.get("MICROBENCH_SOURCES", []):
-        name =  src.rsplit('/',1)[1].split('.')[0] if '/' in src else src.split('.')[0]
-        TARGETS.add_binary(
-            name,
-            [src],
-            [],
-            extra_bench_libs=True
-            )
+        name = src.rsplit("/", 1)[1].split(".")[0] if "/" in src else src.split(".")[0]
+        TARGETS.add_binary(name, [src], [], extra_bench_libs=True)
    print("Extra dependencies:\n{0}".format(json.dumps(deps_map)))

    # Dictionary test executable name -> relative source file path
@@ -219,7 +217,7 @@ def generate_targets(repo_path, deps_map):
    # are more than one .c test file, we need to extend
    # TARGETS.add_c_test() to include other C tests too.
    for test_src in src_mk.get("TEST_MAIN_SOURCES_C", []):
-        if test_src != 'db/c_test.c':
+        if test_src != "db/c_test.c":
            print("Don't know how to deal with " + test_src)
            return False
    TARGETS.add_c_test()
@@ -229,7 +227,7 @@ def generate_targets(repo_path, deps_map):
            fast_fancy_bench_config_list = json.load(json_file)
            for config_dict in fast_fancy_bench_config_list:
                clean_benchmarks = {}
-                benchmarks = config_dict['benchmarks']
+                benchmarks = config_dict["benchmarks"]
                for binary, benchmark_dict in benchmarks.items():
                    clean_benchmarks[binary] = {}
                    for benchmark, overloaded_metric_list in benchmark_dict.items():
@@ -237,13 +235,20 @@ def generate_targets(repo_path, deps_map):
                        for metric in overloaded_metric_list:
                            if not isinstance(metric, dict):
                                clean_benchmarks[binary][benchmark].append(metric)
-                TARGETS.add_fancy_bench_config(config_dict['name'], clean_benchmarks, False, config_dict['expected_runtime_one_iter'], config_dict['sl_iterations'], config_dict['regression_threshold'])
+                TARGETS.add_fancy_bench_config(
+                    config_dict["name"],
+                    clean_benchmarks,
+                    False,
+                    config_dict["expected_runtime_one_iter"],
+                    config_dict["sl_iterations"],
+                    config_dict["regression_threshold"],
+                )

        with open(f"{repo_path}/buckifier/bench-slow.json") as json_file:
            slow_fancy_bench_config_list = json.load(json_file)
            for config_dict in slow_fancy_bench_config_list:
                clean_benchmarks = {}
-                benchmarks = config_dict['benchmarks']
+                benchmarks = config_dict["benchmarks"]
                for binary, benchmark_dict in benchmarks.items():
                    clean_benchmarks[binary] = {}
                    for benchmark, overloaded_metric_list in benchmark_dict.items():
@@ -252,7 +257,14 @@ def generate_targets(repo_path, deps_map):
                            if not isinstance(metric, dict):
                                clean_benchmarks[binary][benchmark].append(metric)
            for config_dict in slow_fancy_bench_config_list:
-                TARGETS.add_fancy_bench_config(config_dict['name']+"_slow", clean_benchmarks, True, config_dict['expected_runtime_one_iter'], config_dict['sl_iterations'], config_dict['regression_threshold'])
+                TARGETS.add_fancy_bench_config(
+                    config_dict["name"] + "_slow",
+                    clean_benchmarks,
+                    True,
+                    config_dict["expected_runtime_one_iter"],
+                    config_dict["sl_iterations"],
+                    config_dict["regression_threshold"],
+                )
    # it is better servicelab experiments break
    # than rocksdb github ci
    except Exception:
@@ -261,7 +273,7 @@ def generate_targets(repo_path, deps_map):
    TARGETS.add_test_header()

    for test_src in src_mk.get("TEST_MAIN_SOURCES", []):
-        test = test_src.split('.c')[0].strip().split('/')[-1].strip()
+        test = test_src.split(".c")[0].strip().split("/")[-1].strip()
        test_source_map[test] = test_src
        print("" + test + " " + test_src)

@@ -271,23 +283,29 @@ def generate_targets(repo_path, deps_map):
                print(ColorString.warning("Failed to get test name for %s" % test_src))
                continue

-            test_target_name = \
-                test if not target_alias else test + "_" + target_alias
+            test_target_name = test if not target_alias else test + "_" + target_alias

            if test in _EXPORTED_TEST_LIBS:
                test_library = "%s_lib" % test_target_name
-                TARGETS.add_library(test_library, [test_src], deps=[":rocksdb_test_lib"], extra_test_libs=True)
+                TARGETS.add_library(
+                    test_library,
+                    [test_src],
+                    deps=[":rocksdb_test_lib"],
+                    extra_test_libs=True,
+                )
                TARGETS.register_test(
                    test_target_name,
                    test_src,
-                    deps = json.dumps(deps['extra_deps'] + [':'+test_library]),
-                    extra_compiler_flags = json.dumps(deps['extra_compiler_flags']))
+                    deps=json.dumps(deps["extra_deps"] + [":" + test_library]),
+                    extra_compiler_flags=json.dumps(deps["extra_compiler_flags"]),
+                )
            else:
                TARGETS.register_test(
                    test_target_name,
                    test_src,
-                    deps = json.dumps(deps['extra_deps'] + [":rocksdb_test_lib"] ),
-                    extra_compiler_flags = json.dumps(deps['extra_compiler_flags']))
+                    deps=json.dumps(deps["extra_deps"] + [":rocksdb_test_lib"]),
+                    extra_compiler_flags=json.dumps(deps["extra_compiler_flags"]),
+                )

    print(ColorString.info("Generated TARGETS Summary:"))
    print(ColorString.info("- %d libs" % TARGETS.total_lib))
@@ -300,8 +318,7 @@ def get_rocksdb_path():
    # rocksdb = {script_dir}/..
    script_dir = os.path.dirname(sys.argv[0])
    script_dir = os.path.abspath(script_dir)
-    rocksdb_path = os.path.abspath(
-        os.path.join(script_dir, "../"))
+    rocksdb_path = os.path.abspath(os.path.join(script_dir, "../"))

    return rocksdb_path

@@ -318,5 +335,6 @@ def main():
    if not ok:
        exit_with_error("Failed to generate TARGETS files")

+
 if __name__ == "__main__":
    main()
@@ -1,113 +1,150 @@
 # Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-from __future__ import unicode_literals
+from __future__ import absolute_import, division, print_function, unicode_literals
+
 try:
-    from builtins import object
-    from builtins import str
+    from builtins import object, str
 except ImportError:
-    from __builtin__ import object
-    from __builtin__ import str
-import targets_cfg
+    from __builtin__ import object, str
 import pprint

+import targets_cfg
+
+
 def pretty_list(lst, indent=8):
    if lst is None or len(lst) == 0:
        return ""

    if len(lst) == 1:
-        return "\"%s\"" % lst[0]
+        return '"%s"' % lst[0]

-    separator = "\",\n%s\"" % (" " * indent)
+    separator = '",\n%s"' % (" " * indent)
    res = separator.join(sorted(lst))
-    res = "\n" + (" " * indent) + "\"" + res + "\",\n" + (" " * (indent - 4))
+    res = "\n" + (" " * indent) + '"' + res + '",\n' + (" " * (indent - 4))
    return res


 class TARGETSBuilder(object):
    def __init__(self, path, extra_argv):
        self.path = path
-        self.targets_file = open(path, 'wb')
        header = targets_cfg.rocksdb_target_header_template.format(
-            extra_argv=extra_argv)
-        self.targets_file.write(header.encode("utf-8"))
+            extra_argv=extra_argv
+        )
+        with open(path, "wb") as targets_file:
+            targets_file.write(header.encode("utf-8"))
        self.total_lib = 0
        self.total_bin = 0
        self.total_test = 0
        self.tests_cfg = ""

-    def __del__(self):
-        self.targets_file.close()
-
-    def add_library(self, name, srcs, deps=None, headers=None,
-                    extra_external_deps="", link_whole=False,
-                     external_dependencies=None, extra_test_libs=False):
+    def add_library(
+        self,
+        name,
+        srcs,
+        deps=None,
+        headers=None,
+        extra_external_deps="",
+        link_whole=False,
+        external_dependencies=None,
+        extra_test_libs=False,
+    ):
        if headers is not None:
            headers = "[" + pretty_list(headers) + "]"
-        self.targets_file.write(targets_cfg.library_template.format(
-            name=name,
-            srcs=pretty_list(srcs),
-            headers=headers,
-            deps=pretty_list(deps),
-            extra_external_deps=extra_external_deps,
-            link_whole=link_whole,
-            external_dependencies=pretty_list(external_dependencies),
-            extra_test_libs=extra_test_libs
-            ).encode("utf-8"))
-        self.total_lib = self.total_lib + 1
-
-    def add_rocksdb_library(self, name, srcs, headers=None,
-                            external_dependencies=None):
-        if headers is not None:
-            headers = "[" + pretty_list(headers) + "]"
-        self.targets_file.write(targets_cfg.rocksdb_library_template.format(
-            name=name,
-            srcs=pretty_list(srcs),
-            headers=headers,
-            external_dependencies=pretty_list(external_dependencies)
-            ).encode("utf-8")
+        with open(self.path, "ab") as targets_file:
+            targets_file.write(
+                targets_cfg.library_template.format(
+                    name=name,
+                    srcs=pretty_list(srcs),
+                    headers=headers,
+                    deps=pretty_list(deps),
+                    extra_external_deps=extra_external_deps,
+                    link_whole=link_whole,
+                    external_dependencies=pretty_list(external_dependencies),
+                    extra_test_libs=extra_test_libs,
+                ).encode("utf-8")
            )
        self.total_lib = self.total_lib + 1

-    def add_binary(self, name, srcs, deps=None, extra_preprocessor_flags=None,extra_bench_libs=False):
-        self.targets_file.write(targets_cfg.binary_template.format(
-            name=name,
-            srcs=pretty_list(srcs),
-            deps=pretty_list(deps),
-            extra_preprocessor_flags=pretty_list(extra_preprocessor_flags),
-            extra_bench_libs=extra_bench_libs,
-            ).encode("utf-8"))
+    def add_rocksdb_library(self, name, srcs, headers=None, external_dependencies=None):
+        if headers is not None:
+            headers = "[" + pretty_list(headers) + "]"
+        with open(self.path, "ab") as targets_file:
+            targets_file.write(
+                targets_cfg.rocksdb_library_template.format(
+                    name=name,
+                    srcs=pretty_list(srcs),
+                    headers=headers,
+                    external_dependencies=pretty_list(external_dependencies),
+                ).encode("utf-8")
+            )
+        self.total_lib = self.total_lib + 1
+
+    def add_binary(
+        self,
+        name,
+        srcs,
+        deps=None,
+        extra_preprocessor_flags=None,
+        extra_bench_libs=False,
+    ):
+        with open(self.path, "ab") as targets_file:
+            targets_file.write(
+                targets_cfg.binary_template.format(
+                    name=name,
+                    srcs=pretty_list(srcs),
+                    deps=pretty_list(deps),
+                    extra_preprocessor_flags=pretty_list(extra_preprocessor_flags),
+                    extra_bench_libs=extra_bench_libs,
+                ).encode("utf-8")
+            )
        self.total_bin = self.total_bin + 1

    def add_c_test(self):
-        self.targets_file.write(b"""
+        with open(self.path, "ab") as targets_file:
+            targets_file.write(
+                b"""
 add_c_test_wrapper()
-""")
+"""
+            )

    def add_test_header(self):
-        self.targets_file.write(b"""
+        with open(self.path, "ab") as targets_file:
+            targets_file.write(
+                b"""
        # Generate a test rule for each entry in ROCKS_TESTS
        # Do not build the tests in opt mode, since SyncPoint and other test code
        # will not be included.
-""")
+"""
+            )

-    def add_fancy_bench_config(self, name, bench_config, slow, expected_runtime, sl_iterations, regression_threshold):
-        self.targets_file.write(targets_cfg.fancy_bench_template.format(
+    def add_fancy_bench_config(
+        self,
+        name,
+        bench_config,
+        slow,
+        expected_runtime,
+        sl_iterations,
+        regression_threshold,
+    ):
+        with open(self.path, "ab") as targets_file:
+            targets_file.write(
+                targets_cfg.fancy_bench_template.format(
                    name=name,
                    bench_config=pprint.pformat(bench_config),
                    slow=slow,
                    expected_runtime=expected_runtime,
                    sl_iterations=sl_iterations,
-                    regression_threshold=regression_threshold
-                    ).encode("utf-8"))
+                    regression_threshold=regression_threshold,
+                ).encode("utf-8")
+            )

-    def register_test(self,
-                      test_name,
-                      src,
-                      deps,
-                      extra_compiler_flags):
-
-        self.targets_file.write(targets_cfg.unittests_template.format(test_name=test_name,test_cc=str(src),deps=deps,
-            extra_compiler_flags=extra_compiler_flags).encode("utf-8"))
+    def register_test(self, test_name, src, deps, extra_compiler_flags):
+        with open(self.path, "ab") as targets_file:
+            targets_file.write(
+                targets_cfg.unittests_template.format(
+                    test_name=test_name,
+                    test_cc=str(src),
+                    deps=deps,
+                    extra_compiler_flags=extra_compiler_flags,
+                ).encode("utf-8")
+            )
        self.total_test = self.total_test + 1
@@ -1,11 +1,7 @@
 # Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-from __future__ import unicode_literals
+from __future__ import absolute_import, division, print_function, unicode_literals

-rocksdb_target_header_template = \
-    """# This file \100generated by:
+rocksdb_target_header_template = """# This file \100generated by:
 #$ python3 buckifier/buckify_rocksdb.py{extra_argv}
 # --> DO NOT EDIT MANUALLY <--
 # This file is a Facebook-specific integration for buck builds, so can
@@ -27,7 +23,6 @@ rocks_cpp_library_wrapper(name="{name}", srcs=[{srcs}], headers={headers})
 """


-
 binary_template = """
 cpp_binary_wrapper(name="{name}", srcs=[{srcs}], deps=[{deps}], extra_preprocessor_flags=[{extra_preprocessor_flags}], extra_bench_libs={extra_bench_libs})
 """
@@ -2,37 +2,35 @@
 """
 This module keeps commonly used components.
 """
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-from __future__ import unicode_literals
+from __future__ import absolute_import, division, print_function, unicode_literals
+
 try:
    from builtins import object
 except ImportError:
    from __builtin__ import object
+import os
 import subprocess
 import sys
-import os
 import time

+
 class ColorString(object):
-    """ Generate colorful strings on terminal """
-    HEADER = '\033[95m'
-    BLUE = '\033[94m'
-    GREEN = '\033[92m'
-    WARNING = '\033[93m'
-    FAIL = '\033[91m'
-    ENDC = '\033[0m'
+    """Generate colorful strings on terminal"""
+
+    HEADER = "\033[95m"
+    BLUE = "\033[94m"
+    GREEN = "\033[92m"
+    WARNING = "\033[93m"
+    FAIL = "\033[91m"
+    ENDC = "\033[0m"

    @staticmethod
    def _make_color_str(text, color):
        # In Python2, default encoding for unicode string is ASCII
        if sys.version_info.major <= 2:
-            return "".join(
-                [color, text.encode('utf-8'), ColorString.ENDC])
+            return "".join([color, text.encode("utf-8"), ColorString.ENDC])
        # From Python3, default encoding for unicode string is UTF-8
-        return "".join(
-            [color, text, ColorString.ENDC])
+        return "".join([color, text, ColorString.ENDC])

    @staticmethod
    def ok(text):
@@ -68,37 +66,38 @@ class ColorString(object):


 def run_shell_command(shell_cmd, cmd_dir=None):
-    """ Run a single shell command.
-        @returns a tuple of shell command return code, stdout, stderr """
+    """Run a single shell command.
+    @returns a tuple of shell command return code, stdout, stderr"""

    if cmd_dir is not None and not os.path.exists(cmd_dir):
        run_shell_command("mkdir -p %s" % cmd_dir)

    start = time.time()
    print("\t>>> Running: " + shell_cmd)
-    p = subprocess.Popen(shell_cmd,
-                         shell=True,
-                         stdout=subprocess.PIPE,
-                         stderr=subprocess.PIPE,
-                         cwd=cmd_dir)
+    p = subprocess.Popen(  # noqa
+        shell_cmd,
+        shell=True,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE,
+        cwd=cmd_dir,
+    )
    stdout, stderr = p.communicate()
    end = time.time()

    # Report time if we spent more than 5 minutes executing a command
    execution_time = end - start
    if execution_time > (60 * 5):
-        mins = (execution_time / 60)
-        secs = (execution_time % 60)
+        mins = execution_time / 60
+        secs = execution_time % 60
        print("\t>time spent: %d minutes %d seconds" % (mins, secs))

-
    return p.returncode, stdout, stderr


 def run_shell_commands(shell_cmds, cmd_dir=None, verbose=False):
-    """ Execute a sequence of shell commands, which is equivalent to
-        running `cmd1 && cmd2 && cmd3`
-        @returns boolean indication if all commands succeeds.
+    """Execute a sequence of shell commands, which is equivalent to
+    running `cmd1 && cmd2 && cmd3`
+    @returns boolean indication if all commands succeeds.
    """

    if cmd_dir:
@@ -28,14 +28,15 @@
 from __future__ import print_function

 import argparse
-from os import path
 import re
 import sys
+from os import path

 include_re = re.compile('^[ \t]*#include[ \t]+"(.*)"[ \t]*$')
 included = set()
 excluded = set()

+
 def find_header(name, abs_path, include_paths):
    samedir = path.join(path.dirname(abs_path), name)
    if path.exists(samedir):
@@ -46,17 +47,31 @@ def find_header(name, abs_path, include_paths):
            return include_path
    return None

-def expand_include(include_path, f, abs_path, source_out, header_out, include_paths, public_include_paths):
+
+def expand_include(
+    include_path,
+    f,
+    abs_path,
+    source_out,
+    header_out,
+    include_paths,
+    public_include_paths,
+):
    if include_path in included:
        return False

    included.add(include_path)
    with open(include_path) as f:
        print('#line 1 "{}"'.format(include_path), file=source_out)
-        process_file(f, include_path, source_out, header_out, include_paths, public_include_paths)
+        process_file(
+            f, include_path, source_out, header_out, include_paths, public_include_paths
+        )
    return True

-def process_file(f, abs_path, source_out, header_out, include_paths, public_include_paths):
+
+def process_file(
+    f, abs_path, source_out, header_out, include_paths, public_include_paths
+):
    for (line, text) in enumerate(f):
        m = include_re.match(text)
        if m:
@@ -68,7 +83,15 @@ def process_file(f, abs_path, source_out, header_out, include_paths, public_incl
                    source_out.write(text)
                    expanded = False
                else:
-                    expanded = expand_include(include_path, f, abs_path, source_out, header_out, include_paths, public_include_paths)
+                    expanded = expand_include(
+                        include_path,
+                        f,
+                        abs_path,
+                        source_out,
+                        header_out,
+                        include_paths,
+                        public_include_paths,
+                    )
            else:
                # now try public headers
                include_path = find_header(filename, abs_path, public_include_paths)
@@ -78,23 +101,52 @@ def process_file(f, abs_path, source_out, header_out, include_paths, public_incl
                    if include_path in excluded:
                        source_out.write(text)
                    else:
-                        expand_include(include_path, f, abs_path, header_out, None, public_include_paths, [])
+                        expand_include(
+                            include_path,
+                            f,
+                            abs_path,
+                            header_out,
+                            None,
+                            public_include_paths,
+                            [],
+                        )
                else:
-                    sys.exit("unable to find {}, included in {} on line {}".format(filename, abs_path, line))
+                    sys.exit(
+                        "unable to find {}, included in {} on line {}".format(
+                            filename, abs_path, line
+                        )
+                    )

            if expanded:
-                print('#line {} "{}"'.format(line+1, abs_path), file=source_out)
+                print('#line {} "{}"'.format(line + 1, abs_path), file=source_out)
        elif text != "#pragma once\n":
            source_out.write(text)

+
 def main():
-    parser = argparse.ArgumentParser(description="Transform a unity build into an amalgamation")
+    parser = argparse.ArgumentParser(
+        description="Transform a unity build into an amalgamation"
+    )
    parser.add_argument("source", help="source file")
-    parser.add_argument("-I", action="append", dest="include_paths", help="include paths for private headers")
-    parser.add_argument("-i", action="append", dest="public_include_paths", help="include paths for public headers")
-    parser.add_argument("-x", action="append", dest="excluded", help="excluded header files")
+    parser.add_argument(
+        "-I",
+        action="append",
+        dest="include_paths",
+        help="include paths for private headers",
+    )
+    parser.add_argument(
+        "-i",
+        action="append",
+        dest="public_include_paths",
+        help="include paths for public headers",
+    )
+    parser.add_argument(
+        "-x", action="append", dest="excluded", help="excluded header files"
+    )
    parser.add_argument("-o", dest="source_out", help="output C++ file", required=True)
-    parser.add_argument("-H", dest="header_out", help="output C++ header file", required=True)
+    parser.add_argument(
+        "-H", dest="header_out", help="output C++ header file", required=True
+    )
    args = parser.parse_args()

    include_paths = list(map(path.abspath, args.include_paths or []))
@@ -102,10 +154,15 @@ def main():
    excluded.update(map(path.abspath, args.excluded or []))
    filename = args.source
    abs_path = path.abspath(filename)
-    with open(filename) as f, open(args.source_out, 'w') as source_out, open(args.header_out, 'w') as header_out:
+    with open(filename) as f, open(args.source_out, "w") as source_out, open(
+        args.header_out, "w"
+    ) as header_out:
        print('#line 1 "{}"'.format(filename), file=source_out)
        print('#include "{}"'.format(header_out.name), file=source_out)
-        process_file(f, abs_path, source_out, header_out, include_paths, public_include_paths)
+        process_file(
+            f, abs_path, source_out, header_out, include_paths, public_include_paths
+        )
+

 if __name__ == "__main__":
    main()
@@ -4,23 +4,27 @@
 #  COPYING file in the root directory) and Apache 2.0 License
 #  (found in the LICENSE.Apache file in the root directory).

-'''Access the results of benchmark runs
+"""Access the results of benchmark runs
 Send these results on to OpenSearch graphing service
-'''
+"""

 import argparse
 import itertools
+import logging
 import os
 import re
 import sys
+
 import requests
 from dateutil import parser
-import logging

 logging.basicConfig(level=logging.DEBUG)
+
+
 class Configuration:
-    opensearch_user = os.environ['ES_USER']
-    opensearch_pass = os.environ['ES_PASS']
+    opensearch_user = os.environ["ES_USER"]
+    opensearch_pass = os.environ["ES_PASS"]
+

 class BenchmarkResultException(Exception):
    def __init__(self, message, content):
@@ -30,45 +34,71 @@ class BenchmarkResultException(Exception):

 class BenchmarkUtils:

-    expected_keys = ['ops_sec', 'mb_sec',  'lsm_sz',  'blob_sz', 'c_wgb',   'w_amp',
-    'c_mbps',  'c_wsecs', 'c_csecs', 'b_rgb',   'b_wgb',   'usec_op',
-    'p50',     'p99',     'p99.9',   'p99.99', 'pmax',
-    'uptime',  'stall%', 'Nstall',  'u_cpu',   's_cpu',   'rss', 'test', 'date', 'version', 'job_id']
+    expected_keys = [
+        "ops_sec",
+        "mb_sec",
+        "lsm_sz",
+        "blob_sz",
+        "c_wgb",
+        "w_amp",
+        "c_mbps",
+        "c_wsecs",
+        "c_csecs",
+        "b_rgb",
+        "b_wgb",
+        "usec_op",
+        "p50",
+        "p99",
+        "p99.9",
+        "p99.99",
+        "pmax",
+        "uptime",
+        "stall%",
+        "Nstall",
+        "u_cpu",
+        "s_cpu",
+        "rss",
+        "test",
+        "date",
+        "version",
+        "job_id",
+    ]

    def sanity_check(row):
-        if not 'test' in row:
+        if "test" not in row:
            logging.debug(f"not 'test' in row: {row}")
            return False
-        if row['test'] == '':
+        if row["test"] == "":
            logging.debug(f"row['test'] == '': {row}")
            return False
-        if not 'date' in row:
+        if "date" not in row:
            logging.debug(f"not 'date' in row: {row}")
            return False
-        if not 'ops_sec' in row:
+        if "ops_sec" not in row:
            logging.debug(f"not 'ops_sec' in row: {row}")
            return False
        try:
-            v = int(row['ops_sec'])
+            _ = int(row["ops_sec"])
        except (ValueError, TypeError):
            logging.debug(f"int(row['ops_sec']): {row}")
            return False
        try:
-            (_, _) = parser.parse(row['date'], fuzzy_with_tokens=True)
+            (_, _) = parser.parse(row["date"], fuzzy_with_tokens=True)
        except (parser.ParserError):
-            logging.error(f"parser.parse((row['date']): not a valid format for date in row: {row}")
+            logging.error(
+                f"parser.parse((row['date']): not a valid format for date in row: {row}"
+            )
            return False
        return True

    def conform_opensearch(row):
-        (dt, _) = parser.parse(row['date'], fuzzy_with_tokens=True)
+        (dt, _) = parser.parse(row["date"], fuzzy_with_tokens=True)
        # create a test_date field, which was previously what was expected
        # repair the date field, which has what can be a WRONG ISO FORMAT, (no leading 0 on single-digit day-of-month)
        # e.g. 2022-07-1T00:14:55 should be 2022-07-01T00:14:55
-        row['test_date'] = dt.isoformat()
-        row['date'] = dt.isoformat()
-        return dict((key.replace('.', '_'), value)
-                    for (key, value) in row.items())
+        row["test_date"] = dt.isoformat()
+        row["date"] = dt.isoformat()
+        return {key.replace(".", "_"): value for key, value in row.items()}


 class ResultParser:
@@ -80,24 +110,24 @@ class ResultParser:
    def ignore(self, l_in: str):
        if len(l_in) == 0:
            return True
-        if l_in[0:1] == '#':
+        if l_in[0:1] == "#":
            return True
        return False

-    def line(self, l_in: str):
-        '''Parse a line into items
+    def line(self, line_in: str):
+        """Parse a line into items
        Being clever about separators
-        '''
-        l = l_in
+        """
+        line = line_in
        row = []
-        while l != '':
-            match_item = self.field.match(l)
+        while line != "":
+            match_item = self.field.match(line)
            if match_item:
                item = match_item.group(0)
                row.append(item)
-                l = l[len(item):]
+                line = line[len(item) :]
            else:
-                match_intra = self.intra.match(l)
+                match_intra = self.intra.match(line)
                if match_intra:
                    intra = match_intra.group(0)
                    # Count the separators
@@ -107,26 +137,29 @@ class ResultParser:
                    sep_count = len(tabbed) - 1
                    if sep_count == 0:
                        sep_count = 1
-                    for i in range(sep_count-1):
-                        row.append('')
-                    l = l[len(intra):]
+                    for _ in range(sep_count - 1):
+                        row.append("")
+                    line = line[len(intra) :]
                else:
                    raise BenchmarkResultException(
-                        'Invalid TSV line', f"{l_in} at {l}")
+                        "Invalid TSV line", f"{line_in} at {line}"
+                    )
        return row

    def parse(self, lines):
-        '''Parse something that iterates lines'''
+        """Parse something that iterates lines"""
        rows = [self.line(line) for line in lines if not self.ignore(line)]
        header = rows[0]
        width = len(header)
-        records = [{k: v for (k, v) in itertools.zip_longest(
-            header, row[:width])} for row in rows[1:]]
+        records = [
+            {k: v for (k, v) in itertools.zip_longest(header, row[:width])}
+            for row in rows[1:]
+        ]
        return records


 def load_report_from_tsv(filename: str):
-    file = open(filename, 'r')
+    file = open(filename, "r")
    contents = file.readlines()
    file.close()
    parser = ResultParser()
@@ -136,52 +169,70 @@ def load_report_from_tsv(filename: str):


 def push_report_to_opensearch(report, esdocument):
-    sanitized = [BenchmarkUtils.conform_opensearch(row)
-                 for row in report if BenchmarkUtils.sanity_check(row)]
-    logging.debug(f"upload {len(sanitized)} sane of {len(report)} benchmarks to opensearch")
+    sanitized = [
+        BenchmarkUtils.conform_opensearch(row)
+        for row in report
+        if BenchmarkUtils.sanity_check(row)
+    ]
+    logging.debug(
+        f"upload {len(sanitized)} sane of {len(report)} benchmarks to opensearch"
+    )
    for single_benchmark in sanitized:
        logging.debug(f"upload benchmark: {single_benchmark}")
        response = requests.post(
            esdocument,
-            json=single_benchmark, auth=(os.environ['ES_USER'], os.environ['ES_PASS']))
+            json=single_benchmark,
+            auth=(os.environ["ES_USER"], os.environ["ES_PASS"]),
+        )
        logging.debug(
-            f"Sent to OpenSearch, status: {response.status_code}, result: {response.text}")
+            f"Sent to OpenSearch, status: {response.status_code}, result: {response.text}"
+        )
        response.raise_for_status()

+
 def push_report_to_null(report):
-    
+
    for row in report:
        if BenchmarkUtils.sanity_check(row):
            logging.debug(f"row {row}")
            conformed = BenchmarkUtils.conform_opensearch(row)
            logging.debug(f"conformed row {conformed}")

+
 def main():
-    '''Tool for fetching, parsing and uploading benchmark results to OpenSearch / ElasticSearch
+    """Tool for fetching, parsing and uploading benchmark results to OpenSearch / ElasticSearch
    This tool will

    (1) Open a local tsv benchmark report file
    (2) Upload to OpenSearch document, via https/JSON
-    '''
+    """

-    parser = argparse.ArgumentParser(
-        description='CircleCI benchmark scraper.')
+    parser = argparse.ArgumentParser(description="CircleCI benchmark scraper.")

    # --tsvfile is the name of the file to read results from
    # --esdocument is the ElasticSearch document to push these results into
    #
-    parser.add_argument('--tsvfile', default='build_tools/circle_api_scraper_input.txt',
-                        help='File from which to read tsv report')
-    parser.add_argument('--esdocument', help='ElasticSearch/OpenSearch document URL to upload report into')
-    parser.add_argument('--upload', choices=['opensearch', 'none'], default='opensearch')
+    parser.add_argument(
+        "--tsvfile",
+        default="build_tools/circle_api_scraper_input.txt",
+        help="File from which to read tsv report",
+    )
+    parser.add_argument(
+        "--esdocument",
+        help="ElasticSearch/OpenSearch document URL to upload report into",
+    )
+    parser.add_argument(
+        "--upload", choices=["opensearch", "none"], default="opensearch"
+    )

    args = parser.parse_args()
    logging.debug(f"Arguments: {args}")
    reports = load_report_from_tsv(args.tsvfile)
-    if (args.upload == 'opensearch'):
+    if args.upload == "opensearch":
        push_report_to_opensearch(reports, args.esdocument)
    else:
        push_report_to_null(reports)

-if __name__ == '__main__':
+
+if __name__ == "__main__":
    sys.exit(main())
@@ -277,6 +277,7 @@ if [ "$CROSS_COMPILE" = "true" -o "$FBCODE_BUILD" = "true" ]; then
    if [ "$FBCODE_BUILD" = "true" ]; then
      # Enable backtrace on fbcode since the necessary libraries are present
      COMMON_FLAGS="$COMMON_FLAGS -DROCKSDB_BACKTRACE"
+      FOLLY_DIR="third-party/folly"
    fi
    true
 else
@@ -601,6 +602,17 @@ EOF
            PLATFORM_LDFLAGS="$PLATFORM_LDFLAGS -lbenchmark"
        fi
    fi
+    if test $USE_FOLLY; then
+        # Test whether libfolly library is installed
+        $CXX $PLATFORM_CXXFLAGS $COMMON_FLAGS -x c++ - -o /dev/null 2>/dev/null  <<EOF
+          #include <folly/synchronization/DistributedMutex.h>
+          int main() {}
+EOF
+        if [ "$?" != 0 ]; then
+          FOLLY_DIR="./third-party/folly"
+        fi
+    fi
+
 fi

 # TODO(tec): Fix -Wshorten-64-to-32 errors on FreeBSD and enable the warning.
@@ -827,6 +839,13 @@ fi

 rm -f test.o test_dl.o

+# Get the path for the folly installation dir
+if [ "$USE_FOLLY" ]; then
+  if [ "$FOLLY_DIR" ]; then
+    FOLLY_PATH=`cd $FOLLY_DIR && $PYTHON build/fbcode_builder/getdeps.py show-inst-dir folly`
+  fi
+fi
+
 PLATFORM_CCFLAGS="$PLATFORM_CCFLAGS $COMMON_FLAGS"
 PLATFORM_CXXFLAGS="$PLATFORM_CXXFLAGS $COMMON_FLAGS"

@@ -866,6 +885,8 @@ echo "CLANG_ANALYZER=$CLANG_ANALYZER" >> "$OUTPUT"
 echo "PROFILING_FLAGS=$PROFILING_FLAGS" >> "$OUTPUT"
 echo "FIND=$FIND" >> "$OUTPUT"
 echo "WATCH=$WATCH" >> "$OUTPUT"
+echo "FOLLY_PATH=$FOLLY_PATH" >> "$OUTPUT"
+
 # This will enable some related identifiers for the preprocessor
 if test -n "$JEMALLOC"; then
  echo "JEMALLOC=1" >> "$OUTPUT"
@@ -19,8 +19,4 @@ BINUTILS_BASE=/mnt/gvfs/third-party2/binutils/08634589372fa5f237bfd374e8c644a836
 VALGRIND_BASE=/mnt/gvfs/third-party2/valgrind/6ae525939ad02e5e676855082fbbc7828dbafeac/3.15.0/platform009/7f3b187
 LUA_BASE=/mnt/gvfs/third-party2/lua/162efd9561a3d21f6869f4814011e9cf1b3ff4dc/5.3.4/platform009/a6271c4
 BENCHMARK_BASE=/mnt/gvfs/third-party2/benchmark/30bf49ad6414325e17f3425b0edcb64239427ae3/1.6.1/platform009/7f3b187
-BOOST_BASE=/mnt/gvfs/third-party2/boost/201b7d74941e54b436dfa364a063aa6d2cd7de4c/1.69.0/platform009/8a7ffdf
-GLOG_BASE=/mnt/gvfs/third-party2/glog/32d751bd5673375b438158717ab6a57c1cc57e3d/0.3.2_fb/platform009/10a364d/
-FMT_BASE=/mnt/gvfs/third-party2/fmt/ce0c25f67165f4d2c22a29b8ef50f5600d7873ca/6.1.1/platform009/7f3b187/
-DBL_CONV_BASE=/mnt/gvfs/third-party2/double_conversion/109b3d9696d71f1048678cd7da1e22505470543d/20141126/platform009/7f3b187/
-LIBEVENT_BASE=/mnt/gvfs/third-party2/libevent/4a4d3a79a76c2439b6bd471bf3586b3481dde75e/1.4.14b_hphp/platform009/7f3b187/
+GLOG_BASE=/mnt/gvfs/third-party2/glog/32d751bd5673375b438158717ab6a57c1cc57e3d/0.3.2_fb/platform009/10a364d
@@ -3,16 +3,13 @@
 #  COPYING file in the root directory) and Apache 2.0 License
 #  (found in the LICENSE.Apache file in the root directory).

-'''Filter for error messages in test output:
+"""Filter for error messages in test output:
    - Receives merged stdout/stderr from test on stdin
    - Finds patterns of known error messages for test name (first argument)
    - Prints those error messages to stdout
-'''
+"""

-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-from __future__ import unicode_literals
+from __future__ import absolute_import, division, print_function, unicode_literals

 import re
 import sys
@@ -20,23 +17,24 @@ import sys

 class ErrorParserBase(object):
    def parse_error(self, line):
-        '''Parses a line of test output. If it contains an error, returns a
+        """Parses a line of test output. If it contains an error, returns a
        formatted message describing the error; otherwise, returns None.
        Subclasses must override this method.
-        '''
+        """
        raise NotImplementedError


 class GTestErrorParser(ErrorParserBase):
-    '''A parser that remembers the last test that began running so it can print
+    """A parser that remembers the last test that began running so it can print
    that test's name upon detecting failure.
-    '''
-    _GTEST_NAME_PATTERN = re.compile(r'\[ RUN      \] (\S+)$')
+    """
+
+    _GTEST_NAME_PATTERN = re.compile(r"\[ RUN      \] (\S+)$")
    # format: '<filename or "unknown file">:<line #>: Failure'
-    _GTEST_FAIL_PATTERN = re.compile(r'(unknown file|\S+:\d+): Failure$')
+    _GTEST_FAIL_PATTERN = re.compile(r"(unknown file|\S+:\d+): Failure$")

    def __init__(self):
-        self._last_gtest_name = 'Unknown test'
+        self._last_gtest_name = "Unknown test"

    def parse_error(self, line):
        gtest_name_match = self._GTEST_NAME_PATTERN.match(line)
@@ -45,14 +43,13 @@ class GTestErrorParser(ErrorParserBase):
            return None
        gtest_fail_match = self._GTEST_FAIL_PATTERN.match(line)
        if gtest_fail_match:
-            return '%s failed: %s' % (
-                    self._last_gtest_name, gtest_fail_match.group(1))
+            return "%s failed: %s" % (self._last_gtest_name, gtest_fail_match.group(1))
        return None


 class MatchErrorParser(ErrorParserBase):
-    '''A simple parser that returns the whole line if it matches the pattern.
-    '''
+    """A simple parser that returns the whole line if it matches the pattern."""
+
    def __init__(self, pattern):
        self._pattern = re.compile(pattern)

@@ -69,97 +66,104 @@ class CompilerErrorParser(MatchErrorParser):
        # format (link error):
        #   '<filename>:<line #>: error: <error msg>'
        # The below regex catches both
-        super(CompilerErrorParser, self).__init__(r'\S+:\d+: error:')
+        super(CompilerErrorParser, self).__init__(r"\S+:\d+: error:")


 class ScanBuildErrorParser(MatchErrorParser):
    def __init__(self):
-        super(ScanBuildErrorParser, self).__init__(
-                r'scan-build: \d+ bugs found.$')
+        super(ScanBuildErrorParser, self).__init__(r"scan-build: \d+ bugs found.$")


 class DbCrashErrorParser(MatchErrorParser):
    def __init__(self):
-        super(DbCrashErrorParser, self).__init__(r'\*\*\*.*\^$|TEST FAILED.')
+        super(DbCrashErrorParser, self).__init__(r"\*\*\*.*\^$|TEST FAILED.")


 class WriteStressErrorParser(MatchErrorParser):
    def __init__(self):
        super(WriteStressErrorParser, self).__init__(
-                r'ERROR: write_stress died with exitcode=\d+')
+            r"ERROR: write_stress died with exitcode=\d+"
+        )


 class AsanErrorParser(MatchErrorParser):
    def __init__(self):
-        super(AsanErrorParser, self).__init__(
-                r'==\d+==ERROR: AddressSanitizer:')
+        super(AsanErrorParser, self).__init__(r"==\d+==ERROR: AddressSanitizer:")


 class UbsanErrorParser(MatchErrorParser):
    def __init__(self):
        # format: '<filename>:<line #>:<column #>: runtime error: <error msg>'
-        super(UbsanErrorParser, self).__init__(r'\S+:\d+:\d+: runtime error:')
+        super(UbsanErrorParser, self).__init__(r"\S+:\d+:\d+: runtime error:")


 class ValgrindErrorParser(MatchErrorParser):
    def __init__(self):
        # just grab the summary, valgrind doesn't clearly distinguish errors
        # from other log messages.
-        super(ValgrindErrorParser, self).__init__(r'==\d+== ERROR SUMMARY:')
+        super(ValgrindErrorParser, self).__init__(r"==\d+== ERROR SUMMARY:")


 class CompatErrorParser(MatchErrorParser):
    def __init__(self):
-        super(CompatErrorParser, self).__init__(r'==== .*[Ee]rror.* ====$')
+        super(CompatErrorParser, self).__init__(r"==== .*[Ee]rror.* ====$")


 class TsanErrorParser(MatchErrorParser):
    def __init__(self):
-        super(TsanErrorParser, self).__init__(r'WARNING: ThreadSanitizer:')
+        super(TsanErrorParser, self).__init__(r"WARNING: ThreadSanitizer:")


 _TEST_NAME_TO_PARSERS = {
-    'punit': [CompilerErrorParser, GTestErrorParser],
-    'unit': [CompilerErrorParser, GTestErrorParser],
-    'release': [CompilerErrorParser, GTestErrorParser],
-    'unit_481': [CompilerErrorParser, GTestErrorParser],
-    'release_481': [CompilerErrorParser, GTestErrorParser],
-    'clang_unit': [CompilerErrorParser, GTestErrorParser],
-    'clang_release': [CompilerErrorParser, GTestErrorParser],
-    'clang_analyze': [CompilerErrorParser, ScanBuildErrorParser],
-    'code_cov': [CompilerErrorParser, GTestErrorParser],
-    'unity': [CompilerErrorParser, GTestErrorParser],
-    'lite': [CompilerErrorParser],
-    'lite_test': [CompilerErrorParser, GTestErrorParser],
-    'stress_crash': [CompilerErrorParser, DbCrashErrorParser],
-    'stress_crash_with_atomic_flush': [CompilerErrorParser, DbCrashErrorParser],
-    'stress_crash_with_txn': [CompilerErrorParser, DbCrashErrorParser],
-    'write_stress': [CompilerErrorParser, WriteStressErrorParser],
-    'asan': [CompilerErrorParser, GTestErrorParser, AsanErrorParser],
-    'asan_crash': [CompilerErrorParser, AsanErrorParser, DbCrashErrorParser],
-    'asan_crash_with_atomic_flush': [CompilerErrorParser, AsanErrorParser, DbCrashErrorParser],
-    'asan_crash_with_txn': [CompilerErrorParser, AsanErrorParser, DbCrashErrorParser],
-    'ubsan': [CompilerErrorParser, GTestErrorParser, UbsanErrorParser],
-    'ubsan_crash': [CompilerErrorParser, UbsanErrorParser, DbCrashErrorParser],
-    'ubsan_crash_with_atomic_flush': [CompilerErrorParser, UbsanErrorParser, DbCrashErrorParser],
-    'ubsan_crash_with_txn': [CompilerErrorParser, UbsanErrorParser, DbCrashErrorParser],
-    'valgrind': [CompilerErrorParser, GTestErrorParser, ValgrindErrorParser],
-    'tsan': [CompilerErrorParser, GTestErrorParser, TsanErrorParser],
-    'format_compatible': [CompilerErrorParser, CompatErrorParser],
-    'run_format_compatible': [CompilerErrorParser, CompatErrorParser],
-    'no_compression': [CompilerErrorParser, GTestErrorParser],
-    'run_no_compression': [CompilerErrorParser, GTestErrorParser],
-    'regression': [CompilerErrorParser],
-    'run_regression': [CompilerErrorParser],
+    "punit": [CompilerErrorParser, GTestErrorParser],
+    "unit": [CompilerErrorParser, GTestErrorParser],
+    "release": [CompilerErrorParser, GTestErrorParser],
+    "unit_481": [CompilerErrorParser, GTestErrorParser],
+    "release_481": [CompilerErrorParser, GTestErrorParser],
+    "clang_unit": [CompilerErrorParser, GTestErrorParser],
+    "clang_release": [CompilerErrorParser, GTestErrorParser],
+    "clang_analyze": [CompilerErrorParser, ScanBuildErrorParser],
+    "code_cov": [CompilerErrorParser, GTestErrorParser],
+    "unity": [CompilerErrorParser, GTestErrorParser],
+    "lite": [CompilerErrorParser],
+    "lite_test": [CompilerErrorParser, GTestErrorParser],
+    "stress_crash": [CompilerErrorParser, DbCrashErrorParser],
+    "stress_crash_with_atomic_flush": [CompilerErrorParser, DbCrashErrorParser],
+    "stress_crash_with_txn": [CompilerErrorParser, DbCrashErrorParser],
+    "write_stress": [CompilerErrorParser, WriteStressErrorParser],
+    "asan": [CompilerErrorParser, GTestErrorParser, AsanErrorParser],
+    "asan_crash": [CompilerErrorParser, AsanErrorParser, DbCrashErrorParser],
+    "asan_crash_with_atomic_flush": [
+        CompilerErrorParser,
+        AsanErrorParser,
+        DbCrashErrorParser,
+    ],
+    "asan_crash_with_txn": [CompilerErrorParser, AsanErrorParser, DbCrashErrorParser],
+    "ubsan": [CompilerErrorParser, GTestErrorParser, UbsanErrorParser],
+    "ubsan_crash": [CompilerErrorParser, UbsanErrorParser, DbCrashErrorParser],
+    "ubsan_crash_with_atomic_flush": [
+        CompilerErrorParser,
+        UbsanErrorParser,
+        DbCrashErrorParser,
+    ],
+    "ubsan_crash_with_txn": [CompilerErrorParser, UbsanErrorParser, DbCrashErrorParser],
+    "valgrind": [CompilerErrorParser, GTestErrorParser, ValgrindErrorParser],
+    "tsan": [CompilerErrorParser, GTestErrorParser, TsanErrorParser],
+    "format_compatible": [CompilerErrorParser, CompatErrorParser],
+    "run_format_compatible": [CompilerErrorParser, CompatErrorParser],
+    "no_compression": [CompilerErrorParser, GTestErrorParser],
+    "run_no_compression": [CompilerErrorParser, GTestErrorParser],
+    "regression": [CompilerErrorParser],
+    "run_regression": [CompilerErrorParser],
 }


 def main():
    if len(sys.argv) != 2:
-        return 'Usage: %s <test name>' % sys.argv[0]
+        return "Usage: %s <test name>" % sys.argv[0]
    test_name = sys.argv[1]
    if test_name not in _TEST_NAME_TO_PARSERS:
-        return 'Unknown test name: %s' % test_name
+        return "Unknown test name: %s" % test_name

    error_parsers = []
    for parser_cls in _TEST_NAME_TO_PARSERS[test_name]:
@@ -173,5 +177,5 @@ def main():
                print(error_msg)


-if __name__ == '__main__':
+if __name__ == "__main__":
    sys.exit(main())
@@ -68,20 +68,9 @@ CFLAGS+=" -DGFLAGS=gflags"
 BENCHMARK_INCLUDE=" -I $BENCHMARK_BASE/include/"
 BENCHMARK_LIBS=" $BENCHMARK_BASE/lib/libbenchmark${MAYBE_PIC}.a"

-BOOST_INCLUDE=" -I $BOOST_BASE/include/"
-
 GLOG_INCLUDE=" -I $GLOG_BASE/include/"
 GLOG_LIBS=" $GLOG_BASE/lib/libglog${MAYBE_PIC}.a"

-FMT_INCLUDE=" -I $FMT_BASE/include/"
-FMT_LIBS=" $FMT_BASE/lib/libfmt${MAYBE_PIC}.a"
-
-DBL_CONV_INCLUDE=" -I $DBL_CONV_BASE/include/"
-DBL_CONV_LIBS=" $DBL_CONV_BASE/lib/libdouble-conversion${MAYBE_PIC}.a"
-
-LIBEVENT_INCLUDE=" -I $LIBEVENT_BASE/include/"
-LIBEVENT_LIBS=" $LIBEVENT_BASE/lib/libevent${MAYBE_PIC}.a"
-
 # location of jemalloc
 JEMALLOC_INCLUDE=" -I $JEMALLOC_BASE/include/"
 JEMALLOC_LIB=" $JEMALLOC_BASE/lib/libjemalloc${MAYBE_PIC}.a"
@@ -113,7 +102,7 @@ BINUTILS="$BINUTILS_BASE/bin"
 AR="$BINUTILS/ar"
 AS="$BINUTILS/as"

-DEPS_INCLUDE="$SNAPPY_INCLUDE $ZLIB_INCLUDE $BZIP_INCLUDE $LZ4_INCLUDE $ZSTD_INCLUDE $GFLAGS_INCLUDE $NUMA_INCLUDE $TBB_INCLUDE $LIBURING_INCLUDE $BENCHMARK_INCLUDE $BOOST_INCLUDE $GLOG_INCLUDE $FMT_INCLUDE $DBL_CONV_INCLUDE $LIBEVENT_INCLUDE"
+DEPS_INCLUDE="$SNAPPY_INCLUDE $ZLIB_INCLUDE $BZIP_INCLUDE $LZ4_INCLUDE $ZSTD_INCLUDE $GFLAGS_INCLUDE $NUMA_INCLUDE $TBB_INCLUDE $LIBURING_INCLUDE $BENCHMARK_INCLUDE $GLOG_INCLUDE"

 STDLIBS="-L $GCC_BASE/lib64"

@@ -1170,7 +1170,7 @@ sub parse_env_var {
    # Check if any variables contain \n
    if(my @v = map { s/BASH_FUNC_(.*)\(\)/$1/; $_ } grep { $ENV{$_}=~/\n/ } @vars) {
 	# \n is bad for csh and will cause it to fail.
-	$Global::envwarn = ::shell_quote_scalar(q{echo $SHELL | egrep "/t?csh" > /dev/null && echo CSH/TCSH DO NOT SUPPORT newlines IN VARIABLES/FUNCTIONS. Unset }."@v".q{ && exec false;}."\n\n") . $Global::envwarn;
+	$Global::envwarn = ::shell_quote_scalar(q{echo $SHELL | grep -E "/t?csh" > /dev/null && echo CSH/TCSH DO NOT SUPPORT newlines IN VARIABLES/FUNCTIONS. Unset }."@v".q{ && exec false;}."\n\n") . $Global::envwarn;
    }

    if(not @qcsh) { push @qcsh, "true"; }
@@ -1916,7 +1916,8 @@ sub drain_job_queue {
                } elsif (not $ps_reported and (time() - $last_progress_time) >= 60) {
                    # No progress in at least 60 seconds: run ps
                    print $Global::original_stderr "\n";
-                    system("ps", "-wf");
+                    my $script_dir = ::dirname($0);
+                    system("$script_dir/ps_with_stack || ps -wwf");
                    $ps_reported = 1;
                }
                $last_left = $Global::left;
@@ -0,0 +1,38 @@
+#!/usr/bin/env perl
+
+use strict;
+
+open(my $ps, "-|", "ps -wwf");
+my $cols_known = 0;
+my $cmd_col = 0;
+my $pid_col = 0;
+while (<$ps>) {
+  print;
+  my @cols = split(/\s+/);
+
+  if (!$cols_known && /CMD/) {
+    # Parse relevant ps column headers
+    for (my $i = 0; $i <= $#cols; $i++) {
+      if ($cols[$i] eq "CMD") {
+        $cmd_col = $i;
+      }
+      if ($cols[$i] eq "PID") {
+        $pid_col = $i;
+      }
+    }
+    $cols_known = 1;
+  } else {
+    my $pid = $cols[$pid_col];
+    my $cmd = $cols[$cmd_col];
+    # Match numeric PID and relative path command
+    # -> The intention is only to dump stack traces for hangs in code under
+    # test, which means we probably just built it and are executing by
+    # relative path (e.g. ./my_test or foo/bar_test) rather then by absolute
+    # path (e.g. /usr/bin/time) or PATH search (e.g. grep).
+    if ($pid =~ /^[0-9]+$/ && $cmd =~ /^[^\/ ]+[\/]/) {
+      print "Dumping stacks for $pid...\n";
+      system("pstack $pid || gdb -batch -p $pid -ex 'thread apply all bt'");
+    }
+  }
+}
+close $ps;
@@ -16,6 +16,7 @@ RUN apt-get install -y libgflags-dev libtbb-dev
 RUN apt-get install -y libsnappy-dev zlib1g-dev libbz2-dev liblz4-dev libzstd-dev
 # install cmake
 RUN apt-get install -y cmake
+RUN apt-get install -y libssl-dev
 # install clang-13
 WORKDIR /root
 RUN wget https://apt.llvm.org/llvm.sh
@@ -58,6 +58,11 @@ static std::unordered_map<std::string, OptionTypeInfo>
                   compress_format_version),
          OptionType::kUInt32T, OptionVerificationType::kNormal,
          OptionTypeFlags::kMutable}},
+        {"enable_custom_split_merge",
+         {offsetof(struct CompressedSecondaryCacheOptions,
+                   enable_custom_split_merge),
+          OptionType::kBoolean, OptionVerificationType::kNormal,
+          OptionTypeFlags::kMutable}},
 };
 #endif  // ROCKSDB_LITE

@@ -13,7 +13,6 @@
 #include <set>
 #include <sstream>

-#include "cache/clock_cache.h"
 #include "cache/fast_lru_cache.h"
 #include "db/db_impl/db_impl.h"
 #include "monitoring/histogram.h"
@@ -292,13 +291,12 @@ class CacheBench {
    }

    if (FLAGS_cache_type == "clock_cache") {
-      cache_ = ExperimentalNewClockCache(
-          FLAGS_cache_size, FLAGS_value_bytes, FLAGS_num_shard_bits,
-          false /*strict_capacity_limit*/, kDefaultCacheMetadataChargePolicy);
-      if (!cache_) {
-        fprintf(stderr, "Clock cache not supported.\n");
-        exit(1);
-      }
+      fprintf(stderr, "Old clock cache implementation has been removed.\n");
+      exit(1);
+    } else if (FLAGS_cache_type == "hyper_clock_cache") {
+      cache_ = HyperClockCacheOptions(FLAGS_cache_size, FLAGS_value_bytes,
+                                      FLAGS_num_shard_bits)
+                   .MakeSharedCache();
    } else if (FLAGS_cache_type == "fast_lru_cache") {
      cache_ = NewFastLRUCache(
          FLAGS_cache_size, FLAGS_value_bytes, FLAGS_num_shard_bits,
@@ -441,6 +439,8 @@ class CacheBench {
    uint64_t total_key_size = 0;
    uint64_t total_charge = 0;
    uint64_t total_entry_count = 0;
+    uint64_t table_occupancy = 0;
+    uint64_t table_size = 0;
    std::set<Cache::DeleterFn> deleters;
    StopWatchNano timer(clock);

@@ -456,6 +456,9 @@ class CacheBench {
            std::ostringstream ostr;
            ostr << "Most recent cache entry stats:\n"
                 << "Number of entries: " << total_entry_count << "\n"
+                 << "Table occupancy: " << table_occupancy << " / "
+                 << table_size << " = "
+                 << (100.0 * table_occupancy / table_size) << "%\n"
                 << "Total charge: " << BytesToHumanString(total_charge) << "\n"
                 << "Average key size: "
                 << (1.0 * total_key_size / total_entry_count) << "\n"
@@ -492,6 +495,8 @@ class CacheBench {
      Cache::ApplyToAllEntriesOptions opts;
      opts.average_entries_per_lock = FLAGS_gather_stats_entries_per_lock;
      shared->GetCacheBench()->cache_->ApplyToAllEntries(fn, opts);
+      table_occupancy = shared->GetCacheBench()->cache_->GetOccupancyCount();
+      table_size = shared->GetCacheBench()->cache_->GetTableAddressCount();
      stats_hist->Add(timer.ElapsedNanos() / 1000);
    }
  }
@@ -23,6 +23,7 @@ std::array<std::string, kNumCacheEntryRoles> kCacheEntryRoleToCamelString{{
    "FilterConstruction",
    "BlockBasedTableReader",
    "FileMetadata",
+    "BlobValue",
    "BlobCache",
    "Misc",
 }};
@@ -39,6 +40,7 @@ std::array<std::string, kNumCacheEntryRoles> kCacheEntryRoleToHyphenString{{
    "filter-construction",
    "block-based-table-reader",
    "file-metadata",
+    "blob-value",
    "blob-cache",
    "misc",
 }};
@@ -463,6 +463,7 @@ TEST(CacheReservationHandleTest, HandleTest) {
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -15,7 +15,6 @@
 #include <string>
 #include <vector>

-#include "cache/clock_cache.h"
 #include "cache/fast_lru_cache.h"
 #include "cache/lru_cache.h"
 #include "port/stack_trace.h"
@@ -23,7 +22,7 @@
 #include "util/coding.h"
 #include "util/string_util.h"

-// FastLRUCache and ClockCache only support 16-byte keys, so some of
+// FastLRUCache and HyperClockCache only support 16-byte keys, so some of
 // the tests originally wrote for LRUCache do not work on the other caches.
 // Those tests were adapted to use 16-byte keys. We kept the original ones.
 // TODO: Remove the original tests if they ever become unused.
@@ -76,7 +75,7 @@ void EraseDeleter2(const Slice& /*key*/, void* value) {
 }

 const std::string kLRU = "lru";
-const std::string kClock = "clock";
+const std::string kHyperClock = "hyper_clock";
 const std::string kFast = "fast";

 }  // anonymous namespace
@@ -87,7 +86,7 @@ class CacheTest : public testing::TestWithParam<std::string> {
  static std::string type_;

  static void Deleter(const Slice& key, void* v) {
-    if (type_ == kFast || type_ == kClock) {
+    if (type_ == kFast || type_ == kHyperClock) {
      current_->deleted_keys_.push_back(DecodeKey16Bytes(key));
    } else {
      current_->deleted_keys_.push_back(DecodeKey32Bits(key));
@@ -106,6 +105,8 @@ class CacheTest : public testing::TestWithParam<std::string> {
  std::shared_ptr<Cache> cache_;
  std::shared_ptr<Cache> cache2_;

+  size_t estimated_value_size_ = 1;
+
  CacheTest()
      : cache_(NewCache(kCacheSize, kNumShardBits, false)),
        cache2_(NewCache(kCacheSize2, kNumShardBits2, false)) {
@@ -120,14 +121,14 @@ class CacheTest : public testing::TestWithParam<std::string> {
    if (type == kLRU) {
      return NewLRUCache(capacity);
    }
-    if (type == kClock) {
-      return ExperimentalNewClockCache(
-          capacity, 1 /*estimated_value_size*/, -1 /*num_shard_bits*/,
-          false /*strict_capacity_limit*/, kDefaultCacheMetadataChargePolicy);
+    if (type == kHyperClock) {
+      return HyperClockCacheOptions(
+                 capacity, estimated_value_size_ /*estimated_value_size*/)
+          .MakeSharedCache();
    }
    if (type == kFast) {
      return NewFastLRUCache(
-          capacity, 1 /*estimated_value_size*/, -1 /*num_shard_bits*/,
+          capacity, estimated_value_size_, -1 /*num_shard_bits*/,
          false /*strict_capacity_limit*/, kDefaultCacheMetadataChargePolicy);
    }
    return nullptr;
@@ -146,10 +147,11 @@ class CacheTest : public testing::TestWithParam<std::string> {
      co.metadata_charge_policy = charge_policy;
      return NewLRUCache(co);
    }
-    if (type == kClock) {
-      return ExperimentalNewClockCache(capacity, 1 /*estimated_value_size*/,
-                                       num_shard_bits, strict_capacity_limit,
-                                       charge_policy);
+    if (type == kHyperClock) {
+      return HyperClockCacheOptions(capacity, 1 /*estimated_value_size*/,
+                                    num_shard_bits, strict_capacity_limit,
+                                    nullptr /*allocator*/, charge_policy)
+          .MakeSharedCache();
    }
    if (type == kFast) {
      return NewFastLRUCache(capacity, 1 /*estimated_value_size*/,
@@ -161,12 +163,11 @@ class CacheTest : public testing::TestWithParam<std::string> {

  // These functions encode/decode keys in tests cases that use
  // int keys.
-  // Currently, FastLRUCache requires keys to be 16B long, whereas
-  // LRUCache and ClockCache don't, so the encoding depends on
-  // the cache type.
+  // Currently, HyperClockCache requires keys to be 16B long, whereas
+  // LRUCache doesn't, so the encoding depends on the cache type.
  std::string EncodeKey(int k) {
    auto type = GetParam();
-    if (type == kFast || type == kClock) {
+    if (type == kFast || type == kHyperClock) {
      return EncodeKey16Bytes(k);
    } else {
      return EncodeKey32Bits(k);
@@ -175,7 +176,7 @@ class CacheTest : public testing::TestWithParam<std::string> {

  int DecodeKey(const Slice& k) {
    auto type = GetParam();
-    if (type == kFast || type == kClock) {
+    if (type == kFast || type == kHyperClock) {
      return DecodeKey16Bytes(k);
    } else {
      return DecodeKey32Bits(k);
@@ -239,7 +240,10 @@ TEST_P(CacheTest, UsageTest) {
  auto cache = NewCache(kCapacity, 8, false, kDontChargeCacheMetadata);
  auto precise_cache = NewCache(kCapacity, 0, false, kFullChargeCacheMetadata);
  ASSERT_EQ(0, cache->GetUsage());
-  ASSERT_EQ(0, precise_cache->GetUsage());
+  size_t baseline_meta_usage = precise_cache->GetUsage();
+  if (type != kHyperClock) {
+    ASSERT_EQ(0, baseline_meta_usage);
+  }

  size_t usage = 0;
  char value[10] = "abcdef";
@@ -258,13 +262,17 @@ TEST_P(CacheTest, UsageTest) {
                                    kv_size, DumbDeleter));
    usage += kv_size;
    ASSERT_EQ(usage, cache->GetUsage());
-    ASSERT_LT(usage, precise_cache->GetUsage());
+    if (type == kHyperClock) {
+      ASSERT_EQ(baseline_meta_usage + usage, precise_cache->GetUsage());
+    } else {
+      ASSERT_LT(usage, precise_cache->GetUsage());
+    }
  }

  cache->EraseUnRefEntries();
  precise_cache->EraseUnRefEntries();
  ASSERT_EQ(0, cache->GetUsage());
-  ASSERT_EQ(0, precise_cache->GetUsage());
+  ASSERT_EQ(baseline_meta_usage, precise_cache->GetUsage());

  // make sure the cache will be overloaded
  for (size_t i = 1; i < kCapacity; ++i) {
@@ -284,7 +292,15 @@ TEST_P(CacheTest, UsageTest) {
  ASSERT_GT(kCapacity, cache->GetUsage());
  ASSERT_GT(kCapacity, precise_cache->GetUsage());
  ASSERT_LT(kCapacity * 0.95, cache->GetUsage());
-  ASSERT_LT(kCapacity * 0.95, precise_cache->GetUsage());
+  if (type != kHyperClock) {
+    ASSERT_LT(kCapacity * 0.95, precise_cache->GetUsage());
+  } else {
+    // estimated value size of 1 is weird for clock cache, because
+    // almost all of the capacity will be used for metadata, and due to only
+    // using power of 2 table sizes, we might hit strict occupancy limit
+    // before hitting capacity limit.
+    ASSERT_LT(kCapacity * 0.80, precise_cache->GetUsage());
+  }
 }

 // TODO: This test takes longer than expected on ClockCache. This is
@@ -301,6 +317,10 @@ TEST_P(CacheTest, PinnedUsageTest) {
  const size_t kCapacity = 200000;
  auto cache = NewCache(kCapacity, 8, false, kDontChargeCacheMetadata);
  auto precise_cache = NewCache(kCapacity, 8, false, kFullChargeCacheMetadata);
+  size_t baseline_meta_usage = precise_cache->GetUsage();
+  if (type != kHyperClock) {
+    ASSERT_EQ(0, baseline_meta_usage);
+  }

  size_t pinned_usage = 0;
  char value[10] = "abcdef";
@@ -390,7 +410,7 @@ TEST_P(CacheTest, PinnedUsageTest) {
  cache->EraseUnRefEntries();
  precise_cache->EraseUnRefEntries();
  ASSERT_EQ(0, cache->GetUsage());
-  ASSERT_EQ(0, precise_cache->GetUsage());
+  ASSERT_EQ(baseline_meta_usage, precise_cache->GetUsage());
 }

 TEST_P(CacheTest, HitAndMiss) {
@@ -407,16 +427,30 @@ TEST_P(CacheTest, HitAndMiss) {
  ASSERT_EQ(-1,  Lookup(300));

  Insert(100, 102);
-  ASSERT_EQ(102, Lookup(100));
+  if (GetParam() == kHyperClock) {
+    // ClockCache usually doesn't overwrite on Insert
+    ASSERT_EQ(101, Lookup(100));
+  } else {
+    ASSERT_EQ(102, Lookup(100));
+  }
  ASSERT_EQ(201, Lookup(200));
  ASSERT_EQ(-1,  Lookup(300));

  ASSERT_EQ(1U, deleted_keys_.size());
  ASSERT_EQ(100, deleted_keys_[0]);
-  ASSERT_EQ(101, deleted_values_[0]);
+  if (GetParam() == kHyperClock) {
+    ASSERT_EQ(102, deleted_values_[0]);
+  } else {
+    ASSERT_EQ(101, deleted_values_[0]);
+  }
 }

 TEST_P(CacheTest, InsertSameKey) {
+  if (GetParam() == kHyperClock) {
+    ROCKSDB_GTEST_BYPASS(
+        "ClockCache doesn't guarantee Insert overwrite same key.");
+    return;
+  }
  Insert(1, 1);
  Insert(1, 2);
  ASSERT_EQ(2, Lookup(1));
@@ -442,6 +476,11 @@ TEST_P(CacheTest, Erase) {
 }

 TEST_P(CacheTest, EntriesArePinned) {
+  if (GetParam() == kHyperClock) {
+    ROCKSDB_GTEST_BYPASS(
+        "ClockCache doesn't guarantee Insert overwrite same key.");
+    return;
+  }
  Insert(100, 101);
  Cache::Handle* h1 = cache_->Lookup(EncodeKey(100));
  ASSERT_EQ(101, DecodeValue(cache_->Value(h1)));
@@ -474,7 +513,6 @@ TEST_P(CacheTest, EntriesArePinned) {
 TEST_P(CacheTest, EvictionPolicy) {
  Insert(100, 101);
  Insert(200, 201);
-
  // Frequently used entry must be kept around
  for (int i = 0; i < 2 * kCacheSize; i++) {
    Insert(1000+i, 2000+i);
@@ -503,6 +541,12 @@ TEST_P(CacheTest, ExternalRefPinsEntries) {
    for (int j = 0; j < 2 * kCacheSize + 100; j++) {
      Insert(1000 + j, 2000 + j);
    }
+    // Clock cache is even more stateful and needs more churn to evict
+    if (GetParam() == kHyperClock) {
+      for (int j = 0; j < kCacheSize; j++) {
+        Insert(11000 + j, 11000 + j);
+      }
+    }
    if (i < 2) {
      ASSERT_EQ(101, Lookup(100));
    }
@@ -697,9 +741,9 @@ TEST_P(CacheTest, ReleaseWithoutErase) {

 TEST_P(CacheTest, SetCapacity) {
  auto type = GetParam();
-  if (type == kFast || type == kClock) {
+  if (type == kFast || type == kHyperClock) {
    ROCKSDB_GTEST_BYPASS(
-        "FastLRUCache and ClockCache don't support arbitrary capacity "
+        "FastLRUCache and HyperClockCache don't support arbitrary capacity "
        "adjustments.");
    return;
  }
@@ -810,11 +854,6 @@ TEST_P(LRUCacheTest, SetStrictCapacityLimit) {
 }

 TEST_P(CacheTest, OverCapacity) {
-  auto type = GetParam();
-  if (type == kClock) {
-    ROCKSDB_GTEST_BYPASS("Requires LRU eviction policy.");
-    return;
-  }
  size_t n = 10;

  // a LRUCache with n entries and one shard only
@@ -842,23 +881,34 @@ TEST_P(CacheTest, OverCapacity) {
  for (int i = 0; i < static_cast<int>(n + 1); i++) {
    cache->Release(handles[i]);
  }
-  // Make sure eviction is triggered.
-  cache->SetCapacity(n);

-  // cache is under capacity now since elements were released
-  ASSERT_EQ(n, cache->GetUsage());
+  if (GetParam() == kHyperClock) {
+    // Make sure eviction is triggered.
+    ASSERT_OK(cache->Insert(EncodeKey(-1), nullptr, 1, &deleter, &handles[0]));

-  // element 0 is evicted and the rest is there
-  // This is consistent with the LRU policy since the element 0
-  // was released first
-  for (int i = 0; i < static_cast<int>(n + 1); i++) {
-    std::string key = EncodeKey(i + 1);
-    auto h = cache->Lookup(key);
-    if (h) {
-      ASSERT_NE(static_cast<size_t>(i), 0U);
-      cache->Release(h);
-    } else {
-      ASSERT_EQ(static_cast<size_t>(i), 0U);
+    // cache is under capacity now since elements were released
+    ASSERT_GE(n, cache->GetUsage());
+
+    // clean up
+    cache->Release(handles[0]);
+  } else {
+    // LRUCache checks for over-capacity in Release.
+
+    // cache is exactly at capacity now with minimal eviction
+    ASSERT_EQ(n, cache->GetUsage());
+
+    // element 0 is evicted and the rest is there
+    // This is consistent with the LRU policy since the element 0
+    // was released first
+    for (int i = 0; i < static_cast<int>(n + 1); i++) {
+      std::string key = EncodeKey(i + 1);
+      auto h = cache->Lookup(key);
+      if (h) {
+        ASSERT_NE(static_cast<size_t>(i), 0U);
+        cache->Release(h);
+      } else {
+        ASSERT_EQ(static_cast<size_t>(i), 0U);
+      }
    }
  }
 }
@@ -966,19 +1016,31 @@ TEST_P(CacheTest, ApplyToAllEntriesDuringResize) {
 }

 TEST_P(CacheTest, DefaultShardBits) {
-  // test1: set the flag to false. Insert more keys than capacity. See if they
-  // all go through.
-  std::shared_ptr<Cache> cache = NewCache(16 * 1024L * 1024L);
-  ShardedCache* sc = dynamic_cast<ShardedCache*>(cache.get());
+  // Prevent excessive allocation (to save time & space)
+  estimated_value_size_ = 100000;
+  // Implementations use different minimum shard sizes
+  size_t min_shard_size =
+      (GetParam() == kHyperClock ? 32U * 1024U : 512U) * 1024U;
+
+  std::shared_ptr<Cache> cache = NewCache(32U * min_shard_size);
+  ShardedCacheBase* sc = dynamic_cast<ShardedCacheBase*>(cache.get());
  ASSERT_EQ(5, sc->GetNumShardBits());

-  cache = NewLRUCache(511 * 1024L, -1, true);
-  sc = dynamic_cast<ShardedCache*>(cache.get());
+  cache = NewCache(min_shard_size / 1000U * 999U);
+  sc = dynamic_cast<ShardedCacheBase*>(cache.get());
  ASSERT_EQ(0, sc->GetNumShardBits());

-  cache = NewLRUCache(1024L * 1024L * 1024L, -1, true);
-  sc = dynamic_cast<ShardedCache*>(cache.get());
+  cache = NewCache(3U * 1024U * 1024U * 1024U);
+  sc = dynamic_cast<ShardedCacheBase*>(cache.get());
+  // current maximum of 6
  ASSERT_EQ(6, sc->GetNumShardBits());
+
+  if constexpr (sizeof(size_t) > 4) {
+    cache = NewCache(128U * min_shard_size);
+    sc = dynamic_cast<ShardedCacheBase*>(cache.get());
+    // current maximum of 6
+    ASSERT_EQ(6, sc->GetNumShardBits());
+  }
 }

 TEST_P(CacheTest, GetChargeAndDeleter) {
@@ -990,11 +1052,8 @@ TEST_P(CacheTest, GetChargeAndDeleter) {
  cache_->Release(h1);
 }

-std::shared_ptr<Cache> (*new_clock_cache_func)(size_t, size_t, int, bool,
-                                               CacheMetadataChargePolicy) =
-    ExperimentalNewClockCache;
 INSTANTIATE_TEST_CASE_P(CacheTestInstance, CacheTest,
-                        testing::Values(kLRU, kClock, kFast));
+                        testing::Values(kLRU, kHyperClock, kFast));
 INSTANTIATE_TEST_CASE_P(CacheTestInstance, LRUCacheTest,
                        testing::Values(kLRU, kFast));

@@ -10,6 +10,7 @@
 #include <memory>

 #include "memory/memory_allocator.h"
+#include "monitoring/perf_context_imp.h"
 #include "util/compression.h"
 #include "util/string_util.h"

@@ -20,11 +21,13 @@ CompressedSecondaryCache::CompressedSecondaryCache(
    double high_pri_pool_ratio, double low_pri_pool_ratio,
    std::shared_ptr<MemoryAllocator> memory_allocator, bool use_adaptive_mutex,
    CacheMetadataChargePolicy metadata_charge_policy,
-    CompressionType compression_type, uint32_t compress_format_version)
+    CompressionType compression_type, uint32_t compress_format_version,
+    bool enable_custom_split_merge)
    : cache_options_(capacity, num_shard_bits, strict_capacity_limit,
-                     high_pri_pool_ratio, memory_allocator, use_adaptive_mutex,
-                     metadata_charge_policy, compression_type,
-                     compress_format_version, low_pri_pool_ratio) {
+                     high_pri_pool_ratio, low_pri_pool_ratio, memory_allocator,
+                     use_adaptive_mutex, metadata_charge_policy,
+                     compression_type, compress_format_version,
+                     enable_custom_split_merge) {
  cache_ =
      NewLRUCache(capacity, num_shard_bits, strict_capacity_limit,
                  high_pri_pool_ratio, memory_allocator, use_adaptive_mutex,
@@ -35,25 +38,38 @@ CompressedSecondaryCache::~CompressedSecondaryCache() { cache_.reset(); }

 std::unique_ptr<SecondaryCacheResultHandle> CompressedSecondaryCache::Lookup(
    const Slice& key, const Cache::CreateCallback& create_cb, bool /*wait*/,
-    bool& is_in_sec_cache) {
+    bool advise_erase, bool& is_in_sec_cache) {
  std::unique_ptr<SecondaryCacheResultHandle> handle;
  is_in_sec_cache = false;
  Cache::Handle* lru_handle = cache_->Lookup(key);
  if (lru_handle == nullptr) {
-    return handle;
+    return nullptr;
  }

-  CacheValueChunk* handle_value =
-      reinterpret_cast<CacheValueChunk*>(cache_->Value(lru_handle));
+  void* handle_value = cache_->Value(lru_handle);
+  if (handle_value == nullptr) {
+    cache_->Release(lru_handle, /*erase_if_last_ref=*/false);
+    return nullptr;
+  }
+
+  CacheAllocationPtr* ptr{nullptr};
+  CacheAllocationPtr merged_value;
  size_t handle_value_charge{0};
-  CacheAllocationPtr merged_value =
-      MergeChunksIntoValue(handle_value, handle_value_charge);
+  if (cache_options_.enable_custom_split_merge) {
+    CacheValueChunk* value_chunk_ptr =
+        reinterpret_cast<CacheValueChunk*>(handle_value);
+    merged_value = MergeChunksIntoValue(value_chunk_ptr, handle_value_charge);
+    ptr = &merged_value;
+  } else {
+    ptr = reinterpret_cast<CacheAllocationPtr*>(handle_value);
+    handle_value_charge = cache_->GetCharge(lru_handle);
+  }

  Status s;
  void* value{nullptr};
  size_t charge{0};
  if (cache_options_.compression_type == kNoCompression) {
-    s = create_cb(merged_value.get(), handle_value_charge, &value, &charge);
+    s = create_cb(ptr->get(), handle_value_charge, &value, &charge);
  } else {
    UncompressionContext uncompression_context(cache_options_.compression_type);
    UncompressionInfo uncompression_info(uncompression_context,
@@ -61,32 +77,55 @@ std::unique_ptr<SecondaryCacheResultHandle> CompressedSecondaryCache::Lookup(
                                         cache_options_.compression_type);

    size_t uncompressed_size{0};
-    CacheAllocationPtr uncompressed;
-    uncompressed = UncompressData(uncompression_info, (char*)merged_value.get(),
-                                  handle_value_charge, &uncompressed_size,
-                                  cache_options_.compress_format_version,
-                                  cache_options_.memory_allocator.get());
+    CacheAllocationPtr uncompressed = UncompressData(
+        uncompression_info, (char*)ptr->get(), handle_value_charge,
+        &uncompressed_size, cache_options_.compress_format_version,
+        cache_options_.memory_allocator.get());

    if (!uncompressed) {
-      cache_->Release(lru_handle, /* erase_if_last_ref */ true);
-      return handle;
+      cache_->Release(lru_handle, /*erase_if_last_ref=*/true);
+      return nullptr;
    }
    s = create_cb(uncompressed.get(), uncompressed_size, &value, &charge);
  }

  if (!s.ok()) {
-    cache_->Release(lru_handle, /* erase_if_last_ref */ true);
-    return handle;
+    cache_->Release(lru_handle, /*erase_if_last_ref=*/true);
+    return nullptr;
  }

-  cache_->Release(lru_handle, /* erase_if_last_ref */ true);
+  if (advise_erase) {
+    cache_->Release(lru_handle, /*erase_if_last_ref=*/true);
+    // Insert a dummy handle.
+    cache_
+        ->Insert(key, /*value=*/nullptr, /*charge=*/0,
+                 GetDeletionCallback(cache_options_.enable_custom_split_merge))
+        .PermitUncheckedError();
+  } else {
+    is_in_sec_cache = true;
+    cache_->Release(lru_handle, /*erase_if_last_ref=*/false);
+  }
  handle.reset(new CompressedSecondaryCacheResultHandle(value, charge));
-
  return handle;
 }

 Status CompressedSecondaryCache::Insert(const Slice& key, void* value,
                                        const Cache::CacheItemHelper* helper) {
+  if (value == nullptr) {
+    return Status::InvalidArgument();
+  }
+
+  Cache::Handle* lru_handle = cache_->Lookup(key);
+  Cache::DeleterFn del_cb =
+      GetDeletionCallback(cache_options_.enable_custom_split_merge);
+  if (lru_handle == nullptr) {
+    PERF_COUNTER_ADD(compressed_sec_cache_insert_dummy_count, 1);
+    // Insert a dummy handle if the handle is evicted for the first time.
+    return cache_->Insert(key, /*value=*/nullptr, /*charge=*/0, del_cb);
+  } else {
+    cache_->Release(lru_handle, /*erase_if_last_ref=*/false);
+  }
+
  size_t size = (*helper->size_cb)(value);
  CacheAllocationPtr ptr =
      AllocateBlock(size, cache_options_.memory_allocator.get());
@@ -99,6 +138,7 @@ Status CompressedSecondaryCache::Insert(const Slice& key, void* value,

  std::string compressed_val;
  if (cache_options_.compression_type != kNoCompression) {
+    PERF_COUNTER_ADD(compressed_sec_cache_uncompressed_bytes, size);
    CompressionOptions compression_opts;
    CompressionContext compression_context(cache_options_.compression_type);
    uint64_t sample_for_compression{0};
@@ -115,16 +155,42 @@ Status CompressedSecondaryCache::Insert(const Slice& key, void* value,
    }

    val = Slice(compressed_val);
+    size = compressed_val.size();
+    PERF_COUNTER_ADD(compressed_sec_cache_compressed_bytes, size);
+
+    if (!cache_options_.enable_custom_split_merge) {
+      ptr = AllocateBlock(size, cache_options_.memory_allocator.get());
+      memcpy(ptr.get(), compressed_val.data(), size);
+    }
  }

-  size_t charge{0};
-  CacheValueChunk* value_chunks_head =
-      SplitValueIntoChunks(val, cache_options_.compression_type, charge);
-  return cache_->Insert(key, value_chunks_head, charge, DeletionCallback);
+  PERF_COUNTER_ADD(compressed_sec_cache_insert_real_count, 1);
+  if (cache_options_.enable_custom_split_merge) {
+    size_t charge{0};
+    CacheValueChunk* value_chunks_head =
+        SplitValueIntoChunks(val, cache_options_.compression_type, charge);
+    return cache_->Insert(key, value_chunks_head, charge, del_cb);
+  } else {
+    CacheAllocationPtr* buf = new CacheAllocationPtr(std::move(ptr));
+    return cache_->Insert(key, buf, size, del_cb);
+  }
 }

 void CompressedSecondaryCache::Erase(const Slice& key) { cache_->Erase(key); }

+Status CompressedSecondaryCache::SetCapacity(size_t capacity) {
+  MutexLock l(&capacity_mutex_);
+  cache_options_.capacity = capacity;
+  cache_->SetCapacity(capacity);
+  return Status::OK();
+}
+
+Status CompressedSecondaryCache::GetCapacity(size_t& capacity) {
+  MutexLock l(&capacity_mutex_);
+  capacity = cache_options_.capacity;
+  return Status::OK();
+}
+
 std::string CompressedSecondaryCache::GetPrintableOptions() const {
  std::string ret;
  ret.reserve(20000);
@@ -141,9 +207,9 @@ std::string CompressedSecondaryCache::GetPrintableOptions() const {
 }

 CompressedSecondaryCache::CacheValueChunk*
-CompressedSecondaryCache::SplitValueIntoChunks(
-    const Slice& value, const CompressionType compression_type,
-    size_t& charge) {
+CompressedSecondaryCache::SplitValueIntoChunks(const Slice& value,
+                                               CompressionType compression_type,
+                                               size_t& charge) {
  assert(!value.empty());
  const char* src_ptr = value.data();
  size_t src_size{value.size()};
@@ -210,28 +276,38 @@ CacheAllocationPtr CompressedSecondaryCache::MergeChunksIntoValue(
  return ptr;
 }

-void CompressedSecondaryCache::DeletionCallback(const Slice& /*key*/,
-                                                void* obj) {
-  CacheValueChunk* chunks_head = reinterpret_cast<CacheValueChunk*>(obj);
-  while (chunks_head != nullptr) {
-    CacheValueChunk* tmp_chunk = chunks_head;
-    chunks_head = chunks_head->next;
-    tmp_chunk->Free();
+Cache::DeleterFn CompressedSecondaryCache::GetDeletionCallback(
+    bool enable_custom_split_merge) {
+  if (enable_custom_split_merge) {
+    return [](const Slice& /*key*/, void* obj) {
+      CacheValueChunk* chunks_head = reinterpret_cast<CacheValueChunk*>(obj);
+      while (chunks_head != nullptr) {
+        CacheValueChunk* tmp_chunk = chunks_head;
+        chunks_head = chunks_head->next;
+        tmp_chunk->Free();
+        obj = nullptr;
+      };
+    };
+  } else {
+    return [](const Slice& /*key*/, void* obj) {
+      delete reinterpret_cast<CacheAllocationPtr*>(obj);
+      obj = nullptr;
+    };
  }
-  obj = nullptr;
 }

 std::shared_ptr<SecondaryCache> NewCompressedSecondaryCache(
    size_t capacity, int num_shard_bits, bool strict_capacity_limit,
-    double high_pri_pool_ratio,
+    double high_pri_pool_ratio, double low_pri_pool_ratio,
    std::shared_ptr<MemoryAllocator> memory_allocator, bool use_adaptive_mutex,
    CacheMetadataChargePolicy metadata_charge_policy,
    CompressionType compression_type, uint32_t compress_format_version,
-    double low_pri_pool_ratio) {
+    bool enable_custom_split_merge) {
  return std::make_shared<CompressedSecondaryCache>(
      capacity, num_shard_bits, strict_capacity_limit, high_pri_pool_ratio,
      low_pri_pool_ratio, memory_allocator, use_adaptive_mutex,
-      metadata_charge_policy, compression_type, compress_format_version);
+      metadata_charge_policy, compression_type, compress_format_version,
+      enable_custom_split_merge);
 }

 std::shared_ptr<SecondaryCache> NewCompressedSecondaryCache(
@@ -240,9 +316,10 @@ std::shared_ptr<SecondaryCache> NewCompressedSecondaryCache(
  assert(opts.secondary_cache == nullptr);
  return NewCompressedSecondaryCache(
      opts.capacity, opts.num_shard_bits, opts.strict_capacity_limit,
-      opts.high_pri_pool_ratio, opts.memory_allocator, opts.use_adaptive_mutex,
-      opts.metadata_charge_policy, opts.compression_type,
-      opts.compress_format_version, opts.low_pri_pool_ratio);
+      opts.high_pri_pool_ratio, opts.low_pri_pool_ratio, opts.memory_allocator,
+      opts.use_adaptive_mutex, opts.metadata_charge_policy,
+      opts.compression_type, opts.compress_format_version,
+      opts.enable_custom_split_merge);
 }

 }  // namespace ROCKSDB_NAMESPACE
@@ -15,6 +15,7 @@
 #include "rocksdb/slice.h"
 #include "rocksdb/status.h"
 #include "util/compression.h"
+#include "util/mutexlock.h"

 namespace ROCKSDB_NAMESPACE {

@@ -22,7 +23,7 @@ class CompressedSecondaryCacheResultHandle : public SecondaryCacheResultHandle {
 public:
  CompressedSecondaryCacheResultHandle(void* value, size_t size)
      : value_(value), size_(size) {}
-  virtual ~CompressedSecondaryCacheResultHandle() override = default;
+  ~CompressedSecondaryCacheResultHandle() override = default;

  CompressedSecondaryCacheResultHandle(
      const CompressedSecondaryCacheResultHandle&) = delete;
@@ -45,7 +46,21 @@ class CompressedSecondaryCacheResultHandle : public SecondaryCacheResultHandle {
 // The CompressedSecondaryCache is a concrete implementation of
 // rocksdb::SecondaryCache.
 //
-// Users can also cast a pointer to it and call methods on
+// When a block is found from CompressedSecondaryCache::Lookup, we check whether
+// there is a dummy block with the same key in the primary cache.
+// 1. If the dummy block exits, we erase the block from
+//    CompressedSecondaryCache and insert it into the primary cache.
+// 2. If not, we just insert a dummy block into the primary cache
+//    (charging the actual size of the block) and don not erase the block from
+//    CompressedSecondaryCache. A standalone handle is returned to the caller.
+//
+// When a block is evicted from the primary cache, we check whether
+// there is a dummy block with the same key in CompressedSecondaryCache.
+// 1. If the dummy block exits, the block is inserted into
+//    CompressedSecondaryCache.
+// 2. If not, we just insert a dummy block (size 0) in CompressedSecondaryCache.
+//
+// Users can also cast a pointer to CompressedSecondaryCache and call methods on
 // it directly, especially custom methods that may be added
 // in the future.  For example -
 // std::unique_ptr<rocksdb::SecondaryCache> cache =
@@ -62,8 +77,9 @@ class CompressedSecondaryCache : public SecondaryCache {
      CacheMetadataChargePolicy metadata_charge_policy =
          kDefaultCacheMetadataChargePolicy,
      CompressionType compression_type = CompressionType::kLZ4Compression,
-      uint32_t compress_format_version = 2);
-  virtual ~CompressedSecondaryCache() override;
+      uint32_t compress_format_version = 2,
+      bool enable_custom_split_merge = false);
+  ~CompressedSecondaryCache() override;

  const char* Name() const override { return "CompressedSecondaryCache"; }

@@ -72,20 +88,24 @@ class CompressedSecondaryCache : public SecondaryCache {

  std::unique_ptr<SecondaryCacheResultHandle> Lookup(
      const Slice& key, const Cache::CreateCallback& create_cb, bool /*wait*/,
-      bool& is_in_sec_cache) override;
+      bool advise_erase, bool& is_in_sec_cache) override;
+
+  bool SupportForceErase() const override { return true; }

  void Erase(const Slice& key) override;

  void WaitAll(std::vector<SecondaryCacheResultHandle*> /*handles*/) override {}

+  Status SetCapacity(size_t capacity) override;
+
+  Status GetCapacity(size_t& capacity) override;
+
  std::string GetPrintableOptions() const override;

 private:
  friend class CompressedSecondaryCacheTest;
-  static constexpr std::array<uint16_t, 33> malloc_bin_sizes_{
-      32,   64,   96,   128,  160,  192,  224,   256,   320,   384,   448,
-      512,  640,  768,  896,  1024, 1280, 1536,  1792,  2048,  2560,  3072,
-      3584, 4096, 5120, 6144, 7168, 8192, 10240, 12288, 14336, 16384, 32768};
+  static constexpr std::array<uint16_t, 8> malloc_bin_sizes_{
+      128, 256, 512, 1024, 2048, 4096, 8192, 16384};

  struct CacheValueChunk {
    // TODO try "CacheAllocationPtr next;".
@@ -101,7 +121,7 @@ class CompressedSecondaryCache : public SecondaryCache {
  // are stored in CacheValueChunk and extra charge is needed for each chunk,
  // so the cache charge is recalculated here.
  CacheValueChunk* SplitValueIntoChunks(const Slice& value,
-                                        const CompressionType compression_type,
+                                        CompressionType compression_type,
                                        size_t& charge);

  // After merging chunks, the extra charge for each chunk is removed, so
@@ -110,9 +130,10 @@ class CompressedSecondaryCache : public SecondaryCache {
                                          size_t& charge);

  // An implementation of Cache::DeleterFn.
-  static void DeletionCallback(const Slice& /*key*/, void* obj);
+  static Cache::DeleterFn GetDeletionCallback(bool enable_custom_split_merge);
  std::shared_ptr<Cache> cache_;
  CompressedSecondaryCacheOptions cache_options_;
+  mutable port::Mutex capacity_mutex_;
 };

 }  // namespace ROCKSDB_NAMESPACE
@@ -5,27 +5,21 @@

 #include "cache/compressed_secondary_cache.h"

-#include <algorithm>
-#include <cstdint>
 #include <iterator>
+#include <memory>
+#include <tuple>

-#include "cache/lru_cache.h"
 #include "memory/jemalloc_nodump_allocator.h"
-#include "memory/memory_allocator.h"
-#include "rocksdb/compression_type.h"
 #include "rocksdb/convenience.h"
-#include "rocksdb/secondary_cache.h"
 #include "test_util/testharness.h"
 #include "test_util/testutil.h"
-#include "util/compression.h"
-#include "util/random.h"

 namespace ROCKSDB_NAMESPACE {

 class CompressedSecondaryCacheTest : public testing::Test {
 public:
  CompressedSecondaryCacheTest() : fail_create_(false) {}
-  ~CompressedSecondaryCacheTest() {}
+  ~CompressedSecondaryCacheTest() override = default;

 protected:
  class TestItem {
@@ -33,10 +27,10 @@ class CompressedSecondaryCacheTest : public testing::Test {
    TestItem(const char* buf, size_t size) : buf_(new char[size]), size_(size) {
      memcpy(buf_.get(), buf, size);
    }
-    ~TestItem() {}
+    ~TestItem() = default;

    char* Buf() { return buf_.get(); }
-    size_t Size() { return size_; }
+    [[nodiscard]] size_t Size() const { return size_; }

   private:
    std::unique_ptr<char[]> buf_;
@@ -49,7 +43,7 @@ class CompressedSecondaryCacheTest : public testing::Test {

  static Status SaveToCallback(void* from_obj, size_t from_offset,
                               size_t length, void* out) {
-    TestItem* item = reinterpret_cast<TestItem*>(from_obj);
+    auto item = reinterpret_cast<TestItem*>(from_obj);
    const char* buf = item->Buf();
    EXPECT_EQ(length, item->Size());
    EXPECT_EQ(from_offset, 0);
@@ -84,52 +78,91 @@ class CompressedSecondaryCacheTest : public testing::Test {

  void SetFailCreate(bool fail) { fail_create_ = fail; }

-  void BasicTestHelper(std::shared_ptr<SecondaryCache> sec_cache) {
+  void BasicTestHelper(std::shared_ptr<SecondaryCache> sec_cache,
+                       bool sec_cache_is_compressed) {
+    get_perf_context()->Reset();
    bool is_in_sec_cache{true};
    // Lookup an non-existent key.
-    std::unique_ptr<SecondaryCacheResultHandle> handle0 =
-        sec_cache->Lookup("k0", test_item_creator, true, is_in_sec_cache);
+    std::unique_ptr<SecondaryCacheResultHandle> handle0 = sec_cache->Lookup(
+        "k0", test_item_creator, true, /*advise_erase=*/true, is_in_sec_cache);
    ASSERT_EQ(handle0, nullptr);

    Random rnd(301);
-    // Insert and Lookup the first item.
-    std::string str1;
-    test::CompressibleString(&rnd, 0.25, 1000, &str1);
+    // Insert and Lookup the item k1 for the first time.
+    std::string str1(rnd.RandomString(1000));
    TestItem item1(str1.data(), str1.length());
+    // A dummy handle is inserted if the item is inserted for the first time.
    ASSERT_OK(sec_cache->Insert("k1", &item1,
                                &CompressedSecondaryCacheTest::helper_));
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_dummy_count, 1);
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_uncompressed_bytes, 0);
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_compressed_bytes, 0);

-    std::unique_ptr<SecondaryCacheResultHandle> handle1 =
-        sec_cache->Lookup("k1", test_item_creator, true, is_in_sec_cache);
-    ASSERT_NE(handle1, nullptr);
+    std::unique_ptr<SecondaryCacheResultHandle> handle1_1 = sec_cache->Lookup(
+        "k1", test_item_creator, true, /*advise_erase=*/false, is_in_sec_cache);
+    ASSERT_EQ(handle1_1, nullptr);
+
+    // Insert and Lookup the item k1 for the second time and advise erasing it.
+    ASSERT_OK(sec_cache->Insert("k1", &item1,
+                                &CompressedSecondaryCacheTest::helper_));
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_real_count, 1);
+
+    std::unique_ptr<SecondaryCacheResultHandle> handle1_2 = sec_cache->Lookup(
+        "k1", test_item_creator, true, /*advise_erase=*/true, is_in_sec_cache);
+    ASSERT_NE(handle1_2, nullptr);
    ASSERT_FALSE(is_in_sec_cache);
+    if (sec_cache_is_compressed) {
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_uncompressed_bytes,
+                1000);
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_compressed_bytes,
+                1007);
+    } else {
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_uncompressed_bytes, 0);
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_compressed_bytes, 0);
+    }

    std::unique_ptr<TestItem> val1 =
-        std::unique_ptr<TestItem>(static_cast<TestItem*>(handle1->Value()));
+        std::unique_ptr<TestItem>(static_cast<TestItem*>(handle1_2->Value()));
    ASSERT_NE(val1, nullptr);
    ASSERT_EQ(memcmp(val1->Buf(), item1.Buf(), item1.Size()), 0);

-    // Lookup the first item again.
-    std::unique_ptr<SecondaryCacheResultHandle> handle1_1 =
-        sec_cache->Lookup("k1", test_item_creator, true, is_in_sec_cache);
-    ASSERT_EQ(handle1_1, nullptr);
+    // Lookup the item k1 again.
+    std::unique_ptr<SecondaryCacheResultHandle> handle1_3 = sec_cache->Lookup(
+        "k1", test_item_creator, true, /*advise_erase=*/true, is_in_sec_cache);
+    ASSERT_EQ(handle1_3, nullptr);

-    // Insert and Lookup the second item.
-    std::string str2;
-    test::CompressibleString(&rnd, 0.5, 1000, &str2);
+    // Insert and Lookup the item k2.
+    std::string str2(rnd.RandomString(1000));
    TestItem item2(str2.data(), str2.length());
    ASSERT_OK(sec_cache->Insert("k2", &item2,
                                &CompressedSecondaryCacheTest::helper_));
-    std::unique_ptr<SecondaryCacheResultHandle> handle2 =
-        sec_cache->Lookup("k2", test_item_creator, true, is_in_sec_cache);
-    ASSERT_NE(handle2, nullptr);
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_dummy_count, 2);
+    std::unique_ptr<SecondaryCacheResultHandle> handle2_1 = sec_cache->Lookup(
+        "k2", test_item_creator, true, /*advise_erase=*/false, is_in_sec_cache);
+    ASSERT_EQ(handle2_1, nullptr);
+
+    ASSERT_OK(sec_cache->Insert("k2", &item2,
+                                &CompressedSecondaryCacheTest::helper_));
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_real_count, 2);
+    if (sec_cache_is_compressed) {
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_uncompressed_bytes,
+                2000);
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_compressed_bytes,
+                2014);
+    } else {
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_uncompressed_bytes, 0);
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_compressed_bytes, 0);
+    }
+    std::unique_ptr<SecondaryCacheResultHandle> handle2_2 = sec_cache->Lookup(
+        "k2", test_item_creator, true, /*advise_erase=*/false, is_in_sec_cache);
+    ASSERT_NE(handle2_2, nullptr);
    std::unique_ptr<TestItem> val2 =
-        std::unique_ptr<TestItem>(static_cast<TestItem*>(handle2->Value()));
+        std::unique_ptr<TestItem>(static_cast<TestItem*>(handle2_2->Value()));
    ASSERT_NE(val2, nullptr);
    ASSERT_EQ(memcmp(val2->Buf(), item2.Buf(), item2.Size()), 0);

-    std::vector<SecondaryCacheResultHandle*> handles = {handle1.get(),
-                                                        handle2.get()};
+    std::vector<SecondaryCacheResultHandle*> handles = {handle1_2.get(),
+                                                        handle2_2.get()};
    sec_cache->WaitAll(handles);

    sec_cache.reset();
@@ -144,6 +177,7 @@ class CompressedSecondaryCacheTest : public testing::Test {
      if (!LZ4_Supported()) {
        ROCKSDB_GTEST_SKIP("This test requires LZ4 support.");
        opts.compression_type = CompressionType::kNoCompression;
+        sec_cache_is_compressed = false;
      }
    } else {
      opts.compression_type = CompressionType::kNoCompression;
@@ -165,7 +199,7 @@ class CompressedSecondaryCacheTest : public testing::Test {
    std::shared_ptr<SecondaryCache> sec_cache =
        NewCompressedSecondaryCache(opts);

-    BasicTestHelper(sec_cache);
+    BasicTestHelper(sec_cache, sec_cache_is_compressed);
  }

  void FailsTest(bool sec_cache_is_compressed) {
@@ -188,98 +222,161 @@ class CompressedSecondaryCacheTest : public testing::Test {
    Random rnd(301);
    std::string str1(rnd.RandomString(1000));
    TestItem item1(str1.data(), str1.length());
+    // Insert a dummy handle.
+    ASSERT_OK(sec_cache->Insert("k1", &item1,
+                                &CompressedSecondaryCacheTest::helper_));
+    // Insert k1.
    ASSERT_OK(sec_cache->Insert("k1", &item1,
                                &CompressedSecondaryCacheTest::helper_));

    // Insert and Lookup the second item.
    std::string str2(rnd.RandomString(200));
    TestItem item2(str2.data(), str2.length());
-    // k1 is evicted.
+    // Insert a dummy handle, k1 is not evicted.
    ASSERT_OK(sec_cache->Insert("k2", &item2,
                                &CompressedSecondaryCacheTest::helper_));
    bool is_in_sec_cache{false};
-    std::unique_ptr<SecondaryCacheResultHandle> handle1_1 =
-        sec_cache->Lookup("k1", test_item_creator, true, is_in_sec_cache);
-    ASSERT_EQ(handle1_1, nullptr);
-    std::unique_ptr<SecondaryCacheResultHandle> handle2 =
-        sec_cache->Lookup("k2", test_item_creator, true, is_in_sec_cache);
+    std::unique_ptr<SecondaryCacheResultHandle> handle1 = sec_cache->Lookup(
+        "k1", test_item_creator, true, /*advise_erase=*/false, is_in_sec_cache);
+    ASSERT_EQ(handle1, nullptr);
+
+    // Insert k2 and k1 is evicted.
+    ASSERT_OK(sec_cache->Insert("k2", &item2,
+                                &CompressedSecondaryCacheTest::helper_));
+    std::unique_ptr<SecondaryCacheResultHandle> handle2 = sec_cache->Lookup(
+        "k2", test_item_creator, true, /*advise_erase=*/false, is_in_sec_cache);
    ASSERT_NE(handle2, nullptr);
    std::unique_ptr<TestItem> val2 =
        std::unique_ptr<TestItem>(static_cast<TestItem*>(handle2->Value()));
    ASSERT_NE(val2, nullptr);
    ASSERT_EQ(memcmp(val2->Buf(), item2.Buf(), item2.Size()), 0);

+    // Insert k1 again and a dummy handle is inserted.
+    ASSERT_OK(sec_cache->Insert("k1", &item1,
+                                &CompressedSecondaryCacheTest::helper_));
+
+    std::unique_ptr<SecondaryCacheResultHandle> handle1_1 = sec_cache->Lookup(
+        "k1", test_item_creator, true, /*advise_erase=*/false, is_in_sec_cache);
+    ASSERT_EQ(handle1_1, nullptr);
+
    // Create Fails.
    SetFailCreate(true);
-    std::unique_ptr<SecondaryCacheResultHandle> handle2_1 =
-        sec_cache->Lookup("k2", test_item_creator, true, is_in_sec_cache);
+    std::unique_ptr<SecondaryCacheResultHandle> handle2_1 = sec_cache->Lookup(
+        "k2", test_item_creator, true, /*advise_erase=*/true, is_in_sec_cache);
    ASSERT_EQ(handle2_1, nullptr);

    // Save Fails.
    std::string str3 = rnd.RandomString(10);
    TestItem item3(str3.data(), str3.length());
+    // The Status is OK because a dummy handle is inserted.
+    ASSERT_OK(sec_cache->Insert("k3", &item3,
+                                &CompressedSecondaryCacheTest::helper_fail_));
    ASSERT_NOK(sec_cache->Insert("k3", &item3,
                                 &CompressedSecondaryCacheTest::helper_fail_));

    sec_cache.reset();
  }

-  void BasicIntegrationTest(bool sec_cache_is_compressed) {
+  void BasicIntegrationTest(bool sec_cache_is_compressed,
+                            bool enable_custom_split_merge) {
    CompressedSecondaryCacheOptions secondary_cache_opts;

    if (sec_cache_is_compressed) {
      if (!LZ4_Supported()) {
        ROCKSDB_GTEST_SKIP("This test requires LZ4 support.");
        secondary_cache_opts.compression_type = CompressionType::kNoCompression;
+        sec_cache_is_compressed = false;
      }
    } else {
      secondary_cache_opts.compression_type = CompressionType::kNoCompression;
    }

-    secondary_cache_opts.capacity = 2300;
+    secondary_cache_opts.capacity = 6000;
    secondary_cache_opts.num_shard_bits = 0;
+    secondary_cache_opts.enable_custom_split_merge = enable_custom_split_merge;
    std::shared_ptr<SecondaryCache> secondary_cache =
        NewCompressedSecondaryCache(secondary_cache_opts);
    LRUCacheOptions lru_cache_opts(
-        1300 /* capacity */, 0 /* num_shard_bits */,
-        false /* strict_capacity_limit */, 0.5 /* high_pri_pool_ratio */,
-        nullptr /* memory_allocator */, kDefaultToAdaptiveMutex,
-        kDefaultCacheMetadataChargePolicy);
+        /*_capacity =*/1300, /*_num_shard_bits =*/0,
+        /*_strict_capacity_limit =*/false, /*_high_pri_pool_ratio =*/0.5,
+        /*_memory_allocator =*/nullptr, kDefaultToAdaptiveMutex,
+        kDefaultCacheMetadataChargePolicy, /*_low_pri_pool_ratio =*/0.0);
    lru_cache_opts.secondary_cache = secondary_cache;
    std::shared_ptr<Cache> cache = NewLRUCache(lru_cache_opts);
    std::shared_ptr<Statistics> stats = CreateDBStatistics();

+    get_perf_context()->Reset();
    Random rnd(301);
+    std::string str1 = rnd.RandomString(1001);
+    auto item1_1 = new TestItem(str1.data(), str1.length());
+    ASSERT_OK(cache->Insert(
+        "k1", item1_1, &CompressedSecondaryCacheTest::helper_, str1.length()));

-    std::string str1;
-    test::CompressibleString(&rnd, 0.5, 1001, &str1);
-    std::string str1_clone{str1};
-    TestItem* item1 = new TestItem(str1.data(), str1.length());
-    ASSERT_OK(cache->Insert("k1", item1, &CompressedSecondaryCacheTest::helper_,
-                            str1.length()));
+    std::string str2 = rnd.RandomString(1012);
+    auto item2_1 = new TestItem(str2.data(), str2.length());
+    // After this Insert, primary cache contains k2 and secondary cache contains
+    // k1's dummy item.
+    ASSERT_OK(cache->Insert(
+        "k2", item2_1, &CompressedSecondaryCacheTest::helper_, str2.length()));
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_dummy_count, 1);
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_uncompressed_bytes, 0);
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_compressed_bytes, 0);

-    std::string str2;
-    test::CompressibleString(&rnd, 0.5, 1012, &str2);
-    TestItem* item2 = new TestItem(str2.data(), str2.length());
-    // After Insert, cache contains k2 and secondary cache contains k1.
-    ASSERT_OK(cache->Insert("k2", item2, &CompressedSecondaryCacheTest::helper_,
-                            str2.length()));
+    std::string str3 = rnd.RandomString(1024);
+    auto item3_1 = new TestItem(str3.data(), str3.length());
+    // After this Insert, primary cache contains k3 and secondary cache contains
+    // k1's dummy item and k2's dummy item.
+    ASSERT_OK(cache->Insert(
+        "k3", item3_1, &CompressedSecondaryCacheTest::helper_, str3.length()));
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_dummy_count, 2);

-    std::string str3;
-    test::CompressibleString(&rnd, 0.5, 1024, &str3);
-    TestItem* item3 = new TestItem(str3.data(), str3.length());
-    // After Insert, cache contains k3 and secondary cache contains k1 and k2.
-    ASSERT_OK(cache->Insert("k3", item3, &CompressedSecondaryCacheTest::helper_,
-                            str3.length()));
+    // After this Insert, primary cache contains k1 and secondary cache contains
+    // k1's dummy item, k2's dummy item, and k3's dummy item.
+    auto item1_2 = new TestItem(str1.data(), str1.length());
+    ASSERT_OK(cache->Insert(
+        "k1", item1_2, &CompressedSecondaryCacheTest::helper_, str1.length()));
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_dummy_count, 3);
+
+    // After this Insert, primary cache contains k2 and secondary cache contains
+    // k1's item, k2's dummy item, and k3's dummy item.
+    auto item2_2 = new TestItem(str2.data(), str2.length());
+    ASSERT_OK(cache->Insert(
+        "k2", item2_2, &CompressedSecondaryCacheTest::helper_, str2.length()));
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_real_count, 1);
+    if (sec_cache_is_compressed) {
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_uncompressed_bytes,
+                str1.length());
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_compressed_bytes,
+                1008);
+    } else {
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_uncompressed_bytes, 0);
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_compressed_bytes, 0);
+    }
+
+    // After this Insert, primary cache contains k3 and secondary cache contains
+    // k1's item and k2's item.
+    auto item3_2 = new TestItem(str3.data(), str3.length());
+    ASSERT_OK(cache->Insert(
+        "k3", item3_2, &CompressedSecondaryCacheTest::helper_, str3.length()));
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_real_count, 2);
+    if (sec_cache_is_compressed) {
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_uncompressed_bytes,
+                str1.length() + str2.length());
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_compressed_bytes,
+                2027);
+    } else {
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_uncompressed_bytes, 0);
+      ASSERT_EQ(get_perf_context()->compressed_sec_cache_compressed_bytes, 0);
+    }

    Cache::Handle* handle;
    handle = cache->Lookup("k3", &CompressedSecondaryCacheTest::helper_,
                           test_item_creator, Cache::Priority::LOW, true,
                           stats.get());
    ASSERT_NE(handle, nullptr);
-    TestItem* val3 = static_cast<TestItem*>(cache->Value(handle));
+    auto val3 = static_cast<TestItem*>(cache->Value(handle));
    ASSERT_NE(val3, nullptr);
-    ASSERT_EQ(memcmp(val3->Buf(), item3->Buf(), item3->Size()), 0);
+    ASSERT_EQ(memcmp(val3->Buf(), item3_2->Buf(), item3_2->Size()), 0);
    cache->Release(handle);

    // Lookup an non-existent key.
@@ -288,23 +385,85 @@ class CompressedSecondaryCacheTest : public testing::Test {
                           stats.get());
    ASSERT_EQ(handle, nullptr);

-    // This Lookup should promote k1 and erase k1 from the secondary cache,
-    // then k3 is demoted. So k2 and k3 are in the secondary cache.
+    // This Lookup should just insert a dummy handle in the primary cache
+    // and the k1 is still in the secondary cache.
    handle = cache->Lookup("k1", &CompressedSecondaryCacheTest::helper_,
                           test_item_creator, Cache::Priority::LOW, true,
                           stats.get());
    ASSERT_NE(handle, nullptr);
-    TestItem* val1_1 = static_cast<TestItem*>(cache->Value(handle));
+    ASSERT_EQ(get_perf_context()->block_cache_standalone_handle_count, 1);
+    auto val1_1 = static_cast<TestItem*>(cache->Value(handle));
    ASSERT_NE(val1_1, nullptr);
-    ASSERT_EQ(memcmp(val1_1->Buf(), str1_clone.data(), str1_clone.size()), 0);
+    ASSERT_EQ(memcmp(val1_1->Buf(), str1.data(), str1.size()), 0);
    cache->Release(handle);

+    // This Lookup should erase k1 from the secondary cache and insert
+    // it into primary cache; then k3 is demoted.
+    // k2 and k3 are in secondary cache.
+    handle = cache->Lookup("k1", &CompressedSecondaryCacheTest::helper_,
+                           test_item_creator, Cache::Priority::LOW, true,
+                           stats.get());
+    ASSERT_NE(handle, nullptr);
+    ASSERT_EQ(get_perf_context()->block_cache_standalone_handle_count, 1);
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_real_count, 3);
+    cache->Release(handle);
+
+    // k2 is still in secondary cache.
    handle = cache->Lookup("k2", &CompressedSecondaryCacheTest::helper_,
                           test_item_creator, Cache::Priority::LOW, true,
                           stats.get());
    ASSERT_NE(handle, nullptr);
+    ASSERT_EQ(get_perf_context()->block_cache_standalone_handle_count, 2);
    cache->Release(handle);

+    // Testing SetCapacity().
+    ASSERT_OK(secondary_cache->SetCapacity(0));
+    handle = cache->Lookup("k3", &CompressedSecondaryCacheTest::helper_,
+                           test_item_creator, Cache::Priority::LOW, true,
+                           stats.get());
+    ASSERT_EQ(handle, nullptr);
+
+    ASSERT_OK(secondary_cache->SetCapacity(7000));
+    size_t capacity;
+    ASSERT_OK(secondary_cache->GetCapacity(capacity));
+    ASSERT_EQ(capacity, 7000);
+    auto item1_3 = new TestItem(str1.data(), str1.length());
+    // After this Insert, primary cache contains k1.
+    ASSERT_OK(cache->Insert(
+        "k1", item1_3, &CompressedSecondaryCacheTest::helper_, str2.length()));
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_dummy_count, 3);
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_real_count, 4);
+
+    auto item2_3 = new TestItem(str2.data(), str2.length());
+    // After this Insert, primary cache contains k2 and secondary cache contains
+    // k1's dummy item.
+    ASSERT_OK(cache->Insert(
+        "k2", item2_3, &CompressedSecondaryCacheTest::helper_, str1.length()));
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_dummy_count, 4);
+
+    auto item1_4 = new TestItem(str1.data(), str1.length());
+    // After this Insert, primary cache contains k1 and secondary cache contains
+    // k1's dummy item and k2's dummy item.
+    ASSERT_OK(cache->Insert(
+        "k1", item1_4, &CompressedSecondaryCacheTest::helper_, str2.length()));
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_dummy_count, 5);
+
+    auto item2_4 = new TestItem(str2.data(), str2.length());
+    // After this Insert, primary cache contains k2 and secondary cache contains
+    // k1's real item and k2's dummy item.
+    ASSERT_OK(cache->Insert(
+        "k2", item2_4, &CompressedSecondaryCacheTest::helper_, str2.length()));
+    ASSERT_EQ(get_perf_context()->compressed_sec_cache_insert_real_count, 5);
+    // This Lookup should just insert a dummy handle in the primary cache
+    // and the k1 is still in the secondary cache.
+    handle = cache->Lookup("k1", &CompressedSecondaryCacheTest::helper_,
+                           test_item_creator, Cache::Priority::LOW, true,
+                           stats.get());
+
+    ASSERT_NE(handle, nullptr);
+    cache->Release(handle);
+    ASSERT_EQ(get_perf_context()->block_cache_standalone_handle_count, 3);
+
    cache.reset();
    secondary_cache.reset();
  }
@@ -321,24 +480,22 @@ class CompressedSecondaryCacheTest : public testing::Test {
      secondary_cache_opts.compression_type = CompressionType::kNoCompression;
    }

-    secondary_cache_opts.capacity = 2300;
+    secondary_cache_opts.capacity = 6000;
    secondary_cache_opts.num_shard_bits = 0;
    std::shared_ptr<SecondaryCache> secondary_cache =
        NewCompressedSecondaryCache(secondary_cache_opts);

    LRUCacheOptions opts(
-        1024 /* capacity */, 0 /* num_shard_bits */,
-        false /* strict_capacity_limit */, 0.5 /* high_pri_pool_ratio */,
-        nullptr /* memory_allocator */, kDefaultToAdaptiveMutex,
-        kDefaultCacheMetadataChargePolicy);
+        /*_capacity=*/1300, /*_num_shard_bits=*/0,
+        /*_strict_capacity_limit=*/false, /*_high_pri_pool_ratio=*/0.5,
+        /*_memory_allocator=*/nullptr, kDefaultToAdaptiveMutex,
+        kDefaultCacheMetadataChargePolicy, /*_low_pri_pool_ratio=*/0.0);
    opts.secondary_cache = secondary_cache;
    std::shared_ptr<Cache> cache = NewLRUCache(opts);

    Random rnd(301);
-    std::string str1;
-    test::CompressibleString(&rnd, 0.5, 1001, &str1);
-    auto item1 =
-        std::unique_ptr<TestItem>(new TestItem(str1.data(), str1.length()));
+    std::string str1 = rnd.RandomString(1001);
+    auto item1 = std::make_unique<TestItem>(str1.data(), str1.length());
    ASSERT_NOK(cache->Insert("k1", item1.get(), nullptr, str1.length()));
    ASSERT_OK(cache->Insert("k1", item1.get(),
                            &CompressedSecondaryCacheTest::helper_,
@@ -369,31 +526,29 @@ class CompressedSecondaryCacheTest : public testing::Test {
      secondary_cache_opts.compression_type = CompressionType::kNoCompression;
    }

-    secondary_cache_opts.capacity = 2300;
+    secondary_cache_opts.capacity = 6000;
    secondary_cache_opts.num_shard_bits = 0;

    std::shared_ptr<SecondaryCache> secondary_cache =
        NewCompressedSecondaryCache(secondary_cache_opts);

    LRUCacheOptions opts(
-        1200 /* capacity */, 0 /* num_shard_bits */,
-        false /* strict_capacity_limit */, 0.5 /* high_pri_pool_ratio */,
-        nullptr /* memory_allocator */, kDefaultToAdaptiveMutex,
-        kDefaultCacheMetadataChargePolicy);
+        /*_capacity=*/1300, /*_num_shard_bits=*/0,
+        /*_strict_capacity_limit=*/false, /*_high_pri_pool_ratio=*/0.5,
+        /*_memory_allocator=*/nullptr, kDefaultToAdaptiveMutex,
+        kDefaultCacheMetadataChargePolicy, /*_low_pri_pool_ratio=*/0.0);
    opts.secondary_cache = secondary_cache;
    std::shared_ptr<Cache> cache = NewLRUCache(opts);

    Random rnd(301);
-    std::string str1;
-    test::CompressibleString(&rnd, 0.5, 1001, &str1);
-    TestItem* item1 = new TestItem(str1.data(), str1.length());
+    std::string str1 = rnd.RandomString(1001);
+    auto item1 = new TestItem(str1.data(), str1.length());
    ASSERT_OK(cache->Insert("k1", item1,
                            &CompressedSecondaryCacheTest::helper_fail_,
                            str1.length()));

-    std::string str2;
-    test::CompressibleString(&rnd, 0.5, 1002, &str2);
-    TestItem* item2 = new TestItem(str2.data(), str2.length());
+    std::string str2 = rnd.RandomString(1002);
+    auto item2 = new TestItem(str2.data(), str2.length());
    // k1 should be demoted to the secondary cache.
    ASSERT_OK(cache->Insert("k2", item2,
                            &CompressedSecondaryCacheTest::helper_fail_,
@@ -404,11 +559,11 @@ class CompressedSecondaryCacheTest : public testing::Test {
                           test_item_creator, Cache::Priority::LOW, true);
    ASSERT_NE(handle, nullptr);
    cache->Release(handle);
-    // This lookup should fail, since k1 demotion would have failed
+    // This lookup should fail, since k1 demotion would have failed.
    handle = cache->Lookup("k1", &CompressedSecondaryCacheTest::helper_fail_,
                           test_item_creator, Cache::Priority::LOW, true);
    ASSERT_EQ(handle, nullptr);
-    // Since k1 didn't get promoted, k2 should still be in cache
+    // Since k1 was not promoted, k2 should still be in cache.
    handle = cache->Lookup("k2", &CompressedSecondaryCacheTest::helper_fail_,
                           test_item_creator, Cache::Priority::LOW, true);
    ASSERT_NE(handle, nullptr);
@@ -430,30 +585,28 @@ class CompressedSecondaryCacheTest : public testing::Test {
      secondary_cache_opts.compression_type = CompressionType::kNoCompression;
    }

-    secondary_cache_opts.capacity = 2300;
+    secondary_cache_opts.capacity = 6000;
    secondary_cache_opts.num_shard_bits = 0;

    std::shared_ptr<SecondaryCache> secondary_cache =
        NewCompressedSecondaryCache(secondary_cache_opts);

    LRUCacheOptions opts(
-        1200 /* capacity */, 0 /* num_shard_bits */,
-        false /* strict_capacity_limit */, 0.5 /* high_pri_pool_ratio */,
-        nullptr /* memory_allocator */, kDefaultToAdaptiveMutex,
-        kDefaultCacheMetadataChargePolicy);
+        /*_capacity=*/1300, /*_num_shard_bits=*/0,
+        /*_strict_capacity_limit=*/false, /*_high_pri_pool_ratio=*/0.5,
+        /*_memory_allocator=*/nullptr, kDefaultToAdaptiveMutex,
+        kDefaultCacheMetadataChargePolicy, /*_low_pri_pool_ratio=*/0.0);
    opts.secondary_cache = secondary_cache;
    std::shared_ptr<Cache> cache = NewLRUCache(opts);

    Random rnd(301);
-    std::string str1;
-    test::CompressibleString(&rnd, 0.5, 1001, &str1);
-    TestItem* item1 = new TestItem(str1.data(), str1.length());
+    std::string str1 = rnd.RandomString(1001);
+    auto item1 = new TestItem(str1.data(), str1.length());
    ASSERT_OK(cache->Insert("k1", item1, &CompressedSecondaryCacheTest::helper_,
                            str1.length()));

-    std::string str2;
-    test::CompressibleString(&rnd, 0.5, 1002, &str2);
-    TestItem* item2 = new TestItem(str2.data(), str2.length());
+    std::string str2 = rnd.RandomString(1002);
+    auto item2 = new TestItem(str2.data(), str2.length());
    // k1 should be demoted to the secondary cache.
    ASSERT_OK(cache->Insert("k2", item2, &CompressedSecondaryCacheTest::helper_,
                            str2.length()));
@@ -490,35 +643,46 @@ class CompressedSecondaryCacheTest : public testing::Test {
      secondary_cache_opts.compression_type = CompressionType::kNoCompression;
    }

-    secondary_cache_opts.capacity = 2300;
+    secondary_cache_opts.capacity = 6000;
    secondary_cache_opts.num_shard_bits = 0;

    std::shared_ptr<SecondaryCache> secondary_cache =
        NewCompressedSecondaryCache(secondary_cache_opts);

    LRUCacheOptions opts(
-        1200 /* capacity */, 0 /* num_shard_bits */,
-        true /* strict_capacity_limit */, 0.5 /* high_pri_pool_ratio */,
-        nullptr /* memory_allocator */, kDefaultToAdaptiveMutex,
-        kDefaultCacheMetadataChargePolicy);
+        /*_capacity=*/1300, /*_num_shard_bits=*/0,
+        /*_strict_capacity_limit=*/false, /*_high_pri_pool_ratio=*/0.5,
+        /*_memory_allocator=*/nullptr, kDefaultToAdaptiveMutex,
+        kDefaultCacheMetadataChargePolicy, /*_low_pri_pool_ratio=*/0.0);
    opts.secondary_cache = secondary_cache;
    std::shared_ptr<Cache> cache = NewLRUCache(opts);

    Random rnd(301);
-    std::string str1;
-    test::CompressibleString(&rnd, 0.5, 1001, &str1);
-    TestItem* item1 = new TestItem(str1.data(), str1.length());
-    ASSERT_OK(cache->Insert("k1", item1, &CompressedSecondaryCacheTest::helper_,
-                            str1.length()));
+    std::string str1 = rnd.RandomString(1001);
+    auto item1_1 = new TestItem(str1.data(), str1.length());
+    ASSERT_OK(cache->Insert(
+        "k1", item1_1, &CompressedSecondaryCacheTest::helper_, str1.length()));

-    std::string str2;
-    test::CompressibleString(&rnd, 0.5, 1002, &str2);
+    std::string str2 = rnd.RandomString(1002);
    std::string str2_clone{str2};
-    TestItem* item2 = new TestItem(str2.data(), str2.length());
-    // k1 should be demoted to the secondary cache.
+    auto item2 = new TestItem(str2.data(), str2.length());
+    // After this Insert, primary cache contains k2 and secondary cache contains
+    // k1's dummy item.
    ASSERT_OK(cache->Insert("k2", item2, &CompressedSecondaryCacheTest::helper_,
                            str2.length()));

+    // After this Insert, primary cache contains k1 and secondary cache contains
+    // k1's dummy item and k2's dummy item.
+    auto item1_2 = new TestItem(str1.data(), str1.length());
+    ASSERT_OK(cache->Insert(
+        "k1", item1_2, &CompressedSecondaryCacheTest::helper_, str1.length()));
+
+    auto item2_2 = new TestItem(str2.data(), str2.length());
+    // After this Insert, primary cache contains k2 and secondary cache contains
+    // k1's item and k2's dummy item.
+    ASSERT_OK(cache->Insert(
+        "k2", item2_2, &CompressedSecondaryCacheTest::helper_, str2.length()));
+
    Cache::Handle* handle2;
    handle2 = cache->Lookup("k2", &CompressedSecondaryCacheTest::helper_,
                            test_item_creator, Cache::Priority::LOW, true);
@@ -527,6 +691,7 @@ class CompressedSecondaryCacheTest : public testing::Test {

    // k1 promotion should fail because cache is at capacity and
    // strict_capacity_limit is true, but the lookup should still succeed.
+    // A k1's dummy item is inserted into primary cache.
    Cache::Handle* handle1;
    handle1 = cache->Lookup("k1", &CompressedSecondaryCacheTest::helper_,
                            test_item_creator, Cache::Priority::LOW, true);
@@ -561,8 +726,8 @@ class CompressedSecondaryCacheTest : public testing::Test {
        std::make_unique<CompressedSecondaryCache>(1000, 0, true, 0.5, 0.0,
                                                   allocator);
    Random rnd(301);
-    // 10000 = 8169 + 1769 + 62 , so there should be 3 chunks after split.
-    size_t str_size{10000};
+    // 8500 = 8169 + 233 + 98, so there should be 3 chunks after split.
+    size_t str_size{8500};
    std::string str = rnd.RandomString(static_cast<int>(str_size));
    size_t charge{0};
    CacheValueChunk* chunks_head =
@@ -572,11 +737,11 @@ class CompressedSecondaryCacheTest : public testing::Test {
    CacheValueChunk* current_chunk = chunks_head;
    ASSERT_EQ(current_chunk->size, 8192 - sizeof(CacheValueChunk) + 1);
    current_chunk = current_chunk->next;
-    ASSERT_EQ(current_chunk->size, 1792 - sizeof(CacheValueChunk) + 1);
+    ASSERT_EQ(current_chunk->size, 256 - sizeof(CacheValueChunk) + 1);
    current_chunk = current_chunk->next;
-    ASSERT_EQ(current_chunk->size, 62);
+    ASSERT_EQ(current_chunk->size, 98);

-    sec_cache->DeletionCallback("dummy", chunks_head);
+    sec_cache->GetDeletionCallback(true)("dummy", chunks_head);
  }

  void MergeChunksIntoValueTest() {
@@ -618,7 +783,11 @@ class CompressedSecondaryCacheTest : public testing::Test {
    std::string value_str{value.get(), charge};
    ASSERT_EQ(strcmp(value_str.data(), str.data()), 0);

-    sec_cache->DeletionCallback("dummy", chunks_head);
+    while (chunks_head != nullptr) {
+      CacheValueChunk* tmp_chunk = chunks_head;
+      chunks_head = chunks_head->next;
+      tmp_chunk->Free();
+    }
  }

  void SplictValueAndMergeChunksTest() {
@@ -639,8 +808,8 @@ class CompressedSecondaryCacheTest : public testing::Test {
        std::make_unique<CompressedSecondaryCache>(1000, 0, true, 0.5, 0.0,
                                                   allocator);
    Random rnd(301);
-    // 10000 = 8169 + 1769 + 62 , so there should be 3 chunks after split.
-    size_t str_size{10000};
+    // 8500 = 8169 + 233 + 98, so there should be 3 chunks after split.
+    size_t str_size{8500};
    std::string str = rnd.RandomString(static_cast<int>(str_size));
    size_t charge{0};
    CacheValueChunk* chunks_head =
@@ -653,7 +822,7 @@ class CompressedSecondaryCacheTest : public testing::Test {
    std::string value_str{value.get(), charge};
    ASSERT_EQ(strcmp(value_str.data(), str.data()), 0);

-    sec_cache->DeletionCallback("dummy", chunks_head);
+    sec_cache->GetDeletionCallback(true)("dummy", chunks_head);
  }

 private:
@@ -670,111 +839,150 @@ Cache::CacheItemHelper CompressedSecondaryCacheTest::helper_fail_(
    CompressedSecondaryCacheTest::SaveToCallbackFail,
    CompressedSecondaryCacheTest::DeletionCallback);

-TEST_F(CompressedSecondaryCacheTest, BasicTestWithNoCompression) {
-  BasicTest(false, false);
+class CompressedSecCacheTestWithCompressAndAllocatorParam
+    : public CompressedSecondaryCacheTest,
+      public ::testing::WithParamInterface<std::tuple<bool, bool>> {
+ public:
+  CompressedSecCacheTestWithCompressAndAllocatorParam() {
+    sec_cache_is_compressed_ = std::get<0>(GetParam());
+    use_jemalloc_ = std::get<1>(GetParam());
+  }
+  bool sec_cache_is_compressed_;
+  bool use_jemalloc_;
+};
+
+TEST_P(CompressedSecCacheTestWithCompressAndAllocatorParam, BasicTes) {
+  BasicTest(sec_cache_is_compressed_, use_jemalloc_);
 }

-TEST_F(CompressedSecondaryCacheTest,
-       BasicTestWithMemoryAllocatorAndNoCompression) {
-  BasicTest(false, true);
-}
+INSTANTIATE_TEST_CASE_P(CompressedSecCacheTests,
+                        CompressedSecCacheTestWithCompressAndAllocatorParam,
+                        ::testing::Combine(testing::Bool(), testing::Bool()));

-TEST_F(CompressedSecondaryCacheTest, BasicTestWithCompression) {
-  BasicTest(true, false);
-}
-
-TEST_F(CompressedSecondaryCacheTest,
-       BasicTestWithMemoryAllocatorAndCompression) {
-  BasicTest(true, true);
-}
+class CompressedSecondaryCacheTestWithCompressionParam
+    : public CompressedSecondaryCacheTest,
+      public ::testing::WithParamInterface<bool> {
+ public:
+  CompressedSecondaryCacheTestWithCompressionParam() {
+    sec_cache_is_compressed_ = GetParam();
+  }
+  bool sec_cache_is_compressed_;
+};

 #ifndef ROCKSDB_LITE

-TEST_F(CompressedSecondaryCacheTest, BasicTestFromStringWithNoCompression) {
-  std::string sec_cache_uri =
-      "compressed_secondary_cache://"
-      "capacity=2048;num_shard_bits=0;compression_type=kNoCompression";
-  std::shared_ptr<SecondaryCache> sec_cache;
-  Status s = SecondaryCache::CreateFromString(ConfigOptions(), sec_cache_uri,
-                                              &sec_cache);
-  EXPECT_OK(s);
-  BasicTestHelper(sec_cache);
-}
-
-TEST_F(CompressedSecondaryCacheTest, BasicTestFromStringWithCompression) {
+TEST_P(CompressedSecondaryCacheTestWithCompressionParam, BasicTestFromString) {
+  std::shared_ptr<SecondaryCache> sec_cache{nullptr};
  std::string sec_cache_uri;
-  if (LZ4_Supported()) {
-    sec_cache_uri =
-        "compressed_secondary_cache://"
-        "capacity=2048;num_shard_bits=0;compression_type=kLZ4Compression;"
-        "compress_format_version=2";
+  if (sec_cache_is_compressed_) {
+    if (LZ4_Supported()) {
+      sec_cache_uri =
+          "compressed_secondary_cache://"
+          "capacity=2048;num_shard_bits=0;compression_type=kLZ4Compression;"
+          "compress_format_version=2";
+    } else {
+      ROCKSDB_GTEST_SKIP("This test requires LZ4 support.");
+      sec_cache_uri =
+          "compressed_secondary_cache://"
+          "capacity=2048;num_shard_bits=0;compression_type=kNoCompression";
+      sec_cache_is_compressed_ = false;
+    }
+    Status s = SecondaryCache::CreateFromString(ConfigOptions(), sec_cache_uri,
+                                                &sec_cache);
+    EXPECT_OK(s);
  } else {
-    ROCKSDB_GTEST_SKIP("This test requires LZ4 support.");
    sec_cache_uri =
        "compressed_secondary_cache://"
        "capacity=2048;num_shard_bits=0;compression_type=kNoCompression";
+    Status s = SecondaryCache::CreateFromString(ConfigOptions(), sec_cache_uri,
+                                                &sec_cache);
+    EXPECT_OK(s);
  }
+  BasicTestHelper(sec_cache, sec_cache_is_compressed_);
+}

-  std::shared_ptr<SecondaryCache> sec_cache;
-  Status s = SecondaryCache::CreateFromString(ConfigOptions(), sec_cache_uri,
-                                              &sec_cache);
-  EXPECT_OK(s);
-  BasicTestHelper(sec_cache);
+TEST_P(CompressedSecondaryCacheTestWithCompressionParam,
+       BasicTestFromStringWithSplit) {
+  std::shared_ptr<SecondaryCache> sec_cache{nullptr};
+  std::string sec_cache_uri;
+  if (sec_cache_is_compressed_) {
+    if (LZ4_Supported()) {
+      sec_cache_uri =
+          "compressed_secondary_cache://"
+          "capacity=2048;num_shard_bits=0;compression_type=kLZ4Compression;"
+          "compress_format_version=2;enable_custom_split_merge=true";
+    } else {
+      ROCKSDB_GTEST_SKIP("This test requires LZ4 support.");
+      sec_cache_uri =
+          "compressed_secondary_cache://"
+          "capacity=2048;num_shard_bits=0;compression_type=kNoCompression;"
+          "enable_custom_split_merge=true";
+      sec_cache_is_compressed_ = false;
+    }
+    Status s = SecondaryCache::CreateFromString(ConfigOptions(), sec_cache_uri,
+                                                &sec_cache);
+    EXPECT_OK(s);
+  } else {
+    sec_cache_uri =
+        "compressed_secondary_cache://"
+        "capacity=2048;num_shard_bits=0;compression_type=kNoCompression;"
+        "enable_custom_split_merge=true";
+    Status s = SecondaryCache::CreateFromString(ConfigOptions(), sec_cache_uri,
+                                                &sec_cache);
+    EXPECT_OK(s);
+  }
+  BasicTestHelper(sec_cache, sec_cache_is_compressed_);
 }

 #endif  // ROCKSDB_LITE

-TEST_F(CompressedSecondaryCacheTest, FailsTestWithNoCompression) {
-  FailsTest(false);
+TEST_P(CompressedSecondaryCacheTestWithCompressionParam, FailsTest) {
+  FailsTest(sec_cache_is_compressed_);
 }

-TEST_F(CompressedSecondaryCacheTest, FailsTestWithCompression) {
-  FailsTest(true);
+TEST_P(CompressedSecondaryCacheTestWithCompressionParam,
+       BasicIntegrationFailTest) {
+  BasicIntegrationFailTest(sec_cache_is_compressed_);
 }

-TEST_F(CompressedSecondaryCacheTest, BasicIntegrationTestWithNoCompression) {
-  BasicIntegrationTest(false);
+TEST_P(CompressedSecondaryCacheTestWithCompressionParam,
+       IntegrationSaveFailTest) {
+  IntegrationSaveFailTest(sec_cache_is_compressed_);
 }

-TEST_F(CompressedSecondaryCacheTest, BasicIntegrationTestWithCompression) {
-  BasicIntegrationTest(true);
+TEST_P(CompressedSecondaryCacheTestWithCompressionParam,
+       IntegrationCreateFailTest) {
+  IntegrationCreateFailTest(sec_cache_is_compressed_);
 }

-TEST_F(CompressedSecondaryCacheTest,
-       BasicIntegrationFailTestWithNoCompression) {
-  BasicIntegrationFailTest(false);
+TEST_P(CompressedSecondaryCacheTestWithCompressionParam,
+       IntegrationFullCapacityTest) {
+  IntegrationFullCapacityTest(sec_cache_is_compressed_);
 }

-TEST_F(CompressedSecondaryCacheTest, BasicIntegrationFailTestWithCompression) {
-  BasicIntegrationFailTest(true);
+INSTANTIATE_TEST_CASE_P(CompressedSecCacheTests,
+                        CompressedSecondaryCacheTestWithCompressionParam,
+                        testing::Bool());
+
+class CompressedSecCacheTestWithCompressAndSplitParam
+    : public CompressedSecondaryCacheTest,
+      public ::testing::WithParamInterface<std::tuple<bool, bool>> {
+ public:
+  CompressedSecCacheTestWithCompressAndSplitParam() {
+    sec_cache_is_compressed_ = std::get<0>(GetParam());
+    enable_custom_split_merge_ = std::get<1>(GetParam());
+  }
+  bool sec_cache_is_compressed_;
+  bool enable_custom_split_merge_;
+};
+
+TEST_P(CompressedSecCacheTestWithCompressAndSplitParam, BasicIntegrationTest) {
+  BasicIntegrationTest(sec_cache_is_compressed_, enable_custom_split_merge_);
 }

-TEST_F(CompressedSecondaryCacheTest, IntegrationSaveFailTestWithNoCompression) {
-  IntegrationSaveFailTest(false);
-}
-
-TEST_F(CompressedSecondaryCacheTest, IntegrationSaveFailTestWithCompression) {
-  IntegrationSaveFailTest(true);
-}
-
-TEST_F(CompressedSecondaryCacheTest,
-       IntegrationCreateFailTestWithNoCompression) {
-  IntegrationCreateFailTest(false);
-}
-
-TEST_F(CompressedSecondaryCacheTest, IntegrationCreateFailTestWithCompression) {
-  IntegrationCreateFailTest(true);
-}
-
-TEST_F(CompressedSecondaryCacheTest,
-       IntegrationFullCapacityTestWithNoCompression) {
-  IntegrationFullCapacityTest(false);
-}
-
-TEST_F(CompressedSecondaryCacheTest,
-       IntegrationFullCapacityTestWithCompression) {
-  IntegrationFullCapacityTest(true);
-}
+INSTANTIATE_TEST_CASE_P(CompressedSecCacheTests,
+                        CompressedSecCacheTestWithCompressAndSplitParam,
+                        ::testing::Combine(testing::Bool(), testing::Bool()));

 TEST_F(CompressedSecondaryCacheTest, SplitValueIntoChunksTest) {
  SplitValueIntoChunksTest();
@@ -791,6 +999,7 @@ TEST_F(CompressedSecondaryCacheTest, SplictValueAndMergeChunksTest) {
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -173,13 +173,13 @@ inline int LRUHandleTable::FindSlot(const Slice& key,
 LRUCacheShard::LRUCacheShard(size_t capacity, size_t estimated_value_size,
                             bool strict_capacity_limit,
                             CacheMetadataChargePolicy metadata_charge_policy)
-    : capacity_(capacity),
+    : CacheShardBase(metadata_charge_policy),
+      capacity_(capacity),
      strict_capacity_limit_(strict_capacity_limit),
      table_(
          CalcHashBits(capacity, estimated_value_size, metadata_charge_policy)),
      usage_(0),
      lru_usage_(0) {
-  set_metadata_charge_policy(metadata_charge_policy);
  // Make empty circular linked list.
  lru_.next = &lru_;
  lru_.prev = &lru_;
@@ -211,27 +211,27 @@ void LRUCacheShard::EraseUnRefEntries() {
 void LRUCacheShard::ApplyToSomeEntries(
    const std::function<void(const Slice& key, void* value, size_t charge,
                             DeleterFn deleter)>& callback,
-    uint32_t average_entries_per_lock, uint32_t* state) {
+    size_t average_entries_per_lock, size_t* state) {
  // The state is essentially going to be the starting hash, which works
  // nicely even if we resize between calls because we use upper-most
  // hash bits for table indexes.
  DMutexLock l(mutex_);
-  uint32_t length_bits = table_.GetLengthBits();
-  uint32_t length = table_.GetTableSize();
+  size_t length_bits = table_.GetLengthBits();
+  size_t length = table_.GetTableSize();

  assert(average_entries_per_lock > 0);
  // Assuming we are called with same average_entries_per_lock repeatedly,
  // this simplifies some logic (index_end will not overflow).
  assert(average_entries_per_lock < length || *state == 0);

-  uint32_t index_begin = *state >> (32 - length_bits);
-  uint32_t index_end = index_begin + average_entries_per_lock;
+  size_t index_begin = *state >> (sizeof(size_t) * 8u - length_bits);
+  size_t index_end = index_begin + average_entries_per_lock;
  if (index_end >= length) {
    // Going to end
    index_end = length;
-    *state = UINT32_MAX;
+    *state = SIZE_MAX;
  } else {
-    *state = index_end << (32 - length_bits);
+    *state = index_end << (sizeof(size_t) * 8u - length_bits);
  }

  table_.ApplyToEntriesRange(
@@ -322,8 +322,7 @@ void LRUCacheShard::SetStrictCapacityLimit(bool strict_capacity_limit) {

 Status LRUCacheShard::Insert(const Slice& key, uint32_t hash, void* value,
                             size_t charge, Cache::DeleterFn deleter,
-                             Cache::Handle** handle,
-                             Cache::Priority /*priority*/) {
+                             LRUHandle** handle, Cache::Priority /*priority*/) {
  if (key.size() != kCacheKeySize) {
    return Status::NotSupported("FastLRUCache only supports key size " +
                                std::to_string(kCacheKeySize) + "B");
@@ -409,7 +408,7 @@ Status LRUCacheShard::Insert(const Slice& key, uint32_t hash, void* value,
        if (!h->HasRefs()) {
          h->Ref();
        }
-        *handle = reinterpret_cast<Cache::Handle*>(h);
+        *handle = h;
      }
    }
  }
@@ -422,7 +421,7 @@ Status LRUCacheShard::Insert(const Slice& key, uint32_t hash, void* value,
  return s;
 }

-Cache::Handle* LRUCacheShard::Lookup(const Slice& key, uint32_t hash) {
+LRUHandle* LRUCacheShard::Lookup(const Slice& key, uint32_t hash) {
  LRUHandle* h = nullptr;
  {
    DMutexLock l(mutex_);
@@ -437,23 +436,21 @@ Cache::Handle* LRUCacheShard::Lookup(const Slice& key, uint32_t hash) {
      h->Ref();
    }
  }
-  return reinterpret_cast<Cache::Handle*>(h);
+  return h;
 }

-bool LRUCacheShard::Ref(Cache::Handle* h) {
-  LRUHandle* e = reinterpret_cast<LRUHandle*>(h);
+bool LRUCacheShard::Ref(LRUHandle* h) {
  DMutexLock l(mutex_);
  // To create another reference - entry must be already externally referenced.
-  assert(e->HasRefs());
-  e->Ref();
+  assert(h->HasRefs());
+  h->Ref();
  return true;
 }

-bool LRUCacheShard::Release(Cache::Handle* handle, bool erase_if_last_ref) {
-  if (handle == nullptr) {
+bool LRUCacheShard::Release(LRUHandle* h, bool erase_if_last_ref) {
+  if (h == nullptr) {
    return false;
  }
-  LRUHandle* h = reinterpret_cast<LRUHandle*>(handle);
  LRUHandle copy;
  bool last_reference = false;
  {
@@ -525,41 +522,28 @@ size_t LRUCacheShard::GetPinnedUsage() const {
  return usage_ - lru_usage_;
 }

-std::string LRUCacheShard::GetPrintableOptions() const { return std::string{}; }
+size_t LRUCacheShard::GetOccupancyCount() const {
+  DMutexLock l(mutex_);
+  return table_.GetOccupancy();
+}
+
+size_t LRUCacheShard::GetTableAddressCount() const {
+  DMutexLock l(mutex_);
+  return table_.GetTableSize();
+}

 LRUCache::LRUCache(size_t capacity, size_t estimated_value_size,
                   int num_shard_bits, bool strict_capacity_limit,
                   CacheMetadataChargePolicy metadata_charge_policy)
-    : ShardedCache(capacity, num_shard_bits, strict_capacity_limit) {
+    : ShardedCache(capacity, num_shard_bits, strict_capacity_limit,
+                   nullptr /*allocator*/) {
  assert(estimated_value_size > 0 ||
         metadata_charge_policy != kDontChargeCacheMetadata);
-  num_shards_ = 1 << num_shard_bits;
-  shards_ = reinterpret_cast<LRUCacheShard*>(
-      port::cacheline_aligned_alloc(sizeof(LRUCacheShard) * num_shards_));
-  size_t per_shard = (capacity + (num_shards_ - 1)) / num_shards_;
-  for (int i = 0; i < num_shards_; i++) {
-    new (&shards_[i])
-        LRUCacheShard(per_shard, estimated_value_size, strict_capacity_limit,
-                      metadata_charge_policy);
-  }
-}
-
-LRUCache::~LRUCache() {
-  if (shards_ != nullptr) {
-    assert(num_shards_ > 0);
-    for (int i = 0; i < num_shards_; i++) {
-      shards_[i].~LRUCacheShard();
-    }
-    port::cacheline_aligned_free(shards_);
-  }
-}
-
-CacheShard* LRUCache::GetShard(uint32_t shard) {
-  return reinterpret_cast<CacheShard*>(&shards_[shard]);
-}
-
-const CacheShard* LRUCache::GetShard(uint32_t shard) const {
-  return reinterpret_cast<CacheShard*>(&shards_[shard]);
+  size_t per_shard = GetPerShardCapacity();
+  InitShards([=](LRUCacheShard* cs) {
+    new (cs) LRUCacheShard(per_shard, estimated_value_size,
+                           strict_capacity_limit, metadata_charge_policy);
+  });
 }

 void* LRUCache::Value(Handle* handle) {
@@ -567,12 +551,8 @@ void* LRUCache::Value(Handle* handle) {
 }

 size_t LRUCache::GetCharge(Handle* handle) const {
-  CacheMetadataChargePolicy metadata_charge_policy = kDontChargeCacheMetadata;
-  if (num_shards_ > 0) {
-    metadata_charge_policy = shards_[0].metadata_charge_policy_;
-  }
  return reinterpret_cast<const LRUHandle*>(handle)->GetCharge(
-      metadata_charge_policy);
+      GetShard(0).metadata_charge_policy_);
 }

 Cache::DeleterFn LRUCache::GetDeleter(Handle* handle) const {
@@ -580,18 +560,6 @@ Cache::DeleterFn LRUCache::GetDeleter(Handle* handle) const {
  return h->deleter;
 }

-uint32_t LRUCache::GetHash(Handle* handle) const {
-  return reinterpret_cast<const LRUHandle*>(handle)->hash;
-}
-
-void LRUCache::DisownData() {
-  // Leak data only if that won't generate an ASAN/valgrind warning.
-  if (!kMustFreeHeapAllocations) {
-    shards_ = nullptr;
-    num_shards_ = 0;
-  }
-}
-
 }  // namespace fast_lru_cache

 std::shared_ptr<Cache> NewFastLRUCache(
@@ -141,6 +141,9 @@ struct LRUHandle {

  Slice key() const { return Slice(key_data.data(), kCacheKeySize); }

+  // For HandleImpl concept
+  uint32_t GetHash() const { return hash; }
+
  // Increase the reference count by 1.
  void Ref() { refs++; }

@@ -260,8 +263,8 @@ class LRUHandleTable {
  void Assign(int slot, LRUHandle* h);

  template <typename T>
-  void ApplyToEntriesRange(T func, uint32_t index_begin, uint32_t index_end) {
-    for (uint32_t i = index_begin; i < index_end; i++) {
+  void ApplyToEntriesRange(T func, size_t index_begin, size_t index_end) {
+    for (size_t i = index_begin; i < index_end; i++) {
      LRUHandle* h = &array_[i];
      if (h->IsVisible()) {
        func(h);
@@ -316,20 +319,30 @@ class LRUHandleTable {
 };

 // A single shard of sharded cache.
-class ALIGN_AS(CACHE_LINE_SIZE) LRUCacheShard final : public CacheShard {
+class ALIGN_AS(CACHE_LINE_SIZE) LRUCacheShard final : public CacheShardBase {
 public:
  LRUCacheShard(size_t capacity, size_t estimated_value_size,
                bool strict_capacity_limit,
                CacheMetadataChargePolicy metadata_charge_policy);
-  ~LRUCacheShard() override = default;
+
+  // For CacheShard concept
+  using HandleImpl = LRUHandle;
+
+  // Keep 32-bit hashing for now (FIXME: upgrade to 64-bit)
+  using HashVal = uint32_t;
+  using HashCref = uint32_t;
+  static inline HashVal ComputeHash(const Slice& key) {
+    return Lower32of64(GetSliceNPHash64(key));
+  }
+  static inline uint32_t HashPieceForSharding(HashCref hash) { return hash; }

  // Separate from constructor so caller can easily make an array of LRUCache
  // if current usage is more than new capacity, the function will attempt to
  // free the needed space.
-  void SetCapacity(size_t capacity) override;
+  void SetCapacity(size_t capacity);

  // Set the flag to reject insertion if cache if full.
-  void SetStrictCapacityLimit(bool strict_capacity_limit) override;
+  void SetStrictCapacityLimit(bool strict_capacity_limit);

  // Like Cache methods, but with an extra "hash" parameter.
  // Insert an item into the hash table and, if handle is null, insert into
@@ -337,46 +350,45 @@ class ALIGN_AS(CACHE_LINE_SIZE) LRUCacheShard final : public CacheShard {
  // and free_handle_on_fail is true, the item is deleted and handle is set to
  // nullptr.
  Status Insert(const Slice& key, uint32_t hash, void* value, size_t charge,
-                Cache::DeleterFn deleter, Cache::Handle** handle,
-                Cache::Priority priority) override;
+                Cache::DeleterFn deleter, LRUHandle** handle,
+                Cache::Priority priority);

  Status Insert(const Slice& key, uint32_t hash, void* value,
                const Cache::CacheItemHelper* helper, size_t charge,
-                Cache::Handle** handle, Cache::Priority priority) override {
+                LRUHandle** handle, Cache::Priority priority) {
    return Insert(key, hash, value, charge, helper->del_cb, handle, priority);
  }

-  Cache::Handle* Lookup(const Slice& key, uint32_t hash,
-                        const Cache::CacheItemHelper* /*helper*/,
-                        const Cache::CreateCallback& /*create_cb*/,
-                        Cache::Priority /*priority*/, bool /*wait*/,
-                        Statistics* /*stats*/) override {
+  LRUHandle* Lookup(const Slice& key, uint32_t hash,
+                    const Cache::CacheItemHelper* /*helper*/,
+                    const Cache::CreateCallback& /*create_cb*/,
+                    Cache::Priority /*priority*/, bool /*wait*/,
+                    Statistics* /*stats*/) {
    return Lookup(key, hash);
  }
-  Cache::Handle* Lookup(const Slice& key, uint32_t hash) override;
+  LRUHandle* Lookup(const Slice& key, uint32_t hash);

-  bool Release(Cache::Handle* handle, bool /*useful*/,
-               bool erase_if_last_ref) override {
+  bool Release(LRUHandle* handle, bool /*useful*/, bool erase_if_last_ref) {
    return Release(handle, erase_if_last_ref);
  }
-  bool IsReady(Cache::Handle* /*handle*/) override { return true; }
-  void Wait(Cache::Handle* /*handle*/) override {}
+  bool IsReady(LRUHandle* /*handle*/) { return true; }
+  void Wait(LRUHandle* /*handle*/) {}

-  bool Ref(Cache::Handle* handle) override;
-  bool Release(Cache::Handle* handle, bool erase_if_last_ref = false) override;
-  void Erase(const Slice& key, uint32_t hash) override;
+  bool Ref(LRUHandle* handle);
+  bool Release(LRUHandle* handle, bool erase_if_last_ref = false);
+  void Erase(const Slice& key, uint32_t hash);

-  size_t GetUsage() const override;
-  size_t GetPinnedUsage() const override;
+  size_t GetUsage() const;
+  size_t GetPinnedUsage() const;
+  size_t GetOccupancyCount() const;
+  size_t GetTableAddressCount() const;

  void ApplyToSomeEntries(
      const std::function<void(const Slice& key, void* value, size_t charge,
                               DeleterFn deleter)>& callback,
-      uint32_t average_entries_per_lock, uint32_t* state) override;
+      size_t average_entries_per_lock, size_t* state);

-  void EraseUnRefEntries() override;
-
-  std::string GetPrintableOptions() const override;
+  void EraseUnRefEntries();

 private:
  friend class LRUCache;
@@ -444,25 +456,16 @@ class LRUCache
 #ifdef NDEBUG
    final
 #endif
-    : public ShardedCache {
+    : public ShardedCache<LRUCacheShard> {
 public:
  LRUCache(size_t capacity, size_t estimated_value_size, int num_shard_bits,
           bool strict_capacity_limit,
           CacheMetadataChargePolicy metadata_charge_policy =
               kDontChargeCacheMetadata);
-  ~LRUCache() override;
  const char* Name() const override { return "LRUCache"; }
-  CacheShard* GetShard(uint32_t shard) override;
-  const CacheShard* GetShard(uint32_t shard) const override;
  void* Value(Handle* handle) override;
  size_t GetCharge(Handle* handle) const override;
-  uint32_t GetHash(Handle* handle) const override;
  DeleterFn GetDeleter(Handle* handle) const override;
-  void DisownData() override;
-
- private:
-  LRUCacheShard* shards_ = nullptr;
-  int num_shards_ = 0;
 };
 }  // namespace fast_lru_cache

@@ -12,6 +12,7 @@
 #include <cassert>
 #include <cstdint>
 #include <cstdio>
+#include <cstdlib>

 #include "monitoring/perf_context_imp.h"
 #include "monitoring/statistics.h"
@@ -21,6 +22,9 @@
 namespace ROCKSDB_NAMESPACE {
 namespace lru_cache {

+// A distinct pointer value for marking "dummy" cache entries
+void* const kDummyValueMarker = const_cast<char*>("kDummyValueMarker");
+
 LRUHandleTable::LRUHandleTable(int max_upper_hash_bits)
    : length_bits_(/* historical starting size*/ 4),
      list_(new LRUHandle* [size_t{1} << length_bits_] {}),
@@ -34,7 +38,7 @@ LRUHandleTable::~LRUHandleTable() {
          h->Free();
        }
      },
-      0, uint32_t{1} << length_bits_);
+      0, size_t{1} << length_bits_);
 }

 LRUHandle* LRUHandleTable::Lookup(const Slice& key, uint32_t hash) {
@@ -109,12 +113,14 @@ void LRUHandleTable::Resize() {
  length_bits_ = new_length_bits;
 }

-LRUCacheShard::LRUCacheShard(
-    size_t capacity, bool strict_capacity_limit, double high_pri_pool_ratio,
-    double low_pri_pool_ratio, bool use_adaptive_mutex,
-    CacheMetadataChargePolicy metadata_charge_policy, int max_upper_hash_bits,
-    const std::shared_ptr<SecondaryCache>& secondary_cache)
-    : capacity_(0),
+LRUCacheShard::LRUCacheShard(size_t capacity, bool strict_capacity_limit,
+                             double high_pri_pool_ratio,
+                             double low_pri_pool_ratio, bool use_adaptive_mutex,
+                             CacheMetadataChargePolicy metadata_charge_policy,
+                             int max_upper_hash_bits,
+                             SecondaryCache* secondary_cache)
+    : CacheShardBase(metadata_charge_policy),
+      capacity_(0),
      high_pri_pool_usage_(0),
      low_pri_pool_usage_(0),
      strict_capacity_limit_(strict_capacity_limit),
@@ -127,7 +133,6 @@ LRUCacheShard::LRUCacheShard(
      lru_usage_(0),
      mutex_(use_adaptive_mutex),
      secondary_cache_(secondary_cache) {
-  set_metadata_charge_policy(metadata_charge_policy);
  // Make empty circular linked list.
  lru_.next = &lru_;
  lru_.prev = &lru_;
@@ -161,27 +166,27 @@ void LRUCacheShard::EraseUnRefEntries() {
 void LRUCacheShard::ApplyToSomeEntries(
    const std::function<void(const Slice& key, void* value, size_t charge,
                             DeleterFn deleter)>& callback,
-    uint32_t average_entries_per_lock, uint32_t* state) {
+    size_t average_entries_per_lock, size_t* state) {
  // The state is essentially going to be the starting hash, which works
  // nicely even if we resize between calls because we use upper-most
  // hash bits for table indexes.
  DMutexLock l(mutex_);
-  uint32_t length_bits = table_.GetLengthBits();
-  uint32_t length = uint32_t{1} << length_bits;
+  int length_bits = table_.GetLengthBits();
+  size_t length = size_t{1} << length_bits;

  assert(average_entries_per_lock > 0);
  // Assuming we are called with same average_entries_per_lock repeatedly,
  // this simplifies some logic (index_end will not overflow).
  assert(average_entries_per_lock < length || *state == 0);

-  uint32_t index_begin = *state >> (32 - length_bits);
-  uint32_t index_end = index_begin + average_entries_per_lock;
+  size_t index_begin = *state >> (sizeof(size_t) * 8u - length_bits);
+  size_t index_end = index_begin + average_entries_per_lock;
  if (index_end >= length) {
    // Going to end
    index_end = length;
-    *state = UINT32_MAX;
+    *state = SIZE_MAX;
  } else {
-    *state = index_end << (32 - length_bits);
+    *state = index_end << (sizeof(size_t) * 8u - length_bits);
  }

  table_.ApplyToEntriesRange(
@@ -329,6 +334,19 @@ void LRUCacheShard::EvictFromLRU(size_t charge,
  }
 }

+void LRUCacheShard::TryInsertIntoSecondaryCache(
+    autovector<LRUHandle*> evicted_handles) {
+  for (auto entry : evicted_handles) {
+    if (secondary_cache_ && entry->IsSecondaryCacheCompatible() &&
+        !entry->IsInSecondaryCache()) {
+      secondary_cache_->Insert(entry->key(), entry->value, entry->info_.helper)
+          .PermitUncheckedError();
+    }
+    // Free the entries here outside of mutex for performance reasons.
+    entry->Free();
+  }
+}
+
 void LRUCacheShard::SetCapacity(size_t capacity) {
  autovector<LRUHandle*> last_reference_list;
  {
@@ -339,16 +357,7 @@ void LRUCacheShard::SetCapacity(size_t capacity) {
    EvictFromLRU(0, &last_reference_list);
  }

-  // Try to insert the evicted entries into tiered cache.
-  // Free the entries outside of mutex for performance reasons.
-  for (auto entry : last_reference_list) {
-    if (secondary_cache_ && entry->IsSecondaryCacheCompatible() &&
-        !entry->IsInSecondaryCache()) {
-      secondary_cache_->Insert(entry->key(), entry->value, entry->info_.helper)
-          .PermitUncheckedError();
-    }
-    entry->Free();
-  }
+  TryInsertIntoSecondaryCache(last_reference_list);
 }

 void LRUCacheShard::SetStrictCapacityLimit(bool strict_capacity_limit) {
@@ -356,7 +365,7 @@ void LRUCacheShard::SetStrictCapacityLimit(bool strict_capacity_limit) {
  strict_capacity_limit_ = strict_capacity_limit;
 }

-Status LRUCacheShard::InsertItem(LRUHandle* e, Cache::Handle** handle,
+Status LRUCacheShard::InsertItem(LRUHandle* e, LRUHandle** handle,
                                 bool free_handle_on_fail) {
  Status s = Status::OK();
  autovector<LRUHandle*> last_reference_list;
@@ -377,7 +386,7 @@ Status LRUCacheShard::InsertItem(LRUHandle* e, Cache::Handle** handle,
        last_reference_list.push_back(e);
      } else {
        if (free_handle_on_fail) {
-          delete[] reinterpret_cast<char*>(e);
+          free(e);
          *handle = nullptr;
        }
        s = Status::MemoryLimit("Insert failed due to LRU cache being full.");
@@ -406,21 +415,12 @@ Status LRUCacheShard::InsertItem(LRUHandle* e, Cache::Handle** handle,
        if (!e->HasRefs()) {
          e->Ref();
        }
-        *handle = reinterpret_cast<Cache::Handle*>(e);
+        *handle = e;
      }
    }
  }

-  // Try to insert the evicted entries into the secondary cache.
-  // Free the entries here outside of mutex for performance reasons.
-  for (auto entry : last_reference_list) {
-    if (secondary_cache_ && entry->IsSecondaryCacheCompatible() &&
-        !entry->IsInSecondaryCache()) {
-      secondary_cache_->Insert(entry->key(), entry->value, entry->info_.helper)
-          .PermitUncheckedError();
-    }
-    entry->Free();
-  }
+  TryInsertIntoSecondaryCache(last_reference_list);

  return s;
 }
@@ -429,58 +429,124 @@ void LRUCacheShard::Promote(LRUHandle* e) {
  SecondaryCacheResultHandle* secondary_handle = e->sec_handle;

  assert(secondary_handle->IsReady());
-  e->SetIncomplete(false);
-  e->SetInCache(true);
+  // e is not thread-shared here; OK to modify "immutable" fields as well as
+  // "mutable" (normally requiring mutex)
+  e->SetIsPending(false);
  e->value = secondary_handle->Value();
-  e->CalcTotalCharge(secondary_handle->Size(), metadata_charge_policy_);
+  assert(e->total_charge == 0);
+  size_t value_size = secondary_handle->Size();
  delete secondary_handle;

-  // This call could fail if the cache is over capacity and
-  // strict_capacity_limit_ is true. In such a case, we don't want
-  // InsertItem() to free the handle, since the item is already in memory
-  // and the caller will most likely just read from disk if we erase it here.
  if (e->value) {
-    Cache::Handle* handle = reinterpret_cast<Cache::Handle*>(e);
-    Status s = InsertItem(e, &handle, /*free_handle_on_fail=*/false);
+    e->CalcTotalCharge(value_size, metadata_charge_policy_);
+    Status s;
+    if (e->IsStandalone()) {
+      assert(secondary_cache_ && secondary_cache_->SupportForceErase());
+
+      // Insert a dummy handle and return a standalone handle to caller.
+      // Charge the standalone handle.
+      autovector<LRUHandle*> last_reference_list;
+      bool free_standalone_handle{false};
+      {
+        DMutexLock l(mutex_);
+
+        // Free the space following strict LRU policy until enough space
+        // is freed or the lru list is empty.
+        EvictFromLRU(e->total_charge, &last_reference_list);
+
+        if ((usage_ + e->total_charge) > capacity_ && strict_capacity_limit_) {
+          free_standalone_handle = true;
+        } else {
+          usage_ += e->total_charge;
+        }
+      }
+
+      TryInsertIntoSecondaryCache(last_reference_list);
+      if (free_standalone_handle) {
+        e->Unref();
+        e->Free();
+        e = nullptr;
+      } else {
+        PERF_COUNTER_ADD(block_cache_standalone_handle_count, 1);
+      }
+
+      // Insert a dummy handle into the primary cache. This dummy handle is
+      // not IsSecondaryCacheCompatible().
+      // FIXME? This should not overwrite an existing non-dummy entry in the
+      // rare case that one exists
+      Cache::Priority priority =
+          e->IsHighPri() ? Cache::Priority::HIGH : Cache::Priority::LOW;
+      s = Insert(e->key(), e->hash, kDummyValueMarker, /*charge=*/0,
+                 /*deleter=*/nullptr, /*helper=*/nullptr, /*handle=*/nullptr,
+                 priority);
+    } else {
+      e->SetInCache(true);
+      LRUHandle* handle = e;
+      // This InsertItem() could fail if the cache is over capacity and
+      // strict_capacity_limit_ is true. In such a case, we don't want
+      // InsertItem() to free the handle, since the item is already in memory
+      // and the caller will most likely just read it from disk if we erase it
+      // here.
+      s = InsertItem(e, &handle, /*free_handle_on_fail=*/false);
+      if (s.ok()) {
+        PERF_COUNTER_ADD(block_cache_real_handle_count, 1);
+      }
+    }
+
    if (!s.ok()) {
      // Item is in memory, but not accounted against the cache capacity.
      // When the handle is released, the item should get deleted.
      assert(!e->InCache());
    }
  } else {
-    // Since the secondary cache lookup failed, mark the item as not in cache
-    // Don't charge the cache as its only metadata that'll shortly be released
-    DMutexLock l(mutex_);
-    // TODO
-    e->CalcTotalCharge(0, metadata_charge_policy_);
-    e->SetInCache(false);
+    // Secondary cache lookup failed. The caller will take care of detecting
+    // this and eventually releasing e.
+    assert(!e->value);
+    assert(!e->InCache());
  }
 }

-Cache::Handle* LRUCacheShard::Lookup(
-    const Slice& key, uint32_t hash,
-    const ShardedCache::CacheItemHelper* helper,
-    const ShardedCache::CreateCallback& create_cb, Cache::Priority priority,
-    bool wait, Statistics* stats) {
+LRUHandle* LRUCacheShard::Lookup(const Slice& key, uint32_t hash,
+                                 const Cache::CacheItemHelper* helper,
+                                 const Cache::CreateCallback& create_cb,
+                                 Cache::Priority priority, bool wait,
+                                 Statistics* stats) {
  LRUHandle* e = nullptr;
+  bool found_dummy_entry{false};
  {
    DMutexLock l(mutex_);
    e = table_.Lookup(key, hash);
    if (e != nullptr) {
      assert(e->InCache());
-      if (!e->HasRefs()) {
-        // The entry is in LRU since it's in hash and has no external references
-        LRU_Remove(e);
+      if (e->value == kDummyValueMarker) {
+        // For a dummy handle, if it was retrieved from secondary cache,
+        // it may still exist in secondary cache.
+        // If the handle exists in secondary cache, the value should be
+        // erased from sec cache and be inserted into primary cache.
+        found_dummy_entry = true;
+        // Let the dummy entry be overwritten
+        e = nullptr;
+      } else {
+        if (!e->HasRefs()) {
+          // The entry is in LRU since it's in hash and has no external
+          // references.
+          LRU_Remove(e);
+        }
+        e->Ref();
+        e->SetHit();
      }
-      e->Ref();
-      e->SetHit();
    }
  }

-  // If handle table lookup failed, then allocate a handle outside the
-  // mutex if we're going to lookup in the secondary cache.
-  // Only support synchronous for now.
-  // TODO: Support asynchronous lookup in secondary cache
+  // If handle table lookup failed or the handle is a dummy one, allocate
+  // a handle outside the mutex if we re going to lookup in the secondary cache.
+  //
+  // When a block is firstly Lookup from CompressedSecondaryCache, we just
+  // insert a dummy block into the primary cache (charging the actual size of
+  // the block) and don't erase the block from CompressedSecondaryCache. A
+  // standalone handle is returned to the caller. Only if the block is hit
+  // again, we erase it from CompressedSecondaryCache and add it into the
+  // primary cache.
  if (!e && secondary_cache_ && helper && helper->saveto_cb) {
    // For objects from the secondary cache, we expect the caller to provide
    // a way to create/delete the primary cache object. The only case where
@@ -490,12 +556,13 @@ Cache::Handle* LRUCacheShard::Lookup(
    assert(create_cb && helper->del_cb);
    bool is_in_sec_cache{false};
    std::unique_ptr<SecondaryCacheResultHandle> secondary_handle =
-        secondary_cache_->Lookup(key, create_cb, wait, is_in_sec_cache);
+        secondary_cache_->Lookup(key, create_cb, wait, found_dummy_entry,
+                                 is_in_sec_cache);
    if (secondary_handle != nullptr) {
-      e = reinterpret_cast<LRUHandle*>(
-          new char[sizeof(LRUHandle) - 1 + key.size()]);
+      e = static_cast<LRUHandle*>(malloc(sizeof(LRUHandle) - 1 + key.size()));

-      e->flags = 0;
+      e->m_flags = 0;
+      e->im_flags = 0;
      e->SetSecondaryCacheCompatible(true);
      e->info_.helper = helper;
      e->key_length = key.size();
@@ -509,37 +576,46 @@ Cache::Handle* LRUCacheShard::Lookup(
      e->total_charge = 0;
      e->Ref();
      e->SetIsInSecondaryCache(is_in_sec_cache);
+      e->SetIsStandalone(secondary_cache_->SupportForceErase() &&
+                         !found_dummy_entry);

      if (wait) {
        Promote(e);
-        if (!e->value) {
-          // The secondary cache returned a handle, but the lookup failed.
-          e->Unref();
-          e->Free();
-          e = nullptr;
-        } else {
-          PERF_COUNTER_ADD(secondary_cache_hit_count, 1);
-          RecordTick(stats, SECONDARY_CACHE_HITS);
+        if (e) {
+          if (!e->value) {
+            // The secondary cache returned a handle, but the lookup failed.
+            e->Unref();
+            e->Free();
+            e = nullptr;
+          } else {
+            PERF_COUNTER_ADD(secondary_cache_hit_count, 1);
+            RecordTick(stats, SECONDARY_CACHE_HITS);
+          }
        }
      } else {
        // If wait is false, we always return a handle and let the caller
        // release the handle after checking for success or failure.
-        e->SetIncomplete(true);
+        e->SetIsPending(true);
        // This may be slightly inaccurate, if the lookup eventually fails.
        // But the probability is very low.
        PERF_COUNTER_ADD(secondary_cache_hit_count, 1);
        RecordTick(stats, SECONDARY_CACHE_HITS);
      }
+    } else {
+      // Caller will most likely overwrite the dummy entry with an Insert
+      // after this Lookup fails
+      assert(e == nullptr);
    }
  }
-  return reinterpret_cast<Cache::Handle*>(e);
+  return e;
 }

-bool LRUCacheShard::Ref(Cache::Handle* h) {
-  LRUHandle* e = reinterpret_cast<LRUHandle*>(h);
+bool LRUCacheShard::Ref(LRUHandle* e) {
  DMutexLock l(mutex_);
  // To create another reference - entry must be already externally referenced.
  assert(e->HasRefs());
+  // Pending handles are not for sharing
+  assert(!e->IsPending());
  e->Ref();
  return true;
 }
@@ -558,12 +634,15 @@ void LRUCacheShard::SetLowPriorityPoolRatio(double low_pri_pool_ratio) {
  MaintainPoolSize();
 }

-bool LRUCacheShard::Release(Cache::Handle* handle, bool erase_if_last_ref) {
-  if (handle == nullptr) {
+bool LRUCacheShard::Release(LRUHandle* e, bool /*useful*/,
+                            bool erase_if_last_ref) {
+  if (e == nullptr) {
    return false;
  }
-  LRUHandle* e = reinterpret_cast<LRUHandle*>(handle);
  bool last_reference = false;
+  // Must Wait or WaitAll first on pending handles. Otherwise, would leak
+  // a secondary cache handle.
+  assert(!e->IsPending());
  {
    DMutexLock l(mutex_);
    last_reference = e->Unref();
@@ -581,12 +660,8 @@ bool LRUCacheShard::Release(Cache::Handle* handle, bool erase_if_last_ref) {
        last_reference = false;
      }
    }
-    // If it was the last reference, and the entry is either not secondary
-    // cache compatible (i.e a dummy entry for accounting), or is secondary
-    // cache compatible and has a non-null value, then decrement the cache
-    // usage. If value is null in the latter case, that means the lookup
-    // failed and we didn't charge the cache.
-    if (last_reference && (!e->IsSecondaryCacheCompatible() || e->value)) {
+    // If it was the last reference, then decrement the cache usage.
+    if (last_reference) {
      assert(usage_ >= e->total_charge);
      usage_ -= e->total_charge;
    }
@@ -603,22 +678,25 @@ Status LRUCacheShard::Insert(const Slice& key, uint32_t hash, void* value,
                             size_t charge,
                             void (*deleter)(const Slice& key, void* value),
                             const Cache::CacheItemHelper* helper,
-                             Cache::Handle** handle, Cache::Priority priority) {
+                             LRUHandle** handle, Cache::Priority priority) {
  // Allocate the memory here outside of the mutex.
  // If the cache is full, we'll have to release it.
  // It shouldn't happen very often though.
-  LRUHandle* e = reinterpret_cast<LRUHandle*>(
-      new char[sizeof(LRUHandle) - 1 + key.size()]);
+  LRUHandle* e =
+      static_cast<LRUHandle*>(malloc(sizeof(LRUHandle) - 1 + key.size()));

  e->value = value;
-  e->flags = 0;
+  e->m_flags = 0;
+  e->im_flags = 0;
  if (helper) {
+    // Use only one of the two parameters
+    assert(deleter == nullptr);
+    // value == nullptr is reserved for indicating failure for when secondary
+    // cache compatible
+    assert(value != nullptr);
    e->SetSecondaryCacheCompatible(true);
    e->info_.helper = helper;
  } else {
-#ifdef __SANITIZE_THREAD__
-    e->is_secondary_cache_compatible_for_tsan = false;
-#endif  // __SANITIZE_THREAD__
    e->info_.deleter = deleter;
  }
  e->key_length = key.size();
@@ -659,9 +737,7 @@ void LRUCacheShard::Erase(const Slice& key, uint32_t hash) {
  }
 }

-bool LRUCacheShard::IsReady(Cache::Handle* handle) {
-  LRUHandle* e = reinterpret_cast<LRUHandle*>(handle);
-  DMutexLock l(mutex_);
+bool LRUCacheShard::IsReady(LRUHandle* e) {
  bool ready = true;
  if (e->IsPending()) {
    assert(secondary_cache_);
@@ -682,7 +758,17 @@ size_t LRUCacheShard::GetPinnedUsage() const {
  return usage_ - lru_usage_;
 }

-std::string LRUCacheShard::GetPrintableOptions() const {
+size_t LRUCacheShard::GetOccupancyCount() const {
+  DMutexLock l(mutex_);
+  return table_.GetOccupancyCount();
+}
+
+size_t LRUCacheShard::GetTableAddressCount() const {
+  DMutexLock l(mutex_);
+  return size_t{1} << table_.GetLengthBits();
+}
+
+void LRUCacheShard::AppendPrintableOptions(std::string& str) const {
  const int kBufferSize = 200;
  char buffer[kBufferSize];
  {
@@ -692,7 +778,7 @@ std::string LRUCacheShard::GetPrintableOptions() const {
    snprintf(buffer + strlen(buffer), kBufferSize - strlen(buffer),
             "    low_pri_pool_ratio: %.3lf\n", low_pri_pool_ratio_);
  }
-  return std::string(buffer);
+  str.append(buffer);
 }

 LRUCache::LRUCache(size_t capacity, int num_shard_bits,
@@ -701,51 +787,30 @@ LRUCache::LRUCache(size_t capacity, int num_shard_bits,
                   std::shared_ptr<MemoryAllocator> allocator,
                   bool use_adaptive_mutex,
                   CacheMetadataChargePolicy metadata_charge_policy,
-                   const std::shared_ptr<SecondaryCache>& secondary_cache)
+                   std::shared_ptr<SecondaryCache> _secondary_cache)
    : ShardedCache(capacity, num_shard_bits, strict_capacity_limit,
-                   std::move(allocator)) {
-  num_shards_ = 1 << num_shard_bits;
-  shards_ = reinterpret_cast<LRUCacheShard*>(
-      port::cacheline_aligned_alloc(sizeof(LRUCacheShard) * num_shards_));
-  size_t per_shard = (capacity + (num_shards_ - 1)) / num_shards_;
-  for (int i = 0; i < num_shards_; i++) {
-    new (&shards_[i]) LRUCacheShard(
+                   std::move(allocator)),
+      secondary_cache_(std::move(_secondary_cache)) {
+  size_t per_shard = GetPerShardCapacity();
+  SecondaryCache* secondary_cache = secondary_cache_.get();
+  InitShards([=](LRUCacheShard* cs) {
+    new (cs) LRUCacheShard(
        per_shard, strict_capacity_limit, high_pri_pool_ratio,
        low_pri_pool_ratio, use_adaptive_mutex, metadata_charge_policy,
        /* max_upper_hash_bits */ 32 - num_shard_bits, secondary_cache);
-  }
-  secondary_cache_ = secondary_cache;
-}
-
-LRUCache::~LRUCache() {
-  if (shards_ != nullptr) {
-    assert(num_shards_ > 0);
-    for (int i = 0; i < num_shards_; i++) {
-      shards_[i].~LRUCacheShard();
-    }
-    port::cacheline_aligned_free(shards_);
-  }
-}
-
-CacheShard* LRUCache::GetShard(uint32_t shard) {
-  return reinterpret_cast<CacheShard*>(&shards_[shard]);
-}
-
-const CacheShard* LRUCache::GetShard(uint32_t shard) const {
-  return reinterpret_cast<CacheShard*>(&shards_[shard]);
+  });
 }

 void* LRUCache::Value(Handle* handle) {
-  return reinterpret_cast<const LRUHandle*>(handle)->value;
+  auto h = reinterpret_cast<const LRUHandle*>(handle);
+  assert(!h->IsPending() || h->value == nullptr);
+  assert(h->value != kDummyValueMarker);
+  return h->value;
 }

 size_t LRUCache::GetCharge(Handle* handle) const {
-  CacheMetadataChargePolicy metadata_charge_policy = kDontChargeCacheMetadata;
-  if (num_shards_ > 0) {
-    metadata_charge_policy = shards_[0].metadata_charge_policy_;
-  }
  return reinterpret_cast<const LRUHandle*>(handle)->GetCharge(
-      metadata_charge_policy);
+      GetShard(0).metadata_charge_policy_);
 }

 Cache::DeleterFn LRUCache::GetDeleter(Handle* handle) const {
@@ -757,32 +822,12 @@ Cache::DeleterFn LRUCache::GetDeleter(Handle* handle) const {
  }
 }

-uint32_t LRUCache::GetHash(Handle* handle) const {
-  return reinterpret_cast<const LRUHandle*>(handle)->hash;
-}
-
-void LRUCache::DisownData() {
-  // Leak data only if that won't generate an ASAN/valgrind warning.
-  if (!kMustFreeHeapAllocations) {
-    shards_ = nullptr;
-    num_shards_ = 0;
-  }
-}
-
 size_t LRUCache::TEST_GetLRUSize() {
-  size_t lru_size_of_all_shards = 0;
-  for (int i = 0; i < num_shards_; i++) {
-    lru_size_of_all_shards += shards_[i].TEST_GetLRUSize();
-  }
-  return lru_size_of_all_shards;
+  return SumOverShards([](LRUCacheShard& cs) { return cs.TEST_GetLRUSize(); });
 }

 double LRUCache::GetHighPriPoolRatio() {
-  double result = 0.0;
-  if (num_shards_ > 0) {
-    result = shards_[0].GetHighPriPoolRatio();
-  }
-  return result;
+  return GetShard(0).GetHighPriPoolRatio();
 }

 void LRUCache::WaitAll(std::vector<Handle*>& handles) {
@@ -808,22 +853,17 @@ void LRUCache::WaitAll(std::vector<Handle*>& handles) {
      if (!lru_handle->IsPending()) {
        continue;
      }
-      uint32_t hash = GetHash(handle);
-      LRUCacheShard* shard = static_cast<LRUCacheShard*>(GetShard(Shard(hash)));
-      shard->Promote(lru_handle);
+      GetShard(lru_handle->hash).Promote(lru_handle);
    }
  }
 }

-std::string LRUCache::GetPrintableOptions() const {
-  std::string ret;
-  ret.reserve(20000);
-  ret.append(ShardedCache::GetPrintableOptions());
+void LRUCache::AppendPrintableOptions(std::string& str) const {
+  ShardedCache::AppendPrintableOptions(str);  // options from shard
  if (secondary_cache_) {
-    ret.append("  secondary_cache:\n");
-    ret.append(secondary_cache_->GetPrintableOptions());
+    str.append("  secondary_cache:\n");
+    str.append(secondary_cache_->GetPrintableOptions());
  }
-  return ret;
 }

 }  // namespace lru_cache
@@ -843,7 +883,7 @@ std::shared_ptr<Cache> NewLRUCache(
    return nullptr;
  }
  if (low_pri_pool_ratio < 0.0 || low_pri_pool_ratio > 1.0) {
-    // Invalid high_pri_pool_ratio
+    // Invalid low_pri_pool_ratio
    return nullptr;
  }
  if (low_pri_pool_ratio + high_pri_pool_ratio > 1.0) {
@@ -38,18 +38,14 @@ namespace lru_cache {
 //    (refs == 0 && in_cache == true)
 // 3. Referenced externally AND not in hash table.
 //    In that case the entry is not in the LRU list and not in hash table.
-//    The entry can be freed when refs becomes 0.
+//    The entry must be freed if refs becomes 0 in this state.
 //    (refs >= 1 && in_cache == false)
-//
-// All newly created LRUHandles are in state 1. If you call
-// LRUCacheShard::Release on entry in state 1, it will go into state 2.
-// To move from state 1 to state 3, either call LRUCacheShard::Erase or
-// LRUCacheShard::Insert with the same key (but possibly different value).
-// To move from state 2 to state 1, use LRUCacheShard::Lookup.
-// Before destruction, make sure that no handles are in state 1. This means
-// that any successful LRUCacheShard::Lookup/LRUCacheShard::Insert have a
-// matching LRUCache::Release (to move into state 2) or LRUCacheShard::Erase
-// (to move into state 3).
+// If you call LRUCacheShard::Release enough times on an entry in state 1, it
+// will go into state 2. To move from state 1 to state 3, either call
+// LRUCacheShard::Erase or LRUCacheShard::Insert with the same key (but
+// possibly different value). To move from state 2 to state 1, use
+// LRUCacheShard::Lookup.
+// While refs > 0, public properties like value and deleter must not change.

 struct LRUHandle {
  void* value;
@@ -57,7 +53,7 @@ struct LRUHandle {
    Info() {}
    ~Info() {}
    Cache::DeleterFn deleter;
-    const ShardedCache::CacheItemHelper* helper;
+    const Cache::CacheItemHelper* helper;
  } info_;
  // An entry is not added to the LRUHandleTable until the secondary cache
  // lookup is complete, so its safe to have this union.
@@ -74,42 +70,47 @@ struct LRUHandle {
  // The number of external refs to this entry. The cache itself is not counted.
  uint32_t refs;

-  enum Flags : uint16_t {
+  // Mutable flags - access controlled by mutex
+  // The m_ and M_ prefixes (and im_ and IM_ later) are to hopefully avoid
+  // checking an M_ flag on im_flags or an IM_ flag on m_flags.
+  uint8_t m_flags;
+  enum MFlags : uint8_t {
    // Whether this entry is referenced by the hash table.
-    IN_CACHE = (1 << 0),
-    // Whether this entry is high priority entry.
-    IS_HIGH_PRI = (1 << 1),
-    // Whether this entry is in high-pri pool.
-    IN_HIGH_PRI_POOL = (1 << 2),
+    M_IN_CACHE = (1 << 0),
    // Whether this entry has had any lookups (hits).
-    HAS_HIT = (1 << 3),
-    // Can this be inserted into the secondary cache.
-    IS_SECONDARY_CACHE_COMPATIBLE = (1 << 4),
-    // Is the handle still being read from a lower tier.
-    IS_PENDING = (1 << 5),
-    // Whether this handle is still in a lower tier
-    IS_IN_SECONDARY_CACHE = (1 << 6),
-    // Whether this entry is low priority entry.
-    IS_LOW_PRI = (1 << 7),
+    M_HAS_HIT = (1 << 1),
+    // Whether this entry is in high-pri pool.
+    M_IN_HIGH_PRI_POOL = (1 << 2),
    // Whether this entry is in low-pri pool.
-    IN_LOW_PRI_POOL = (1 << 8),
+    M_IN_LOW_PRI_POOL = (1 << 3),
  };

-  uint16_t flags;
-
-#ifdef __SANITIZE_THREAD__
-  // TSAN can report a false data race on flags, where one thread is writing
-  // to one of the mutable bits and another thread is reading this immutable
-  // bit. So precisely suppress that TSAN warning, we separate out this bit
-  // during TSAN runs.
-  bool is_secondary_cache_compatible_for_tsan;
-#endif  // __SANITIZE_THREAD__
+  // "Immutable" flags - only set in single-threaded context and then
+  // can be accessed without mutex
+  uint8_t im_flags;
+  enum ImFlags : uint8_t {
+    // Whether this entry is high priority entry.
+    IM_IS_HIGH_PRI = (1 << 0),
+    // Whether this entry is low priority entry.
+    IM_IS_LOW_PRI = (1 << 1),
+    // Can this be inserted into the secondary cache.
+    IM_IS_SECONDARY_CACHE_COMPATIBLE = (1 << 2),
+    // Is the handle still being read from a lower tier.
+    IM_IS_PENDING = (1 << 3),
+    // Whether this handle is still in a lower tier
+    IM_IS_IN_SECONDARY_CACHE = (1 << 4),
+    // Marks result handles that should not be inserted into cache
+    IM_IS_STANDALONE = (1 << 5),
+  };

  // Beginning of the key (MUST BE THE LAST FIELD IN THIS STRUCT!)
  char key_data[1];

  Slice key() const { return Slice(key_data, key_length); }

+  // For HandleImpl concept
+  uint32_t GetHash() const { return hash; }
+
  // Increase the reference count by 1.
  void Ref() { refs++; }

@@ -123,95 +124,95 @@ struct LRUHandle {
  // Return true if there are external refs, false otherwise.
  bool HasRefs() const { return refs > 0; }

-  bool InCache() const { return flags & IN_CACHE; }
-  bool IsHighPri() const { return flags & IS_HIGH_PRI; }
-  bool InHighPriPool() const { return flags & IN_HIGH_PRI_POOL; }
-  bool IsLowPri() const { return flags & IS_LOW_PRI; }
-  bool InLowPriPool() const { return flags & IN_LOW_PRI_POOL; }
-  bool HasHit() const { return flags & HAS_HIT; }
+  bool InCache() const { return m_flags & M_IN_CACHE; }
+  bool IsHighPri() const { return im_flags & IM_IS_HIGH_PRI; }
+  bool InHighPriPool() const { return m_flags & M_IN_HIGH_PRI_POOL; }
+  bool IsLowPri() const { return im_flags & IM_IS_LOW_PRI; }
+  bool InLowPriPool() const { return m_flags & M_IN_LOW_PRI_POOL; }
+  bool HasHit() const { return m_flags & M_HAS_HIT; }
  bool IsSecondaryCacheCompatible() const {
-#ifdef __SANITIZE_THREAD__
-    return is_secondary_cache_compatible_for_tsan;
-#else
-    return flags & IS_SECONDARY_CACHE_COMPATIBLE;
-#endif  // __SANITIZE_THREAD__
+    return im_flags & IM_IS_SECONDARY_CACHE_COMPATIBLE;
  }
-  bool IsPending() const { return flags & IS_PENDING; }
-  bool IsInSecondaryCache() const { return flags & IS_IN_SECONDARY_CACHE; }
+  bool IsPending() const { return im_flags & IM_IS_PENDING; }
+  bool IsInSecondaryCache() const {
+    return im_flags & IM_IS_IN_SECONDARY_CACHE;
+  }
+  bool IsStandalone() const { return im_flags & IM_IS_STANDALONE; }

  void SetInCache(bool in_cache) {
    if (in_cache) {
-      flags |= IN_CACHE;
+      m_flags |= M_IN_CACHE;
    } else {
-      flags &= ~IN_CACHE;
+      m_flags &= ~M_IN_CACHE;
    }
  }

  void SetPriority(Cache::Priority priority) {
    if (priority == Cache::Priority::HIGH) {
-      flags |= IS_HIGH_PRI;
-      flags &= ~IS_LOW_PRI;
+      im_flags |= IM_IS_HIGH_PRI;
+      im_flags &= ~IM_IS_LOW_PRI;
    } else if (priority == Cache::Priority::LOW) {
-      flags &= ~IS_HIGH_PRI;
-      flags |= IS_LOW_PRI;
+      im_flags &= ~IM_IS_HIGH_PRI;
+      im_flags |= IM_IS_LOW_PRI;
    } else {
-      flags &= ~IS_HIGH_PRI;
-      flags &= ~IS_LOW_PRI;
+      im_flags &= ~IM_IS_HIGH_PRI;
+      im_flags &= ~IM_IS_LOW_PRI;
    }
  }

  void SetInHighPriPool(bool in_high_pri_pool) {
    if (in_high_pri_pool) {
-      flags |= IN_HIGH_PRI_POOL;
+      m_flags |= M_IN_HIGH_PRI_POOL;
    } else {
-      flags &= ~IN_HIGH_PRI_POOL;
+      m_flags &= ~M_IN_HIGH_PRI_POOL;
    }
  }

  void SetInLowPriPool(bool in_low_pri_pool) {
    if (in_low_pri_pool) {
-      flags |= IN_LOW_PRI_POOL;
+      m_flags |= M_IN_LOW_PRI_POOL;
    } else {
-      flags &= ~IN_LOW_PRI_POOL;
+      m_flags &= ~M_IN_LOW_PRI_POOL;
    }
  }

-  void SetHit() { flags |= HAS_HIT; }
+  void SetHit() { m_flags |= M_HAS_HIT; }

  void SetSecondaryCacheCompatible(bool compat) {
    if (compat) {
-      flags |= IS_SECONDARY_CACHE_COMPATIBLE;
+      im_flags |= IM_IS_SECONDARY_CACHE_COMPATIBLE;
    } else {
-      flags &= ~IS_SECONDARY_CACHE_COMPATIBLE;
+      im_flags &= ~IM_IS_SECONDARY_CACHE_COMPATIBLE;
    }
-#ifdef __SANITIZE_THREAD__
-    is_secondary_cache_compatible_for_tsan = compat;
-#endif  // __SANITIZE_THREAD__
  }

-  void SetIncomplete(bool incomp) {
-    if (incomp) {
-      flags |= IS_PENDING;
+  void SetIsPending(bool pending) {
+    if (pending) {
+      im_flags |= IM_IS_PENDING;
    } else {
-      flags &= ~IS_PENDING;
+      im_flags &= ~IM_IS_PENDING;
    }
  }

  void SetIsInSecondaryCache(bool is_in_secondary_cache) {
    if (is_in_secondary_cache) {
-      flags |= IS_IN_SECONDARY_CACHE;
+      im_flags |= IM_IS_IN_SECONDARY_CACHE;
    } else {
-      flags &= ~IS_IN_SECONDARY_CACHE;
+      im_flags &= ~IM_IS_IN_SECONDARY_CACHE;
+    }
+  }
+
+  void SetIsStandalone(bool is_standalone) {
+    if (is_standalone) {
+      im_flags |= IM_IS_STANDALONE;
+    } else {
+      im_flags &= ~IM_IS_STANDALONE;
    }
  }

  void Free() {
    assert(refs == 0);
-#ifdef __SANITIZE_THREAD__
-    // Here we can safely assert they are the same without a data race reported
-    assert(((flags & IS_SECONDARY_CACHE_COMPATIBLE) != 0) ==
-           is_secondary_cache_compatible_for_tsan);
-#endif  // __SANITIZE_THREAD__
+
    if (!IsSecondaryCacheCompatible() && info_.deleter) {
      (*info_.deleter)(key(), value);
    } else if (IsSecondaryCacheCompatible()) {
@@ -226,7 +227,8 @@ struct LRUHandle {
        (*info_.helper->del_cb)(key(), value);
      }
    }
-    delete[] reinterpret_cast<char*>(this);
+
+    free(this);
  }

  inline size_t CalcuMetaCharge(
@@ -265,9 +267,6 @@ struct LRUHandle {
 // 4.4.3's builtin hashtable.
 class LRUHandleTable {
 public:
-  // If the table uses more hash bits than `max_upper_hash_bits`,
-  // it will eat into the bits used for sharding, which are constant
-  // for a given LRUHandleTable.
  explicit LRUHandleTable(int max_upper_hash_bits);
  ~LRUHandleTable();

@@ -276,8 +275,8 @@ class LRUHandleTable {
  LRUHandle* Remove(const Slice& key, uint32_t hash);

  template <typename T>
-  void ApplyToEntriesRange(T func, uint32_t index_begin, uint32_t index_end) {
-    for (uint32_t i = index_begin; i < index_end; i++) {
+  void ApplyToEntriesRange(T func, size_t index_begin, size_t index_end) {
+    for (size_t i = index_begin; i < index_end; i++) {
      LRUHandle* h = list_[i];
      while (h != nullptr) {
        auto n = h->next_hash;
@@ -290,6 +289,8 @@ class LRUHandleTable {

  int GetLengthBits() const { return length_bits_; }

+  size_t GetOccupancyCount() const { return elems_; }
+
 private:
  // Return a pointer to slot that points to a cache entry that
  // matches key/hash.  If there is no such cache entry, return a
@@ -314,23 +315,31 @@ class LRUHandleTable {
 };

 // A single shard of sharded cache.
-class ALIGN_AS(CACHE_LINE_SIZE) LRUCacheShard final : public CacheShard {
+class ALIGN_AS(CACHE_LINE_SIZE) LRUCacheShard final : public CacheShardBase {
 public:
  LRUCacheShard(size_t capacity, bool strict_capacity_limit,
                double high_pri_pool_ratio, double low_pri_pool_ratio,
                bool use_adaptive_mutex,
                CacheMetadataChargePolicy metadata_charge_policy,
-                int max_upper_hash_bits,
-                const std::shared_ptr<SecondaryCache>& secondary_cache);
-  virtual ~LRUCacheShard() override = default;
+                int max_upper_hash_bits, SecondaryCache* secondary_cache);
+
+ public:  // Type definitions expected as parameter to ShardedCache
+  using HandleImpl = LRUHandle;
+  using HashVal = uint32_t;
+  using HashCref = uint32_t;
+
+ public:  // Function definitions expected as parameter to ShardedCache
+  static inline HashVal ComputeHash(const Slice& key) {
+    return Lower32of64(GetSliceNPHash64(key));
+  }

  // Separate from constructor so caller can easily make an array of LRUCache
  // if current usage is more than new capacity, the function will attempt to
  // free the needed space.
-  virtual void SetCapacity(size_t capacity) override;
+  void SetCapacity(size_t capacity);

  // Set the flag to reject insertion if cache if full.
-  virtual void SetStrictCapacityLimit(bool strict_capacity_limit) override;
+  void SetStrictCapacityLimit(bool strict_capacity_limit);

  // Set percentage of capacity reserved for high-pri cache entries.
  void SetHighPriorityPoolRatio(double high_pri_pool_ratio);
@@ -339,56 +348,49 @@ class ALIGN_AS(CACHE_LINE_SIZE) LRUCacheShard final : public CacheShard {
  void SetLowPriorityPoolRatio(double low_pri_pool_ratio);

  // Like Cache methods, but with an extra "hash" parameter.
-  virtual Status Insert(const Slice& key, uint32_t hash, void* value,
-                        size_t charge, Cache::DeleterFn deleter,
-                        Cache::Handle** handle,
-                        Cache::Priority priority) override {
+  inline Status Insert(const Slice& key, uint32_t hash, void* value,
+                       size_t charge, Cache::DeleterFn deleter,
+                       LRUHandle** handle, Cache::Priority priority) {
    return Insert(key, hash, value, charge, deleter, nullptr, handle, priority);
  }
-  virtual Status Insert(const Slice& key, uint32_t hash, void* value,
-                        const Cache::CacheItemHelper* helper, size_t charge,
-                        Cache::Handle** handle,
-                        Cache::Priority priority) override {
+  inline Status Insert(const Slice& key, uint32_t hash, void* value,
+                       const Cache::CacheItemHelper* helper, size_t charge,
+                       LRUHandle** handle, Cache::Priority priority) {
    assert(helper);
    return Insert(key, hash, value, charge, nullptr, helper, handle, priority);
  }
  // If helper_cb is null, the values of the following arguments don't matter.
-  virtual Cache::Handle* Lookup(const Slice& key, uint32_t hash,
-                                const ShardedCache::CacheItemHelper* helper,
-                                const ShardedCache::CreateCallback& create_cb,
-                                ShardedCache::Priority priority, bool wait,
-                                Statistics* stats) override;
-  virtual Cache::Handle* Lookup(const Slice& key, uint32_t hash) override {
+  LRUHandle* Lookup(const Slice& key, uint32_t hash,
+                    const Cache::CacheItemHelper* helper,
+                    const Cache::CreateCallback& create_cb,
+                    Cache::Priority priority, bool wait, Statistics* stats);
+  inline LRUHandle* Lookup(const Slice& key, uint32_t hash) {
    return Lookup(key, hash, nullptr, nullptr, Cache::Priority::LOW, true,
                  nullptr);
  }
-  virtual bool Release(Cache::Handle* handle, bool /*useful*/,
-                       bool erase_if_last_ref) override {
-    return Release(handle, erase_if_last_ref);
-  }
-  virtual bool IsReady(Cache::Handle* /*handle*/) override;
-  virtual void Wait(Cache::Handle* /*handle*/) override {}
-  virtual bool Ref(Cache::Handle* handle) override;
-  virtual bool Release(Cache::Handle* handle,
-                       bool erase_if_last_ref = false) override;
-  virtual void Erase(const Slice& key, uint32_t hash) override;
+  bool Release(LRUHandle* handle, bool useful, bool erase_if_last_ref);
+  bool IsReady(LRUHandle* /*handle*/);
+  void Wait(LRUHandle* /*handle*/) {}
+  bool Ref(LRUHandle* handle);
+  void Erase(const Slice& key, uint32_t hash);

  // Although in some platforms the update of size_t is atomic, to make sure
  // GetUsage() and GetPinnedUsage() work correctly under any platform, we'll
  // protect them with mutex_.

-  virtual size_t GetUsage() const override;
-  virtual size_t GetPinnedUsage() const override;
+  size_t GetUsage() const;
+  size_t GetPinnedUsage() const;
+  size_t GetOccupancyCount() const;
+  size_t GetTableAddressCount() const;

-  virtual void ApplyToSomeEntries(
+  void ApplyToSomeEntries(
      const std::function<void(const Slice& key, void* value, size_t charge,
                               DeleterFn deleter)>& callback,
-      uint32_t average_entries_per_lock, uint32_t* state) override;
+      size_t average_entries_per_lock, size_t* state);

-  virtual void EraseUnRefEntries() override;
-
-  virtual std::string GetPrintableOptions() const override;
+  void EraseUnRefEntries();

+ public:  // other function definitions
  void TEST_GetLRUList(LRUHandle** lru, LRUHandle** lru_low_pri,
                       LRUHandle** lru_bottom_pri);

@@ -402,17 +404,19 @@ class ALIGN_AS(CACHE_LINE_SIZE) LRUCacheShard final : public CacheShard {
  // Retrieves low pri pool ratio
  double GetLowPriPoolRatio();

+  void AppendPrintableOptions(std::string& /*str*/) const;
+
 private:
  friend class LRUCache;
  // Insert an item into the hash table and, if handle is null, insert into
  // the LRU list. Older items are evicted as necessary. If the cache is full
  // and free_handle_on_fail is true, the item is deleted and handle is set to
  // nullptr.
-  Status InsertItem(LRUHandle* item, Cache::Handle** handle,
+  Status InsertItem(LRUHandle* item, LRUHandle** handle,
                    bool free_handle_on_fail);
  Status Insert(const Slice& key, uint32_t hash, void* value, size_t charge,
                DeleterFn deleter, const Cache::CacheItemHelper* helper,
-                Cache::Handle** handle, Cache::Priority priority);
+                LRUHandle** handle, Cache::Priority priority);
  // Promote an item looked up from the secondary cache to the LRU cache.
  // The item may be still in the secondary cache.
  // It is only inserted into the hash table and not the LRU list, and only
@@ -435,6 +439,9 @@ class ALIGN_AS(CACHE_LINE_SIZE) LRUCacheShard final : public CacheShard {
  // holding the mutex_.
  void EvictFromLRU(size_t charge, autovector<LRUHandle*>* deleted);

+  // Try to insert the evicted handles into the secondary cache.
+  void TryInsertIntoSecondaryCache(autovector<LRUHandle*> evicted_handles);
+
  // Initialized before use.
  size_t capacity_;

@@ -496,14 +503,15 @@ class ALIGN_AS(CACHE_LINE_SIZE) LRUCacheShard final : public CacheShard {
  // don't mind mutex_ invoking the non-const actions.
  mutable DMutex mutex_;

-  std::shared_ptr<SecondaryCache> secondary_cache_;
+  // Owned by LRUCache
+  SecondaryCache* secondary_cache_;
 };

 class LRUCache
 #ifdef NDEBUG
    final
 #endif
-    : public ShardedCache {
+    : public ShardedCache<LRUCacheShard> {
 public:
  LRUCache(size_t capacity, int num_shard_bits, bool strict_capacity_limit,
           double high_pri_pool_ratio, double low_pri_pool_ratio,
@@ -511,27 +519,21 @@ class LRUCache
           bool use_adaptive_mutex = kDefaultToAdaptiveMutex,
           CacheMetadataChargePolicy metadata_charge_policy =
               kDontChargeCacheMetadata,
-           const std::shared_ptr<SecondaryCache>& secondary_cache = nullptr);
-  virtual ~LRUCache();
-  virtual const char* Name() const override { return "LRUCache"; }
-  virtual CacheShard* GetShard(uint32_t shard) override;
-  virtual const CacheShard* GetShard(uint32_t shard) const override;
-  virtual void* Value(Handle* handle) override;
-  virtual size_t GetCharge(Handle* handle) const override;
-  virtual uint32_t GetHash(Handle* handle) const override;
-  virtual DeleterFn GetDeleter(Handle* handle) const override;
-  virtual void DisownData() override;
-  virtual void WaitAll(std::vector<Handle*>& handles) override;
-  std::string GetPrintableOptions() const override;
+           std::shared_ptr<SecondaryCache> secondary_cache = nullptr);
+  const char* Name() const override { return "LRUCache"; }
+  void* Value(Handle* handle) override;
+  size_t GetCharge(Handle* handle) const override;
+  DeleterFn GetDeleter(Handle* handle) const override;
+  void WaitAll(std::vector<Handle*>& handles) override;

  // Retrieves number of elements in LRU, for unit test purpose only.
  size_t TEST_GetLRUSize();
  // Retrieves high pri pool ratio.
  double GetHighPriPoolRatio();

+  void AppendPrintableOptions(std::string& str) const override;
+
 private:
-  LRUCacheShard* shards_ = nullptr;
-  int num_shards_ = 0;
  std::shared_ptr<SecondaryCache> secondary_cache_;
 };

@@ -46,10 +46,11 @@ class LRUCacheTest : public testing::Test {
    DeleteCache();
    cache_ = reinterpret_cast<LRUCacheShard*>(
        port::cacheline_aligned_alloc(sizeof(LRUCacheShard)));
-    new (cache_) LRUCacheShard(
-        capacity, false /*strict_capcity_limit*/, high_pri_pool_ratio,
-        low_pri_pool_ratio, use_adaptive_mutex, kDontChargeCacheMetadata,
-        24 /*max_upper_hash_bits*/, nullptr /*secondary_cache*/);
+    new (cache_) LRUCacheShard(capacity, /*strict_capacity_limit=*/false,
+                               high_pri_pool_ratio, low_pri_pool_ratio,
+                               use_adaptive_mutex, kDontChargeCacheMetadata,
+                               /*max_upper_hash_bits=*/24,
+                               /*secondary_cache=*/nullptr);
  }

  void Insert(const std::string& key,
@@ -66,7 +67,7 @@ class LRUCacheTest : public testing::Test {
  bool Lookup(const std::string& key) {
    auto handle = cache_->Lookup(key, 0 /*hash*/);
    if (handle) {
-      cache_->Release(handle);
+      cache_->Release(handle, true /*useful*/, false /*erase*/);
      return true;
    }
    return false;
@@ -505,7 +506,7 @@ TEST_F(FastLRUCacheTest, CalcHashBitsTest) {

 }  // namespace fast_lru_cache

-namespace clock_cache {
+namespace hyper_clock_cache {

 class ClockCacheTest : public testing::Test {
 public:
@@ -520,171 +521,485 @@ class ClockCacheTest : public testing::Test {
    }
  }

-  void NewShard(size_t capacity) {
+  void NewShard(size_t capacity, bool strict_capacity_limit = true) {
    DeleteShard();
    shard_ = reinterpret_cast<ClockCacheShard*>(
        port::cacheline_aligned_alloc(sizeof(ClockCacheShard)));
-    new (shard_) ClockCacheShard(capacity, 1, true /*strict_capacity_limit*/,
+    new (shard_) ClockCacheShard(capacity, 1, strict_capacity_limit,
                                 kDontChargeCacheMetadata);
  }

-  Status Insert(const std::string& key,
+  Status Insert(const UniqueId64x2& hashed_key,
                Cache::Priority priority = Cache::Priority::LOW) {
-    return shard_->Insert(key, 0 /*hash*/, nullptr /*value*/, 1 /*charge*/,
-                          nullptr /*deleter*/, nullptr /*handle*/, priority);
+    return shard_->Insert(TestKey(hashed_key), hashed_key, nullptr /*value*/,
+                          1 /*charge*/, nullptr /*deleter*/, nullptr /*handle*/,
+                          priority);
  }

  Status Insert(char key, Cache::Priority priority = Cache::Priority::LOW) {
-    return Insert(std::string(kCacheKeySize, key), priority);
+    return Insert(TestHashedKey(key), priority);
  }

-  Status Insert(char key, size_t len) { return Insert(std::string(len, key)); }
+  Status InsertWithLen(char key, size_t len) {
+    std::string skey(len, key);
+    return shard_->Insert(skey, TestHashedKey(key), nullptr /*value*/,
+                          1 /*charge*/, nullptr /*deleter*/, nullptr /*handle*/,
+                          Cache::Priority::LOW);
+  }

-  bool Lookup(const std::string& key) {
-    auto handle = shard_->Lookup(key, 0 /*hash*/);
+  bool Lookup(const Slice& key, const UniqueId64x2& hashed_key,
+              bool useful = true) {
+    auto handle = shard_->Lookup(key, hashed_key);
    if (handle) {
-      shard_->Release(handle);
+      shard_->Release(handle, useful, /*erase_if_last_ref=*/false);
      return true;
    }
    return false;
  }

-  bool Lookup(char key) { return Lookup(std::string(kCacheKeySize, key)); }
-
-  void Erase(const std::string& key) { shard_->Erase(key, 0 /*hash*/); }
-
-  size_t CalcEstimatedHandleChargeWrapper(
-      size_t estimated_value_size,
-      CacheMetadataChargePolicy metadata_charge_policy) {
-    return ClockCacheShard::CalcEstimatedHandleCharge(estimated_value_size,
-                                                      metadata_charge_policy);
+  bool Lookup(const UniqueId64x2& hashed_key, bool useful = true) {
+    return Lookup(TestKey(hashed_key), hashed_key, useful);
  }

-  int CalcHashBitsWrapper(size_t capacity, size_t estimated_value_size,
-                          CacheMetadataChargePolicy metadata_charge_policy) {
-    return ClockCacheShard::CalcHashBits(capacity, estimated_value_size,
-                                         metadata_charge_policy);
+  bool Lookup(char key, bool useful = true) {
+    return Lookup(TestHashedKey(key), useful);
  }

-  // Maximum number of items that a shard can hold.
-  double CalcMaxOccupancy(size_t capacity, size_t estimated_value_size,
-                          CacheMetadataChargePolicy metadata_charge_policy) {
-    size_t handle_charge = ClockCacheShard::CalcEstimatedHandleCharge(
-        estimated_value_size, metadata_charge_policy);
-    return capacity / (kLoadFactor * handle_charge);
+  void Erase(char key) {
+    UniqueId64x2 hashed_key = TestHashedKey(key);
+    shard_->Erase(TestKey(hashed_key), hashed_key);
  }

-  bool TableSizeIsAppropriate(int hash_bits, double max_occupancy) {
-    if (hash_bits == 0) {
-      return max_occupancy <= 1;
-    } else {
-      return (1 << hash_bits >= max_occupancy) &&
-             (1 << (hash_bits - 1) <= max_occupancy);
-    }
+  static inline Slice TestKey(const UniqueId64x2& hashed_key) {
+    return Slice(reinterpret_cast<const char*>(&hashed_key), 16U);
+  }
+
+  static inline UniqueId64x2 TestHashedKey(char key) {
+    // For testing hash near-collision behavior, put the variance in
+    // hashed_key in bits that are unlikely to be used as hash bits.
+    return {(static_cast<uint64_t>(key) << 56) + 1234U, 5678U};
  }

- private:
  ClockCacheShard* shard_ = nullptr;
 };

-TEST_F(ClockCacheTest, Validate) {
+TEST_F(ClockCacheTest, Misc) {
  NewShard(3);
-  EXPECT_OK(Insert('a', 16));
-  EXPECT_NOK(Insert('b', 15));
-  EXPECT_OK(Insert('b', 16));
-  EXPECT_NOK(Insert('c', 17));
-  EXPECT_NOK(Insert('d', 1000));
-  EXPECT_NOK(Insert('e', 11));
-  EXPECT_NOK(Insert('f', 0));
+
+  // Key size stuff
+  EXPECT_OK(InsertWithLen('a', 16));
+  EXPECT_NOK(InsertWithLen('b', 15));
+  EXPECT_OK(InsertWithLen('b', 16));
+  EXPECT_NOK(InsertWithLen('c', 17));
+  EXPECT_NOK(InsertWithLen('d', 1000));
+  EXPECT_NOK(InsertWithLen('e', 11));
+  EXPECT_NOK(InsertWithLen('f', 0));
+
+  // Some of this is motivated by code coverage
+  std::string wrong_size_key(15, 'x');
+  EXPECT_FALSE(Lookup(wrong_size_key, TestHashedKey('x')));
+  EXPECT_FALSE(shard_->Ref(nullptr));
+  EXPECT_FALSE(shard_->Release(nullptr));
+  shard_->Erase(wrong_size_key, TestHashedKey('x'));  // no-op
 }

-TEST_F(ClockCacheTest, ClockPriorityTest) {
-  ClockHandle handle;
-  EXPECT_EQ(handle.GetClockPriority(), ClockHandle::ClockPriority::NONE);
-  handle.SetClockPriority(ClockHandle::ClockPriority::HIGH);
-  EXPECT_EQ(handle.GetClockPriority(), ClockHandle::ClockPriority::HIGH);
-  handle.DecreaseClockPriority();
-  EXPECT_EQ(handle.GetClockPriority(), ClockHandle::ClockPriority::MEDIUM);
-  handle.DecreaseClockPriority();
-  EXPECT_EQ(handle.GetClockPriority(), ClockHandle::ClockPriority::LOW);
-  handle.SetClockPriority(ClockHandle::ClockPriority::MEDIUM);
-  EXPECT_EQ(handle.GetClockPriority(), ClockHandle::ClockPriority::MEDIUM);
-  handle.SetClockPriority(ClockHandle::ClockPriority::NONE);
-  EXPECT_EQ(handle.GetClockPriority(), ClockHandle::ClockPriority::NONE);
-  handle.SetClockPriority(ClockHandle::ClockPriority::MEDIUM);
-  EXPECT_EQ(handle.GetClockPriority(), ClockHandle::ClockPriority::MEDIUM);
-  handle.DecreaseClockPriority();
-  handle.DecreaseClockPriority();
-  EXPECT_EQ(handle.GetClockPriority(), ClockHandle::ClockPriority::NONE);
+TEST_F(ClockCacheTest, Limits) {
+  NewShard(3, false /*strict_capacity_limit*/);
+  for (bool strict_capacity_limit : {false, true, false}) {
+    SCOPED_TRACE("strict_capacity_limit = " +
+                 std::to_string(strict_capacity_limit));
+
+    // Also tests switching between strict limit and not
+    shard_->SetStrictCapacityLimit(strict_capacity_limit);
+
+    UniqueId64x2 hkey = TestHashedKey('x');
+
+    // Single entry charge beyond capacity
+    {
+      Status s = shard_->Insert(TestKey(hkey), hkey, nullptr /*value*/,
+                                5 /*charge*/, nullptr /*deleter*/,
+                                nullptr /*handle*/, Cache::Priority::LOW);
+      if (strict_capacity_limit) {
+        EXPECT_TRUE(s.IsMemoryLimit());
+      } else {
+        EXPECT_OK(s);
+      }
+    }
+
+    // Single entry fills capacity
+    {
+      ClockHandle* h;
+      ASSERT_OK(shard_->Insert(TestKey(hkey), hkey, nullptr /*value*/,
+                               3 /*charge*/, nullptr /*deleter*/, &h,
+                               Cache::Priority::LOW));
+      // Try to insert more
+      Status s = Insert('a');
+      if (strict_capacity_limit) {
+        EXPECT_TRUE(s.IsMemoryLimit());
+      } else {
+        EXPECT_OK(s);
+      }
+      // Release entry filling capacity.
+      // Cover useful = false case.
+      shard_->Release(h, false /*useful*/, false /*erase_if_last_ref*/);
+    }
+
+    // Insert more than table size can handle (cleverly using zero-charge
+    // entries) to exceed occupancy limit.
+    {
+      size_t n = shard_->GetTableAddressCount() + 1;
+      std::unique_ptr<ClockHandle* []> ha { new ClockHandle* [n] {} };
+      Status s;
+      for (size_t i = 0; i < n && s.ok(); ++i) {
+        hkey[1] = i;
+        s = shard_->Insert(TestKey(hkey), hkey, nullptr /*value*/, 0 /*charge*/,
+                           nullptr /*deleter*/, &ha[i], Cache::Priority::LOW);
+        if (i == 0) {
+          EXPECT_OK(s);
+        }
+      }
+      if (strict_capacity_limit) {
+        EXPECT_TRUE(s.IsMemoryLimit());
+      } else {
+        EXPECT_OK(s);
+      }
+      // Same result if not keeping a reference
+      s = Insert('a');
+      if (strict_capacity_limit) {
+        EXPECT_TRUE(s.IsMemoryLimit());
+      } else {
+        EXPECT_OK(s);
+      }
+
+      // Regardless, we didn't allow table to actually get full
+      EXPECT_LT(shard_->GetOccupancyCount(), shard_->GetTableAddressCount());
+
+      // Release handles
+      for (size_t i = 0; i < n; ++i) {
+        if (ha[i]) {
+          shard_->Release(ha[i]);
+        }
+      }
+    }
+  }
 }

-TEST_F(ClockCacheTest, CalcHashBitsTest) {
-  size_t capacity;
-  size_t estimated_value_size;
-  double max_occupancy;
-  int hash_bits;
-  CacheMetadataChargePolicy metadata_charge_policy;
+TEST_F(ClockCacheTest, ClockEvictionTest) {
+  for (bool strict_capacity_limit : {false, true}) {
+    SCOPED_TRACE("strict_capacity_limit = " +
+                 std::to_string(strict_capacity_limit));

-  // Vary the cache capacity, fix the element charge.
-  for (int i = 0; i < 2048; i++) {
-    capacity = i;
-    estimated_value_size = 0;
-    metadata_charge_policy = kFullChargeCacheMetadata;
-    max_occupancy = CalcMaxOccupancy(capacity, estimated_value_size,
-                                     metadata_charge_policy);
-    hash_bits = CalcHashBitsWrapper(capacity, estimated_value_size,
-                                    metadata_charge_policy);
-    EXPECT_TRUE(TableSizeIsAppropriate(hash_bits, max_occupancy));
+    NewShard(6, strict_capacity_limit);
+    EXPECT_OK(Insert('a', Cache::Priority::BOTTOM));
+    EXPECT_OK(Insert('b', Cache::Priority::LOW));
+    EXPECT_OK(Insert('c', Cache::Priority::HIGH));
+    EXPECT_OK(Insert('d', Cache::Priority::BOTTOM));
+    EXPECT_OK(Insert('e', Cache::Priority::LOW));
+    EXPECT_OK(Insert('f', Cache::Priority::HIGH));
+
+    EXPECT_TRUE(Lookup('a', /*use*/ false));
+    EXPECT_TRUE(Lookup('b', /*use*/ false));
+    EXPECT_TRUE(Lookup('c', /*use*/ false));
+    EXPECT_TRUE(Lookup('d', /*use*/ false));
+    EXPECT_TRUE(Lookup('e', /*use*/ false));
+    EXPECT_TRUE(Lookup('f', /*use*/ false));
+
+    // Ensure bottom are evicted first, even if new entries are low
+    EXPECT_OK(Insert('g', Cache::Priority::LOW));
+    EXPECT_OK(Insert('h', Cache::Priority::LOW));
+
+    EXPECT_FALSE(Lookup('a', /*use*/ false));
+    EXPECT_TRUE(Lookup('b', /*use*/ false));
+    EXPECT_TRUE(Lookup('c', /*use*/ false));
+    EXPECT_FALSE(Lookup('d', /*use*/ false));
+    EXPECT_TRUE(Lookup('e', /*use*/ false));
+    EXPECT_TRUE(Lookup('f', /*use*/ false));
+    // Mark g & h useful
+    EXPECT_TRUE(Lookup('g', /*use*/ true));
+    EXPECT_TRUE(Lookup('h', /*use*/ true));
+
+    // Then old LOW entries
+    EXPECT_OK(Insert('i', Cache::Priority::LOW));
+    EXPECT_OK(Insert('j', Cache::Priority::LOW));
+
+    EXPECT_FALSE(Lookup('b', /*use*/ false));
+    EXPECT_TRUE(Lookup('c', /*use*/ false));
+    EXPECT_FALSE(Lookup('e', /*use*/ false));
+    EXPECT_TRUE(Lookup('f', /*use*/ false));
+    // Mark g & h useful once again
+    EXPECT_TRUE(Lookup('g', /*use*/ true));
+    EXPECT_TRUE(Lookup('h', /*use*/ true));
+    EXPECT_TRUE(Lookup('i', /*use*/ false));
+    EXPECT_TRUE(Lookup('j', /*use*/ false));
+
+    // Then old HIGH entries
+    EXPECT_OK(Insert('k', Cache::Priority::LOW));
+    EXPECT_OK(Insert('l', Cache::Priority::LOW));
+
+    EXPECT_FALSE(Lookup('c', /*use*/ false));
+    EXPECT_FALSE(Lookup('f', /*use*/ false));
+    EXPECT_TRUE(Lookup('g', /*use*/ false));
+    EXPECT_TRUE(Lookup('h', /*use*/ false));
+    EXPECT_TRUE(Lookup('i', /*use*/ false));
+    EXPECT_TRUE(Lookup('j', /*use*/ false));
+    EXPECT_TRUE(Lookup('k', /*use*/ false));
+    EXPECT_TRUE(Lookup('l', /*use*/ false));
+
+    // Then the (roughly) least recently useful
+    EXPECT_OK(Insert('m', Cache::Priority::HIGH));
+    EXPECT_OK(Insert('n', Cache::Priority::HIGH));
+
+    EXPECT_TRUE(Lookup('g', /*use*/ false));
+    EXPECT_TRUE(Lookup('h', /*use*/ false));
+    EXPECT_FALSE(Lookup('i', /*use*/ false));
+    EXPECT_FALSE(Lookup('j', /*use*/ false));
+    EXPECT_TRUE(Lookup('k', /*use*/ false));
+    EXPECT_TRUE(Lookup('l', /*use*/ false));
+
+    // Now try changing capacity down
+    shard_->SetCapacity(4);
+    // Insert to ensure evictions happen
+    EXPECT_OK(Insert('o', Cache::Priority::LOW));
+    EXPECT_OK(Insert('p', Cache::Priority::LOW));
+
+    EXPECT_FALSE(Lookup('g', /*use*/ false));
+    EXPECT_FALSE(Lookup('h', /*use*/ false));
+    EXPECT_FALSE(Lookup('k', /*use*/ false));
+    EXPECT_FALSE(Lookup('l', /*use*/ false));
+    EXPECT_TRUE(Lookup('m', /*use*/ false));
+    EXPECT_TRUE(Lookup('n', /*use*/ false));
+    EXPECT_TRUE(Lookup('o', /*use*/ false));
+    EXPECT_TRUE(Lookup('p', /*use*/ false));
+
+    // Now try changing capacity up
+    EXPECT_TRUE(Lookup('m', /*use*/ true));
+    EXPECT_TRUE(Lookup('n', /*use*/ true));
+    shard_->SetCapacity(6);
+    EXPECT_OK(Insert('q', Cache::Priority::HIGH));
+    EXPECT_OK(Insert('r', Cache::Priority::HIGH));
+    EXPECT_OK(Insert('s', Cache::Priority::HIGH));
+    EXPECT_OK(Insert('t', Cache::Priority::HIGH));
+
+    EXPECT_FALSE(Lookup('o', /*use*/ false));
+    EXPECT_FALSE(Lookup('p', /*use*/ false));
+    EXPECT_TRUE(Lookup('m', /*use*/ false));
+    EXPECT_TRUE(Lookup('n', /*use*/ false));
+    EXPECT_TRUE(Lookup('q', /*use*/ false));
+    EXPECT_TRUE(Lookup('r', /*use*/ false));
+    EXPECT_TRUE(Lookup('s', /*use*/ false));
+    EXPECT_TRUE(Lookup('t', /*use*/ false));
+  }
+}
+
+void IncrementIntDeleter(const Slice& /*key*/, void* value) {
+  *reinterpret_cast<int*>(value) += 1;
+}
+
+// Testing calls to CorrectNearOverflow in Release
+TEST_F(ClockCacheTest, ClockCounterOverflowTest) {
+  NewShard(6, /*strict_capacity_limit*/ false);
+  ClockHandle* h;
+  int deleted = 0;
+  UniqueId64x2 hkey = TestHashedKey('x');
+  ASSERT_OK(shard_->Insert(TestKey(hkey), hkey, &deleted, 1,
+                           IncrementIntDeleter, &h, Cache::Priority::HIGH));
+
+  // Some large number outstanding
+  shard_->TEST_RefN(h, 123456789);
+  // Simulate many lookup/ref + release, plenty to overflow counters
+  for (int i = 0; i < 10000; ++i) {
+    shard_->TEST_RefN(h, 1234567);
+    shard_->TEST_ReleaseN(h, 1234567);
+  }
+  // Mark it invisible (to reach a different CorrectNearOverflow() in Release)
+  shard_->Erase(TestKey(hkey), hkey);
+  // Simulate many more lookup/ref + release (one-by-one would be too
+  // expensive for unit test)
+  for (int i = 0; i < 10000; ++i) {
+    shard_->TEST_RefN(h, 1234567);
+    shard_->TEST_ReleaseN(h, 1234567);
+  }
+  // Free all but last 1
+  shard_->TEST_ReleaseN(h, 123456789);
+  // Still alive
+  ASSERT_EQ(deleted, 0);
+  // Free last ref, which will finalize erasure
+  shard_->Release(h);
+  // Deleted
+  ASSERT_EQ(deleted, 1);
+}
+
+// This test is mostly to exercise some corner case logic, by forcing two
+// keys to have the same hash, and more
+TEST_F(ClockCacheTest, CollidingInsertEraseTest) {
+  NewShard(6, /*strict_capacity_limit*/ false);
+  int deleted = 0;
+  UniqueId64x2 hkey1 = TestHashedKey('x');
+  Slice key1 = TestKey(hkey1);
+  UniqueId64x2 hkey2 = TestHashedKey('y');
+  Slice key2 = TestKey(hkey2);
+  UniqueId64x2 hkey3 = TestHashedKey('z');
+  Slice key3 = TestKey(hkey3);
+  ClockHandle* h1;
+  ASSERT_OK(shard_->Insert(key1, hkey1, &deleted, 1, IncrementIntDeleter, &h1,
+                           Cache::Priority::HIGH));
+  ClockHandle* h2;
+  ASSERT_OK(shard_->Insert(key2, hkey2, &deleted, 1, IncrementIntDeleter, &h2,
+                           Cache::Priority::HIGH));
+  ClockHandle* h3;
+  ASSERT_OK(shard_->Insert(key3, hkey3, &deleted, 1, IncrementIntDeleter, &h3,
+                           Cache::Priority::HIGH));
+
+  // Can repeatedly lookup+release despite the hash collision
+  ClockHandle* tmp_h;
+  for (bool erase_if_last_ref : {true, false}) {  // but not last ref
+    tmp_h = shard_->Lookup(key1, hkey1);
+    ASSERT_EQ(h1, tmp_h);
+    ASSERT_FALSE(shard_->Release(tmp_h, erase_if_last_ref));
+
+    tmp_h = shard_->Lookup(key2, hkey2);
+    ASSERT_EQ(h2, tmp_h);
+    ASSERT_FALSE(shard_->Release(tmp_h, erase_if_last_ref));
+
+    tmp_h = shard_->Lookup(key3, hkey3);
+    ASSERT_EQ(h3, tmp_h);
+    ASSERT_FALSE(shard_->Release(tmp_h, erase_if_last_ref));
  }

-  // Fix the cache capacity, vary the element charge.
-  for (int i = 0; i < 1024; i++) {
-    capacity = 1024;
-    estimated_value_size = i;
-    metadata_charge_policy = kFullChargeCacheMetadata;
-    max_occupancy = CalcMaxOccupancy(capacity, estimated_value_size,
-                                     metadata_charge_policy);
-    hash_bits = CalcHashBitsWrapper(capacity, estimated_value_size,
-                                    metadata_charge_policy);
-    EXPECT_TRUE(TableSizeIsAppropriate(hash_bits, max_occupancy));
+  // Make h1 invisible
+  shard_->Erase(key1, hkey1);
+  // Redundant erase
+  shard_->Erase(key1, hkey1);
+
+  // All still alive
+  ASSERT_EQ(deleted, 0);
+
+  // Invisible to Lookup
+  tmp_h = shard_->Lookup(key1, hkey1);
+  ASSERT_EQ(nullptr, tmp_h);
+
+  // Can still find h2, h3
+  for (bool erase_if_last_ref : {true, false}) {  // but not last ref
+    tmp_h = shard_->Lookup(key2, hkey2);
+    ASSERT_EQ(h2, tmp_h);
+    ASSERT_FALSE(shard_->Release(tmp_h, erase_if_last_ref));
+
+    tmp_h = shard_->Lookup(key3, hkey3);
+    ASSERT_EQ(h3, tmp_h);
+    ASSERT_FALSE(shard_->Release(tmp_h, erase_if_last_ref));
  }

-  // Zero-capacity cache, and only values have charge.
-  capacity = 0;
-  estimated_value_size = 1;
-  metadata_charge_policy = kDontChargeCacheMetadata;
-  hash_bits = CalcHashBitsWrapper(capacity, estimated_value_size,
-                                  metadata_charge_policy);
-  EXPECT_TRUE(TableSizeIsAppropriate(hash_bits, 0 /* max_occupancy */));
+  // Also Insert with invisible entry there
+  ASSERT_OK(shard_->Insert(key1, hkey1, &deleted, 1, IncrementIntDeleter,
+                           nullptr, Cache::Priority::HIGH));
+  tmp_h = shard_->Lookup(key1, hkey1);
+  // Found but distinct handle
+  ASSERT_NE(nullptr, tmp_h);
+  ASSERT_NE(h1, tmp_h);
+  ASSERT_TRUE(shard_->Release(tmp_h, /*erase_if_last_ref*/ true));

-  // Zero-capacity cache, and only metadata has charge.
-  capacity = 0;
-  estimated_value_size = 0;
-  metadata_charge_policy = kFullChargeCacheMetadata;
-  hash_bits = CalcHashBitsWrapper(capacity, estimated_value_size,
-                                  metadata_charge_policy);
-  EXPECT_TRUE(TableSizeIsAppropriate(hash_bits, 0 /* max_occupancy */));
+  // tmp_h deleted
+  ASSERT_EQ(deleted--, 1);

-  // Small cache, large elements.
-  capacity = 1024;
-  estimated_value_size = 8192;
-  metadata_charge_policy = kFullChargeCacheMetadata;
-  hash_bits = CalcHashBitsWrapper(capacity, estimated_value_size,
-                                  metadata_charge_policy);
-  EXPECT_TRUE(TableSizeIsAppropriate(hash_bits, 0 /* max_occupancy */));
+  // Release last ref on h1 (already invisible)
+  ASSERT_TRUE(shard_->Release(h1, /*erase_if_last_ref*/ false));

-  // Large capacity.
-  capacity = 31924172;
-  estimated_value_size = 8192;
-  metadata_charge_policy = kFullChargeCacheMetadata;
-  max_occupancy =
-      CalcMaxOccupancy(capacity, estimated_value_size, metadata_charge_policy);
-  hash_bits = CalcHashBitsWrapper(capacity, estimated_value_size,
-                                  metadata_charge_policy);
-  EXPECT_TRUE(TableSizeIsAppropriate(hash_bits, max_occupancy));
+  // h1 deleted
+  ASSERT_EQ(deleted--, 1);
+  h1 = nullptr;
+
+  // Can still find h2, h3
+  for (bool erase_if_last_ref : {true, false}) {  // but not last ref
+    tmp_h = shard_->Lookup(key2, hkey2);
+    ASSERT_EQ(h2, tmp_h);
+    ASSERT_FALSE(shard_->Release(tmp_h, erase_if_last_ref));
+
+    tmp_h = shard_->Lookup(key3, hkey3);
+    ASSERT_EQ(h3, tmp_h);
+    ASSERT_FALSE(shard_->Release(tmp_h, erase_if_last_ref));
+  }
+
+  // Release last ref on h2
+  ASSERT_FALSE(shard_->Release(h2, /*erase_if_last_ref*/ false));
+
+  // h2 still not deleted (unreferenced in cache)
+  ASSERT_EQ(deleted, 0);
+
+  // Can still find it
+  tmp_h = shard_->Lookup(key2, hkey2);
+  ASSERT_EQ(h2, tmp_h);
+
+  // Release last ref on h2, with erase
+  ASSERT_TRUE(shard_->Release(h2, /*erase_if_last_ref*/ true));
+
+  // h2 deleted
+  ASSERT_EQ(deleted--, 1);
+  tmp_h = shard_->Lookup(key2, hkey2);
+  ASSERT_EQ(nullptr, tmp_h);
+
+  // Can still find h3
+  for (bool erase_if_last_ref : {true, false}) {  // but not last ref
+    tmp_h = shard_->Lookup(key3, hkey3);
+    ASSERT_EQ(h3, tmp_h);
+    ASSERT_FALSE(shard_->Release(tmp_h, erase_if_last_ref));
+  }
+
+  // Release last ref on h3, without erase
+  ASSERT_FALSE(shard_->Release(h3, /*erase_if_last_ref*/ false));
+
+  // h3 still not deleted (unreferenced in cache)
+  ASSERT_EQ(deleted, 0);
+
+  // Explicit erase
+  shard_->Erase(key3, hkey3);
+
+  // h3 deleted
+  ASSERT_EQ(deleted--, 1);
+  tmp_h = shard_->Lookup(key3, hkey3);
+  ASSERT_EQ(nullptr, tmp_h);
 }

-}  // namespace clock_cache
+// This uses the public API to effectively test CalcHashBits etc.
+TEST_F(ClockCacheTest, TableSizesTest) {
+  for (size_t est_val_size : {1U, 5U, 123U, 2345U, 345678U}) {
+    SCOPED_TRACE("est_val_size = " + std::to_string(est_val_size));
+    for (double est_count : {1.1, 2.2, 511.9, 512.1, 2345.0}) {
+      SCOPED_TRACE("est_count = " + std::to_string(est_count));
+      size_t capacity = static_cast<size_t>(est_val_size * est_count);
+      // kDontChargeCacheMetadata
+      auto cache = HyperClockCacheOptions(
+                       capacity, est_val_size, /*num shard_bits*/ -1,
+                       /*strict_capacity_limit*/ false,
+                       /*memory_allocator*/ nullptr, kDontChargeCacheMetadata)
+                       .MakeSharedCache();
+      // Table sizes are currently only powers of two
+      EXPECT_GE(cache->GetTableAddressCount(), est_count / kLoadFactor);
+      EXPECT_LE(cache->GetTableAddressCount(), est_count / kLoadFactor * 2.0);
+      EXPECT_EQ(cache->GetUsage(), 0);
+
+      // kFullChargeMetaData
+      // Because table sizes are currently only powers of two, sizes get
+      // really weird when metadata is a huge portion of capacity. For example,
+      // doubling the table size could cut by 90% the space available to
+      // values. Therefore, we omit those weird cases for now.
+      if (est_val_size >= 512) {
+        cache = HyperClockCacheOptions(
+                    capacity, est_val_size, /*num shard_bits*/ -1,
+                    /*strict_capacity_limit*/ false,
+                    /*memory_allocator*/ nullptr, kFullChargeCacheMetadata)
+                    .MakeSharedCache();
+        double est_count_after_meta =
+            (capacity - cache->GetUsage()) * 1.0 / est_val_size;
+        EXPECT_GE(cache->GetTableAddressCount(),
+                  est_count_after_meta / kLoadFactor);
+        EXPECT_LE(cache->GetTableAddressCount(),
+                  est_count_after_meta / kLoadFactor * 2.0);
+      }
+    }
+  }
+}
+
+}  // namespace hyper_clock_cache

 class TestSecondaryCache : public SecondaryCache {
 public:
@@ -742,7 +1057,7 @@ class TestSecondaryCache : public SecondaryCache {

  std::unique_ptr<SecondaryCacheResultHandle> Lookup(
      const Slice& key, const Cache::CreateCallback& create_cb, bool /*wait*/,
-      bool& is_in_sec_cache) override {
+      bool /*advise_erase*/, bool& is_in_sec_cache) override {
    std::string key_str = key.ToString();
    TEST_SYNC_POINT_CALLBACK("TestSecondaryCache::Lookup", &key_str);

@@ -780,6 +1095,8 @@ class TestSecondaryCache : public SecondaryCache {
    return secondary_handle;
  }

+  bool SupportForceErase() const override { return false; }
+
  void Erase(const Slice& /*key*/) override {}

  void WaitAll(std::vector<SecondaryCacheResultHandle*> handles) override {
@@ -960,7 +1277,7 @@ TEST_F(LRUCacheSecondaryCacheTest, BasicTest) {
  TestItem* item1 = new TestItem(str1.data(), str1.length());
  ASSERT_OK(cache->Insert(k1.AsSlice(), item1,
                          &LRUCacheSecondaryCacheTest::helper_, str1.length()));
-  std::string str2 = rnd.RandomString(1020);
+  std::string str2 = rnd.RandomString(1021);
  TestItem* item2 = new TestItem(str2.data(), str2.length());
  // k1 should be demoted to NVM
  ASSERT_OK(cache->Insert(k2.AsSlice(), item2,
@@ -1046,9 +1363,11 @@ TEST_F(LRUCacheSecondaryCacheTest, SaveFailTest) {
  std::string str2 = rnd.RandomString(1020);
  TestItem* item2 = new TestItem(str2.data(), str2.length());
  // k1 should be demoted to NVM
+  ASSERT_EQ(secondary_cache->num_inserts(), 0u);
  ASSERT_OK(cache->Insert(k2.AsSlice(), item2,
                          &LRUCacheSecondaryCacheTest::helper_fail_,
                          str2.length()));
+  ASSERT_EQ(secondary_cache->num_inserts(), 1u);

  Cache::Handle* handle;
  handle =
@@ -2378,6 +2697,7 @@ TEST_F(DBSecondaryCacheTest, TestSecondaryCacheOptionTwoDB) {
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -19,184 +19,49 @@

 namespace ROCKSDB_NAMESPACE {

-namespace {
-
-inline uint32_t HashSlice(const Slice& s) {
-  return Lower32of64(GetSliceNPHash64(s));
-}
-
-}  // namespace
-
-ShardedCache::ShardedCache(size_t capacity, int num_shard_bits,
-                           bool strict_capacity_limit,
-                           std::shared_ptr<MemoryAllocator> allocator)
+ShardedCacheBase::ShardedCacheBase(size_t capacity, int num_shard_bits,
+                                   bool strict_capacity_limit,
+                                   std::shared_ptr<MemoryAllocator> allocator)
    : Cache(std::move(allocator)),
+      last_id_(1),
      shard_mask_((uint32_t{1} << num_shard_bits) - 1),
-      capacity_(capacity),
      strict_capacity_limit_(strict_capacity_limit),
-      last_id_(1) {}
+      capacity_(capacity) {}

-void ShardedCache::SetCapacity(size_t capacity) {
+size_t ShardedCacheBase::ComputePerShardCapacity(size_t capacity) const {
  uint32_t num_shards = GetNumShards();
-  const size_t per_shard = (capacity + (num_shards - 1)) / num_shards;
-  MutexLock l(&capacity_mutex_);
-  for (uint32_t s = 0; s < num_shards; s++) {
-    GetShard(s)->SetCapacity(per_shard);
-  }
-  capacity_ = capacity;
+  return (capacity + (num_shards - 1)) / num_shards;
 }

-void ShardedCache::SetStrictCapacityLimit(bool strict_capacity_limit) {
-  uint32_t num_shards = GetNumShards();
-  MutexLock l(&capacity_mutex_);
-  for (uint32_t s = 0; s < num_shards; s++) {
-    GetShard(s)->SetStrictCapacityLimit(strict_capacity_limit);
-  }
-  strict_capacity_limit_ = strict_capacity_limit;
+size_t ShardedCacheBase::GetPerShardCapacity() const {
+  return ComputePerShardCapacity(GetCapacity());
 }

-Status ShardedCache::Insert(const Slice& key, void* value, size_t charge,
-                            DeleterFn deleter, Handle** handle,
-                            Priority priority) {
-  uint32_t hash = HashSlice(key);
-  return GetShard(Shard(hash))
-      ->Insert(key, hash, value, charge, deleter, handle, priority);
-}
-
-Status ShardedCache::Insert(const Slice& key, void* value,
-                            const CacheItemHelper* helper, size_t charge,
-                            Handle** handle, Priority priority) {
-  uint32_t hash = HashSlice(key);
-  if (!helper) {
-    return Status::InvalidArgument();
-  }
-  return GetShard(Shard(hash))
-      ->Insert(key, hash, value, helper, charge, handle, priority);
-}
-
-Cache::Handle* ShardedCache::Lookup(const Slice& key, Statistics* /*stats*/) {
-  uint32_t hash = HashSlice(key);
-  return GetShard(Shard(hash))->Lookup(key, hash);
-}
-
-Cache::Handle* ShardedCache::Lookup(const Slice& key,
-                                    const CacheItemHelper* helper,
-                                    const CreateCallback& create_cb,
-                                    Priority priority, bool wait,
-                                    Statistics* stats) {
-  uint32_t hash = HashSlice(key);
-  return GetShard(Shard(hash))
-      ->Lookup(key, hash, helper, create_cb, priority, wait, stats);
-}
-
-bool ShardedCache::IsReady(Handle* handle) {
-  uint32_t hash = GetHash(handle);
-  return GetShard(Shard(hash))->IsReady(handle);
-}
-
-void ShardedCache::Wait(Handle* handle) {
-  uint32_t hash = GetHash(handle);
-  GetShard(Shard(hash))->Wait(handle);
-}
-
-bool ShardedCache::Ref(Handle* handle) {
-  uint32_t hash = GetHash(handle);
-  return GetShard(Shard(hash))->Ref(handle);
-}
-
-bool ShardedCache::Release(Handle* handle, bool erase_if_last_ref) {
-  uint32_t hash = GetHash(handle);
-  return GetShard(Shard(hash))->Release(handle, erase_if_last_ref);
-}
-
-bool ShardedCache::Release(Handle* handle, bool useful,
-                           bool erase_if_last_ref) {
-  uint32_t hash = GetHash(handle);
-  return GetShard(Shard(hash))->Release(handle, useful, erase_if_last_ref);
-}
-
-void ShardedCache::Erase(const Slice& key) {
-  uint32_t hash = HashSlice(key);
-  GetShard(Shard(hash))->Erase(key, hash);
-}
-
-uint64_t ShardedCache::NewId() {
+uint64_t ShardedCacheBase::NewId() {
  return last_id_.fetch_add(1, std::memory_order_relaxed);
 }

-size_t ShardedCache::GetCapacity() const {
-  MutexLock l(&capacity_mutex_);
+size_t ShardedCacheBase::GetCapacity() const {
+  MutexLock l(&config_mutex_);
  return capacity_;
 }

-bool ShardedCache::HasStrictCapacityLimit() const {
-  MutexLock l(&capacity_mutex_);
+bool ShardedCacheBase::HasStrictCapacityLimit() const {
+  MutexLock l(&config_mutex_);
  return strict_capacity_limit_;
 }

-size_t ShardedCache::GetUsage() const {
-  // We will not lock the cache when getting the usage from shards.
-  uint32_t num_shards = GetNumShards();
-  size_t usage = 0;
-  for (uint32_t s = 0; s < num_shards; s++) {
-    usage += GetShard(s)->GetUsage();
-  }
-  return usage;
-}
-
-size_t ShardedCache::GetUsage(Handle* handle) const {
+size_t ShardedCacheBase::GetUsage(Handle* handle) const {
  return GetCharge(handle);
 }

-size_t ShardedCache::GetPinnedUsage() const {
-  // We will not lock the cache when getting the usage from shards.
-  uint32_t num_shards = GetNumShards();
-  size_t usage = 0;
-  for (uint32_t s = 0; s < num_shards; s++) {
-    usage += GetShard(s)->GetPinnedUsage();
-  }
-  return usage;
-}
-
-void ShardedCache::ApplyToAllEntries(
-    const std::function<void(const Slice& key, void* value, size_t charge,
-                             DeleterFn deleter)>& callback,
-    const ApplyToAllEntriesOptions& opts) {
-  uint32_t num_shards = GetNumShards();
-  // Iterate over part of each shard, rotating between shards, to
-  // minimize impact on latency of concurrent operations.
-  std::unique_ptr<uint32_t[]> states(new uint32_t[num_shards]{});
-
-  uint32_t aepl_in_32 = static_cast<uint32_t>(
-      std::min(size_t{UINT32_MAX}, opts.average_entries_per_lock));
-  aepl_in_32 = std::min(aepl_in_32, uint32_t{1});
-
-  bool remaining_work;
-  do {
-    remaining_work = false;
-    for (uint32_t s = 0; s < num_shards; s++) {
-      if (states[s] != UINT32_MAX) {
-        GetShard(s)->ApplyToSomeEntries(callback, aepl_in_32, &states[s]);
-        remaining_work |= states[s] != UINT32_MAX;
-      }
-    }
-  } while (remaining_work);
-}
-
-void ShardedCache::EraseUnRefEntries() {
-  uint32_t num_shards = GetNumShards();
-  for (uint32_t s = 0; s < num_shards; s++) {
-    GetShard(s)->EraseUnRefEntries();
-  }
-}
-
-std::string ShardedCache::GetPrintableOptions() const {
+std::string ShardedCacheBase::GetPrintableOptions() const {
  std::string ret;
  ret.reserve(20000);
  const int kBufferSize = 200;
  char buffer[kBufferSize];
  {
-    MutexLock l(&capacity_mutex_);
+    MutexLock l(&config_mutex_);
    snprintf(buffer, kBufferSize, "    capacity : %" ROCKSDB_PRIszt "\n",
             capacity_);
    ret.append(buffer);
@@ -210,12 +75,12 @@ std::string ShardedCache::GetPrintableOptions() const {
  snprintf(buffer, kBufferSize, "    memory_allocator : %s\n",
           memory_allocator() ? memory_allocator()->Name() : "None");
  ret.append(buffer);
-  ret.append(GetShard(0)->GetPrintableOptions());
+  AppendPrintableOptions(ret);
  return ret;
 }
-int GetDefaultCacheShardBits(size_t capacity) {
+
+int GetDefaultCacheShardBits(size_t capacity, size_t min_shard_size) {
  int num_shard_bits = 0;
-  size_t min_shard_size = 512L * 1024L;  // Every shard is at least 512KB.
  size_t num_shards = capacity / min_shard_size;
  while (num_shards >>= 1) {
    if (++num_shard_bits >= 6) {
@@ -226,8 +91,10 @@ int GetDefaultCacheShardBits(size_t capacity) {
  return num_shard_bits;
 }

-int ShardedCache::GetNumShardBits() const { return BitsSetToOne(shard_mask_); }
+int ShardedCacheBase::GetNumShardBits() const {
+  return BitsSetToOne(shard_mask_);
+}

-uint32_t ShardedCache::GetNumShards() const { return shard_mask_ + 1; }
+uint32_t ShardedCacheBase::GetNumShards() const { return shard_mask_ + 1; }

 }  // namespace ROCKSDB_NAMESPACE
@@ -10,123 +10,313 @@
 #pragma once

 #include <atomic>
+#include <cstdint>
 #include <string>

+#include "port/lang.h"
 #include "port/port.h"
 #include "rocksdb/cache.h"
+#include "util/hash.h"
+#include "util/mutexlock.h"

 namespace ROCKSDB_NAMESPACE {

-// Single cache shard interface.
-class CacheShard {
+// Optional base class for classes implementing the CacheShard concept
+class CacheShardBase {
 public:
-  CacheShard() = default;
-  virtual ~CacheShard() = default;
+  explicit CacheShardBase(CacheMetadataChargePolicy metadata_charge_policy)
+      : metadata_charge_policy_(metadata_charge_policy) {}

  using DeleterFn = Cache::DeleterFn;
-  virtual Status Insert(const Slice& key, uint32_t hash, void* value,
-                        size_t charge, DeleterFn deleter,
-                        Cache::Handle** handle, Cache::Priority priority) = 0;
-  virtual Status Insert(const Slice& key, uint32_t hash, void* value,
-                        const Cache::CacheItemHelper* helper, size_t charge,
-                        Cache::Handle** handle, Cache::Priority priority) = 0;
-  virtual Cache::Handle* Lookup(const Slice& key, uint32_t hash) = 0;
-  virtual Cache::Handle* Lookup(const Slice& key, uint32_t hash,
-                                const Cache::CacheItemHelper* helper,
-                                const Cache::CreateCallback& create_cb,
-                                Cache::Priority priority, bool wait,
-                                Statistics* stats) = 0;
-  virtual bool Release(Cache::Handle* handle, bool useful,
-                       bool erase_if_last_ref) = 0;
-  virtual bool IsReady(Cache::Handle* handle) = 0;
-  virtual void Wait(Cache::Handle* handle) = 0;
-  virtual bool Ref(Cache::Handle* handle) = 0;
-  virtual bool Release(Cache::Handle* handle, bool erase_if_last_ref) = 0;
-  virtual void Erase(const Slice& key, uint32_t hash) = 0;
-  virtual void SetCapacity(size_t capacity) = 0;
-  virtual void SetStrictCapacityLimit(bool strict_capacity_limit) = 0;
-  virtual size_t GetUsage() const = 0;
-  virtual size_t GetPinnedUsage() const = 0;
+
+  // Expected by concept CacheShard (TODO with C++20 support)
+  // Some Defaults
+  std::string GetPrintableOptions() const { return ""; }
+  using HashVal = uint64_t;
+  using HashCref = uint64_t;
+  static inline HashVal ComputeHash(const Slice& key) {
+    return GetSliceNPHash64(key);
+  }
+  static inline uint32_t HashPieceForSharding(HashCref hash) {
+    return Lower32of64(hash);
+  }
+  void AppendPrintableOptions(std::string& /*str*/) const {}
+
+  // Must be provided for concept CacheShard (TODO with C++20 support)
+  /*
+  struct HandleImpl {  // for concept HandleImpl
+    HashVal hash;
+    HashCref GetHash() const;
+    ...
+  };
+  Status Insert(const Slice& key, HashCref hash, void* value, size_t charge,
+                DeleterFn deleter, HandleImpl** handle,
+                Cache::Priority priority) = 0;
+  Status Insert(const Slice& key, HashCref hash, void* value,
+                const Cache::CacheItemHelper* helper, size_t charge,
+                HandleImpl** handle, Cache::Priority priority) = 0;
+  HandleImpl* Lookup(const Slice& key, HashCref hash) = 0;
+  HandleImpl* Lookup(const Slice& key, HashCref hash,
+                        const Cache::CacheItemHelper* helper,
+                        const Cache::CreateCallback& create_cb,
+                        Cache::Priority priority, bool wait,
+                        Statistics* stats) = 0;
+  bool Release(HandleImpl* handle, bool useful, bool erase_if_last_ref) = 0;
+  bool IsReady(HandleImpl* handle) = 0;
+  void Wait(HandleImpl* handle) = 0;
+  bool Ref(HandleImpl* handle) = 0;
+  void Erase(const Slice& key, HashCref hash) = 0;
+  void SetCapacity(size_t capacity) = 0;
+  void SetStrictCapacityLimit(bool strict_capacity_limit) = 0;
+  size_t GetUsage() const = 0;
+  size_t GetPinnedUsage() const = 0;
+  size_t GetOccupancyCount() const = 0;
+  size_t GetTableAddressCount() const = 0;
  // Handles iterating over roughly `average_entries_per_lock` entries, using
  // `state` to somehow record where it last ended up. Caller initially uses
-  // *state == 0 and implementation sets *state = UINT32_MAX to indicate
+  // *state == 0 and implementation sets *state = SIZE_MAX to indicate
  // completion.
-  virtual void ApplyToSomeEntries(
+  void ApplyToSomeEntries(
      const std::function<void(const Slice& key, void* value, size_t charge,
                               DeleterFn deleter)>& callback,
-      uint32_t average_entries_per_lock, uint32_t* state) = 0;
-  virtual void EraseUnRefEntries() = 0;
-  virtual std::string GetPrintableOptions() const { return ""; }
-  void set_metadata_charge_policy(
-      CacheMetadataChargePolicy metadata_charge_policy) {
-    metadata_charge_policy_ = metadata_charge_policy;
-  }
+      size_t average_entries_per_lock, size_t* state) = 0;
+  void EraseUnRefEntries() = 0;
+  */

 protected:
-  CacheMetadataChargePolicy metadata_charge_policy_ = kDontChargeCacheMetadata;
+  const CacheMetadataChargePolicy metadata_charge_policy_;
 };

-// Generic cache interface which shards cache by hash of keys. 2^num_shard_bits
-// shards will be created, with capacity split evenly to each of the shards.
-// Keys are sharded by the highest num_shard_bits bits of hash value.
-class ShardedCache : public Cache {
+// Portions of ShardedCache that do not depend on the template parameter
+class ShardedCacheBase : public Cache {
 public:
-  ShardedCache(size_t capacity, int num_shard_bits, bool strict_capacity_limit,
-               std::shared_ptr<MemoryAllocator> memory_allocator = nullptr);
-  virtual ~ShardedCache() = default;
-  virtual CacheShard* GetShard(uint32_t shard) = 0;
-  virtual const CacheShard* GetShard(uint32_t shard) const = 0;
-
-  virtual uint32_t GetHash(Handle* handle) const = 0;
-
-  virtual void SetCapacity(size_t capacity) override;
-  virtual void SetStrictCapacityLimit(bool strict_capacity_limit) override;
-
-  virtual Status Insert(const Slice& key, void* value, size_t charge,
-                        DeleterFn deleter, Handle** handle,
-                        Priority priority) override;
-  virtual Status Insert(const Slice& key, void* value,
-                        const CacheItemHelper* helper, size_t charge,
-                        Handle** handle = nullptr,
-                        Priority priority = Priority::LOW) override;
-  virtual Handle* Lookup(const Slice& key, Statistics* stats) override;
-  virtual Handle* Lookup(const Slice& key, const CacheItemHelper* helper,
-                         const CreateCallback& create_cb, Priority priority,
-                         bool wait, Statistics* stats = nullptr) override;
-  virtual bool Release(Handle* handle, bool useful,
-                       bool erase_if_last_ref = false) override;
-  virtual bool IsReady(Handle* handle) override;
-  virtual void Wait(Handle* handle) override;
-  virtual bool Ref(Handle* handle) override;
-  virtual bool Release(Handle* handle, bool erase_if_last_ref = false) override;
-  virtual void Erase(const Slice& key) override;
-  virtual uint64_t NewId() override;
-  virtual size_t GetCapacity() const override;
-  virtual bool HasStrictCapacityLimit() const override;
-  virtual size_t GetUsage() const override;
-  virtual size_t GetUsage(Handle* handle) const override;
-  virtual size_t GetPinnedUsage() const override;
-  virtual void ApplyToAllEntries(
-      const std::function<void(const Slice& key, void* value, size_t charge,
-                               DeleterFn deleter)>& callback,
-      const ApplyToAllEntriesOptions& opts) override;
-  virtual void EraseUnRefEntries() override;
-  virtual std::string GetPrintableOptions() const override;
+  ShardedCacheBase(size_t capacity, int num_shard_bits,
+                   bool strict_capacity_limit,
+                   std::shared_ptr<MemoryAllocator> memory_allocator);
+  virtual ~ShardedCacheBase() = default;

  int GetNumShardBits() const;
  uint32_t GetNumShards() const;

- protected:
-  inline uint32_t Shard(uint32_t hash) { return hash & shard_mask_; }
+  uint64_t NewId() override;

- private:
+  bool HasStrictCapacityLimit() const override;
+  size_t GetCapacity() const override;
+
+  using Cache::GetUsage;
+  size_t GetUsage(Handle* handle) const override;
+  std::string GetPrintableOptions() const override;
+
+ protected:  // fns
+  virtual void AppendPrintableOptions(std::string& str) const = 0;
+  size_t GetPerShardCapacity() const;
+  size_t ComputePerShardCapacity(size_t capacity) const;
+
+ protected:                        // data
+  std::atomic<uint64_t> last_id_;  // For NewId
  const uint32_t shard_mask_;
-  mutable port::Mutex capacity_mutex_;
-  size_t capacity_;
+
+  // Dynamic configuration parameters, guarded by config_mutex_
  bool strict_capacity_limit_;
-  std::atomic<uint64_t> last_id_;
+  size_t capacity_;
+  mutable port::Mutex config_mutex_;
 };

-extern int GetDefaultCacheShardBits(size_t capacity);
+// Generic cache interface that shards cache by hash of keys. 2^num_shard_bits
+// shards will be created, with capacity split evenly to each of the shards.
+// Keys are typically sharded by the lowest num_shard_bits bits of hash value
+// so that the upper bits of the hash value can keep a stable ordering of
+// table entries even as the table grows (using more upper hash bits).
+// See CacheShardBase above for what is expected of the CacheShard parameter.
+template <class CacheShard>
+class ShardedCache : public ShardedCacheBase {
+ public:
+  using HashVal = typename CacheShard::HashVal;
+  using HashCref = typename CacheShard::HashCref;
+  using HandleImpl = typename CacheShard::HandleImpl;
+
+  ShardedCache(size_t capacity, int num_shard_bits, bool strict_capacity_limit,
+               std::shared_ptr<MemoryAllocator> allocator)
+      : ShardedCacheBase(capacity, num_shard_bits, strict_capacity_limit,
+                         allocator),
+        shards_(reinterpret_cast<CacheShard*>(port::cacheline_aligned_alloc(
+            sizeof(CacheShard) * GetNumShards()))),
+        destroy_shards_in_dtor_(false) {}
+
+  virtual ~ShardedCache() {
+    if (destroy_shards_in_dtor_) {
+      ForEachShard([](CacheShard* cs) { cs->~CacheShard(); });
+    }
+    port::cacheline_aligned_free(shards_);
+  }
+
+  CacheShard& GetShard(HashCref hash) {
+    return shards_[CacheShard::HashPieceForSharding(hash) & shard_mask_];
+  }
+
+  const CacheShard& GetShard(HashCref hash) const {
+    return shards_[CacheShard::HashPieceForSharding(hash) & shard_mask_];
+  }
+
+  void SetCapacity(size_t capacity) override {
+    MutexLock l(&config_mutex_);
+    capacity_ = capacity;
+    auto per_shard = ComputePerShardCapacity(capacity);
+    ForEachShard([=](CacheShard* cs) { cs->SetCapacity(per_shard); });
+  }
+
+  void SetStrictCapacityLimit(bool s_c_l) override {
+    MutexLock l(&config_mutex_);
+    strict_capacity_limit_ = s_c_l;
+    ForEachShard(
+        [s_c_l](CacheShard* cs) { cs->SetStrictCapacityLimit(s_c_l); });
+  }
+
+  Status Insert(const Slice& key, void* value, size_t charge, DeleterFn deleter,
+                Handle** handle, Priority priority) override {
+    HashVal hash = CacheShard::ComputeHash(key);
+    auto h_out = reinterpret_cast<HandleImpl**>(handle);
+    return GetShard(hash).Insert(key, hash, value, charge, deleter, h_out,
+                                 priority);
+  }
+  Status Insert(const Slice& key, void* value, const CacheItemHelper* helper,
+                size_t charge, Handle** handle = nullptr,
+                Priority priority = Priority::LOW) override {
+    if (!helper) {
+      return Status::InvalidArgument();
+    }
+    HashVal hash = CacheShard::ComputeHash(key);
+    auto h_out = reinterpret_cast<HandleImpl**>(handle);
+    return GetShard(hash).Insert(key, hash, value, helper, charge, h_out,
+                                 priority);
+  }
+
+  Handle* Lookup(const Slice& key, Statistics* /*stats*/) override {
+    HashVal hash = CacheShard::ComputeHash(key);
+    HandleImpl* result = GetShard(hash).Lookup(key, hash);
+    return reinterpret_cast<Handle*>(result);
+  }
+  Handle* Lookup(const Slice& key, const CacheItemHelper* helper,
+                 const CreateCallback& create_cb, Priority priority, bool wait,
+                 Statistics* stats = nullptr) override {
+    HashVal hash = CacheShard::ComputeHash(key);
+    HandleImpl* result = GetShard(hash).Lookup(key, hash, helper, create_cb,
+                                               priority, wait, stats);
+    return reinterpret_cast<Handle*>(result);
+  }
+
+  void Erase(const Slice& key) override {
+    HashVal hash = CacheShard::ComputeHash(key);
+    GetShard(hash).Erase(key, hash);
+  }
+
+  bool Release(Handle* handle, bool useful,
+               bool erase_if_last_ref = false) override {
+    auto h = reinterpret_cast<HandleImpl*>(handle);
+    return GetShard(h->GetHash()).Release(h, useful, erase_if_last_ref);
+  }
+  bool IsReady(Handle* handle) override {
+    auto h = reinterpret_cast<HandleImpl*>(handle);
+    return GetShard(h->GetHash()).IsReady(h);
+  }
+  void Wait(Handle* handle) override {
+    auto h = reinterpret_cast<HandleImpl*>(handle);
+    GetShard(h->GetHash()).Wait(h);
+  }
+  bool Ref(Handle* handle) override {
+    auto h = reinterpret_cast<HandleImpl*>(handle);
+    return GetShard(h->GetHash()).Ref(h);
+  }
+  bool Release(Handle* handle, bool erase_if_last_ref = false) override {
+    return Release(handle, true /*useful*/, erase_if_last_ref);
+  }
+  using ShardedCacheBase::GetUsage;
+  size_t GetUsage() const override {
+    return SumOverShards2(&CacheShard::GetUsage);
+  }
+  size_t GetPinnedUsage() const override {
+    return SumOverShards2(&CacheShard::GetPinnedUsage);
+  }
+  size_t GetOccupancyCount() const override {
+    return SumOverShards2(&CacheShard::GetPinnedUsage);
+  }
+  size_t GetTableAddressCount() const override {
+    return SumOverShards2(&CacheShard::GetTableAddressCount);
+  }
+  void ApplyToAllEntries(
+      const std::function<void(const Slice& key, void* value, size_t charge,
+                               DeleterFn deleter)>& callback,
+      const ApplyToAllEntriesOptions& opts) override {
+    uint32_t num_shards = GetNumShards();
+    // Iterate over part of each shard, rotating between shards, to
+    // minimize impact on latency of concurrent operations.
+    std::unique_ptr<size_t[]> states(new size_t[num_shards]{});
+
+    size_t aepl = opts.average_entries_per_lock;
+    aepl = std::min(aepl, size_t{1});
+
+    bool remaining_work;
+    do {
+      remaining_work = false;
+      for (uint32_t i = 0; i < num_shards; i++) {
+        if (states[i] != SIZE_MAX) {
+          shards_[i].ApplyToSomeEntries(callback, aepl, &states[i]);
+          remaining_work |= states[i] != SIZE_MAX;
+        }
+      }
+    } while (remaining_work);
+  }
+
+  virtual void EraseUnRefEntries() override {
+    ForEachShard([](CacheShard* cs) { cs->EraseUnRefEntries(); });
+  }
+
+  void DisownData() override {
+    // Leak data only if that won't generate an ASAN/valgrind warning.
+    if (!kMustFreeHeapAllocations) {
+      destroy_shards_in_dtor_ = false;
+    }
+  }
+
+ protected:
+  inline void ForEachShard(const std::function<void(CacheShard*)>& fn) {
+    uint32_t num_shards = GetNumShards();
+    for (uint32_t i = 0; i < num_shards; i++) {
+      fn(shards_ + i);
+    }
+  }
+
+  inline size_t SumOverShards(
+      const std::function<size_t(CacheShard&)>& fn) const {
+    uint32_t num_shards = GetNumShards();
+    size_t result = 0;
+    for (uint32_t i = 0; i < num_shards; i++) {
+      result += fn(shards_[i]);
+    }
+    return result;
+  }
+
+  inline size_t SumOverShards2(size_t (CacheShard::*fn)() const) const {
+    return SumOverShards([fn](CacheShard& cs) { return (cs.*fn)(); });
+  }
+
+  // Must be called exactly once by derived class constructor
+  void InitShards(const std::function<void(CacheShard*)>& placement_new) {
+    ForEachShard(placement_new);
+    destroy_shards_in_dtor_ = true;
+  }
+
+  void AppendPrintableOptions(std::string& str) const override {
+    shards_[0].AppendPrintableOptions(str);
+  }
+
+ private:
+  CacheShard* const shards_;
+  bool destroy_shards_in_dtor_;
+};
+
+// 512KB is traditional minimum shard size.
+int GetDefaultCacheShardBits(size_t capacity,
+                             size_t min_shard_size = 512U * 1024U);

 }  // namespace ROCKSDB_NAMESPACE
@@ -47,35 +47,39 @@ def parse_gcov_report(gcov_input):

    return per_file_coverage, total_coverage

+
 def get_option_parser():
-    usage = "Parse the gcov output and generate more human-readable code " +\
-            "coverage report."
+    usage = (
+        "Parse the gcov output and generate more human-readable code "
+        + "coverage report."
+    )
    parser = optparse.OptionParser(usage)

    parser.add_option(
-        "--interested-files", "-i",
+        "--interested-files",
+        "-i",
        dest="filenames",
-        help="Comma separated files names. if specified, we will display " +
-             "the coverage report only for interested source files. " +
-             "Otherwise we will display the coverage report for all " +
-             "source files."
+        help="Comma separated files names. if specified, we will display "
+        + "the coverage report only for interested source files. "
+        + "Otherwise we will display the coverage report for all "
+        + "source files.",
    )
    return parser

+
 def display_file_coverage(per_file_coverage, total_coverage):
    # To print out auto-adjustable column, we need to know the longest
    # length of file names.
-    max_file_name_length = max(
-        len(fname) for fname in per_file_coverage.keys()
-    )
+    max_file_name_length = max(len(fname) for fname in per_file_coverage.keys())

    # -- Print header
    # size of separator is determined by 3 column sizes:
    # file name, coverage percentage and lines.
-    header_template = \
-        "%" + str(max_file_name_length) + "s\t%s\t%s"
+    header_template = "%" + str(max_file_name_length) + "s\t%s\t%s"
    separator = "-" * (max_file_name_length + 10 + 20)
-    print(header_template % ("Filename", "Coverage", "Lines"))  # noqa: E999 T25377293 Grandfathered in
+    print(
+        header_template % ("Filename", "Coverage", "Lines")
+    )  # noqa: E999 T25377293 Grandfathered in
    print(separator)

    # -- Print body
@@ -91,13 +95,14 @@ def display_file_coverage(per_file_coverage, total_coverage):
        print(separator)
        print(record_template % ("Total", total_coverage[0], total_coverage[1]))

+
 def report_coverage():
    parser = get_option_parser()
    (options, args) = parser.parse_args()

    interested_files = set()
    if options.filenames is not None:
-        interested_files = set(f.strip() for f in options.filenames.split(','))
+        interested_files = {f.strip() for f in options.filenames.split(",")}

    # To make things simple, right now we only read gcov report from the input
    per_file_coverage, total_coverage = parse_gcov_report(sys.stdin)
@@ -105,7 +110,8 @@ def report_coverage():
    # Check if we need to display coverage info for interested files.
    if len(interested_files):
        per_file_coverage = dict(
-            (fname, per_file_coverage[fname]) for fname in interested_files
+            (fname, per_file_coverage[fname])
+            for fname in interested_files
            if fname in per_file_coverage
        )
        # If we only interested in several files, it makes no sense to report
@@ -117,5 +123,6 @@ def report_coverage():
        return
    display_file_coverage(per_file_coverage, total_coverage)

+
 if __name__ == "__main__":
    report_coverage()
@@ -78,7 +78,7 @@ blackbox_crash_test_with_multiops_wp_txn: $(DB_STRESS_CMD)
 	$(CRASHTEST_PY) --test_multiops_txn --write_policy write_prepared blackbox $(CRASH_TEST_EXT_ARGS)

 blackbox_crash_test_with_tiered_storage: $(DB_STRESS_CMD)
-	$(CRASHTEST_PY) --enable_tiered_storage blackbox $(CRASH_TEST_EXT_ARGS)
+	$(CRASHTEST_PY) --test_tiered_storage blackbox $(CRASH_TEST_EXT_ARGS)

 ifeq ($(CRASH_TEST_KILL_ODD),)
  CRASH_TEST_KILL_ODD=888887
@@ -103,5 +103,5 @@ whitebox_crash_test_with_ts: $(DB_STRESS_CMD)
      $(CRASH_TEST_KILL_ODD) $(CRASH_TEST_EXT_ARGS)

 whitebox_crash_test_with_tiered_storage: $(DB_STRESS_CMD)
-	$(CRASHTEST_PY) --enable_tiered_storage whitebox --random_kill_odd \
+	$(CRASHTEST_PY) --test_tiered_storage whitebox --random_kill_odd \
      $(CRASH_TEST_KILL_ODD) $(CRASH_TEST_EXT_ARGS)
@@ -45,6 +45,7 @@ void ArenaWrappedDBIter::Init(
  sv_number_ = version_number;
  read_options_ = read_options;
  allow_refresh_ = allow_refresh;
+  memtable_range_tombstone_iter_ = nullptr;
 }

 Status ArenaWrappedDBIter::Refresh() {
@@ -58,41 +59,68 @@ Status ArenaWrappedDBIter::Refresh() {
  uint64_t cur_sv_number = cfd_->GetSuperVersionNumber();
  TEST_SYNC_POINT("ArenaWrappedDBIter::Refresh:1");
  TEST_SYNC_POINT("ArenaWrappedDBIter::Refresh:2");
+  auto reinit_internal_iter = [&]() {
+    Env* env = db_iter_->env();
+    db_iter_->~DBIter();
+    arena_.~Arena();
+    new (&arena_) Arena();
+
+    SuperVersion* sv = cfd_->GetReferencedSuperVersion(db_impl_);
+    SequenceNumber latest_seq = db_impl_->GetLatestSequenceNumber();
+    if (read_callback_) {
+      read_callback_->Refresh(latest_seq);
+    }
+    Init(env, read_options_, *(cfd_->ioptions()), sv->mutable_cf_options,
+         sv->current, latest_seq,
+         sv->mutable_cf_options.max_sequential_skip_in_iterations,
+         cur_sv_number, read_callback_, db_impl_, cfd_, expose_blob_index_,
+         allow_refresh_);
+
+    InternalIterator* internal_iter = db_impl_->NewInternalIterator(
+        read_options_, cfd_, sv, &arena_, latest_seq,
+        /* allow_unprepared_value */ true, /* db_iter */ this);
+    SetIterUnderDBIter(internal_iter);
+  };
  while (true) {
    if (sv_number_ != cur_sv_number) {
-      Env* env = db_iter_->env();
-      db_iter_->~DBIter();
-      arena_.~Arena();
-      new (&arena_) Arena();
-
-      SuperVersion* sv = cfd_->GetReferencedSuperVersion(db_impl_);
-      SequenceNumber latest_seq = db_impl_->GetLatestSequenceNumber();
-      if (read_callback_) {
-        read_callback_->Refresh(latest_seq);
-      }
-      Init(env, read_options_, *(cfd_->ioptions()), sv->mutable_cf_options,
-           sv->current, latest_seq,
-           sv->mutable_cf_options.max_sequential_skip_in_iterations,
-           cur_sv_number, read_callback_, db_impl_, cfd_, expose_blob_index_,
-           allow_refresh_);
-
-      InternalIterator* internal_iter = db_impl_->NewInternalIterator(
-          read_options_, cfd_, sv, &arena_, db_iter_->GetRangeDelAggregator(),
-          latest_seq, /* allow_unprepared_value */ true);
-      SetIterUnderDBIter(internal_iter);
+      reinit_internal_iter();
      break;
    } else {
      SequenceNumber latest_seq = db_impl_->GetLatestSequenceNumber();
      // Refresh range-tombstones in MemTable
      if (!read_options_.ignore_range_deletions) {
        SuperVersion* sv = cfd_->GetThreadLocalSuperVersion(db_impl_);
-        ReadRangeDelAggregator* range_del_agg =
-            db_iter_->GetRangeDelAggregator();
-        std::unique_ptr<FragmentedRangeTombstoneIterator> range_del_iter;
-        range_del_iter.reset(sv->mem->NewRangeTombstoneIterator(
-            read_options_, latest_seq, false /* immutable_memtable */));
-        range_del_agg->AddTombstones(std::move(range_del_iter));
-        cfd_->ReturnThreadLocalSuperVersion(sv);
+        TEST_SYNC_POINT_CALLBACK("ArenaWrappedDBIter::Refresh:SV", nullptr);
+        auto t = sv->mem->NewRangeTombstoneIterator(
+            read_options_, latest_seq, false /* immutable_memtable */);
+        if (!t || t->empty()) {
+          // If memtable_range_tombstone_iter_ points to a non-empty tombstone
+          // iterator, then it means sv->mem is not the memtable that
+          // memtable_range_tombstone_iter_ points to, so SV must have changed
+          // after the sv_number_ != cur_sv_number check above. We will fall
+          // back to re-init the InternalIterator, and the tombstone iterator
+          // will be freed during db_iter destruction there.
+          if (memtable_range_tombstone_iter_) {
+            assert(!*memtable_range_tombstone_iter_ ||
+                   sv_number_ != cfd_->GetSuperVersionNumber());
+          }
+          delete t;
+        } else {  // current mutable memtable has range tombstones
+          if (!memtable_range_tombstone_iter_) {
+            delete t;
+            db_impl_->ReturnAndCleanupSuperVersion(cfd_, sv);
+            // The memtable under DBIter did not have range tombstone before
+            // refresh.
+            reinit_internal_iter();
+            break;
+          } else {
+            delete *memtable_range_tombstone_iter_;
+            *memtable_range_tombstone_iter_ = new TruncatedRangeDelIterator(
+                std::unique_ptr<FragmentedRangeTombstoneIterator>(t),
+                &cfd_->internal_comparator(), nullptr, nullptr);
+          }
+        }
+        db_impl_->ReturnAndCleanupSuperVersion(cfd_, sv);
      }
      // Refresh latest sequence number
      db_iter_->set_sequence(latest_seq);
@@ -44,9 +44,7 @@ class ArenaWrappedDBIter : public Iterator {
  // Get the arena to be used to allocate memory for DBIter to be wrapped,
  // as well as child iterators in it.
  virtual Arena* GetArena() { return &arena_; }
-  virtual ReadRangeDelAggregator* GetRangeDelAggregator() {
-    return db_iter_->GetRangeDelAggregator();
-  }
+
  const ReadOptions& GetReadOptions() { return read_options_; }

  // Set the internal iterator wrapped inside the DB Iterator. Usually it is
@@ -55,6 +53,10 @@ class ArenaWrappedDBIter : public Iterator {
    db_iter_->SetIter(iter);
  }

+  void SetMemtableRangetombstoneIter(TruncatedRangeDelIterator** iter) {
+    memtable_range_tombstone_iter_ = iter;
+  }
+
  bool Valid() const override { return db_iter_->Valid(); }
  void SeekToFirst() override { db_iter_->SeekToFirst(); }
  void SeekToLast() override { db_iter_->SeekToLast(); }
@@ -68,6 +70,7 @@ class ArenaWrappedDBIter : public Iterator {
  void Prev() override { db_iter_->Prev(); }
  Slice key() const override { return db_iter_->key(); }
  Slice value() const override { return db_iter_->value(); }
+  const WideColumns& columns() const override { return db_iter_->columns(); }
  Status status() const override { return db_iter_->status(); }
  Slice timestamp() const override { return db_iter_->timestamp(); }
  bool IsBlob() const { return db_iter_->IsBlob(); }
@@ -104,6 +107,9 @@ class ArenaWrappedDBIter : public Iterator {
  ReadCallback* read_callback_;
  bool expose_blob_index_ = false;
  bool allow_refresh_ = true;
+  // If this is nullptr, it means the mutable memtable does not contain range
+  // tombstone when added under this DBIter.
+  TruncatedRangeDelIterator** memtable_range_tombstone_iter_ = nullptr;
 };

 // Generate the arena wrapped iterator class.
@@ -0,0 +1,90 @@
+//  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/blob/blob_contents.h"
+
+#include <cassert>
+
+#include "cache/cache_entry_roles.h"
+#include "cache/cache_helpers.h"
+#include "port/malloc.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+std::unique_ptr<BlobContents> BlobContents::Create(
+    CacheAllocationPtr&& allocation, size_t size) {
+  return std::unique_ptr<BlobContents>(
+      new BlobContents(std::move(allocation), size));
+}
+
+size_t BlobContents::ApproximateMemoryUsage() const {
+  size_t usage = 0;
+
+  if (allocation_) {
+    MemoryAllocator* const allocator = allocation_.get_deleter().allocator;
+
+    if (allocator) {
+      usage += allocator->UsableSize(allocation_.get(), data_.size());
+    } else {
+#ifdef ROCKSDB_MALLOC_USABLE_SIZE
+      usage += malloc_usable_size(allocation_.get());
+#else
+      usage += data_.size();
+#endif
+    }
+  }
+
+#ifdef ROCKSDB_MALLOC_USABLE_SIZE
+  usage += malloc_usable_size(const_cast<BlobContents*>(this));
+#else
+  usage += sizeof(*this);
+#endif
+
+  return usage;
+}
+
+size_t BlobContents::SizeCallback(void* obj) {
+  assert(obj);
+
+  return static_cast<const BlobContents*>(obj)->size();
+}
+
+Status BlobContents::SaveToCallback(void* from_obj, size_t from_offset,
+                                    size_t length, void* out) {
+  assert(from_obj);
+
+  const BlobContents* buf = static_cast<const BlobContents*>(from_obj);
+  assert(buf->size() >= from_offset + length);
+
+  memcpy(out, buf->data().data() + from_offset, length);
+
+  return Status::OK();
+}
+
+Cache::CacheItemHelper* BlobContents::GetCacheItemHelper() {
+  static Cache::CacheItemHelper cache_helper(
+      &SizeCallback, &SaveToCallback,
+      GetCacheEntryDeleterForRole<BlobContents, CacheEntryRole::kBlobValue>());
+
+  return &cache_helper;
+}
+
+Status BlobContents::CreateCallback(CacheAllocationPtr&& allocation,
+                                    const void* buf, size_t size,
+                                    void** out_obj, size_t* charge) {
+  assert(allocation);
+
+  memcpy(allocation.get(), buf, size);
+
+  std::unique_ptr<BlobContents> obj = Create(std::move(allocation), size);
+  BlobContents* const contents = obj.release();
+
+  *out_obj = contents;
+  *charge = contents->ApproximateMemoryUsage();
+
+  return Status::OK();
+}
+
+}  // namespace ROCKSDB_NAMESPACE
@@ -0,0 +1,56 @@
+//  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 <memory>
+
+#include "memory/memory_allocator.h"
+#include "rocksdb/cache.h"
+#include "rocksdb/rocksdb_namespace.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/status.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// A class representing a single uncompressed value read from a blob file.
+class BlobContents {
+ public:
+  static std::unique_ptr<BlobContents> Create(CacheAllocationPtr&& allocation,
+                                              size_t size);
+
+  BlobContents(const BlobContents&) = delete;
+  BlobContents& operator=(const BlobContents&) = delete;
+
+  BlobContents(BlobContents&&) = default;
+  BlobContents& operator=(BlobContents&&) = default;
+
+  ~BlobContents() = default;
+
+  const Slice& data() const { return data_; }
+  size_t size() const { return data_.size(); }
+
+  size_t ApproximateMemoryUsage() const;
+
+  // Callbacks for secondary cache
+  static size_t SizeCallback(void* obj);
+
+  static Status SaveToCallback(void* from_obj, size_t from_offset,
+                               size_t length, void* out);
+
+  static Cache::CacheItemHelper* GetCacheItemHelper();
+
+  static Status CreateCallback(CacheAllocationPtr&& allocation, const void* buf,
+                               size_t size, void** out_obj, size_t* charge);
+
+ private:
+  BlobContents(CacheAllocationPtr&& allocation, size_t size)
+      : allocation_(std::move(allocation)), data_(allocation_.get(), size) {}
+
+  CacheAllocationPtr allocation_;
+  Slice data_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
@@ -321,6 +321,7 @@ TEST(BlobCountingIteratorTest, CorruptBlobIndex) {
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -205,6 +205,7 @@ TEST_F(BlobFileAdditionTest, ForwardIncompatibleCustomField) {
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -7,6 +7,7 @@

 #include <cassert>

+#include "db/blob/blob_contents.h"
 #include "db/blob/blob_file_addition.h"
 #include "db/blob/blob_file_completion_callback.h"
 #include "db/blob/blob_index.h"
@@ -408,16 +409,28 @@ Status BlobFileBuilder::PutBlobIntoCacheIfNeeded(const Slice& blob,

    // Objects to be put into the cache have to be heap-allocated and
    // self-contained, i.e. own their contents. The Cache has to be able to
-    // take unique ownership of them. Therefore, we copy the blob into a
-    // string directly, and insert that into the cache.
-    std::unique_ptr<std::string> buf = std::make_unique<std::string>();
-    buf->assign(blob.data(), blob.size());
+    // take unique ownership of them.
+    CacheAllocationPtr allocation =
+        AllocateBlock(blob.size(), blob_cache->memory_allocator());
+    memcpy(allocation.get(), blob.data(), blob.size());
+    std::unique_ptr<BlobContents> buf =
+        BlobContents::Create(std::move(allocation), blob.size());
+
+    Cache::CacheItemHelper* const cache_item_helper =
+        BlobContents::GetCacheItemHelper();
+    assert(cache_item_helper);
+
+    if (immutable_options_->lowest_used_cache_tier ==
+        CacheTier::kNonVolatileBlockTier) {
+      s = blob_cache->Insert(key, buf.get(), cache_item_helper,
+                             buf->ApproximateMemoryUsage(),
+                             nullptr /* cache_handle */, priority);
+    } else {
+      s = blob_cache->Insert(key, buf.get(), buf->ApproximateMemoryUsage(),
+                             cache_item_helper->del_cb,
+                             nullptr /* cache_handle */, priority);
+    }

-    // TODO: support custom allocators and provide a better estimated memory
-    // usage using malloc_usable_size.
-    s = blob_cache->Insert(key, buf.get(), buf->size(),
-                           &DeleteCacheEntry<std::string>,
-                           nullptr /* cache_handle */, priority);
    if (s.ok()) {
      RecordTick(statistics, BLOB_DB_CACHE_ADD);
      RecordTick(statistics, BLOB_DB_CACHE_BYTES_WRITE, buf->size());
@@ -674,6 +674,7 @@ TEST_P(BlobFileBuilderIOErrorTest, IOError) {
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -263,6 +263,7 @@ TEST_F(BlobFileCacheTest, GetBlobFileReader_CacheFull) {
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -168,6 +168,7 @@ TEST_F(BlobFileGarbageTest, ForwardIncompatibleCustomField) {
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -8,6 +8,7 @@
 #include <cassert>
 #include <string>

+#include "db/blob/blob_contents.h"
 #include "db/blob/blob_log_format.h"
 #include "file/file_prefetch_buffer.h"
 #include "file/filename.h"
@@ -283,14 +284,12 @@ BlobFileReader::BlobFileReader(

 BlobFileReader::~BlobFileReader() = default;

-Status BlobFileReader::GetBlob(const ReadOptions& read_options,
-                               const Slice& user_key, uint64_t offset,
-                               uint64_t value_size,
-                               CompressionType compression_type,
-                               FilePrefetchBuffer* prefetch_buffer,
-                               PinnableSlice* value,
-                               uint64_t* bytes_read) const {
-  assert(value);
+Status BlobFileReader::GetBlob(
+    const ReadOptions& read_options, const Slice& user_key, uint64_t offset,
+    uint64_t value_size, CompressionType compression_type,
+    FilePrefetchBuffer* prefetch_buffer, MemoryAllocator* allocator,
+    std::unique_ptr<BlobContents>* result, uint64_t* bytes_read) const {
+  assert(result);

  const uint64_t key_size = user_key.size();

@@ -361,8 +360,8 @@ Status BlobFileReader::GetBlob(const ReadOptions& read_options,
  const Slice value_slice(record_slice.data() + adjustment, value_size);

  {
-    const Status s = UncompressBlobIfNeeded(value_slice, compression_type,
-                                            clock_, statistics_, value);
+    const Status s = UncompressBlobIfNeeded(
+        value_slice, compression_type, allocator, clock_, statistics_, result);
    if (!s.ok()) {
      return s;
    }
@@ -375,16 +374,18 @@ Status BlobFileReader::GetBlob(const ReadOptions& read_options,
  return Status::OK();
 }

-void BlobFileReader::MultiGetBlob(const ReadOptions& read_options,
-                                  autovector<BlobReadRequest*>& blob_reqs,
-                                  uint64_t* bytes_read) const {
+void BlobFileReader::MultiGetBlob(
+    const ReadOptions& read_options, MemoryAllocator* allocator,
+    autovector<std::pair<BlobReadRequest*, std::unique_ptr<BlobContents>>>&
+        blob_reqs,
+    uint64_t* bytes_read) const {
  const size_t num_blobs = blob_reqs.size();
  assert(num_blobs > 0);
  assert(num_blobs <= MultiGetContext::MAX_BATCH_SIZE);

 #ifndef NDEBUG
  for (size_t i = 0; i < num_blobs - 1; ++i) {
-    assert(blob_reqs[i]->offset <= blob_reqs[i + 1]->offset);
+    assert(blob_reqs[i].first->offset <= blob_reqs[i + 1].first->offset);
  }
 #endif  // !NDEBUG

@@ -393,16 +394,21 @@ void BlobFileReader::MultiGetBlob(const ReadOptions& read_options,
  uint64_t total_len = 0;
  read_reqs.reserve(num_blobs);
  for (size_t i = 0; i < num_blobs; ++i) {
-    const size_t key_size = blob_reqs[i]->user_key->size();
-    const uint64_t offset = blob_reqs[i]->offset;
-    const uint64_t value_size = blob_reqs[i]->len;
+    BlobReadRequest* const req = blob_reqs[i].first;
+    assert(req);
+    assert(req->user_key);
+    assert(req->status);
+
+    const size_t key_size = req->user_key->size();
+    const uint64_t offset = req->offset;
+    const uint64_t value_size = req->len;

    if (!IsValidBlobOffset(offset, key_size, value_size, file_size_)) {
-      *blob_reqs[i]->status = Status::Corruption("Invalid blob offset");
+      *req->status = Status::Corruption("Invalid blob offset");
      continue;
    }
-    if (blob_reqs[i]->compression != compression_type_) {
-      *blob_reqs[i]->status =
+    if (req->compression != compression_type_) {
+      *req->status =
          Status::Corruption("Compression type mismatch when reading a blob");
      continue;
    }
@@ -411,12 +417,12 @@ void BlobFileReader::MultiGetBlob(const ReadOptions& read_options,
        read_options.verify_checksums
            ? BlobLogRecord::CalculateAdjustmentForRecordHeader(key_size)
            : 0;
-    assert(blob_reqs[i]->offset >= adjustment);
+    assert(req->offset >= adjustment);
    adjustments.push_back(adjustment);

    FSReadRequest read_req = {};
-    read_req.offset = blob_reqs[i]->offset - adjustment;
-    read_req.len = blob_reqs[i]->len + adjustment;
+    read_req.offset = req->offset - adjustment;
+    read_req.len = req->len + adjustment;
    read_reqs.emplace_back(read_req);
    total_len += read_req.len;
  }
@@ -450,8 +456,11 @@ void BlobFileReader::MultiGetBlob(const ReadOptions& read_options,
    for (auto& req : read_reqs) {
      req.status.PermitUncheckedError();
    }
-    for (auto& req : blob_reqs) {
+    for (auto& blob_req : blob_reqs) {
+      BlobReadRequest* const req = blob_req.first;
+      assert(req);
      assert(req->status);
+
      if (!req->status->IsCorruption()) {
        // Avoid overwriting corruption status.
        *req->status = s;
@@ -464,38 +473,42 @@ void BlobFileReader::MultiGetBlob(const ReadOptions& read_options,

  uint64_t total_bytes = 0;
  for (size_t i = 0, j = 0; i < num_blobs; ++i) {
-    assert(blob_reqs[i]->status);
-    if (!blob_reqs[i]->status->ok()) {
+    BlobReadRequest* const req = blob_reqs[i].first;
+    assert(req);
+    assert(req->user_key);
+    assert(req->status);
+
+    if (!req->status->ok()) {
      continue;
    }

    assert(j < read_reqs.size());
-    auto& req = read_reqs[j++];
-    const auto& record_slice = req.result;
-    if (req.status.ok() && record_slice.size() != req.len) {
-      req.status = IOStatus::Corruption("Failed to read data from blob file");
+    auto& read_req = read_reqs[j++];
+    const auto& record_slice = read_req.result;
+    if (read_req.status.ok() && record_slice.size() != read_req.len) {
+      read_req.status =
+          IOStatus::Corruption("Failed to read data from blob file");
    }

-    *blob_reqs[i]->status = req.status;
-    if (!blob_reqs[i]->status->ok()) {
+    *req->status = read_req.status;
+    if (!req->status->ok()) {
      continue;
    }

    // Verify checksums if enabled
    if (read_options.verify_checksums) {
-      *blob_reqs[i]->status =
-          VerifyBlob(record_slice, *blob_reqs[i]->user_key, blob_reqs[i]->len);
-      if (!blob_reqs[i]->status->ok()) {
+      *req->status = VerifyBlob(record_slice, *req->user_key, req->len);
+      if (!req->status->ok()) {
        continue;
      }
    }

    // Uncompress blob if needed
-    Slice value_slice(record_slice.data() + adjustments[i], blob_reqs[i]->len);
-    *blob_reqs[i]->status =
-        UncompressBlobIfNeeded(value_slice, compression_type_, clock_,
-                               statistics_, blob_reqs[i]->result);
-    if (blob_reqs[i]->status->ok()) {
+    Slice value_slice(record_slice.data() + adjustments[i], req->len);
+    *req->status =
+        UncompressBlobIfNeeded(value_slice, compression_type_, allocator,
+                               clock_, statistics_, &blob_reqs[i].second);
+    if (req->status->ok()) {
      total_bytes += record_slice.size();
    }
  }
@@ -549,15 +562,18 @@ Status BlobFileReader::VerifyBlob(const Slice& record_slice,
  return Status::OK();
 }

-Status BlobFileReader::UncompressBlobIfNeeded(const Slice& value_slice,
-                                              CompressionType compression_type,
-                                              SystemClock* clock,
-                                              Statistics* statistics,
-                                              PinnableSlice* value) {
-  assert(value);
+Status BlobFileReader::UncompressBlobIfNeeded(
+    const Slice& value_slice, CompressionType compression_type,
+    MemoryAllocator* allocator, SystemClock* clock, Statistics* statistics,
+    std::unique_ptr<BlobContents>* result) {
+  assert(result);

  if (compression_type == kNoCompression) {
-    SaveValue(value_slice, value);
+    CacheAllocationPtr allocation =
+        AllocateBlock(value_slice.size(), allocator);
+    memcpy(allocation.get(), value_slice.data(), value_slice.size());
+
+    *result = BlobContents::Create(std::move(allocation), value_slice.size());

    return Status::OK();
  }
@@ -568,7 +584,6 @@ Status BlobFileReader::UncompressBlobIfNeeded(const Slice& value_slice,

  size_t uncompressed_size = 0;
  constexpr uint32_t compression_format_version = 2;
-  constexpr MemoryAllocator* allocator = nullptr;

  CacheAllocationPtr output;

@@ -587,19 +602,9 @@ Status BlobFileReader::UncompressBlobIfNeeded(const Slice& value_slice,
    return Status::Corruption("Unable to uncompress blob");
  }

-  SaveValue(Slice(output.get(), uncompressed_size), value);
+  *result = BlobContents::Create(std::move(output), uncompressed_size);

  return Status::OK();
 }

-void BlobFileReader::SaveValue(const Slice& src, PinnableSlice* dst) {
-  assert(dst);
-
-  if (dst->IsPinned()) {
-    dst->Reset();
-  }
-
-  dst->PinSelf(src);
-}
-
 }  // namespace ROCKSDB_NAMESPACE
@@ -23,7 +23,7 @@ class HistogramImpl;
 struct ReadOptions;
 class Slice;
 class FilePrefetchBuffer;
-class PinnableSlice;
+class BlobContents;
 class Statistics;

 class BlobFileReader {
@@ -44,13 +44,17 @@ class BlobFileReader {
  Status GetBlob(const ReadOptions& read_options, const Slice& user_key,
                 uint64_t offset, uint64_t value_size,
                 CompressionType compression_type,
-                 FilePrefetchBuffer* prefetch_buffer, PinnableSlice* value,
+                 FilePrefetchBuffer* prefetch_buffer,
+                 MemoryAllocator* allocator,
+                 std::unique_ptr<BlobContents>* result,
                 uint64_t* bytes_read) const;

  // offsets must be sorted in ascending order by caller.
-  void MultiGetBlob(const ReadOptions& read_options,
-                    autovector<BlobReadRequest*>& blob_reqs,
-                    uint64_t* bytes_read) const;
+  void MultiGetBlob(
+      const ReadOptions& read_options, MemoryAllocator* allocator,
+      autovector<std::pair<BlobReadRequest*, std::unique_ptr<BlobContents>>>&
+          blob_reqs,
+      uint64_t* bytes_read) const;

  CompressionType GetCompressionType() const { return compression_type_; }

@@ -89,11 +93,10 @@ class BlobFileReader {

  static Status UncompressBlobIfNeeded(const Slice& value_slice,
                                       CompressionType compression_type,
+                                       MemoryAllocator* allocator,
                                       SystemClock* clock,
                                       Statistics* statistics,
-                                       PinnableSlice* value);
-
-  static void SaveValue(const Slice& src, PinnableSlice* dst);
+                                       std::unique_ptr<BlobContents>* result);

  std::unique_ptr<RandomAccessFileReader> file_reader_;
  uint64_t file_size_;
@@ -8,6 +8,7 @@
 #include <cassert>
 #include <string>

+#include "db/blob/blob_contents.h"
 #include "db/blob/blob_log_format.h"
 #include "db/blob/blob_log_writer.h"
 #include "env/mock_env.h"
@@ -180,15 +181,17 @@ TEST_F(BlobFileReaderTest, CreateReaderAndGetBlob) {
  read_options.verify_checksums = false;

  constexpr FilePrefetchBuffer* prefetch_buffer = nullptr;
+  constexpr MemoryAllocator* allocator = nullptr;

  {
-    PinnableSlice value;
+    std::unique_ptr<BlobContents> value;
    uint64_t bytes_read = 0;

    ASSERT_OK(reader->GetBlob(read_options, keys[0], blob_offsets[0],
                              blob_sizes[0], kNoCompression, prefetch_buffer,
-                              &value, &bytes_read));
-    ASSERT_EQ(value, blobs[0]);
+                              allocator, &value, &bytes_read));
+    ASSERT_NE(value, nullptr);
+    ASSERT_EQ(value->data(), blobs[0]);
    ASSERT_EQ(bytes_read, blob_sizes[0]);

    // MultiGetBlob
@@ -196,22 +199,25 @@ TEST_F(BlobFileReaderTest, CreateReaderAndGetBlob) {
    size_t total_size = 0;

    std::array<Status, num_blobs> statuses_buf;
-    std::array<PinnableSlice, num_blobs> value_buf;
    std::array<BlobReadRequest, num_blobs> requests_buf;
-    autovector<BlobReadRequest*> blob_reqs;
+    autovector<std::pair<BlobReadRequest*, std::unique_ptr<BlobContents>>>
+        blob_reqs;

    for (size_t i = 0; i < num_blobs; ++i) {
      requests_buf[i] =
          BlobReadRequest(keys[i], blob_offsets[i], blob_sizes[i],
-                          kNoCompression, &value_buf[i], &statuses_buf[i]);
-      blob_reqs.push_back(&requests_buf[i]);
+                          kNoCompression, nullptr, &statuses_buf[i]);
+      blob_reqs.emplace_back(&requests_buf[i], std::unique_ptr<BlobContents>());
    }

-    reader->MultiGetBlob(read_options, blob_reqs, &bytes_read);
+    reader->MultiGetBlob(read_options, allocator, blob_reqs, &bytes_read);

    for (size_t i = 0; i < num_blobs; ++i) {
+      const auto& result = blob_reqs[i].second;
+
      ASSERT_OK(statuses_buf[i]);
-      ASSERT_EQ(value_buf[i], blobs[i]);
+      ASSERT_NE(result, nullptr);
+      ASSERT_EQ(result->data(), blobs[i]);
      total_size += blob_sizes[i];
    }
    ASSERT_EQ(bytes_read, total_size);
@@ -220,13 +226,14 @@ TEST_F(BlobFileReaderTest, CreateReaderAndGetBlob) {
  read_options.verify_checksums = true;

  {
-    PinnableSlice value;
+    std::unique_ptr<BlobContents> value;
    uint64_t bytes_read = 0;

    ASSERT_OK(reader->GetBlob(read_options, keys[1], blob_offsets[1],
                              blob_sizes[1], kNoCompression, prefetch_buffer,
-                              &value, &bytes_read));
-    ASSERT_EQ(value, blobs[1]);
+                              allocator, &value, &bytes_read));
+    ASSERT_NE(value, nullptr);
+    ASSERT_EQ(value->data(), blobs[1]);

    const uint64_t key_size = keys[1].size();
    ASSERT_EQ(bytes_read,
@@ -236,47 +243,50 @@ TEST_F(BlobFileReaderTest, CreateReaderAndGetBlob) {

  // Invalid offset (too close to start of file)
  {
-    PinnableSlice value;
+    std::unique_ptr<BlobContents> value;
    uint64_t bytes_read = 0;

    ASSERT_TRUE(reader
                    ->GetBlob(read_options, keys[0], blob_offsets[0] - 1,
                              blob_sizes[0], kNoCompression, prefetch_buffer,
-                              &value, &bytes_read)
+                              allocator, &value, &bytes_read)
                    .IsCorruption());
+    ASSERT_EQ(value, nullptr);
    ASSERT_EQ(bytes_read, 0);
  }

  // Invalid offset (too close to end of file)
  {
-    PinnableSlice value;
+    std::unique_ptr<BlobContents> value;
    uint64_t bytes_read = 0;

    ASSERT_TRUE(reader
                    ->GetBlob(read_options, keys[2], blob_offsets[2] + 1,
                              blob_sizes[2], kNoCompression, prefetch_buffer,
-                              &value, &bytes_read)
+                              allocator, &value, &bytes_read)
                    .IsCorruption());
+    ASSERT_EQ(value, nullptr);
    ASSERT_EQ(bytes_read, 0);
  }

  // Incorrect compression type
  {
-    PinnableSlice value;
+    std::unique_ptr<BlobContents> value;
    uint64_t bytes_read = 0;

    ASSERT_TRUE(reader
                    ->GetBlob(read_options, keys[0], blob_offsets[0],
-                              blob_sizes[0], kZSTD, prefetch_buffer, &value,
-                              &bytes_read)
+                              blob_sizes[0], kZSTD, prefetch_buffer, allocator,
+                              &value, &bytes_read)
                    .IsCorruption());
+    ASSERT_EQ(value, nullptr);
    ASSERT_EQ(bytes_read, 0);
  }

  // Incorrect key size
  {
    constexpr char shorter_key[] = "k";
-    PinnableSlice value;
+    std::unique_ptr<BlobContents> value;
    uint64_t bytes_read = 0;

    ASSERT_TRUE(reader
@@ -284,8 +294,9 @@ TEST_F(BlobFileReaderTest, CreateReaderAndGetBlob) {
                              blob_offsets[0] -
                                  (keys[0].size() - sizeof(shorter_key) + 1),
                              blob_sizes[0], kNoCompression, prefetch_buffer,
-                              &value, &bytes_read)
+                              allocator, &value, &bytes_read)
                    .IsCorruption());
+    ASSERT_EQ(value, nullptr);
    ASSERT_EQ(bytes_read, 0);

    // MultiGetBlob
@@ -302,18 +313,18 @@ TEST_F(BlobFileReaderTest, CreateReaderAndGetBlob) {
        blob_offsets[2]};

    std::array<Status, num_blobs> statuses_buf;
-    std::array<PinnableSlice, num_blobs> value_buf;
    std::array<BlobReadRequest, num_blobs> requests_buf;
-    autovector<BlobReadRequest*> blob_reqs;
+    autovector<std::pair<BlobReadRequest*, std::unique_ptr<BlobContents>>>
+        blob_reqs;

    for (size_t i = 0; i < num_blobs; ++i) {
      requests_buf[i] =
          BlobReadRequest(key_refs[i], offsets[i], blob_sizes[i],
-                          kNoCompression, &value_buf[i], &statuses_buf[i]);
-      blob_reqs.push_back(&requests_buf[i]);
+                          kNoCompression, nullptr, &statuses_buf[i]);
+      blob_reqs.emplace_back(&requests_buf[i], std::unique_ptr<BlobContents>());
    }

-    reader->MultiGetBlob(read_options, blob_reqs, &bytes_read);
+    reader->MultiGetBlob(read_options, allocator, blob_reqs, &bytes_read);

    for (size_t i = 0; i < num_blobs; ++i) {
      if (i == 1) {
@@ -327,14 +338,15 @@ TEST_F(BlobFileReaderTest, CreateReaderAndGetBlob) {
  // Incorrect key
  {
    constexpr char incorrect_key[] = "foo1";
-    PinnableSlice value;
+    std::unique_ptr<BlobContents> value;
    uint64_t bytes_read = 0;

    ASSERT_TRUE(reader
                    ->GetBlob(read_options, incorrect_key, blob_offsets[0],
                              blob_sizes[0], kNoCompression, prefetch_buffer,
-                              &value, &bytes_read)
+                              allocator, &value, &bytes_read)
                    .IsCorruption());
+    ASSERT_EQ(value, nullptr);
    ASSERT_EQ(bytes_read, 0);

    // MultiGetBlob
@@ -346,19 +358,18 @@ TEST_F(BlobFileReaderTest, CreateReaderAndGetBlob) {
    key_refs[2] = std::cref(wrong_key_slice);

    std::array<Status, num_blobs> statuses_buf;
-    std::array<PinnableSlice, num_blobs> value_buf;
    std::array<BlobReadRequest, num_blobs> requests_buf;
+    autovector<std::pair<BlobReadRequest*, std::unique_ptr<BlobContents>>>
+        blob_reqs;

    for (size_t i = 0; i < num_blobs; ++i) {
      requests_buf[i] =
          BlobReadRequest(key_refs[i], blob_offsets[i], blob_sizes[i],
-                          kNoCompression, &value_buf[i], &statuses_buf[i]);
+                          kNoCompression, nullptr, &statuses_buf[i]);
+      blob_reqs.emplace_back(&requests_buf[i], std::unique_ptr<BlobContents>());
    }

-    autovector<BlobReadRequest*> blob_reqs = {
-        &requests_buf[0], &requests_buf[1], &requests_buf[2]};
-
-    reader->MultiGetBlob(read_options, blob_reqs, &bytes_read);
+    reader->MultiGetBlob(read_options, allocator, blob_reqs, &bytes_read);

    for (size_t i = 0; i < num_blobs; ++i) {
      if (i == num_blobs - 1) {
@@ -371,14 +382,15 @@ TEST_F(BlobFileReaderTest, CreateReaderAndGetBlob) {

  // Incorrect value size
  {
-    PinnableSlice value;
+    std::unique_ptr<BlobContents> value;
    uint64_t bytes_read = 0;

    ASSERT_TRUE(reader
                    ->GetBlob(read_options, keys[1], blob_offsets[1],
                              blob_sizes[1] + 1, kNoCompression,
-                              prefetch_buffer, &value, &bytes_read)
+                              prefetch_buffer, allocator, &value, &bytes_read)
                    .IsCorruption());
+    ASSERT_EQ(value, nullptr);
    ASSERT_EQ(bytes_read, 0);

    // MultiGetBlob
@@ -388,23 +400,26 @@ TEST_F(BlobFileReaderTest, CreateReaderAndGetBlob) {
    }

    std::array<Status, num_blobs> statuses_buf;
-    std::array<PinnableSlice, num_blobs> value_buf;
    std::array<BlobReadRequest, num_blobs> requests_buf;

    requests_buf[0] =
        BlobReadRequest(key_refs[0], blob_offsets[0], blob_sizes[0],
-                        kNoCompression, &value_buf[0], &statuses_buf[0]);
+                        kNoCompression, nullptr, &statuses_buf[0]);
    requests_buf[1] =
        BlobReadRequest(key_refs[1], blob_offsets[1], blob_sizes[1] + 1,
-                        kNoCompression, &value_buf[1], &statuses_buf[1]);
+                        kNoCompression, nullptr, &statuses_buf[1]);
    requests_buf[2] =
        BlobReadRequest(key_refs[2], blob_offsets[2], blob_sizes[2],
-                        kNoCompression, &value_buf[2], &statuses_buf[2]);
+                        kNoCompression, nullptr, &statuses_buf[2]);

-    autovector<BlobReadRequest*> blob_reqs = {
-        &requests_buf[0], &requests_buf[1], &requests_buf[2]};
+    autovector<std::pair<BlobReadRequest*, std::unique_ptr<BlobContents>>>
+        blob_reqs;

-    reader->MultiGetBlob(read_options, blob_reqs, &bytes_read);
+    for (size_t i = 0; i < num_blobs; ++i) {
+      blob_reqs.emplace_back(&requests_buf[i], std::unique_ptr<BlobContents>());
+    }
+
+    reader->MultiGetBlob(read_options, allocator, blob_reqs, &bytes_read);

    for (size_t i = 0; i < num_blobs; ++i) {
      if (i != 1) {
@@ -665,14 +680,17 @@ TEST_F(BlobFileReaderTest, BlobCRCError) {
  SyncPoint::GetInstance()->EnableProcessing();

  constexpr FilePrefetchBuffer* prefetch_buffer = nullptr;
-  PinnableSlice value;
+  constexpr MemoryAllocator* allocator = nullptr;
+
+  std::unique_ptr<BlobContents> value;
  uint64_t bytes_read = 0;

  ASSERT_TRUE(reader
                  ->GetBlob(ReadOptions(), key, blob_offset, blob_size,
-                            kNoCompression, prefetch_buffer, &value,
+                            kNoCompression, prefetch_buffer, allocator, &value,
                            &bytes_read)
                  .IsCorruption());
+  ASSERT_EQ(value, nullptr);
  ASSERT_EQ(bytes_read, 0);

  SyncPoint::GetInstance()->DisableProcessing();
@@ -720,28 +738,31 @@ TEST_F(BlobFileReaderTest, Compression) {
  read_options.verify_checksums = false;

  constexpr FilePrefetchBuffer* prefetch_buffer = nullptr;
+  constexpr MemoryAllocator* allocator = nullptr;

  {
-    PinnableSlice value;
+    std::unique_ptr<BlobContents> value;
    uint64_t bytes_read = 0;

    ASSERT_OK(reader->GetBlob(read_options, key, blob_offset, blob_size,
-                              kSnappyCompression, prefetch_buffer, &value,
-                              &bytes_read));
-    ASSERT_EQ(value, blob);
+                              kSnappyCompression, prefetch_buffer, allocator,
+                              &value, &bytes_read));
+    ASSERT_NE(value, nullptr);
+    ASSERT_EQ(value->data(), blob);
    ASSERT_EQ(bytes_read, blob_size);
  }

  read_options.verify_checksums = true;

  {
-    PinnableSlice value;
+    std::unique_ptr<BlobContents> value;
    uint64_t bytes_read = 0;

    ASSERT_OK(reader->GetBlob(read_options, key, blob_offset, blob_size,
-                              kSnappyCompression, prefetch_buffer, &value,
-                              &bytes_read));
-    ASSERT_EQ(value, blob);
+                              kSnappyCompression, prefetch_buffer, allocator,
+                              &value, &bytes_read));
+    ASSERT_NE(value, nullptr);
+    ASSERT_EQ(value->data(), blob);

    constexpr uint64_t key_size = sizeof(key) - 1;
    ASSERT_EQ(bytes_read,
@@ -799,14 +820,17 @@ TEST_F(BlobFileReaderTest, UncompressionError) {
  SyncPoint::GetInstance()->EnableProcessing();

  constexpr FilePrefetchBuffer* prefetch_buffer = nullptr;
-  PinnableSlice value;
+  constexpr MemoryAllocator* allocator = nullptr;
+
+  std::unique_ptr<BlobContents> value;
  uint64_t bytes_read = 0;

  ASSERT_TRUE(reader
                  ->GetBlob(ReadOptions(), key, blob_offset, blob_size,
-                            kSnappyCompression, prefetch_buffer, &value,
-                            &bytes_read)
+                            kSnappyCompression, prefetch_buffer, allocator,
+                            &value, &bytes_read)
                  .IsCorruption());
+  ASSERT_EQ(value, nullptr);
  ASSERT_EQ(bytes_read, 0);

  SyncPoint::GetInstance()->DisableProcessing();
@@ -885,14 +909,17 @@ TEST_P(BlobFileReaderIOErrorTest, IOError) {
    ASSERT_OK(s);

    constexpr FilePrefetchBuffer* prefetch_buffer = nullptr;
-    PinnableSlice value;
+    constexpr MemoryAllocator* allocator = nullptr;
+
+    std::unique_ptr<BlobContents> value;
    uint64_t bytes_read = 0;

    ASSERT_TRUE(reader
                    ->GetBlob(ReadOptions(), key, blob_offset, blob_size,
-                              kNoCompression, prefetch_buffer, &value,
-                              &bytes_read)
+                              kNoCompression, prefetch_buffer, allocator,
+                              &value, &bytes_read)
                    .IsIOError());
+    ASSERT_EQ(value, nullptr);
    ASSERT_EQ(bytes_read, 0);
  }

@@ -970,14 +997,17 @@ TEST_P(BlobFileReaderDecodingErrorTest, DecodingError) {
    ASSERT_OK(s);

    constexpr FilePrefetchBuffer* prefetch_buffer = nullptr;
-    PinnableSlice value;
+    constexpr MemoryAllocator* allocator = nullptr;
+
+    std::unique_ptr<BlobContents> value;
    uint64_t bytes_read = 0;

    ASSERT_TRUE(reader
                    ->GetBlob(ReadOptions(), key, blob_offset, blob_size,
-                              kNoCompression, prefetch_buffer, &value,
-                              &bytes_read)
+                              kNoCompression, prefetch_buffer, allocator,
+                              &value, &bytes_read)
                    .IsCorruption());
+    ASSERT_EQ(value, nullptr);
    ASSERT_EQ(bytes_read, 0);
  }

@@ -988,6 +1018,7 @@ TEST_P(BlobFileReaderDecodingErrorTest, DecodingError) {
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -191,6 +191,7 @@ TEST(BlobGarbageMeterTest, InlinedTTLBlobIndex) {
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -10,6 +10,7 @@

 #include "cache/cache_reservation_manager.h"
 #include "cache/charged_cache.h"
+#include "db/blob/blob_contents.h"
 #include "db/blob/blob_file_reader.h"
 #include "db/blob/blob_log_format.h"
 #include "monitoring/statistics.h"
@@ -43,52 +44,63 @@ BlobSource::BlobSource(const ImmutableOptions* immutable_options,

 BlobSource::~BlobSource() = default;

-Status BlobSource::GetBlobFromCache(const Slice& cache_key,
-                                    CacheHandleGuard<std::string>* blob) const {
-  assert(blob);
-  assert(blob->IsEmpty());
+Status BlobSource::GetBlobFromCache(
+    const Slice& cache_key, CacheHandleGuard<BlobContents>* cached_blob) const {
  assert(blob_cache_);
  assert(!cache_key.empty());
+  assert(cached_blob);
+  assert(cached_blob->IsEmpty());

  Cache::Handle* cache_handle = nullptr;
  cache_handle = GetEntryFromCache(cache_key);
  if (cache_handle != nullptr) {
-    *blob = CacheHandleGuard<std::string>(blob_cache_.get(), cache_handle);
+    *cached_blob =
+        CacheHandleGuard<BlobContents>(blob_cache_.get(), cache_handle);
+
+    assert(cached_blob->GetValue());
+
+    PERF_COUNTER_ADD(blob_cache_hit_count, 1);
+    RecordTick(statistics_, BLOB_DB_CACHE_HIT);
+    RecordTick(statistics_, BLOB_DB_CACHE_BYTES_READ,
+               cached_blob->GetValue()->size());
+
    return Status::OK();
  }

-  assert(blob->IsEmpty());
+  RecordTick(statistics_, BLOB_DB_CACHE_MISS);

  return Status::NotFound("Blob not found in cache");
 }

-Status BlobSource::PutBlobIntoCache(const Slice& cache_key,
-                                    CacheHandleGuard<std::string>* cached_blob,
-                                    PinnableSlice* blob) const {
-  assert(blob);
-  assert(!cache_key.empty());
+Status BlobSource::PutBlobIntoCache(
+    const Slice& cache_key, std::unique_ptr<BlobContents>* blob,
+    CacheHandleGuard<BlobContents>* cached_blob) const {
  assert(blob_cache_);
+  assert(!cache_key.empty());
+  assert(blob);
+  assert(*blob);
+  assert(cached_blob);
+  assert(cached_blob->IsEmpty());

-  Status s;
-  const Cache::Priority priority = Cache::Priority::BOTTOM;
-
-  // Objects to be put into the cache have to be heap-allocated and
-  // self-contained, i.e. own their contents. The Cache has to be able to take
-  // unique ownership of them. Therefore, we copy the blob into a string
-  // directly, and insert that into the cache.
-  std::unique_ptr<std::string> buf = std::make_unique<std::string>();
-  buf->assign(blob->data(), blob->size());
-
-  // TODO: support custom allocators and provide a better estimated memory
-  // usage using malloc_usable_size.
  Cache::Handle* cache_handle = nullptr;
-  s = InsertEntryIntoCache(cache_key, buf.get(), buf->size(), &cache_handle,
-                           priority);
+  const Status s = InsertEntryIntoCache(cache_key, blob->get(),
+                                        (*blob)->ApproximateMemoryUsage(),
+                                        &cache_handle, Cache::Priority::BOTTOM);
  if (s.ok()) {
-    buf.release();
+    blob->release();
+
    assert(cache_handle != nullptr);
    *cached_blob =
-        CacheHandleGuard<std::string>(blob_cache_.get(), cache_handle);
+        CacheHandleGuard<BlobContents>(blob_cache_.get(), cache_handle);
+
+    assert(cached_blob->GetValue());
+
+    RecordTick(statistics_, BLOB_DB_CACHE_ADD);
+    RecordTick(statistics_, BLOB_DB_CACHE_BYTES_WRITE,
+               cached_blob->GetValue()->size());
+
+  } else {
+    RecordTick(statistics_, BLOB_DB_CACHE_ADD_FAILURES);
  }

  return s;
@@ -98,54 +110,77 @@ Cache::Handle* BlobSource::GetEntryFromCache(const Slice& key) const {
  Cache::Handle* cache_handle = nullptr;

  if (lowest_used_cache_tier_ == CacheTier::kNonVolatileBlockTier) {
-    Cache::CreateCallback create_cb = [&](const void* buf, size_t size,
-                                          void** out_obj,
-                                          size_t* charge) -> Status {
-      std::string* blob = new std::string();
-      blob->assign(static_cast<const char*>(buf), size);
-      *out_obj = blob;
-      *charge = size;
-      return Status::OK();
+    Cache::CreateCallback create_cb =
+        [allocator = blob_cache_->memory_allocator()](
+            const void* buf, size_t size, void** out_obj,
+            size_t* charge) -> Status {
+      return BlobContents::CreateCallback(AllocateBlock(size, allocator), buf,
+                                          size, out_obj, charge);
    };
-    cache_handle = blob_cache_->Lookup(key, GetCacheItemHelper(), create_cb,
-                                       Cache::Priority::BOTTOM,
+
+    cache_handle = blob_cache_->Lookup(key, BlobContents::GetCacheItemHelper(),
+                                       create_cb, Cache::Priority::BOTTOM,
                                       true /* wait_for_cache */, statistics_);
  } else {
    cache_handle = blob_cache_->Lookup(key, statistics_);
  }

-  if (cache_handle != nullptr) {
-    PERF_COUNTER_ADD(blob_cache_hit_count, 1);
-    RecordTick(statistics_, BLOB_DB_CACHE_HIT);
-    RecordTick(statistics_, BLOB_DB_CACHE_BYTES_READ,
-               blob_cache_->GetUsage(cache_handle));
-  } else {
-    RecordTick(statistics_, BLOB_DB_CACHE_MISS);
-  }
  return cache_handle;
 }

-Status BlobSource::InsertEntryIntoCache(const Slice& key, std::string* value,
+void BlobSource::PinCachedBlob(CacheHandleGuard<BlobContents>* cached_blob,
+                               PinnableSlice* value) {
+  assert(cached_blob);
+  assert(cached_blob->GetValue());
+  assert(value);
+
+  // To avoid copying the cached blob into the buffer provided by the
+  // application, we can simply transfer ownership of the cache handle to
+  // the target PinnableSlice. This has the potential to save a lot of
+  // CPU, especially with large blob values.
+
+  value->Reset();
+
+  constexpr Cleanable* cleanable = nullptr;
+  value->PinSlice(cached_blob->GetValue()->data(), cleanable);
+
+  cached_blob->TransferTo(value);
+}
+
+void BlobSource::PinOwnedBlob(std::unique_ptr<BlobContents>* owned_blob,
+                              PinnableSlice* value) {
+  assert(owned_blob);
+  assert(*owned_blob);
+  assert(value);
+
+  BlobContents* const blob = owned_blob->release();
+  assert(blob);
+
+  value->Reset();
+  value->PinSlice(
+      blob->data(),
+      [](void* arg1, void* /* arg2 */) {
+        delete static_cast<BlobContents*>(arg1);
+      },
+      blob, nullptr);
+}
+
+Status BlobSource::InsertEntryIntoCache(const Slice& key, BlobContents* value,
                                        size_t charge,
                                        Cache::Handle** cache_handle,
                                        Cache::Priority priority) const {
  Status s;

-  if (lowest_used_cache_tier_ == CacheTier::kNonVolatileBlockTier) {
-    s = blob_cache_->Insert(key, value, GetCacheItemHelper(), charge,
-                            cache_handle, priority);
-  } else {
-    s = blob_cache_->Insert(key, value, charge, &DeleteCacheEntry<std::string>,
-                            cache_handle, priority);
-  }
+  Cache::CacheItemHelper* const cache_item_helper =
+      BlobContents::GetCacheItemHelper();
+  assert(cache_item_helper);

-  if (s.ok()) {
-    assert(*cache_handle != nullptr);
-    RecordTick(statistics_, BLOB_DB_CACHE_ADD);
-    RecordTick(statistics_, BLOB_DB_CACHE_BYTES_WRITE,
-               blob_cache_->GetUsage(*cache_handle));
+  if (lowest_used_cache_tier_ == CacheTier::kNonVolatileBlockTier) {
+    s = blob_cache_->Insert(key, value, cache_item_helper, charge, cache_handle,
+                            priority);
  } else {
-    RecordTick(statistics_, BLOB_DB_CACHE_ADD_FAILURES);
+    s = blob_cache_->Insert(key, value, charge, cache_item_helper->del_cb,
+                            cache_handle, priority);
  }

  return s;
@@ -164,7 +199,7 @@ Status BlobSource::GetBlob(const ReadOptions& read_options,

  const CacheKey cache_key = GetCacheKey(file_number, file_size, offset);

-  CacheHandleGuard<std::string> blob_handle;
+  CacheHandleGuard<BlobContents> blob_handle;

  // First, try to get the blob from the cache
  //
@@ -172,24 +207,8 @@ Status BlobSource::GetBlob(const ReadOptions& read_options,
  if (blob_cache_) {
    Slice key = cache_key.AsSlice();
    s = GetBlobFromCache(key, &blob_handle);
-    if (s.ok() && blob_handle.GetValue()) {
-      {
-        value->Reset();
-        // To avoid copying the cached blob into the buffer provided by the
-        // application, we can simply transfer ownership of the cache handle to
-        // the target PinnableSlice. This has the potential to save a lot of
-        // CPU, especially with large blob values.
-        value->PinSlice(
-            *blob_handle.GetValue(),
-            [](void* arg1, void* arg2) {
-              Cache* const cache = static_cast<Cache*>(arg1);
-              Cache::Handle* const handle = static_cast<Cache::Handle*>(arg2);
-              cache->Release(handle);
-            },
-            blob_handle.GetCache(), blob_handle.GetCacheHandle());
-        // Make the CacheHandleGuard relinquish ownership of the handle.
-        blob_handle.TransferTo(nullptr);
-      }
+    if (s.ok()) {
+      PinCachedBlob(&blob_handle, value);

      // For consistency, the size of on-disk (possibly compressed) blob record
      // is assigned to bytes_read.
@@ -218,6 +237,8 @@ Status BlobSource::GetBlob(const ReadOptions& read_options,

  // Can't find the blob from the cache. Since I/O is allowed, read from the
  // file.
+  std::unique_ptr<BlobContents> blob_contents;
+
  {
    CacheHandleGuard<BlobFileReader> blob_file_reader;
    s = blob_file_cache_->GetBlobFileReader(file_number, &blob_file_reader);
@@ -231,10 +252,14 @@ Status BlobSource::GetBlob(const ReadOptions& read_options,
      return Status::Corruption("Compression type mismatch when reading blob");
    }

+    MemoryAllocator* const allocator = (blob_cache_ && read_options.fill_cache)
+                                           ? blob_cache_->memory_allocator()
+                                           : nullptr;
+
    uint64_t read_size = 0;
    s = blob_file_reader.GetValue()->GetBlob(
        read_options, user_key, offset, value_size, compression_type,
-        prefetch_buffer, value, &read_size);
+        prefetch_buffer, allocator, &blob_contents, &read_size);
    if (!s.ok()) {
      return s;
    }
@@ -247,10 +272,14 @@ Status BlobSource::GetBlob(const ReadOptions& read_options,
    // If filling cache is allowed and a cache is configured, try to put the
    // blob to the cache.
    Slice key = cache_key.AsSlice();
-    s = PutBlobIntoCache(key, &blob_handle, value);
+    s = PutBlobIntoCache(key, &blob_contents, &blob_handle);
    if (!s.ok()) {
      return s;
    }
+
+    PinCachedBlob(&blob_handle, value);
+  } else {
+    PinOwnedBlob(&blob_contents, value);
  }

  assert(s.ok());
@@ -310,33 +339,17 @@ void BlobSource::MultiGetBlobFromOneFile(const ReadOptions& read_options,
    for (size_t i = 0; i < num_blobs; ++i) {
      auto& req = blob_reqs[i];

-      CacheHandleGuard<std::string> blob_handle;
+      CacheHandleGuard<BlobContents> blob_handle;
      const CacheKey cache_key = base_cache_key.WithOffset(req.offset);
      const Slice key = cache_key.AsSlice();

      const Status s = GetBlobFromCache(key, &blob_handle);

-      if (s.ok() && blob_handle.GetValue()) {
+      if (s.ok()) {
        assert(req.status);
        *req.status = s;

-        {
-          req.result->Reset();
-          // To avoid copying the cached blob into the buffer provided by the
-          // application, we can simply transfer ownership of the cache handle
-          // to the target PinnableSlice. This has the potential to save a lot
-          // of CPU, especially with large blob values.
-          req.result->PinSlice(
-              *blob_handle.GetValue(),
-              [](void* arg1, void* arg2) {
-                Cache* const cache = static_cast<Cache*>(arg1);
-                Cache::Handle* const handle = static_cast<Cache::Handle*>(arg2);
-                cache->Release(handle);
-              },
-              blob_handle.GetCache(), blob_handle.GetCacheHandle());
-          // Make the CacheHandleGuard relinquish ownership of the handle.
-          blob_handle.TransferTo(nullptr);
-        }
+        PinCachedBlob(&blob_handle, req.result);

        // Update the counter for the number of valid blobs read from the cache.
        ++cached_blob_count;
@@ -367,8 +380,10 @@ void BlobSource::MultiGetBlobFromOneFile(const ReadOptions& read_options,
  if (no_io) {
    for (size_t i = 0; i < num_blobs; ++i) {
      if (!(cache_hit_mask & (Mask{1} << i))) {
-        assert(blob_reqs[i].status);
-        *blob_reqs[i].status =
+        BlobReadRequest& req = blob_reqs[i];
+        assert(req.status);
+
+        *req.status =
            Status::Incomplete("Cannot read blob(s): no disk I/O allowed");
      }
    }
@@ -377,12 +392,13 @@ void BlobSource::MultiGetBlobFromOneFile(const ReadOptions& read_options,

  {
    // Find the rest of blobs from the file since I/O is allowed.
-    autovector<BlobReadRequest*> _blob_reqs;
+    autovector<std::pair<BlobReadRequest*, std::unique_ptr<BlobContents>>>
+        _blob_reqs;
    uint64_t _bytes_read = 0;

    for (size_t i = 0; i < num_blobs; ++i) {
      if (!(cache_hit_mask & (Mask{1} << i))) {
-        _blob_reqs.push_back(&blob_reqs[i]);
+        _blob_reqs.emplace_back(&blob_reqs[i], std::unique_ptr<BlobContents>());
      }
    }

@@ -391,32 +407,50 @@ void BlobSource::MultiGetBlobFromOneFile(const ReadOptions& read_options,
        blob_file_cache_->GetBlobFileReader(file_number, &blob_file_reader);
    if (!s.ok()) {
      for (size_t i = 0; i < _blob_reqs.size(); ++i) {
-        assert(_blob_reqs[i]->status);
-        *_blob_reqs[i]->status = s;
+        BlobReadRequest* const req = _blob_reqs[i].first;
+        assert(req);
+        assert(req->status);
+
+        *req->status = s;
      }
      return;
    }

    assert(blob_file_reader.GetValue());

-    blob_file_reader.GetValue()->MultiGetBlob(read_options, _blob_reqs,
-                                              &_bytes_read);
+    MemoryAllocator* const allocator = (blob_cache_ && read_options.fill_cache)
+                                           ? blob_cache_->memory_allocator()
+                                           : nullptr;
+
+    blob_file_reader.GetValue()->MultiGetBlob(read_options, allocator,
+                                              _blob_reqs, &_bytes_read);

    if (blob_cache_ && read_options.fill_cache) {
      // If filling cache is allowed and a cache is configured, try to put
      // the blob(s) to the cache.
-      for (size_t i = 0; i < _blob_reqs.size(); ++i) {
-        if (_blob_reqs[i]->status->ok()) {
-          CacheHandleGuard<std::string> blob_handle;
-          const CacheKey cache_key =
-              base_cache_key.WithOffset(_blob_reqs[i]->offset);
+      for (auto& [req, blob_contents] : _blob_reqs) {
+        assert(req);
+
+        if (req->status->ok()) {
+          CacheHandleGuard<BlobContents> blob_handle;
+          const CacheKey cache_key = base_cache_key.WithOffset(req->offset);
          const Slice key = cache_key.AsSlice();
-          s = PutBlobIntoCache(key, &blob_handle, _blob_reqs[i]->result);
+          s = PutBlobIntoCache(key, &blob_contents, &blob_handle);
          if (!s.ok()) {
-            *_blob_reqs[i]->status = s;
+            *req->status = s;
+          } else {
+            PinCachedBlob(&blob_handle, req->result);
          }
        }
      }
+    } else {
+      for (auto& [req, blob_contents] : _blob_reqs) {
+        assert(req);
+
+        if (req->status->ok()) {
+          PinOwnedBlob(&blob_contents, req->result);
+        }
+      }
    }

    total_bytes += _bytes_read;
@@ -427,39 +461,28 @@ void BlobSource::MultiGetBlobFromOneFile(const ReadOptions& read_options,
 }

 bool BlobSource::TEST_BlobInCache(uint64_t file_number, uint64_t file_size,
-                                  uint64_t offset) const {
+                                  uint64_t offset, size_t* charge) const {
  const CacheKey cache_key = GetCacheKey(file_number, file_size, offset);
  const Slice key = cache_key.AsSlice();

-  CacheHandleGuard<std::string> blob_handle;
+  CacheHandleGuard<BlobContents> blob_handle;
  const Status s = GetBlobFromCache(key, &blob_handle);

  if (s.ok() && blob_handle.GetValue() != nullptr) {
+    if (charge) {
+      const Cache* const cache = blob_handle.GetCache();
+      assert(cache);
+
+      Cache::Handle* const handle = blob_handle.GetCacheHandle();
+      assert(handle);
+
+      *charge = cache->GetUsage(handle);
+    }
+
    return true;
  }

  return false;
 }

-// Callbacks for secondary blob cache
-size_t BlobSource::SizeCallback(void* obj) {
-  assert(obj != nullptr);
-  return static_cast<const std::string*>(obj)->size();
-}
-
-Status BlobSource::SaveToCallback(void* from_obj, size_t from_offset,
-                                  size_t length, void* out) {
-  assert(from_obj != nullptr);
-  const std::string* buf = static_cast<const std::string*>(from_obj);
-  assert(buf->size() >= from_offset + length);
-  memcpy(out, buf->data() + from_offset, length);
-  return Status::OK();
-}
-
-Cache::CacheItemHelper* BlobSource::GetCacheItemHelper() {
-  static Cache::CacheItemHelper cache_helper(SizeCallback, SaveToCallback,
-                                             &DeleteCacheEntry<std::string>);
-  return &cache_helper;
-}
-
 }  // namespace ROCKSDB_NAMESPACE
@@ -6,6 +6,7 @@
 #pragma once

 #include <cinttypes>
+#include <memory>

 #include "cache/cache_helpers.h"
 #include "cache/cache_key.h"
@@ -22,6 +23,7 @@ struct ImmutableOptions;
 class Status;
 class FilePrefetchBuffer;
 class Slice;
+class BlobContents;

 // BlobSource is a class that provides universal access to blobs, regardless of
 // whether they are in the blob cache, secondary cache, or (remote) storage.
@@ -102,19 +104,25 @@ class BlobSource {
  inline Cache* GetBlobCache() const { return blob_cache_.get(); }

  bool TEST_BlobInCache(uint64_t file_number, uint64_t file_size,
-                        uint64_t offset) const;
+                        uint64_t offset, size_t* charge = nullptr) const;

 private:
  Status GetBlobFromCache(const Slice& cache_key,
-                          CacheHandleGuard<std::string>* blob) const;
+                          CacheHandleGuard<BlobContents>* cached_blob) const;

  Status PutBlobIntoCache(const Slice& cache_key,
-                          CacheHandleGuard<std::string>* cached_blob,
-                          PinnableSlice* blob) const;
+                          std::unique_ptr<BlobContents>* blob,
+                          CacheHandleGuard<BlobContents>* cached_blob) const;
+
+  static void PinCachedBlob(CacheHandleGuard<BlobContents>* cached_blob,
+                            PinnableSlice* value);
+
+  static void PinOwnedBlob(std::unique_ptr<BlobContents>* owned_blob,
+                           PinnableSlice* value);

  Cache::Handle* GetEntryFromCache(const Slice& key) const;

-  Status InsertEntryIntoCache(const Slice& key, std::string* value,
+  Status InsertEntryIntoCache(const Slice& key, BlobContents* value,
                              size_t charge, Cache::Handle** cache_handle,
                              Cache::Priority priority) const;

@@ -124,14 +132,6 @@ class BlobSource {
    return base_cache_key.WithOffset(offset);
  }

-  // Callbacks for secondary blob cache
-  static size_t SizeCallback(void* obj);
-
-  static Status SaveToCallback(void* from_obj, size_t from_offset,
-                               size_t length, void* out);
-
-  static Cache::CacheItemHelper* GetCacheItemHelper();
-
  const std::string& db_id_;
  const std::string& db_session_id_;

@@ -13,6 +13,7 @@

 #include "cache/charged_cache.h"
 #include "cache/compressed_secondary_cache.h"
+#include "db/blob/blob_contents.h"
 #include "db/blob/blob_file_cache.h"
 #include "db/blob/blob_file_reader.h"
 #include "db/blob/blob_log_format.h"
@@ -218,6 +219,7 @@ TEST_F(BlobSourceTest, GetBlobsFromCache) {
                                    kNoCompression, prefetch_buffer, &values[i],
                                    &bytes_read));
      ASSERT_EQ(values[i], blobs[i]);
+      ASSERT_TRUE(values[i].IsPinned());
      ASSERT_EQ(bytes_read,
                BlobLogRecord::kHeaderSize + keys[i].size() + blob_sizes[i]);

@@ -255,6 +257,7 @@ TEST_F(BlobSourceTest, GetBlobsFromCache) {
                                    kNoCompression, prefetch_buffer, &values[i],
                                    &bytes_read));
      ASSERT_EQ(values[i], blobs[i]);
+      ASSERT_TRUE(values[i].IsPinned());
      ASSERT_EQ(bytes_read,
                BlobLogRecord::kHeaderSize + keys[i].size() + blob_sizes[i]);

@@ -298,6 +301,7 @@ TEST_F(BlobSourceTest, GetBlobsFromCache) {
                                    kNoCompression, prefetch_buffer, &values[i],
                                    &bytes_read));
      ASSERT_EQ(values[i], blobs[i]);
+      ASSERT_TRUE(values[i].IsPinned());
      ASSERT_EQ(bytes_read,
                BlobLogRecord::kHeaderSize + keys[i].size() + blob_sizes[i]);

@@ -336,6 +340,7 @@ TEST_F(BlobSourceTest, GetBlobsFromCache) {
                                    kNoCompression, prefetch_buffer, &values[i],
                                    &bytes_read));
      ASSERT_EQ(values[i], blobs[i]);
+      ASSERT_TRUE(values[i].IsPinned());
      ASSERT_EQ(bytes_read,
                BlobLogRecord::kHeaderSize + keys[i].size() + blob_sizes[i]);

@@ -382,6 +387,7 @@ TEST_F(BlobSourceTest, GetBlobsFromCache) {
                               &bytes_read)
                      .IsIncomplete());
      ASSERT_TRUE(values[i].empty());
+      ASSERT_FALSE(values[i].IsPinned());
      ASSERT_EQ(bytes_read, 0);

      ASSERT_FALSE(blob_source.TEST_BlobInCache(blob_file_number, file_size,
@@ -423,6 +429,7 @@ TEST_F(BlobSourceTest, GetBlobsFromCache) {
                               &bytes_read)
                      .IsIOError());
      ASSERT_TRUE(values[i].empty());
+      ASSERT_FALSE(values[i].IsPinned());
      ASSERT_EQ(bytes_read, 0);

      ASSERT_FALSE(blob_source.TEST_BlobInCache(file_number, file_size,
@@ -855,6 +862,7 @@ TEST_F(BlobSourceTest, MultiGetBlobsFromCache) {
      if (i % 2 == 0) {
        ASSERT_OK(statuses_buf[i]);
        ASSERT_EQ(value_buf[i], blobs[i]);
+        ASSERT_TRUE(value_buf[i].IsPinned());
        fs_read_bytes +=
            blob_sizes[i] + keys[i].size() + BlobLogRecord::kHeaderSize;
        ASSERT_TRUE(blob_source.TEST_BlobInCache(blob_file_number, file_size,
@@ -863,6 +871,7 @@ TEST_F(BlobSourceTest, MultiGetBlobsFromCache) {
      } else {
        statuses_buf[i].PermitUncheckedError();
        ASSERT_TRUE(value_buf[i].empty());
+        ASSERT_FALSE(value_buf[i].IsPinned());
        ASSERT_FALSE(blob_source.TEST_BlobInCache(blob_file_number, file_size,
                                                  blob_offsets[i]));
      }
@@ -895,6 +904,7 @@ TEST_F(BlobSourceTest, MultiGetBlobsFromCache) {
                                    kNoCompression, prefetch_buffer,
                                    &value_buf[i], &bytes_read));
      ASSERT_EQ(value_buf[i], blobs[i]);
+      ASSERT_TRUE(value_buf[i].IsPinned());
      ASSERT_EQ(bytes_read,
                BlobLogRecord::kHeaderSize + keys[i].size() + blob_sizes[i]);

@@ -920,6 +930,7 @@ TEST_F(BlobSourceTest, MultiGetBlobsFromCache) {
    for (size_t i = 0; i < num_blobs; ++i) {
      ASSERT_OK(statuses_buf[i]);
      ASSERT_EQ(value_buf[i], blobs[i]);
+      ASSERT_TRUE(value_buf[i].IsPinned());
      ASSERT_TRUE(blob_source.TEST_BlobInCache(blob_file_number, file_size,
                                               blob_offsets[i]));
      blob_bytes += blob_sizes[i];
@@ -968,6 +979,7 @@ TEST_F(BlobSourceTest, MultiGetBlobsFromCache) {
    for (size_t i = 0; i < num_blobs; ++i) {
      ASSERT_TRUE(statuses_buf[i].IsIncomplete());
      ASSERT_TRUE(value_buf[i].empty());
+      ASSERT_FALSE(value_buf[i].IsPinned());
      ASSERT_FALSE(blob_source.TEST_BlobInCache(blob_file_number, file_size,
                                                blob_offsets[i]));
    }
@@ -1011,6 +1023,7 @@ TEST_F(BlobSourceTest, MultiGetBlobsFromCache) {
    for (size_t i = 0; i < num_blobs; ++i) {
      ASSERT_TRUE(statuses_buf[i].IsIOError());
      ASSERT_TRUE(value_buf[i].empty());
+      ASSERT_FALSE(value_buf[i].IsPinned());
      ASSERT_FALSE(blob_source.TEST_BlobInCache(non_existing_file_number,
                                                file_size, blob_offsets[i]));
    }
@@ -1040,16 +1053,17 @@ class BlobSecondaryCacheTest : public DBTestBase {

    // Set a small cache capacity to evict entries from the cache, and to test
    // that secondary cache is used properly.
-    lru_cache_ops_.capacity = 1024;
-    lru_cache_ops_.num_shard_bits = 0;
-    lru_cache_ops_.strict_capacity_limit = true;
-    lru_cache_ops_.metadata_charge_policy = kDontChargeCacheMetadata;
-    lru_cache_ops_.high_pri_pool_ratio = 0.2;
-    lru_cache_ops_.low_pri_pool_ratio = 0.2;
+    lru_cache_opts_.capacity = 1024;
+    lru_cache_opts_.num_shard_bits = 0;
+    lru_cache_opts_.strict_capacity_limit = true;
+    lru_cache_opts_.metadata_charge_policy = kDontChargeCacheMetadata;
+    lru_cache_opts_.high_pri_pool_ratio = 0.2;
+    lru_cache_opts_.low_pri_pool_ratio = 0.2;

    secondary_cache_opts_.capacity = 8 << 20;  // 8 MB
    secondary_cache_opts_.num_shard_bits = 0;
-    secondary_cache_opts_.metadata_charge_policy = kDontChargeCacheMetadata;
+    secondary_cache_opts_.metadata_charge_policy =
+        kDefaultCacheMetadataChargePolicy;

    // Read blobs from the secondary cache if they are not in the primary cache
    options_.lowest_used_cache_tier = CacheTier::kNonVolatileBlockTier;
@@ -1060,7 +1074,7 @@ class BlobSecondaryCacheTest : public DBTestBase {

  Options options_;

-  LRUCacheOptions lru_cache_ops_;
+  LRUCacheOptions lru_cache_opts_;
  CompressedSecondaryCacheOptions secondary_cache_opts_;

  std::string db_id_;
@@ -1073,9 +1087,9 @@ TEST_F(BlobSecondaryCacheTest, GetBlobsFromSecondaryCache) {
  }

  secondary_cache_opts_.compression_type = kSnappyCompression;
-  lru_cache_ops_.secondary_cache =
+  lru_cache_opts_.secondary_cache =
      NewCompressedSecondaryCache(secondary_cache_opts_);
-  options_.blob_cache = NewLRUCache(lru_cache_ops_);
+  options_.blob_cache = NewLRUCache(lru_cache_opts_);

  options_.cf_paths.emplace_back(
      test::PerThreadDBPath(
@@ -1098,8 +1112,8 @@ TEST_F(BlobSecondaryCacheTest, GetBlobsFromSecondaryCache) {
  Random rnd(301);

  std::vector<std::string> key_strs{"key0", "key1"};
-  std::vector<std::string> blob_strs{rnd.RandomString(1010),
-                                     rnd.RandomString(1020)};
+  std::vector<std::string> blob_strs{rnd.RandomString(512),
+                                     rnd.RandomString(768)};

  std::vector<Slice> keys{key_strs[0], key_strs[1]};
  std::vector<Slice> blobs{blob_strs[0], blob_strs[1]};
@@ -1134,16 +1148,15 @@ TEST_F(BlobSecondaryCacheTest, GetBlobsFromSecondaryCache) {
  read_options.verify_checksums = true;

  auto blob_cache = options_.blob_cache;
-  auto secondary_cache = lru_cache_ops_.secondary_cache;
+  auto secondary_cache = lru_cache_opts_.secondary_cache;

-  Cache::CreateCallback create_cb = [&](const void* buf, size_t size,
-                                        void** out_obj,
-                                        size_t* charge) -> Status {
-    std::string* blob = new std::string();
-    blob->assign(static_cast<const char*>(buf), size);
-    *out_obj = blob;
-    *charge = size;
-    return Status::OK();
+  Cache::CreateCallback create_cb = [](const void* buf, size_t size,
+                                       void** out_obj,
+                                       size_t* charge) -> Status {
+    CacheAllocationPtr allocation(new char[size]);
+
+    return BlobContents::CreateCallback(std::move(allocation), buf, size,
+                                        out_obj, charge);
  };

  {
@@ -1154,6 +1167,25 @@ TEST_F(BlobSecondaryCacheTest, GetBlobsFromSecondaryCache) {
    get_perf_context()->Reset();

    // key0 should be filled to the primary cache from the blob file.
+    ASSERT_OK(blob_source.GetBlob(read_options, keys[0], file_number,
+                                  blob_offsets[0], file_size, blob_sizes[0],
+                                  kNoCompression, nullptr /* prefetch_buffer */,
+                                  &values[0], nullptr /* bytes_read */));
+    // Release cache handle
+    values[0].Reset();
+
+    // key0 should be evicted and key0's dummy item is inserted into secondary
+    // cache. key1 should be filled to the primary cache from the blob file.
+    ASSERT_OK(blob_source.GetBlob(read_options, keys[1], file_number,
+                                  blob_offsets[1], file_size, blob_sizes[1],
+                                  kNoCompression, nullptr /* prefetch_buffer */,
+                                  &values[1], nullptr /* bytes_read */));
+
+    // Release cache handle
+    values[1].Reset();
+
+    // key0 should be filled to the primary cache from the blob file. key1
+    // should be evicted and key1's dummy item is inserted into secondary cache.
    ASSERT_OK(blob_source.GetBlob(read_options, keys[0], file_number,
                                  blob_offsets[0], file_size, blob_sizes[0],
                                  kNoCompression, nullptr /* prefetch_buffer */,
@@ -1162,8 +1194,11 @@ TEST_F(BlobSecondaryCacheTest, GetBlobsFromSecondaryCache) {
    ASSERT_TRUE(
        blob_source.TEST_BlobInCache(file_number, file_size, blob_offsets[0]));

-    // key0 should be demoted to the secondary cache, and key1 should be filled
-    // to the primary cache from the blob file.
+    // Release cache handle
+    values[0].Reset();
+
+    // key0 should be evicted and is inserted into secondary cache.
+    // key1 should be filled to the primary cache from the blob file.
    ASSERT_OK(blob_source.GetBlob(read_options, keys[1], file_number,
                                  blob_offsets[1], file_size, blob_sizes[1],
                                  kNoCompression, nullptr /* prefetch_buffer */,
@@ -1172,6 +1207,9 @@ TEST_F(BlobSecondaryCacheTest, GetBlobsFromSecondaryCache) {
    ASSERT_TRUE(
        blob_source.TEST_BlobInCache(file_number, file_size, blob_offsets[1]));

+    // Release cache handle
+    values[1].Reset();
+
    OffsetableCacheKey base_cache_key(db_id_, db_session_id_, file_number);

    // blob_cache here only looks at the primary cache since we didn't provide
@@ -1184,24 +1222,27 @@ TEST_F(BlobSecondaryCacheTest, GetBlobsFromSecondaryCache) {
      auto handle0 = blob_cache->Lookup(key0, statistics);
      ASSERT_EQ(handle0, nullptr);

-      // key0 should be in the secondary cache. After looking up key0 in the
-      // secondary cache, it will be erased from the secondary cache.
+      // key0's item should be in the secondary cache.
      bool is_in_sec_cache = false;
      auto sec_handle0 =
-          secondary_cache->Lookup(key0, create_cb, true, is_in_sec_cache);
+          secondary_cache->Lookup(key0, create_cb, true,
+                                  /*advise_erase=*/true, is_in_sec_cache);
      ASSERT_FALSE(is_in_sec_cache);
      ASSERT_NE(sec_handle0, nullptr);
      ASSERT_TRUE(sec_handle0->IsReady());
-      auto value = static_cast<std::string*>(sec_handle0->Value());
-      ASSERT_EQ(*value, blobs[0]);
+      auto value = static_cast<BlobContents*>(sec_handle0->Value());
+      ASSERT_NE(value, nullptr);
+      ASSERT_EQ(value->data(), blobs[0]);
      delete value;

-      // key0 doesn't exist in the blob cache
+      // key0 doesn't exist in the blob cache although key0's dummy
+      // item exist in the secondary cache.
      ASSERT_FALSE(blob_source.TEST_BlobInCache(file_number, file_size,
                                                blob_offsets[0]));
    }

-    // key1 should exist in the primary cache.
+    // key1 should exists in the primary cache. key1's dummy item exists
+    // in the secondary cache.
    {
      CacheKey cache_key = base_cache_key.WithOffset(blob_offsets[1]);
      const Slice key1 = cache_key.AsSlice();
@@ -1211,7 +1252,8 @@ TEST_F(BlobSecondaryCacheTest, GetBlobsFromSecondaryCache) {

      bool is_in_sec_cache = false;
      auto sec_handle1 =
-          secondary_cache->Lookup(key1, create_cb, true, is_in_sec_cache);
+          secondary_cache->Lookup(key1, create_cb, true,
+                                  /*advise_erase=*/true, is_in_sec_cache);
      ASSERT_FALSE(is_in_sec_cache);
      ASSERT_EQ(sec_handle1, nullptr);

@@ -1221,23 +1263,27 @@ TEST_F(BlobSecondaryCacheTest, GetBlobsFromSecondaryCache) {

    {
      // fetch key0 from the blob file to the primary cache.
+      // key1 is evicted and inserted into the secondary cache.
      ASSERT_OK(blob_source.GetBlob(
          read_options, keys[0], file_number, blob_offsets[0], file_size,
          blob_sizes[0], kNoCompression, nullptr /* prefetch_buffer */,
          &values[0], nullptr /* bytes_read */));
      ASSERT_EQ(values[0], blobs[0]);

+      // Release cache handle
+      values[0].Reset();
+
      // key0 should be in the primary cache.
      CacheKey cache_key0 = base_cache_key.WithOffset(blob_offsets[0]);
      const Slice key0 = cache_key0.AsSlice();
      auto handle0 = blob_cache->Lookup(key0, statistics);
      ASSERT_NE(handle0, nullptr);
-      auto value = static_cast<std::string*>(blob_cache->Value(handle0));
-      ASSERT_EQ(*value, blobs[0]);
+      auto value = static_cast<BlobContents*>(blob_cache->Value(handle0));
+      ASSERT_NE(value, nullptr);
+      ASSERT_EQ(value->data(), blobs[0]);
      blob_cache->Release(handle0);

-      // key1 is not in the primary cache, and it should be demoted to the
-      // secondary cache.
+      // key1 is not in the primary cache and is in the secondary cache.
      CacheKey cache_key1 = base_cache_key.WithOffset(blob_offsets[1]);
      const Slice key1 = cache_key1.AsSlice();
      auto handle1 = blob_cache->Lookup(key1, statistics);
@@ -1251,16 +1297,27 @@ TEST_F(BlobSecondaryCacheTest, GetBlobsFromSecondaryCache) {
      // key1 promotion should succeed due to the primary cache being empty. we
      // did't call secondary cache's Lookup() here, because it will remove the
      // key but it won't be able to promote the key to the primary cache.
-      // Instead we use the end-to-end blob source API to promote the key to
-      // the primary cache.
+      // Instead we use the end-to-end blob source API to read key1.
+      // In function TEST_BlobInCache, key1's dummy item is inserted into the
+      // primary cache and a standalone handle is checked by GetValue().
      ASSERT_TRUE(blob_source.TEST_BlobInCache(file_number, file_size,
                                               blob_offsets[1]));

-      // key1 should be in the primary cache.
+      // key1's dummy handle is in the primary cache and key1's item is still
+      // in the secondary cache. So, the primary cache's Lookup() without
+      // secondary cache support cannot see it. (NOTE: The dummy handle used
+      // to be a leaky abstraction but not anymore.)
+      handle1 = blob_cache->Lookup(key1, statistics);
+      ASSERT_EQ(handle1, nullptr);
+
+      // But after another access, it is promoted to primary cache
+      ASSERT_TRUE(blob_source.TEST_BlobInCache(file_number, file_size,
+                                               blob_offsets[1]));
+
+      // And Lookup() can find it (without secondary cache support)
      handle1 = blob_cache->Lookup(key1, statistics);
      ASSERT_NE(handle1, nullptr);
-      value = static_cast<std::string*>(blob_cache->Value(handle1));
-      ASSERT_EQ(*value, blobs[1]);
+      ASSERT_NE(blob_cache->Value(handle1), nullptr);
      blob_cache->Release(handle1);
    }
  }
@@ -1385,11 +1442,11 @@ TEST_F(BlobSourceCacheReservationTest, SimpleCacheReservation) {
  ReadOptions read_options;
  read_options.verify_checksums = true;

-  std::vector<PinnableSlice> values(keys_.size());
-
  {
    read_options.fill_cache = false;

+    std::vector<PinnableSlice> values(keys_.size());
+
    for (size_t i = 0; i < kNumBlobs; ++i) {
      ASSERT_OK(blob_source.GetBlob(
          read_options, keys_[i], kBlobFileNumber, blob_offsets[i],
@@ -1403,6 +1460,8 @@ TEST_F(BlobSourceCacheReservationTest, SimpleCacheReservation) {
  {
    read_options.fill_cache = true;

+    std::vector<PinnableSlice> values(keys_.size());
+
    // num_blobs is 16, so the total blob cache usage is less than a single
    // dummy entry. Therefore, cache reservation manager only reserves one dummy
    // entry here.
@@ -1412,7 +1471,12 @@ TEST_F(BlobSourceCacheReservationTest, SimpleCacheReservation) {
          read_options, keys_[i], kBlobFileNumber, blob_offsets[i],
          blob_file_size_, blob_sizes[i], kNoCompression,
          nullptr /* prefetch_buffer */, &values[i], nullptr /* bytes_read */));
-      blob_bytes += blob_sizes[i];
+
+      size_t charge = 0;
+      ASSERT_TRUE(blob_source.TEST_BlobInCache(kBlobFileNumber, blob_file_size_,
+                                               blob_offsets[i], &charge));
+
+      blob_bytes += charge;
      ASSERT_EQ(cache_res_mgr->GetTotalReservedCacheSize(), kSizeDummyEntry);
      ASSERT_EQ(cache_res_mgr->GetTotalMemoryUsed(), blob_bytes);
      ASSERT_EQ(cache_res_mgr->GetTotalMemoryUsed(),
@@ -1425,19 +1489,23 @@ TEST_F(BlobSourceCacheReservationTest, SimpleCacheReservation) {
    size_t blob_bytes = options_.blob_cache->GetUsage();

    for (size_t i = 0; i < kNumBlobs; ++i) {
+      size_t charge = 0;
+      ASSERT_TRUE(blob_source.TEST_BlobInCache(kBlobFileNumber, blob_file_size_,
+                                               blob_offsets[i], &charge));
+
      CacheKey cache_key = base_cache_key.WithOffset(blob_offsets[i]);
      // We didn't call options_.blob_cache->Erase() here, this is because
      // the cache wrapper's Erase() method must be called to update the
      // cache usage after erasing the cache entry.
      blob_source.GetBlobCache()->Erase(cache_key.AsSlice());
      if (i == kNumBlobs - 1) {
-        // The last blob is not in the cache. cache_res_mgr should not reserve
-        // any space for it.
+        // All the blobs got removed from the cache. cache_res_mgr should not
+        // reserve any space for them.
        ASSERT_EQ(cache_res_mgr->GetTotalReservedCacheSize(), 0);
      } else {
        ASSERT_EQ(cache_res_mgr->GetTotalReservedCacheSize(), kSizeDummyEntry);
      }
-      blob_bytes -= blob_sizes[i];
+      blob_bytes -= charge;
      ASSERT_EQ(cache_res_mgr->GetTotalMemoryUsed(), blob_bytes);
      ASSERT_EQ(cache_res_mgr->GetTotalMemoryUsed(),
                options_.blob_cache->GetUsage());
@@ -1495,11 +1563,11 @@ TEST_F(BlobSourceCacheReservationTest, IncreaseCacheReservationOnFullCache) {
  ReadOptions read_options;
  read_options.verify_checksums = true;

-  std::vector<PinnableSlice> values(keys_.size());
-
  {
    read_options.fill_cache = false;

+    std::vector<PinnableSlice> values(keys_.size());
+
    for (size_t i = 0; i < kNumBlobs; ++i) {
      ASSERT_OK(blob_source.GetBlob(
          read_options, keys_[i], kBlobFileNumber, blob_offsets[i],
@@ -1513,21 +1581,33 @@ TEST_F(BlobSourceCacheReservationTest, IncreaseCacheReservationOnFullCache) {
  {
    read_options.fill_cache = true;

-    // Since we resized each blob to be kSizeDummyEntry / (num_blobs/ 2), we
-    // should observe cache eviction for the second half blobs.
+    std::vector<PinnableSlice> values(keys_.size());
+
+    // Since we resized each blob to be kSizeDummyEntry / (num_blobs / 2), we
+    // can't fit all the blobs in the cache at the same time, which means we
+    // should observe cache evictions once we reach the cache's capacity.
+    // Due to the overhead of the cache and the BlobContents objects, as well as
+    // jemalloc bin sizes, this happens after inserting seven blobs.
    uint64_t blob_bytes = 0;
    for (size_t i = 0; i < kNumBlobs; ++i) {
      ASSERT_OK(blob_source.GetBlob(
          read_options, keys_[i], kBlobFileNumber, blob_offsets[i],
          blob_file_size_, blob_sizes[i], kNoCompression,
          nullptr /* prefetch_buffer */, &values[i], nullptr /* bytes_read */));
-      blob_bytes += blob_sizes[i];
-      ASSERT_EQ(cache_res_mgr->GetTotalReservedCacheSize(), kSizeDummyEntry);
-      if (i >= kNumBlobs / 2) {
-        ASSERT_EQ(cache_res_mgr->GetTotalMemoryUsed(), kSizeDummyEntry);
-      } else {
-        ASSERT_EQ(cache_res_mgr->GetTotalMemoryUsed(), blob_bytes);
+
+      // Release cache handle
+      values[i].Reset();
+
+      if (i < kNumBlobs / 2 - 1) {
+        size_t charge = 0;
+        ASSERT_TRUE(blob_source.TEST_BlobInCache(
+            kBlobFileNumber, blob_file_size_, blob_offsets[i], &charge));
+
+        blob_bytes += charge;
      }
+
+      ASSERT_EQ(cache_res_mgr->GetTotalReservedCacheSize(), kSizeDummyEntry);
+      ASSERT_EQ(cache_res_mgr->GetTotalMemoryUsed(), blob_bytes);
      ASSERT_EQ(cache_res_mgr->GetTotalMemoryUsed(),
                options_.blob_cache->GetUsage());
    }
@@ -1538,6 +1618,7 @@ TEST_F(BlobSourceCacheReservationTest, IncreaseCacheReservationOnFullCache) {
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -7,6 +7,7 @@
 #include <sstream>
 #include <string>

+#include "cache/compressed_secondary_cache.h"
 #include "db/blob/blob_index.h"
 #include "db/blob/blob_log_format.h"
 #include "db/db_test_util.h"
@@ -53,7 +54,7 @@ TEST_F(DBBlobBasicTest, GetBlobFromCache) {
  Options options = GetDefaultOptions();

  LRUCacheOptions co;
-  co.capacity = 2048;
+  co.capacity = 2 << 20;  // 2MB
  co.num_shard_bits = 2;
  co.metadata_charge_policy = kDontChargeCacheMetadata;
  auto backing_cache = NewLRUCache(co);
@@ -121,7 +122,7 @@ TEST_F(DBBlobBasicTest, IterateBlobsFromCache) {
  Options options = GetDefaultOptions();

  LRUCacheOptions co;
-  co.capacity = 2048;
+  co.capacity = 2 << 20;  // 2MB
  co.num_shard_bits = 2;
  co.metadata_charge_policy = kDontChargeCacheMetadata;
  auto backing_cache = NewLRUCache(co);
@@ -452,7 +453,7 @@ TEST_F(DBBlobBasicTest, MultiGetBlobsFromCache) {
  Options options = GetDefaultOptions();

  LRUCacheOptions co;
-  co.capacity = 2048;
+  co.capacity = 2 << 20;  // 2MB
  co.num_shard_bits = 2;
  co.metadata_charge_policy = kDontChargeCacheMetadata;
  auto backing_cache = NewLRUCache(co);
@@ -1701,6 +1702,83 @@ TEST_F(DBBlobBasicTest, DynamicallyWarmCacheDuringFlush) {
 }
 #endif  // !ROCKSDB_LITE

+TEST_F(DBBlobBasicTest, WarmCacheWithBlobsSecondary) {
+  CompressedSecondaryCacheOptions secondary_cache_opts;
+  secondary_cache_opts.capacity = 1 << 20;
+  secondary_cache_opts.num_shard_bits = 0;
+  secondary_cache_opts.metadata_charge_policy = kDontChargeCacheMetadata;
+  secondary_cache_opts.compression_type = kNoCompression;
+
+  LRUCacheOptions primary_cache_opts;
+  primary_cache_opts.capacity = 1024;
+  primary_cache_opts.num_shard_bits = 0;
+  primary_cache_opts.metadata_charge_policy = kDontChargeCacheMetadata;
+  primary_cache_opts.secondary_cache =
+      NewCompressedSecondaryCache(secondary_cache_opts);
+
+  Options options = GetDefaultOptions();
+  options.create_if_missing = true;
+  options.statistics = CreateDBStatistics();
+  options.enable_blob_files = true;
+  options.blob_cache = NewLRUCache(primary_cache_opts);
+  options.prepopulate_blob_cache = PrepopulateBlobCache::kFlushOnly;
+
+  DestroyAndReopen(options);
+
+  // Note: only one of the two blobs fit in the primary cache at any given time.
+  constexpr char first_key[] = "foo";
+  constexpr size_t first_blob_size = 512;
+  const std::string first_blob(first_blob_size, 'a');
+
+  constexpr char second_key[] = "bar";
+  constexpr size_t second_blob_size = 768;
+  const std::string second_blob(second_blob_size, 'b');
+
+  // First blob is inserted into primary cache during flush.
+  ASSERT_OK(Put(first_key, first_blob));
+  ASSERT_OK(Flush());
+  ASSERT_EQ(options.statistics->getAndResetTickerCount(BLOB_DB_CACHE_ADD), 1);
+
+  // Second blob is inserted into primary cache during flush,
+  // First blob is evicted but only a dummy handle is inserted into secondary
+  // cache.
+  ASSERT_OK(Put(second_key, second_blob));
+  ASSERT_OK(Flush());
+  ASSERT_EQ(options.statistics->getAndResetTickerCount(BLOB_DB_CACHE_ADD), 1);
+
+  // First blob is inserted into primary cache.
+  // Second blob is evicted but only a dummy handle is inserted into secondary
+  // cache.
+  ASSERT_EQ(Get(first_key), first_blob);
+  ASSERT_EQ(options.statistics->getAndResetTickerCount(BLOB_DB_CACHE_MISS), 1);
+  ASSERT_EQ(options.statistics->getAndResetTickerCount(BLOB_DB_CACHE_HIT), 0);
+  ASSERT_EQ(options.statistics->getAndResetTickerCount(SECONDARY_CACHE_HITS),
+            0);
+  // Second blob is inserted into primary cache,
+  // First blob is evicted and is inserted into secondary cache.
+  ASSERT_EQ(Get(second_key), second_blob);
+  ASSERT_EQ(options.statistics->getAndResetTickerCount(BLOB_DB_CACHE_MISS), 1);
+  ASSERT_EQ(options.statistics->getAndResetTickerCount(BLOB_DB_CACHE_HIT), 0);
+  ASSERT_EQ(options.statistics->getAndResetTickerCount(SECONDARY_CACHE_HITS),
+            0);
+
+  // First blob's dummy item is inserted into primary cache b/c of lookup.
+  // Second blob is still in primary cache.
+  ASSERT_EQ(Get(first_key), first_blob);
+  ASSERT_EQ(options.statistics->getAndResetTickerCount(BLOB_DB_CACHE_MISS), 0);
+  ASSERT_EQ(options.statistics->getAndResetTickerCount(BLOB_DB_CACHE_HIT), 1);
+  ASSERT_EQ(options.statistics->getAndResetTickerCount(SECONDARY_CACHE_HITS),
+            1);
+
+  // First blob's item is inserted into primary cache b/c of lookup.
+  // Second blob is evicted and inserted into secondary cache.
+  ASSERT_EQ(Get(first_key), first_blob);
+  ASSERT_EQ(options.statistics->getAndResetTickerCount(BLOB_DB_CACHE_MISS), 0);
+  ASSERT_EQ(options.statistics->getAndResetTickerCount(BLOB_DB_CACHE_HIT), 1);
+  ASSERT_EQ(options.statistics->getAndResetTickerCount(SECONDARY_CACHE_HITS),
+            1);
+}
+
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
@@ -90,7 +90,7 @@ Status BuildTable(
  iter->SeekToFirst();
  std::unique_ptr<CompactionRangeDelAggregator> range_del_agg(
      new CompactionRangeDelAggregator(&tboptions.internal_comparator,
-                                       snapshots));
+                                       snapshots, full_history_ts_low));
  uint64_t num_unfragmented_tombstones = 0;
  uint64_t total_tombstone_payload_bytes = 0;
  for (auto& range_del_iter : range_del_iters) {
@@ -1054,6 +1054,18 @@ void rocksdb_drop_column_family(
  SaveError(errptr, db->rep->DropColumnFamily(handle->rep));
 }

+uint32_t rocksdb_column_family_handle_get_id(
+    rocksdb_column_family_handle_t* handle) {
+  return handle->rep->GetID();
+}
+
+char* rocksdb_column_family_handle_get_name(
+    rocksdb_column_family_handle_t* handle, size_t* name_len) {
+  auto name = handle->rep->GetName();
+  *name_len = name.size();
+  return CopyString(name);
+}
+
 void rocksdb_column_family_handle_destroy(rocksdb_column_family_handle_t* handle) {
  delete handle->rep;
  delete handle;
@@ -2606,6 +2618,11 @@ void rocksdb_block_based_options_destroy(
  delete options;
 }

+void rocksdb_block_based_options_set_checksum(
+    rocksdb_block_based_table_options_t* opt, char v) {
+  opt->rep.checksum = static_cast<ROCKSDB_NAMESPACE::ChecksumType>(v);
+}
+
 void rocksdb_block_based_options_set_block_size(
    rocksdb_block_based_table_options_t* options, size_t block_size) {
  options->rep.block_size = block_size;
@@ -4099,6 +4116,8 @@ uint64_t rocksdb_perfcontext_metric(rocksdb_perfcontext_t* context,
      return rep->blob_checksum_time;
    case rocksdb_blob_decompress_time:
      return rep->blob_decompress_time;
+    case rocksdb_internal_range_del_reseek_count:
+      return rep->internal_range_del_reseek_count;
    default:
      break;
  }
@@ -6449,6 +6468,16 @@ void rocksdb_options_set_memtable_whole_key_filtering(rocksdb_options_t* opt,
  opt->rep.memtable_whole_key_filtering = val;
 }

+void rocksdb_options_set_avoid_unnecessary_blocking_io(rocksdb_options_t* opt,
+                                                       unsigned char val) {
+  opt->rep.avoid_unnecessary_blocking_io = val;
+}
+
+unsigned char rocksdb_options_get_avoid_unnecessary_blocking_io(
+    rocksdb_options_t* opt) {
+  return opt->rep.avoid_unnecessary_blocking_io;
+}
+
 // deletes container with memory usage estimates
 void rocksdb_approximate_memory_usage_destroy(rocksdb_memory_usage_t* usage) {
  delete usage;
@@ -233,6 +233,21 @@ ColumnFamilyOptions SanitizeOptions(const ImmutableDBOptions& db_options,
    result.min_write_buffer_number_to_merge = 1;
  }

+  if (db_options.atomic_flush && result.min_write_buffer_number_to_merge > 1) {
+    ROCKS_LOG_WARN(
+        db_options.logger,
+        "Currently, if atomic_flush is true, then triggering flush for any "
+        "column family internally (non-manual flush) will trigger flushing "
+        "all column families even if the number of memtables is smaller "
+        "min_write_buffer_number_to_merge. Therefore, configuring "
+        "min_write_buffer_number_to_merge > 1 is not compatible and should "
+        "be satinized to 1. Not doing so will lead to data loss and "
+        "inconsistent state across multiple column families when WAL is "
+        "disabled, which is a common setting for atomic flush");
+
+    result.min_write_buffer_number_to_merge = 1;
+  }
+
  if (result.num_levels < 1) {
    result.num_levels = 1;
  }
@@ -276,7 +291,6 @@ ColumnFamilyOptions SanitizeOptions(const ImmutableDBOptions& db_options,
  }

  if (result.compaction_style == kCompactionStyleFIFO) {
-    result.num_levels = 1;
    // since we delete level0 files in FIFO compaction when there are too many
    // of them, these options don't really mean anything
    result.level0_slowdown_writes_trigger = std::numeric_limits<int>::max();
@@ -1144,7 +1158,8 @@ Status ColumnFamilyData::RangesOverlapWithMemtables(
  MergeIteratorBuilder merge_iter_builder(&internal_comparator_, &arena);
  merge_iter_builder.AddIterator(
      super_version->mem->NewIterator(read_opts, &arena));
-  super_version->imm->AddIterators(read_opts, &merge_iter_builder);
+  super_version->imm->AddIterators(read_opts, &merge_iter_builder,
+                                   false /* add_range_tombstone_iter */);
  ScopedArenaIterator memtable_iter(merge_iter_builder.Finish());

  auto read_seq = super_version->current->version_set()->LastSequence();
@@ -1415,6 +1430,14 @@ Status ColumnFamilyData::ValidateOptions(
        "FIFO compaction only supported with max_open_files = -1.");
  }

+  std::vector<uint32_t> supported{0, 1, 2, 4, 8};
+  if (std::find(supported.begin(), supported.end(),
+                cf_options.memtable_protection_bytes_per_key) ==
+      supported.end()) {
+    return Status::NotSupported(
+        "Memtable per key-value checksum protection only supports 0, 1, 2, 4 "
+        "or 8 bytes per key.");
+  }
  return s;
 }

@@ -3433,6 +3433,27 @@ TEST(ColumnFamilyTest, ValidateBlobGCForceThreshold) {
                  .IsInvalidArgument());
 }

+TEST(ColumnFamilyTest, ValidateMemtableKVChecksumOption) {
+  DBOptions db_options;
+
+  ColumnFamilyOptions cf_options;
+  ASSERT_OK(ColumnFamilyData::ValidateOptions(db_options, cf_options));
+
+  cf_options.memtable_protection_bytes_per_key = 5;
+  ASSERT_TRUE(ColumnFamilyData::ValidateOptions(db_options, cf_options)
+                  .IsNotSupported());
+
+  cf_options.memtable_protection_bytes_per_key = 1;
+  ASSERT_OK(ColumnFamilyData::ValidateOptions(db_options, cf_options));
+
+  cf_options.memtable_protection_bytes_per_key = 16;
+  ASSERT_TRUE(ColumnFamilyData::ValidateOptions(db_options, cf_options)
+                  .IsNotSupported());
+
+  cf_options.memtable_protection_bytes_per_key = 0;
+  ASSERT_OK(ColumnFamilyData::ValidateOptions(db_options, cf_options));
+}
+
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
@@ -490,6 +490,7 @@ TEST_F(CompactFilesTest, GetCompactionJobInfo) {
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -253,6 +253,7 @@ INSTANTIATE_TEST_CASE_P(
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -221,7 +221,7 @@ Compaction::Compaction(
    : input_vstorage_(vstorage),
      start_level_(_inputs[0].level),
      output_level_(_output_level),
-      max_output_file_size_(_target_file_size),
+      target_output_file_size_(_target_file_size),
      max_compaction_bytes_(_max_compaction_bytes),
      max_subcompactions_(_max_subcompactions),
      immutable_options_(_immutable_options),
@@ -259,7 +259,7 @@ Compaction::Compaction(
              ? mutable_cf_options()->blob_garbage_collection_age_cutoff
              : _blob_garbage_collection_age_cutoff),
      penultimate_level_(EvaluatePenultimateLevel(
-          immutable_options_, start_level_, output_level_)) {
+          vstorage, immutable_options_, start_level_, output_level_)) {
  MarkFilesBeingCompacted(true);
  if (is_manual_compaction_) {
    compaction_reason_ = CompactionReason::kManualCompaction;
@@ -268,6 +268,14 @@ Compaction::Compaction(
    max_subcompactions_ = _mutable_db_options.max_subcompactions;
  }

+  // for the non-bottommost levels, it tries to build files match the target
+  // file size, but not guaranteed. It could be 2x the size of the target size.
+  max_output_file_size_ =
+      bottommost_level_ || grandparents_.empty() ||
+              !_immutable_options.level_compaction_dynamic_file_size
+          ? target_output_file_size_
+          : 2 * target_output_file_size_;
+
 #ifndef NDEBUG
  for (size_t i = 1; i < inputs_.size(); ++i) {
    assert(inputs_[i].level > inputs_[i - 1].level);
@@ -314,9 +322,67 @@ void Compaction::PopulatePenultimateLevelOutputRange() {
    return;
  }

+  // exclude the last level, the range of all input levels is the safe range
+  // of keys that can be moved up.
+  int exclude_level = number_levels_ - 1;
+  penultimate_output_range_type_ = PenultimateOutputRangeType::kNonLastRange;
+
+  // For universal compaction, the penultimate_output_range could be extended if
+  // all penultimate level files are included in the compaction (which includes
+  // the case that the penultimate level is empty).
+  if (immutable_options_.compaction_style == kCompactionStyleUniversal) {
+    exclude_level = kInvalidLevel;
+    std::set<uint64_t> penultimate_inputs;
+    for (const auto& input_lvl : inputs_) {
+      if (input_lvl.level == penultimate_level_) {
+        for (const auto& file : input_lvl.files) {
+          penultimate_inputs.emplace(file->fd.GetNumber());
+        }
+      }
+    }
+    auto penultimate_files = input_vstorage_->LevelFiles(penultimate_level_);
+    for (const auto& file : penultimate_files) {
+      if (penultimate_inputs.find(file->fd.GetNumber()) ==
+          penultimate_inputs.end()) {
+        exclude_level = number_levels_ - 1;
+        penultimate_output_range_type_ = PenultimateOutputRangeType::kFullRange;
+        break;
+      }
+    }
+  }
+
  GetBoundaryKeys(input_vstorage_, inputs_,
                  &penultimate_level_smallest_user_key_,
-                  &penultimate_level_largest_user_key_, number_levels_ - 1);
+                  &penultimate_level_largest_user_key_, exclude_level);
+
+  // If there's a case that the penultimate level output range is overlapping
+  // with the existing files, disable the penultimate level output by setting
+  // the range to empty. One example is the range delete could have overlap
+  // boundary with the next file. (which is actually a false overlap)
+  // TODO: Exclude such false overlap, so it won't disable the penultimate
+  //  output.
+  std::set<uint64_t> penultimate_inputs;
+  for (const auto& input_lvl : inputs_) {
+    if (input_lvl.level == penultimate_level_) {
+      for (const auto& file : input_lvl.files) {
+        penultimate_inputs.emplace(file->fd.GetNumber());
+      }
+    }
+  }
+
+  auto penultimate_files = input_vstorage_->LevelFiles(penultimate_level_);
+  for (const auto& file : penultimate_files) {
+    if (penultimate_inputs.find(file->fd.GetNumber()) ==
+            penultimate_inputs.end() &&
+        OverlapPenultimateLevelOutputRange(file->smallest.user_key(),
+                                           file->largest.user_key())) {
+      // basically disable the penultimate range output. which should be rare
+      // or a false overlap caused by range del
+      penultimate_level_smallest_user_key_ = "";
+      penultimate_level_largest_user_key_ = "";
+      penultimate_output_range_type_ = PenultimateOutputRangeType::kDisabled;
+    }
+  }
 }

 Compaction::~Compaction() {
@@ -334,6 +400,8 @@ bool Compaction::SupportsPerKeyPlacement() const {

 int Compaction::GetPenultimateLevel() const { return penultimate_level_; }

+// smallest_key and largest_key include timestamps if user-defined timestamp is
+// enabled.
 bool Compaction::OverlapPenultimateLevelOutputRange(
    const Slice& smallest_key, const Slice& largest_key) const {
  if (!SupportsPerKeyPlacement()) {
@@ -342,21 +410,30 @@ bool Compaction::OverlapPenultimateLevelOutputRange(
  const Comparator* ucmp =
      input_vstorage_->InternalComparator()->user_comparator();

-  return ucmp->Compare(smallest_key, penultimate_level_largest_user_key_) <=
-             0 &&
-         ucmp->Compare(largest_key, penultimate_level_smallest_user_key_) >= 0;
+  return ucmp->CompareWithoutTimestamp(
+             smallest_key, penultimate_level_largest_user_key_) <= 0 &&
+         ucmp->CompareWithoutTimestamp(
+             largest_key, penultimate_level_smallest_user_key_) >= 0;
 }

+// key includes timestamp if user-defined timestamp is enabled.
 bool Compaction::WithinPenultimateLevelOutputRange(const Slice& key) const {
  if (!SupportsPerKeyPlacement()) {
    return false;
  }

+  if (penultimate_level_smallest_user_key_.empty() ||
+      penultimate_level_largest_user_key_.empty()) {
+    return false;
+  }
+
  const Comparator* ucmp =
      input_vstorage_->InternalComparator()->user_comparator();

-  return ucmp->Compare(key, penultimate_level_smallest_user_key_) >= 0 &&
-         ucmp->Compare(key, penultimate_level_largest_user_key_) <= 0;
+  return ucmp->CompareWithoutTimestamp(
+             key, penultimate_level_smallest_user_key_) >= 0 &&
+         ucmp->CompareWithoutTimestamp(
+             key, penultimate_level_largest_user_key_) <= 0;
 }

 bool Compaction::InputCompressionMatchesOutput() const {
@@ -731,6 +808,7 @@ uint64_t Compaction::MinInputFileOldestAncesterTime(
 }

 int Compaction::EvaluatePenultimateLevel(
+    const VersionStorageInfo* vstorage,
    const ImmutableOptions& immutable_options, const int start_level,
    const int output_level) {
  // TODO: currently per_key_placement feature only support level and universal
@@ -745,7 +823,19 @@ int Compaction::EvaluatePenultimateLevel(

  int penultimate_level = output_level - 1;
  assert(penultimate_level < immutable_options.num_levels);
-  if (penultimate_level <= 0 || penultimate_level < start_level) {
+  if (penultimate_level <= 0) {
+    return kInvalidLevel;
+  }
+
+  // If the penultimate level is not within input level -> output level range
+  // check if the penultimate output level is empty, if it's empty, it could
+  // also be locked for the penultimate output.
+  // TODO: ideally, it only needs to check if there's a file within the
+  //  compaction output key range. For simplicity, it just check if there's any
+  //  file on the penultimate level.
+  if (start_level == immutable_options.num_levels - 1 &&
+      (immutable_options.compaction_style != kCompactionStyleUniversal ||
+       !vstorage->LevelFiles(penultimate_level).empty())) {
    return kInvalidLevel;
  }

@@ -87,6 +87,15 @@ class Compaction {
                 BlobGarbageCollectionPolicy::kUseDefault,
             double blob_garbage_collection_age_cutoff = -1);

+  // The type of the penultimate level output range
+  enum class PenultimateOutputRangeType : int {
+    kNotSupported,  // it cannot output to the penultimate level
+    kFullRange,     // any data could be output to the penultimate level
+    kNonLastRange,  // only the keys within non_last_level compaction inputs can
+                    // be outputted to the penultimate level
+    kDisabled,      // no data can be outputted to the penultimate level
+  };
+
  // No copying allowed
  Compaction(const Compaction&) = delete;
  void operator=(const Compaction&) = delete;
@@ -163,6 +172,9 @@ class Compaction {
  // Maximum size of files to build during this compaction.
  uint64_t max_output_file_size() const { return max_output_file_size_; }

+  // Target output file size for this compaction
+  uint64_t target_output_file_size() const { return target_output_file_size_; }
+
  // What compression for output
  CompressionType output_compression() const { return output_compression_; }

@@ -307,6 +319,18 @@ class Compaction {

  Slice GetLargestUserKey() const { return largest_user_key_; }

+  Slice GetPenultimateLevelSmallestUserKey() const {
+    return penultimate_level_smallest_user_key_;
+  }
+
+  Slice GetPenultimateLevelLargestUserKey() const {
+    return penultimate_level_largest_user_key_;
+  }
+
+  PenultimateOutputRangeType GetPenultimateOutputRangeType() const {
+    return penultimate_output_range_type_;
+  }
+
  // Return true if the compaction supports per_key_placement
  bool SupportsPerKeyPlacement() const;

@@ -316,6 +340,8 @@ class Compaction {

  // Return true if the given range is overlap with penultimate level output
  // range.
+  // Both smallest_key and largest_key include timestamps if user-defined
+  // timestamp is enabled.
  bool OverlapPenultimateLevelOutputRange(const Slice& smallest_key,
                                          const Slice& largest_key) const;

@@ -325,6 +351,7 @@ class Compaction {
  // If per_key_placement is not supported, always return false.
  // TODO: currently it doesn't support moving data from the last level to the
  //  penultimate level
+  //  key includes timestamp if user-defined timestamp is enabled.
  bool WithinPenultimateLevelOutputRange(const Slice& key) const;

  CompactionReason compaction_reason() const { return compaction_reason_; }
@@ -363,11 +390,18 @@ class Compaction {
  }

  static constexpr int kInvalidLevel = -1;
+
  // Evaluate penultimate output level. If the compaction supports
  // per_key_placement feature, it returns the penultimate level number.
  // Otherwise, it's set to kInvalidLevel (-1), which means
  // output_to_penultimate_level is not supported.
-  static int EvaluatePenultimateLevel(const ImmutableOptions& immutable_options,
+  // Note: even the penultimate level output is supported (PenultimateLevel !=
+  // kInvalidLevel), some key range maybe unsafe to be outputted to the
+  // penultimate level. The safe key range is populated by
+  // `PopulatePenultimateLevelOutputRange()`.
+  // Which could potentially disable all penultimate level output.
+  static int EvaluatePenultimateLevel(const VersionStorageInfo* vstorage,
+                                      const ImmutableOptions& immutable_options,
                                      const int start_level,
                                      const int output_level);

@@ -384,11 +418,6 @@ class Compaction {
  // populate penultimate level output range, which will be used to determine if
  // a key is safe to output to the penultimate level (details see
  // `Compaction::WithinPenultimateLevelOutputRange()`.
-  // TODO: Currently the penultimate level output range is the min/max keys of
-  //  non-last-level input files. Which is only good if there's no key moved
-  //  from the last level to the penultimate level. For a more complicated per
-  //  key placement which may move data from the last level to the penultimate
-  //  level, it needs extra check.
  void PopulatePenultimateLevelOutputRange();

  // Get the atomic file boundaries for all files in the compaction. Necessary
@@ -412,6 +441,7 @@ class Compaction {

  const int start_level_;   // the lowest level to be compacted
  const int output_level_;  // levels to which output files are stored
+  uint64_t target_output_file_size_;
  uint64_t max_output_file_size_;
  uint64_t max_compaction_bytes_;
  uint32_t max_subcompactions_;
@@ -470,9 +500,11 @@ class Compaction {
  TablePropertiesCollection output_table_properties_;

  // smallest user keys in compaction
+  // includes timestamp if user-defined timestamp is enabled.
  Slice smallest_user_key_;

  // largest user keys in compaction
+  // includes timestamp if user-defined timestamp is enabled.
  Slice largest_user_key_;

  // Reason for compaction
@@ -493,8 +525,12 @@ class Compaction {
  const int penultimate_level_;

  // Key range for penultimate level output
+  // includes timestamp if user-defined timestamp is enabled.
+  // penultimate_output_range_type_ shows the range type
  Slice penultimate_level_smallest_user_key_;
  Slice penultimate_level_largest_user_key_;
+  PenultimateOutputRangeType penultimate_output_range_type_ =
+      PenultimateOutputRangeType::kNotSupported;
 };

 #ifndef NDEBUG
@@ -34,7 +34,8 @@ CompactionIterator::CompactionIterator(
    const std::atomic<bool>* shutting_down,
    const std::shared_ptr<Logger> info_log,
    const std::string* full_history_ts_low,
-    const SequenceNumber penultimate_level_cutoff_seqno)
+    const SequenceNumber preserve_time_min_seqno,
+    const SequenceNumber preclude_last_level_min_seqno)
    : CompactionIterator(
          input, cmp, merge_helper, last_sequence, snapshots,
          earliest_write_conflict_snapshot, job_snapshot, snapshot_checker, env,
@@ -44,7 +45,7 @@ CompactionIterator::CompactionIterator(
          std::unique_ptr<CompactionProxy>(
              compaction ? new RealCompaction(compaction) : nullptr),
          compaction_filter, shutting_down, info_log, full_history_ts_low,
-          penultimate_level_cutoff_seqno) {}
+          preserve_time_min_seqno, preclude_last_level_min_seqno) {}

 CompactionIterator::CompactionIterator(
    InternalIterator* input, const Comparator* cmp, MergeHelper* merge_helper,
@@ -61,7 +62,8 @@ CompactionIterator::CompactionIterator(
    const std::atomic<bool>* shutting_down,
    const std::shared_ptr<Logger> info_log,
    const std::string* full_history_ts_low,
-    const SequenceNumber penultimate_level_cutoff_seqno)
+    const SequenceNumber preserve_time_min_seqno,
+    const SequenceNumber preclude_last_level_min_seqno)
    : input_(input, cmp,
             !compaction || compaction->DoesInputReferenceBlobFiles()),
      cmp_(cmp),
@@ -105,8 +107,10 @@ CompactionIterator::CompactionIterator(
      current_key_committed_(false),
      cmp_with_history_ts_low_(0),
      level_(compaction_ == nullptr ? 0 : compaction_->level()),
-      penultimate_level_cutoff_seqno_(penultimate_level_cutoff_seqno) {
+      preserve_time_min_seqno_(preserve_time_min_seqno),
+      preclude_last_level_min_seqno_(preclude_last_level_min_seqno) {
  assert(snapshots_ != nullptr);
+  assert(preserve_time_min_seqno_ <= preclude_last_level_min_seqno_);

  if (compaction_ != nullptr) {
    level_ptrs_ = std::vector<size_t>(compaction_->number_levels(), 0);
@@ -914,8 +918,17 @@ void CompactionIterator::NextFromInput() {
    } else {
      // 1. new user key -OR-
      // 2. different snapshot stripe
-      bool should_delete = range_del_agg_->ShouldDelete(
-          key_, RangeDelPositioningMode::kForwardTraversal);
+      // If user-defined timestamp is enabled, we consider keys for GC if they
+      // are below history_ts_low_. CompactionRangeDelAggregator::ShouldDelete()
+      // only considers range deletions that are at or below history_ts_low_ and
+      // trim_ts_. We drop keys here that are below history_ts_low_ and are
+      // covered by a range tombstone that is at or below history_ts_low_ and
+      // trim_ts.
+      bool should_delete = false;
+      if (!timestamp_size_ || cmp_with_history_ts_low_ < 0) {
+        should_delete = range_del_agg_->ShouldDelete(
+            key_, RangeDelPositioningMode::kForwardTraversal);
+      }
      if (should_delete) {
        ++iter_stats_.num_record_drop_hidden;
        ++iter_stats_.num_record_drop_range_del;
@@ -1079,6 +1092,7 @@ void CompactionIterator::GarbageCollectBlobIfNeeded() {
 }

 void CompactionIterator::DecideOutputLevel() {
+  assert(compaction_->SupportsPerKeyPlacement());
 #ifndef NDEBUG
  // Could be overridden by unittest
  PerKeyPlacementContext context(level_, ikey_.user_key, value_,
@@ -1086,11 +1100,13 @@ void CompactionIterator::DecideOutputLevel() {
  TEST_SYNC_POINT_CALLBACK("CompactionIterator::PrepareOutput.context",
                           &context);
  output_to_penultimate_level_ = context.output_to_penultimate_level;
-#endif /* !NDEBUG */
+#else
+  output_to_penultimate_level_ = false;
+#endif  // NDEBUG

  // if the key is newer than the cutoff sequence or within the earliest
  // snapshot, it should output to the penultimate level.
-  if (ikey_.sequence > penultimate_level_cutoff_seqno_ ||
+  if (ikey_.sequence > preclude_last_level_min_seqno_ ||
      ikey_.sequence > earliest_snapshot_) {
    output_to_penultimate_level_ = true;
  }
@@ -1152,7 +1168,7 @@ void CompactionIterator::PrepareOutput() {
        DefinitelyInSnapshot(ikey_.sequence, earliest_snapshot_) &&
        ikey_.type != kTypeMerge && current_key_committed_ &&
        !output_to_penultimate_level_ &&
-        ikey_.sequence < penultimate_level_cutoff_seqno_) {
+        ikey_.sequence < preserve_time_min_seqno_) {
      if (ikey_.type == kTypeDeletion ||
          (ikey_.type == kTypeSingleDeletion && timestamp_size_ == 0)) {
        ROCKS_LOG_FATAL(
@@ -88,6 +88,7 @@ class CompactionIterator {

    virtual int number_levels() const = 0;

+    // Result includes timestamp if user-defined timestamp is enabled.
    virtual Slice GetLargestUserKey() const = 0;

    virtual bool allow_ingest_behind() const = 0;
@@ -108,6 +109,7 @@ class CompactionIterator {

    virtual bool SupportsPerKeyPlacement() const = 0;

+    // `key` includes timestamp if user-defined timestamp is enabled.
    virtual bool WithinPenultimateLevelOutputRange(const Slice& key) const = 0;
  };

@@ -133,6 +135,7 @@ class CompactionIterator {

    int number_levels() const override { return compaction_->number_levels(); }

+    // Result includes timestamp if user-defined timestamp is enabled.
    Slice GetLargestUserKey() const override {
      return compaction_->GetLargestUserKey();
    }
@@ -173,6 +176,7 @@ class CompactionIterator {

    // Check if key is within penultimate level output range, to see if it's
    // safe to output to the penultimate level for per_key_placement feature.
+    // `key` includes timestamp if user-defined timestamp is enabled.
    bool WithinPenultimateLevelOutputRange(const Slice& key) const override {
      return compaction_->WithinPenultimateLevelOutputRange(key);
    }
@@ -196,7 +200,8 @@ class CompactionIterator {
      const std::atomic<bool>* shutting_down = nullptr,
      const std::shared_ptr<Logger> info_log = nullptr,
      const std::string* full_history_ts_low = nullptr,
-      const SequenceNumber penultimate_level_cutoff_seqno = kMaxSequenceNumber);
+      const SequenceNumber preserve_time_min_seqno = kMaxSequenceNumber,
+      const SequenceNumber preclude_last_level_min_seqno = kMaxSequenceNumber);

  // Constructor with custom CompactionProxy, used for tests.
  CompactionIterator(
@@ -214,7 +219,8 @@ class CompactionIterator {
      const std::atomic<bool>* shutting_down = nullptr,
      const std::shared_ptr<Logger> info_log = nullptr,
      const std::string* full_history_ts_low = nullptr,
-      const SequenceNumber penultimate_level_cutoff_seqno = kMaxSequenceNumber);
+      const SequenceNumber preserve_time_min_seqno = kMaxSequenceNumber,
+      const SequenceNumber preclude_last_level_min_seqno = kMaxSequenceNumber);

  ~CompactionIterator();

@@ -466,9 +472,12 @@ class CompactionIterator {
  // output to.
  bool output_to_penultimate_level_{false};

-  // any key later than this sequence number should have
-  // output_to_penultimate_level_ set to true
-  const SequenceNumber penultimate_level_cutoff_seqno_ = kMaxSequenceNumber;
+  // min seqno for preserving the time information.
+  const SequenceNumber preserve_time_min_seqno_ = kMaxSequenceNumber;
+
+  // min seqno to preclude the data from the last level, if the key seqno larger
+  // than this, it will be output to penultimate level
+  const SequenceNumber preclude_last_level_min_seqno_ = kMaxSequenceNumber;

  void AdvanceInputIter() { input_.Next(); }

@@ -1497,6 +1497,7 @@ INSTANTIATE_TEST_CASE_P(CompactionIteratorTsGcTestInstance,
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -97,6 +97,8 @@ const char* GetCompactionReasonString(CompactionReason compaction_reason) {
      return "ChangeTemperature";
    case CompactionReason::kForcedBlobGC:
      return "ForcedBlobGC";
+    case CompactionReason::kRoundRobinTtl:
+      return "RoundRobinTtl";
    case CompactionReason::kNumOfReasons:
      // fall through
    default:
@@ -105,6 +107,23 @@ const char* GetCompactionReasonString(CompactionReason compaction_reason) {
  }
 }

+const char* GetCompactionPenultimateOutputRangeTypeString(
+    Compaction::PenultimateOutputRangeType range_type) {
+  switch (range_type) {
+    case Compaction::PenultimateOutputRangeType::kNotSupported:
+      return "NotSupported";
+    case Compaction::PenultimateOutputRangeType::kFullRange:
+      return "FullRange";
+    case Compaction::PenultimateOutputRangeType::kNonLastRange:
+      return "NonLastRange";
+    case Compaction::PenultimateOutputRangeType::kDisabled:
+      return "Disabled";
+    default:
+      assert(false);
+      return "Invalid";
+  }
+}
+
 CompactionJob::CompactionJob(
    int job_id, Compaction* compaction, const ImmutableDBOptions& db_options,
    const MutableDBOptions& mutable_db_options, const FileOptions& file_options,
@@ -254,7 +273,7 @@ void CompactionJob::Prepare() {
      // timestamp part).
      assert(i == 0 || i == boundaries_.size() ||
             cfd->user_comparator()->CompareWithoutTimestamp(
-                 boundaries_[i - 1], true, boundaries_[i], true) < 0);
+                 boundaries_[i - 1], boundaries_[i]) < 0);
    }
    RecordInHistogram(stats_, NUM_SUBCOMPACTIONS_SCHEDULED,
                      compact_->sub_compact_states.size());
@@ -263,11 +282,15 @@ void CompactionJob::Prepare() {
                                              /*sub_job_id*/ 0);
  }

-  if (c->immutable_options()->preclude_last_level_data_seconds > 0) {
-    // TODO(zjay): move to a function
-    seqno_time_mapping_.SetMaxTimeDuration(
-        c->immutable_options()->preclude_last_level_data_seconds);
+  // 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);
+
+  if (preserve_time_duration > 0) {
    // setup seqno_time_mapping_
+    seqno_time_mapping_.SetMaxTimeDuration(preserve_time_duration);
    for (const auto& each_level : *c->inputs()) {
      for (const auto& fmd : each_level.files) {
        std::shared_ptr<const TableProperties> tp;
@@ -293,10 +316,27 @@ void CompactionJob::Prepare() {
      ROCKS_LOG_WARN(db_options_.info_log,
                     "Failed to get current time in compaction: Status: %s",
                     status.ToString().c_str());
-      penultimate_level_cutoff_seqno_ = 0;
+      // preserve all time information
+      preserve_time_min_seqno_ = 0;
+      preclude_last_level_min_seqno_ = 0;
    } else {
-      penultimate_level_cutoff_seqno_ =
-          seqno_time_mapping_.TruncateOldEntries(_current_time);
+      seqno_time_mapping_.TruncateOldEntries(_current_time);
+      uint64_t preserve_time =
+          static_cast<uint64_t>(_current_time) > preserve_time_duration
+              ? _current_time - preserve_time_duration
+              : 0;
+      preserve_time_min_seqno_ =
+          seqno_time_mapping_.GetOldestSequenceNum(preserve_time);
+      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_time_mapping_.GetOldestSequenceNum(preclude_last_level_time);
+      }
    }
  }
 }
@@ -321,6 +361,7 @@ void CompactionJob::AcquireSubcompactionResources(
              ->write_controller()
              ->NeedSpeedupCompaction())
          .max_compactions;
+  InstrumentedMutexLock l(db_mutex_);
  // Apply min function first since We need to compute the extra subcompaction
  // against compaction limits. And then try to reserve threads for extra
  // subcompactions. The actual number of reserved threads could be less than
@@ -329,7 +370,6 @@ void CompactionJob::AcquireSubcompactionResources(
      std::max(max_db_compactions - *bg_compaction_scheduled_ -
                   *bg_bottom_compaction_scheduled_,
               0);
-  db_mutex_->Lock();
  // Reservation only supports backgrdoun threads of which the priority is
  // between BOTTOM and HIGH. Need to degrade the priority to HIGH if the
  // origin thread_pri_ is higher than that. Similar to ReleaseThreads().
@@ -346,7 +386,6 @@ void CompactionJob::AcquireSubcompactionResources(
  } else {
    *bg_compaction_scheduled_ += extra_num_subcompaction_threads_reserved_;
  }
-  db_mutex_->Unlock();
 }

 void CompactionJob::ShrinkSubcompactionResources(uint64_t num_extra_resources) {
@@ -380,17 +419,20 @@ void CompactionJob::ReleaseSubcompactionResources() {
  if (extra_num_subcompaction_threads_reserved_ == 0) {
    return;
  }
-  // The number of reserved threads becomes larger than 0 only if the
-  // compaction prioity is round robin and there is no sufficient
-  // sub-compactions available
+  {
+    InstrumentedMutexLock l(db_mutex_);
+    // The number of reserved threads becomes larger than 0 only if the
+    // compaction prioity is round robin and there is no sufficient
+    // sub-compactions available

-  // The scheduled compaction must be no less than 1 + extra number
-  // subcompactions using acquired resources since this compaction job has not
-  // finished yet
-  assert(*bg_bottom_compaction_scheduled_ >=
-             1 + extra_num_subcompaction_threads_reserved_ ||
-         *bg_compaction_scheduled_ >=
-             1 + extra_num_subcompaction_threads_reserved_);
+    // The scheduled compaction must be no less than 1 + extra number
+    // subcompactions using acquired resources since this compaction job has not
+    // finished yet
+    assert(*bg_bottom_compaction_scheduled_ >=
+               1 + extra_num_subcompaction_threads_reserved_ ||
+           *bg_compaction_scheduled_ >=
+               1 + extra_num_subcompaction_threads_reserved_);
+  }
  ShrinkSubcompactionResources(extra_num_subcompaction_threads_reserved_);
 }

@@ -462,7 +504,7 @@ void CompactionJob::GenSubcompactionBoundaries() {
        FileMetaData* f = flevel->files[i].file_metadata;
        std::vector<TableReader::Anchor> my_anchors;
        Status s = cfd->table_cache()->ApproximateKeyAnchors(
-            ReadOptions(), icomp, f->fd, my_anchors);
+            ReadOptions(), icomp, *f, my_anchors);
        if (!s.ok() || my_anchors.empty()) {
          my_anchors.emplace_back(f->largest.user_key(), f->fd.GetFileSize());
        }
@@ -484,8 +526,8 @@ void CompactionJob::GenSubcompactionBoundaries() {
  std::sort(
      all_anchors.begin(), all_anchors.end(),
      [cfd_comparator](TableReader::Anchor& a, TableReader::Anchor& b) -> bool {
-        return cfd_comparator->CompareWithoutTimestamp(a.user_key, true,
-                                                       b.user_key, true) < 0;
+        return cfd_comparator->CompareWithoutTimestamp(a.user_key, b.user_key) <
+               0;
      });

  // Remove duplicated entries from boundaries.
@@ -494,7 +536,7 @@ void CompactionJob::GenSubcompactionBoundaries() {
                  [cfd_comparator](TableReader::Anchor& a,
                                   TableReader::Anchor& b) -> bool {
                    return cfd_comparator->CompareWithoutTimestamp(
-                               a.user_key, true, b.user_key, true) == 0;
+                               a.user_key, b.user_key) == 0;
                  }),
      all_anchors.end());

@@ -1033,7 +1075,8 @@ void CompactionJob::ProcessKeyValueCompaction(SubcompactionState* sub_compact) {
  NotifyOnSubcompactionBegin(sub_compact);

  auto range_del_agg = std::make_unique<CompactionRangeDelAggregator>(
-      &cfd->internal_comparator(), existing_snapshots_);
+      &cfd->internal_comparator(), existing_snapshots_, &full_history_ts_low_,
+      &trim_ts_);

  // TODO: since we already use C++17, should use
  // std::optional<const Slice> instead.
@@ -1083,13 +1126,34 @@ void CompactionJob::ProcessKeyValueCompaction(SubcompactionState* sub_compact) {
  Slice start_slice;
  Slice end_slice;

+  static constexpr char kMaxTs[] =
+      "\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff";
+  Slice ts_slice;
+  std::string max_ts;
+  if (ts_sz > 0) {
+    if (ts_sz <= strlen(kMaxTs)) {
+      ts_slice = Slice(kMaxTs, ts_sz);
+    } else {
+      max_ts = std::string(ts_sz, '\xff');
+      ts_slice = Slice(max_ts);
+    }
+  }
+
  if (start.has_value()) {
    start_ikey.SetInternalKey(start.value(), kMaxSequenceNumber,
                              kValueTypeForSeek);
+    if (ts_sz > 0) {
+      start_ikey.UpdateInternalKey(kMaxSequenceNumber, kValueTypeForSeek,
+                                   &ts_slice);
+    }
    start_slice = start_ikey.GetInternalKey();
  }
  if (end.has_value()) {
    end_ikey.SetInternalKey(end.value(), kMaxSequenceNumber, kValueTypeForSeek);
+    if (ts_sz > 0) {
+      end_ikey.UpdateInternalKey(kMaxSequenceNumber, kValueTypeForSeek,
+                                 &ts_slice);
+    }
    end_slice = end_ikey.GetInternalKey();
  }

@@ -1110,7 +1174,7 @@ void CompactionJob::ProcessKeyValueCompaction(SubcompactionState* sub_compact) {
  }

  std::unique_ptr<InternalIterator> trim_history_iter;
-  if (cfd->user_comparator()->timestamp_size() > 0 && !trim_ts_.empty()) {
+  if (ts_sz > 0 && !trim_ts_.empty()) {
    trim_history_iter = std::make_unique<HistoryTrimmingIterator>(
        input, cfd->user_comparator(), trim_ts_);
    input = trim_history_iter.get();
@@ -1190,21 +1254,14 @@ void CompactionJob::ProcessKeyValueCompaction(SubcompactionState* sub_compact) {
      blob_file_builder.get(), db_options_.allow_data_in_errors,
      db_options_.enforce_single_del_contracts, manual_compaction_canceled_,
      sub_compact->compaction, compaction_filter, shutting_down_,
-      db_options_.info_log, full_history_ts_low,
-      penultimate_level_cutoff_seqno_);
+      db_options_.info_log, full_history_ts_low, preserve_time_min_seqno_,
+      preclude_last_level_min_seqno_);
  c_iter->SeekToFirst();

  // Assign range delete aggregator to the target output level, which makes sure
  // it only output to single level
  sub_compact->AssignRangeDelAggregator(std::move(range_del_agg));

-  if (c_iter->Valid() && sub_compact->compaction->output_level() != 0) {
-    sub_compact->FillFilesToCutForTtl();
-    // ShouldStopBefore() maintains state based on keys processed so far. The
-    // compaction loop always calls it on the "next" key, thus won't tell it the
-    // first key. So we do that here.
-    sub_compact->ShouldStopBefore(c_iter->key());
-  }
  const auto& c_iter_stats = c_iter->iter_stats();

  // define the open and close functions for the compaction files, which will be
@@ -1221,6 +1278,10 @@ void CompactionJob::ProcessKeyValueCompaction(SubcompactionState* sub_compact) {
      };

  Status status;
+  TEST_SYNC_POINT_CALLBACK(
+      "CompactionJob::ProcessKeyValueCompaction()::Processing",
+      reinterpret_cast<void*>(
+          const_cast<Compaction*>(sub_compact->compaction)));
  while (status.ok() && !cfd->IsDropped() && c_iter->Valid()) {
    // Invariant: c_iter.status() is guaranteed to be OK if c_iter->Valid()
    // returns true.
@@ -1253,16 +1314,6 @@ void CompactionJob::ProcessKeyValueCompaction(SubcompactionState* sub_compact) {
    if (c_iter->status().IsManualCompactionPaused()) {
      break;
    }
-
-    // TODO: Support earlier file cut for the penultimate level files. Maybe by
-    //  moving `ShouldStopBefore()` to `CompactionOutputs` class. Currently
-    //  the penultimate level output is only cut when it reaches the size limit.
-    if (!sub_compact->Current().IsPendingClose() &&
-        sub_compact->compaction->output_level() != 0 &&
-        !sub_compact->compaction->SupportsPerKeyPlacement() &&
-        sub_compact->ShouldStopBefore(c_iter->key())) {
-      sub_compact->Current().SetPendingClose();
-    }
  }

  sub_compact->compaction_job_stats.num_blobs_read =
@@ -1449,7 +1500,7 @@ Status CompactionJob::FinishCompactionOutputFile(
                                         : nullptr,
          sub_compact->end.has_value() ? &(sub_compact->end.value()) : nullptr,
          range_del_out_stats, bottommost_level_, cfd->internal_comparator(),
-          earliest_snapshot, next_table_min_key);
+          earliest_snapshot, next_table_min_key, full_history_ts_low_);
    }
    RecordDroppedKeys(range_del_out_stats, &sub_compact->compaction_job_stats);
    TEST_SYNC_POINT("CompactionJob::FinishCompactionOutputFile1");
@@ -1655,7 +1706,9 @@ Status CompactionJob::InstallCompactionResults(
                             stats.GetBytes());
  }

-  if (compaction->compaction_reason() == CompactionReason::kLevelMaxLevelSize &&
+  if ((compaction->compaction_reason() ==
+           CompactionReason::kLevelMaxLevelSize ||
+       compaction->compaction_reason() == CompactionReason::kRoundRobinTtl) &&
      compaction->immutable_options()->compaction_pri == kRoundRobin) {
    int start_level = compaction->start_level();
    if (start_level > 0) {
@@ -1944,7 +1997,7 @@ void CompactionJob::LogCompaction() {
        compaction->InputLevelSummary(&inputs_summary), compaction->score());
    char scratch[2345];
    compaction->Summary(scratch, sizeof(scratch));
-    ROCKS_LOG_INFO(db_options_.info_log, "[%s] Compaction start summary: %s\n",
+    ROCKS_LOG_INFO(db_options_.info_log, "[%s]: Compaction start summary: %s\n",
                   cfd->GetName().c_str(), scratch);
    // build event logger report
    auto stream = event_logger_->Log();
@@ -1965,6 +2018,23 @@ void CompactionJob::LogCompaction() {
           << (existing_snapshots_.empty()
                   ? int64_t{-1}  // Use -1 for "none"
                   : static_cast<int64_t>(existing_snapshots_[0]));
+    if (compaction->SupportsPerKeyPlacement()) {
+      stream << "preclude_last_level_min_seqno"
+             << preclude_last_level_min_seqno_;
+      stream << "penultimate_output_level" << compaction->GetPenultimateLevel();
+      stream << "penultimate_output_range"
+             << GetCompactionPenultimateOutputRangeTypeString(
+                    compaction->GetPenultimateOutputRangeType());
+
+      if (compaction->GetPenultimateOutputRangeType() ==
+          Compaction::PenultimateOutputRangeType::kDisabled) {
+        ROCKS_LOG_WARN(
+            db_options_.info_log,
+            "[%s] [JOB %d] Penultimate level output is disabled, likely "
+            "because of the range conflict in the penultimate level",
+            cfd->GetName().c_str(), job_id_);
+      }
+    }
  }
 }

@@ -335,12 +335,15 @@ class CompactionJob {
  // it also collects the smallest_seqno -> oldest_ancester_time from the SST.
  SeqnoToTimeMapping seqno_time_mapping_;

-  // cutoff sequence number for penultimate level, only set when
-  // per_key_placement feature is enabled.
-  // If a key with sequence number larger than penultimate_level_cutoff_seqno_,
-  // it will be placed on the penultimate_level and seqnuence number won't be
-  // zeroed out.
-  SequenceNumber penultimate_level_cutoff_seqno_ = kMaxSequenceNumber;
+  // Minimal sequence number for preserving the time information. The time info
+  // older than this sequence number won't be preserved after the compaction and
+  // if it's bottommost compaction, the seq num will be zeroed out.
+  SequenceNumber preserve_time_min_seqno_ = kMaxSequenceNumber;
+
+  // Minimal sequence number to preclude the data from the last level. If the
+  // key has bigger (newer) sequence number than this, it will be precluded from
+  // the last level (output to penultimate level).
+  SequenceNumber preclude_last_level_min_seqno_ = kMaxSequenceNumber;

  // Get table file name in where it's outputting to, which should also be in
  // `output_directory_`.
@@ -76,6 +76,258 @@ IOStatus CompactionOutputs::WriterSyncClose(const Status& input_status,
  return io_s;
 }

+size_t CompactionOutputs::UpdateGrandparentBoundaryInfo(
+    const Slice& internal_key) {
+  size_t curr_key_boundary_switched_num = 0;
+  const std::vector<FileMetaData*>& grandparents = compaction_->grandparents();
+
+  if (grandparents.empty()) {
+    return curr_key_boundary_switched_num;
+  }
+  assert(!internal_key.empty());
+  InternalKey ikey;
+  ikey.DecodeFrom(internal_key);
+  assert(ikey.Valid());
+
+  const Comparator* ucmp = compaction_->column_family_data()->user_comparator();
+
+  // Move the grandparent_index_ to the file containing the current user_key.
+  // If there are multiple files containing the same user_key, make sure the
+  // index points to the last file containing the key.
+  while (grandparent_index_ < grandparents.size()) {
+    if (being_grandparent_gap_) {
+      if (sstableKeyCompare(ucmp, ikey,
+                            grandparents[grandparent_index_]->smallest) < 0) {
+        break;
+      }
+      if (seen_key_) {
+        curr_key_boundary_switched_num++;
+        grandparent_overlapped_bytes_ +=
+            grandparents[grandparent_index_]->fd.GetFileSize();
+        grandparent_boundary_switched_num_++;
+      }
+      being_grandparent_gap_ = false;
+    } else {
+      int cmp_result = sstableKeyCompare(
+          ucmp, ikey, grandparents[grandparent_index_]->largest);
+      // If it's same key, make sure grandparent_index_ is pointing to the last
+      // one.
+      if (cmp_result < 0 ||
+          (cmp_result == 0 &&
+           (grandparent_index_ == grandparents.size() - 1 ||
+            sstableKeyCompare(ucmp, ikey,
+                              grandparents[grandparent_index_ + 1]->smallest) <
+                0))) {
+        break;
+      }
+      if (seen_key_) {
+        curr_key_boundary_switched_num++;
+        grandparent_boundary_switched_num_++;
+      }
+      being_grandparent_gap_ = true;
+      grandparent_index_++;
+    }
+  }
+
+  // If the first key is in the middle of a grandparent file, adding it to the
+  // overlap
+  if (!seen_key_ && !being_grandparent_gap_) {
+    assert(grandparent_overlapped_bytes_ == 0);
+    grandparent_overlapped_bytes_ =
+        GetCurrentKeyGrandparentOverlappedBytes(internal_key);
+  }
+
+  seen_key_ = true;
+  return curr_key_boundary_switched_num;
+}
+
+uint64_t CompactionOutputs::GetCurrentKeyGrandparentOverlappedBytes(
+    const Slice& internal_key) const {
+  // no overlap with any grandparent file
+  if (being_grandparent_gap_) {
+    return 0;
+  }
+  uint64_t overlapped_bytes = 0;
+
+  const std::vector<FileMetaData*>& grandparents = compaction_->grandparents();
+  const Comparator* ucmp = compaction_->column_family_data()->user_comparator();
+  InternalKey ikey;
+  ikey.DecodeFrom(internal_key);
+#ifndef NDEBUG
+  // make sure the grandparent_index_ is pointing to the last files containing
+  // the current key.
+  int cmp_result =
+      sstableKeyCompare(ucmp, ikey, grandparents[grandparent_index_]->largest);
+  assert(
+      cmp_result < 0 ||
+      (cmp_result == 0 &&
+       (grandparent_index_ == grandparents.size() - 1 ||
+        sstableKeyCompare(
+            ucmp, ikey, grandparents[grandparent_index_ + 1]->smallest) < 0)));
+  assert(sstableKeyCompare(ucmp, ikey,
+                           grandparents[grandparent_index_]->smallest) >= 0);
+#endif
+  overlapped_bytes += grandparents[grandparent_index_]->fd.GetFileSize();
+
+  // go backwards to find all overlapped files, one key can overlap multiple
+  // files. In the following example, if the current output key is `c`, and one
+  // compaction file was cut before `c`, current `c` can overlap with 3 files:
+  //  [a b]               [c...
+  // [b, b] [c, c] [c, c] [c, d]
+  for (int64_t i = static_cast<int64_t>(grandparent_index_) - 1;
+       i >= 0 && sstableKeyCompare(ucmp, ikey, grandparents[i]->largest) == 0;
+       i--) {
+    overlapped_bytes += grandparents[i]->fd.GetFileSize();
+  }
+
+  return overlapped_bytes;
+}
+
+bool CompactionOutputs::ShouldStopBefore(const CompactionIterator& c_iter) {
+  assert(c_iter.Valid());
+
+  // always update grandparent information like overlapped file number, size
+  // etc.
+  const Slice& internal_key = c_iter.key();
+  const uint64_t previous_overlapped_bytes = grandparent_overlapped_bytes_;
+  size_t num_grandparent_boundaries_crossed =
+      UpdateGrandparentBoundaryInfo(internal_key);
+
+  if (!HasBuilder()) {
+    return false;
+  }
+
+  // If there's user defined partitioner, check that first
+  if (partitioner_ && partitioner_->ShouldPartition(PartitionerRequest(
+                          last_key_for_partitioner_, c_iter.user_key(),
+                          current_output_file_size_)) == kRequired) {
+    return true;
+  }
+
+  // files output to Level 0 won't be split
+  if (compaction_->output_level() == 0) {
+    return false;
+  }
+
+  // reach the max file size
+  if (current_output_file_size_ >= compaction_->max_output_file_size()) {
+    return true;
+  }
+
+  const InternalKeyComparator* icmp =
+      &compaction_->column_family_data()->internal_comparator();
+
+  // Check if it needs to split for RoundRobin
+  // Invalid local_output_split_key indicates that we do not need to split
+  if (local_output_split_key_ != nullptr && !is_split_) {
+    // Split occurs when the next key is larger than/equal to the cursor
+    if (icmp->Compare(internal_key, local_output_split_key_->Encode()) >= 0) {
+      is_split_ = true;
+      return true;
+    }
+  }
+
+  // only check if the current key is going to cross the grandparents file
+  // boundary (either the file beginning or ending).
+  if (num_grandparent_boundaries_crossed > 0) {
+    // Cut the file before the current key if the size of the current output
+    // file + its overlapped grandparent files is bigger than
+    // max_compaction_bytes. Which is to prevent future bigger than
+    // max_compaction_bytes compaction from the current output level.
+    if (grandparent_overlapped_bytes_ + current_output_file_size_ >
+        compaction_->max_compaction_bytes()) {
+      return true;
+    }
+
+    // Cut the file if including the key is going to add a skippable file on
+    // the grandparent level AND its size is reasonably big (1/8 of target file
+    // size). For example, if it's compacting the files L0 + L1:
+    //  L0:  [1,   21]
+    //  L1:    [3,   23]
+    //  L2: [2, 4] [11, 15] [22, 24]
+    // Without this break, it will output as:
+    //  L1: [1,3, 21,23]
+    // With this break, it will output as (assuming [11, 15] at L2 is bigger
+    // than 1/8 of target size):
+    //  L1: [1,3] [21,23]
+    // Then for the future compactions, [11,15] won't be included.
+    // For random datasets (either evenly distributed or skewed), it rarely
+    // triggers this condition, but if the user is adding 2 different datasets
+    // without any overlap, it may likely happen.
+    // More details, check PR #1963
+    const size_t num_skippable_boundaries_crossed =
+        being_grandparent_gap_ ? 2 : 3;
+    if (compaction_->immutable_options()->compaction_style ==
+            kCompactionStyleLevel &&
+        compaction_->immutable_options()->level_compaction_dynamic_file_size &&
+        num_grandparent_boundaries_crossed >=
+            num_skippable_boundaries_crossed &&
+        grandparent_overlapped_bytes_ - previous_overlapped_bytes >
+            compaction_->target_output_file_size() / 8) {
+      return true;
+    }
+
+    // Pre-cut the output file if it's reaching a certain size AND it's at the
+    // boundary of a grandparent file. It can reduce the future compaction size,
+    // the cost is having smaller files.
+    // The pre-cut size threshold is based on how many grandparent boundaries
+    // it has seen before. Basically, if it has seen no boundary at all, then it
+    // will pre-cut at 50% target file size. Every boundary it has seen
+    // increases the threshold by 5%, max at 90%, which it will always cut.
+    // The idea is based on if it has seen more boundaries before, it will more
+    // likely to see another boundary (file cutting opportunity) before the
+    // target file size. The test shows it can generate larger files than a
+    // static threshold like 75% and has a similar write amplification
+    // improvement.
+    if (compaction_->immutable_options()->compaction_style ==
+            kCompactionStyleLevel &&
+        compaction_->immutable_options()->level_compaction_dynamic_file_size &&
+        current_output_file_size_ >=
+            ((compaction_->target_output_file_size() + 99) / 100) *
+                (50 + std::min(grandparent_boundary_switched_num_ * 5,
+                               size_t{40}))) {
+      return true;
+    }
+  }
+
+  // check ttl file boundaries if there's any
+  if (!files_to_cut_for_ttl_.empty()) {
+    if (cur_files_to_cut_for_ttl_ != -1) {
+      // Previous key is inside the range of a file
+      if (icmp->Compare(internal_key,
+                        files_to_cut_for_ttl_[cur_files_to_cut_for_ttl_]
+                            ->largest.Encode()) > 0) {
+        next_files_to_cut_for_ttl_ = cur_files_to_cut_for_ttl_ + 1;
+        cur_files_to_cut_for_ttl_ = -1;
+        return true;
+      }
+    } else {
+      // Look for the key position
+      while (next_files_to_cut_for_ttl_ <
+             static_cast<int>(files_to_cut_for_ttl_.size())) {
+        if (icmp->Compare(internal_key,
+                          files_to_cut_for_ttl_[next_files_to_cut_for_ttl_]
+                              ->smallest.Encode()) >= 0) {
+          if (icmp->Compare(internal_key,
+                            files_to_cut_for_ttl_[next_files_to_cut_for_ttl_]
+                                ->largest.Encode()) <= 0) {
+            // With in the current file
+            cur_files_to_cut_for_ttl_ = next_files_to_cut_for_ttl_;
+            return true;
+          }
+          // Beyond the current file
+          next_files_to_cut_for_ttl_++;
+        } else {
+          // Still fall into the gap
+          break;
+        }
+      }
+    }
+  }
+
+  return false;
+}
+
 Status CompactionOutputs::AddToOutput(
    const CompactionIterator& c_iter,
    const CompactionFileOpenFunc& open_file_func,
@@ -83,33 +335,28 @@ Status CompactionOutputs::AddToOutput(
  Status s;
  const Slice& key = c_iter.key();

-  if (!pending_close_ && c_iter.Valid() && partitioner_ && HasBuilder() &&
-      partitioner_->ShouldPartition(
-          PartitionerRequest(last_key_for_partitioner_, c_iter.user_key(),
-                             current_output_file_size_)) == kRequired) {
-    pending_close_ = true;
-  }
-
-  if (pending_close_) {
+  if (ShouldStopBefore(c_iter) && HasBuilder()) {
    s = close_file_func(*this, c_iter.InputStatus(), key);
-    pending_close_ = false;
-  }
-  if (!s.ok()) {
-    return s;
+    if (!s.ok()) {
+      return s;
+    }
+    // reset grandparent information
+    grandparent_boundary_switched_num_ = 0;
+    grandparent_overlapped_bytes_ =
+        GetCurrentKeyGrandparentOverlappedBytes(key);
  }

  // Open output file if necessary
  if (!HasBuilder()) {
    s = open_file_func(*this);
-  }
-  if (!s.ok()) {
-    return s;
+    if (!s.ok()) {
+      return s;
+    }
  }

-  Output& curr = current_output();
  assert(builder_ != nullptr);
  const Slice& value = c_iter.value();
-  s = curr.validator.Add(key, value);
+  s = current_output().validator.Add(key, value);
  if (!s.ok()) {
    return s;
  }
@@ -130,19 +377,6 @@ Status CompactionOutputs::AddToOutput(
  s = current_output().meta.UpdateBoundaries(key, value, ikey.sequence,
                                             ikey.type);

-  // Close output file if it is big enough. Two possibilities determine it's
-  // time to close it: (1) the current key should be this file's last key, (2)
-  // the next key should not be in this file.
-  //
-  // TODO(aekmekji): determine if file should be closed earlier than this
-  // during subcompactions (i.e. if output size, estimated by input size, is
-  // going to be 1.2MB and max_output_file_size = 1MB, prefer to have 0.6MB
-  // and 0.6MB instead of 1MB and 0.2MB)
-  if (compaction_->output_level() != 0 &&
-      current_output_file_size_ >= compaction_->max_output_file_size()) {
-    pending_close_ = true;
-  }
-
  if (partitioner_) {
    last_key_for_partitioner_.assign(c_iter.user_key().data_,
                                     c_iter.user_key().size_);
@@ -152,10 +386,10 @@ Status CompactionOutputs::AddToOutput(
 }

 Status CompactionOutputs::AddRangeDels(
-    const Slice* comp_start, const Slice* comp_end,
+    const Slice* comp_start_user_key, const Slice* comp_end_user_key,
    CompactionIterationStats& range_del_out_stats, bool bottommost_level,
    const InternalKeyComparator& icmp, SequenceNumber earliest_snapshot,
-    const Slice& next_table_min_key) {
+    const Slice& next_table_min_key, const std::string& full_history_ts_low) {
  assert(HasRangeDel());
  FileMetaData& meta = current_output().meta;
  const Comparator* ucmp = icmp.user_comparator();
@@ -169,7 +403,7 @@ Status CompactionOutputs::AddRangeDels(
  if (output_size == 1) {
    // For the first output table, include range tombstones before the min
    // key but after the subcompaction boundary.
-    lower_bound = comp_start;
+    lower_bound = comp_start_user_key;
    lower_bound_from_sub_compact = true;
  } else if (meta.smallest.size() > 0) {
    // For subsequent output tables, only include range tombstones from min
@@ -189,21 +423,22 @@ Status CompactionOutputs::AddRangeDels(
    // use the smaller key as the upper bound of the output file, to ensure
    // that there is no overlapping between different output files.
    upper_bound_guard = ExtractUserKey(next_table_min_key);
-    if (comp_end != nullptr &&
-        ucmp->Compare(upper_bound_guard, *comp_end) >= 0) {
-      upper_bound = comp_end;
+    if (comp_end_user_key != nullptr &&
+        ucmp->CompareWithoutTimestamp(upper_bound_guard, *comp_end_user_key) >=
+            0) {
+      upper_bound = comp_end_user_key;
    } else {
      upper_bound = &upper_bound_guard;
    }
  } else {
    // This is the last file in the subcompaction, so extend until the
    // subcompaction ends.
-    upper_bound = comp_end;
+    upper_bound = comp_end_user_key;
  }
  bool has_overlapping_endpoints;
  if (upper_bound != nullptr && meta.largest.size() > 0) {
-    has_overlapping_endpoints =
-        ucmp->Compare(meta.largest.user_key(), *upper_bound) == 0;
+    has_overlapping_endpoints = ucmp->CompareWithoutTimestamp(
+                                    meta.largest.user_key(), *upper_bound) == 0;
  } else {
    has_overlapping_endpoints = false;
  }
@@ -212,8 +447,8 @@ Status CompactionOutputs::AddRangeDels(
  // bound. If the end of subcompaction is null or the upper bound is null,
  // it means that this file is the last file in the compaction. So there
  // will be no overlapping between this file and others.
-  assert(comp_end == nullptr || upper_bound == nullptr ||
-         ucmp->Compare(*upper_bound, *comp_end) <= 0);
+  assert(comp_end_user_key == nullptr || upper_bound == nullptr ||
+         ucmp->CompareWithoutTimestamp(*upper_bound, *comp_end_user_key) <= 0);
  auto it = range_del_agg_->NewIterator(lower_bound, upper_bound,
                                        has_overlapping_endpoints);
  // Position the range tombstone output iterator. There may be tombstone
@@ -227,7 +462,8 @@ Status CompactionOutputs::AddRangeDels(
  for (; it->Valid(); it->Next()) {
    auto tombstone = it->Tombstone();
    if (upper_bound != nullptr) {
-      int cmp = ucmp->Compare(*upper_bound, tombstone.start_key_);
+      int cmp =
+          ucmp->CompareWithoutTimestamp(*upper_bound, tombstone.start_key_);
      if ((has_overlapping_endpoints && cmp < 0) ||
          (!has_overlapping_endpoints && cmp <= 0)) {
        // Tombstones starting after upper_bound only need to be included in
@@ -240,7 +476,17 @@ Status CompactionOutputs::AddRangeDels(
      }
    }

-    if (bottommost_level && tombstone.seq_ <= earliest_snapshot) {
+    const size_t ts_sz = ucmp->timestamp_size();
+    // Garbage collection for range tombstones.
+    // If user-defined timestamp is enabled, range tombstones are dropped if
+    // they are at bottommost_level, below full_history_ts_low and not visible
+    // in any snapshot. trim_ts_ is passed to the constructor for
+    // range_del_agg_, and range_del_agg_ internally drops tombstones above
+    // trim_ts_.
+    if (bottommost_level && tombstone.seq_ <= earliest_snapshot &&
+        (ts_sz == 0 ||
+         (!full_history_ts_low.empty() &&
+          ucmp->CompareTimestamp(tombstone.ts_, full_history_ts_low) < 0))) {
      // TODO(andrewkr): tombstones that span multiple output files are
      // counted for each compaction output file, so lots of double
      // counting.
@@ -251,12 +497,13 @@ Status CompactionOutputs::AddRangeDels(

    auto kv = tombstone.Serialize();
    assert(lower_bound == nullptr ||
-           ucmp->Compare(*lower_bound, kv.second) < 0);
+           ucmp->CompareWithoutTimestamp(*lower_bound, kv.second) < 0);
    // Range tombstone is not supported by output validator yet.
    builder_->Add(kv.first.Encode(), kv.second);
    InternalKey smallest_candidate = std::move(kv.first);
    if (lower_bound != nullptr &&
-        ucmp->Compare(smallest_candidate.user_key(), *lower_bound) <= 0) {
+        ucmp->CompareWithoutTimestamp(smallest_candidate.user_key(),
+                                      *lower_bound) <= 0) {
      // Pretend the smallest key has the same user key as lower_bound
      // (the max key in the previous table or subcompaction) in order for
      // files to appear key-space partitioned.
@@ -276,13 +523,23 @@ Status CompactionOutputs::AddRangeDels(
      // choose lowest seqnum so this file's smallest internal key comes
      // after the previous file's largest. The fake seqnum is OK because
      // the read path's file-picking code only considers user key.
-      smallest_candidate = InternalKey(
-          *lower_bound, lower_bound_from_sub_compact ? tombstone.seq_ : 0,
-          kTypeRangeDeletion);
+      if (lower_bound_from_sub_compact) {
+        if (ts_sz) {
+          assert(tombstone.ts_.size() == ts_sz);
+          smallest_candidate = InternalKey(*lower_bound, tombstone.seq_,
+                                           kTypeRangeDeletion, tombstone.ts_);
+        } else {
+          smallest_candidate =
+              InternalKey(*lower_bound, tombstone.seq_, kTypeRangeDeletion);
+        }
+      } else {
+        smallest_candidate = InternalKey(*lower_bound, 0, kTypeRangeDeletion);
+      }
    }
    InternalKey largest_candidate = tombstone.SerializeEndKey();
    if (upper_bound != nullptr &&
-        ucmp->Compare(*upper_bound, largest_candidate.user_key()) <= 0) {
+        ucmp->CompareWithoutTimestamp(*upper_bound,
+                                      largest_candidate.user_key()) <= 0) {
      // Pretend the largest key has the same user key as upper_bound (the
      // min key in the following table or subcompaction) in order for files
      // to appear key-space partitioned.
@@ -296,9 +553,22 @@ Status CompactionOutputs::AddRangeDels(
      // kMaxSequenceNumber), but with kTypeDeletion (0x7) instead of
      // kTypeRangeDeletion (0xF), so the range tombstone comes before the
      // Seek() key in InternalKey's ordering. So Seek() will look in the
-      // next file for the user key.
-      largest_candidate =
-          InternalKey(*upper_bound, kMaxSequenceNumber, kTypeRangeDeletion);
+      // next file for the user key
+      if (ts_sz) {
+        static constexpr char kTsMax[] = "\xff\xff\xff\xff\xff\xff\xff\xff\xff";
+        if (ts_sz <= strlen(kTsMax)) {
+          largest_candidate =
+              InternalKey(*upper_bound, kMaxSequenceNumber, kTypeRangeDeletion,
+                          Slice(kTsMax, ts_sz));
+        } else {
+          largest_candidate =
+              InternalKey(*upper_bound, kMaxSequenceNumber, kTypeRangeDeletion,
+                          std::string(ts_sz, '\xff'));
+        }
+      } else {
+        largest_candidate =
+            InternalKey(*upper_bound, kMaxSequenceNumber, kTypeRangeDeletion);
+      }
    }
 #ifndef NDEBUG
    SequenceNumber smallest_ikey_seqnum = kMaxSequenceNumber;
@@ -318,4 +588,59 @@ Status CompactionOutputs::AddRangeDels(
  }
  return Status::OK();
 }
+
+void CompactionOutputs::FillFilesToCutForTtl() {
+  if (compaction_->immutable_options()->compaction_style !=
+          kCompactionStyleLevel ||
+      compaction_->immutable_options()->compaction_pri !=
+          kMinOverlappingRatio ||
+      compaction_->mutable_cf_options()->ttl == 0 ||
+      compaction_->num_input_levels() < 2 || compaction_->bottommost_level()) {
+    return;
+  }
+
+  // We define new file with the oldest ancestor time to be younger than 1/4
+  // TTL, and an old one to be older than 1/2 TTL time.
+  int64_t temp_current_time;
+  auto get_time_status =
+      compaction_->immutable_options()->clock->GetCurrentTime(
+          &temp_current_time);
+  if (!get_time_status.ok()) {
+    return;
+  }
+
+  auto current_time = static_cast<uint64_t>(temp_current_time);
+  if (current_time < compaction_->mutable_cf_options()->ttl) {
+    return;
+  }
+
+  uint64_t old_age_thres =
+      current_time - compaction_->mutable_cf_options()->ttl / 2;
+  const std::vector<FileMetaData*>& olevel =
+      *(compaction_->inputs(compaction_->num_input_levels() - 1));
+  for (FileMetaData* file : olevel) {
+    // Worth filtering out by start and end?
+    uint64_t oldest_ancester_time = file->TryGetOldestAncesterTime();
+    // We put old files if they are not too small to prevent a flood
+    // of small files.
+    if (oldest_ancester_time < old_age_thres &&
+        file->fd.GetFileSize() >
+            compaction_->mutable_cf_options()->target_file_size_base / 2) {
+      files_to_cut_for_ttl_.push_back(file);
+    }
+  }
+}
+
+CompactionOutputs::CompactionOutputs(const Compaction* compaction,
+                                     const bool is_penultimate_level)
+    : compaction_(compaction), is_penultimate_level_(is_penultimate_level) {
+  partitioner_ = compaction->output_level() == 0
+                     ? nullptr
+                     : compaction->CreateSstPartitioner();
+
+  if (compaction->output_level() != 0) {
+    FillFilesToCutForTtl();
+  }
+}
+
 }  // namespace ROCKSDB_NAMESPACE
@@ -45,12 +45,7 @@ class CompactionOutputs {
  CompactionOutputs() = delete;

  explicit CompactionOutputs(const Compaction* compaction,
-                             const bool is_penultimate_level)
-      : compaction_(compaction), is_penultimate_level_(is_penultimate_level) {
-    partitioner_ = compaction->output_level() == 0
-                       ? nullptr
-                       : compaction->CreateSstPartitioner();
-  }
+                             const bool is_penultimate_level);

  // Add generated output to the list
  void AddOutput(FileMetaData&& meta, const InternalKeyComparator& icmp,
@@ -173,17 +168,16 @@ class CompactionOutputs {
  }

  // Add range-dels from the aggregator to the current output file
-  Status AddRangeDels(const Slice* comp_start, const Slice* comp_end,
+  // @param comp_start_user_key and comp_end_user_key include timestamp if
+  // user-defined timestamp is enabled.
+  // @param full_history_ts_low used for range tombstone garbage collection.
+  Status AddRangeDels(const Slice* comp_start_user_key,
+                      const Slice* comp_end_user_key,
                      CompactionIterationStats& range_del_out_stats,
                      bool bottommost_level, const InternalKeyComparator& icmp,
                      SequenceNumber earliest_snapshot,
-                      const Slice& next_table_min_key);
-
-  // Is the current file is already pending for close
-  bool IsPendingClose() const { return pending_close_; }
-
-  // Current file should close before adding a new key
-  void SetPendingClose() { pending_close_ = true; }
+                      const Slice& next_table_min_key,
+                      const std::string& full_history_ts_low);

  // if the outputs have range delete, range delete is also data
  bool HasRangeDel() const {
@@ -193,6 +187,32 @@ class CompactionOutputs {
 private:
  friend class SubcompactionState;

+  void FillFilesToCutForTtl();
+
+  void SetOutputSlitKey(const std::optional<Slice> start,
+                        const std::optional<Slice> end) {
+    const InternalKeyComparator* icmp =
+        &compaction_->column_family_data()->internal_comparator();
+
+    const InternalKey* output_split_key = compaction_->GetOutputSplitKey();
+    // Invalid output_split_key indicates that we do not need to split
+    if (output_split_key != nullptr) {
+      // We may only split the output when the cursor is in the range. Split
+      if ((!end.has_value() ||
+           icmp->user_comparator()->Compare(
+               ExtractUserKey(output_split_key->Encode()), end.value()) < 0) &&
+          (!start.has_value() || icmp->user_comparator()->Compare(
+                                     ExtractUserKey(output_split_key->Encode()),
+                                     start.value()) > 0)) {
+        local_output_split_key_ = output_split_key;
+      }
+    }
+  }
+
+  // Returns true iff we should stop building the current output
+  // before processing the current key in compaction iterator.
+  bool ShouldStopBefore(const CompactionIterator& c_iter);
+
  void Cleanup() {
    if (builder_ != nullptr) {
      // May happen if we get a shutdown call in the middle of compaction
@@ -201,11 +221,18 @@ class CompactionOutputs {
    }
  }

-  uint64_t GetCurrentOutputFileSize() const {
-    return current_output_file_size_;
-  }
+  // update tracked grandparents information like grandparent index, if it's
+  // in the gap between 2 grandparent files, accumulated grandparent files size
+  // etc.
+  // It returns how many boundaries it crosses by including current key.
+  size_t UpdateGrandparentBoundaryInfo(const Slice& internal_key);

-  // Add curent key from compaction_iterator to the output file. If needed
+  // helper function to get the overlapped grandparent files size, it's only
+  // used for calculating the first key's overlap.
+  uint64_t GetCurrentKeyGrandparentOverlappedBytes(
+      const Slice& internal_key) const;
+
+  // Add current key from compaction_iterator to the output file. If needed
  // close and open new compaction output with the functions provided.
  Status AddToOutput(const CompactionIterator& c_iter,
                     const CompactionFileOpenFunc& open_file_func,
@@ -255,10 +282,6 @@ class CompactionOutputs {

  const Compaction* compaction_;

-  // The current file is pending close, which needs to run `close_file_func()`
-  // first to add a new key.
-  bool pending_close_ = false;
-
  // current output builder and writer
  std::unique_ptr<TableBuilder> builder_;
  std::unique_ptr<WritableFileWriter> file_writer_;
@@ -282,6 +305,39 @@ class CompactionOutputs {
  // partitioner information
  std::string last_key_for_partitioner_;
  std::unique_ptr<SstPartitioner> partitioner_;
+
+  // A flag determines if this subcompaction has been split by the cursor
+  bool is_split_ = false;
+
+  // We also maintain the output split key for each subcompaction to avoid
+  // repetitive comparison in ShouldStopBefore()
+  const InternalKey* local_output_split_key_ = nullptr;
+
+  // Some identified files with old oldest ancester time and the range should be
+  // isolated out so that the output file(s) in that range can be merged down
+  // for TTL and clear the timestamps for the range.
+  std::vector<FileMetaData*> files_to_cut_for_ttl_;
+  int cur_files_to_cut_for_ttl_ = -1;
+  int next_files_to_cut_for_ttl_ = 0;
+
+  // An index that used to speed up ShouldStopBefore().
+  size_t grandparent_index_ = 0;
+
+  // if the output key is being grandparent files gap, so:
+  //  key > grandparents[grandparent_index_ - 1].largest &&
+  //  key < grandparents[grandparent_index_].smallest
+  bool being_grandparent_gap_ = true;
+
+  // The number of bytes overlapping between the current output and
+  // grandparent files used in ShouldStopBefore().
+  uint64_t grandparent_overlapped_bytes_ = 0;
+
+  // A flag determines whether the key has been seen in ShouldStopBefore()
+  bool seen_key_ = false;
+
+  // for the current output file, how many file boundaries has it crossed,
+  // basically number of files overlapped * 2
+  size_t grandparent_boundary_switched_num_ = 0;
 };

 // helper struct to concatenate the last level and penultimate level outputs
@@ -27,16 +27,6 @@

 namespace ROCKSDB_NAMESPACE {

-namespace {
-uint64_t TotalCompensatedFileSize(const std::vector<FileMetaData*>& files) {
-  uint64_t sum = 0;
-  for (size_t i = 0; i < files.size() && files[i]; i++) {
-    sum += files[i]->compensated_file_size;
-  }
-  return sum;
-}
-}  // anonymous namespace
-
 bool FindIntraL0Compaction(const std::vector<FileMetaData*>& level_files,
                           size_t min_files_to_compact,
                           uint64_t max_compact_bytes_per_del_file,
@@ -63,8 +53,6 @@ bool FindIntraL0Compaction(const std::vector<FileMetaData*>& level_files,
    return false;
  }
  size_t compact_bytes = static_cast<size_t>(level_files[start]->fd.file_size);
-  uint64_t compensated_compact_bytes =
-      level_files[start]->compensated_file_size;
  size_t compact_bytes_per_del_file = std::numeric_limits<size_t>::max();
  // Compaction range will be [start, limit).
  size_t limit;
@@ -73,11 +61,10 @@ bool FindIntraL0Compaction(const std::vector<FileMetaData*>& level_files,
  size_t new_compact_bytes_per_del_file = 0;
  for (limit = start + 1; limit < level_files.size(); ++limit) {
    compact_bytes += static_cast<size_t>(level_files[limit]->fd.file_size);
-    compensated_compact_bytes += level_files[limit]->compensated_file_size;
    new_compact_bytes_per_del_file = compact_bytes / (limit - start);
    if (level_files[limit]->being_compacted ||
        new_compact_bytes_per_del_file > compact_bytes_per_del_file ||
-        compensated_compact_bytes > max_compaction_bytes) {
+        compact_bytes > max_compaction_bytes) {
      break;
    }
    compact_bytes_per_del_file = new_compact_bytes_per_del_file;
@@ -288,8 +275,10 @@ bool CompactionPicker::RangeOverlapWithCompaction(
  const Comparator* ucmp = icmp_->user_comparator();
  for (Compaction* c : compactions_in_progress_) {
    if (c->output_level() == level &&
-        ucmp->Compare(smallest_user_key, c->GetLargestUserKey()) <= 0 &&
-        ucmp->Compare(largest_user_key, c->GetSmallestUserKey()) >= 0) {
+        ucmp->CompareWithoutTimestamp(smallest_user_key,
+                                      c->GetLargestUserKey()) <= 0 &&
+        ucmp->CompareWithoutTimestamp(largest_user_key,
+                                      c->GetSmallestUserKey()) >= 0) {
      // Overlap
      return true;
    }
@@ -305,13 +294,12 @@ bool CompactionPicker::RangeOverlapWithCompaction(
 }

 bool CompactionPicker::FilesRangeOverlapWithCompaction(
-    const std::vector<CompactionInputFiles>& inputs, int level) const {
+    const std::vector<CompactionInputFiles>& inputs, int level,
+    int penultimate_level) const {
  bool is_empty = true;
-  int start_level = -1;
  for (auto& in : inputs) {
    if (!in.empty()) {
      is_empty = false;
-      start_level = in.level;  // inputs are sorted by level
      break;
    }
  }
@@ -320,17 +308,24 @@ bool CompactionPicker::FilesRangeOverlapWithCompaction(
    return false;
  }

+  // TODO: Intra L0 compactions can have the ranges overlapped, but the input
+  //  files cannot be overlapped in the order of L0 files.
  InternalKey smallest, largest;
  GetRange(inputs, &smallest, &largest, Compaction::kInvalidLevel);
-  int penultimate_level =
-      Compaction::EvaluatePenultimateLevel(ioptions_, start_level, level);
  if (penultimate_level != Compaction::kInvalidLevel) {
-    InternalKey penultimate_smallest, penultimate_largest;
-    GetRange(inputs, &penultimate_smallest, &penultimate_largest, level);
-    if (RangeOverlapWithCompaction(penultimate_smallest.user_key(),
-                                   penultimate_largest.user_key(),
-                                   penultimate_level)) {
-      return true;
+    if (ioptions_.compaction_style == kCompactionStyleUniversal) {
+      if (RangeOverlapWithCompaction(smallest.user_key(), largest.user_key(),
+                                     penultimate_level)) {
+        return true;
+      }
+    } else {
+      InternalKey penultimate_smallest, penultimate_largest;
+      GetRange(inputs, &penultimate_smallest, &penultimate_largest, level);
+      if (RangeOverlapWithCompaction(penultimate_smallest.user_key(),
+                                     penultimate_largest.user_key(),
+                                     penultimate_level)) {
+        return true;
+      }
    }
  }

@@ -354,11 +349,25 @@ Compaction* CompactionPicker::CompactFiles(
    const std::vector<CompactionInputFiles>& input_files, int output_level,
    VersionStorageInfo* vstorage, const MutableCFOptions& mutable_cf_options,
    const MutableDBOptions& mutable_db_options, uint32_t output_path_id) {
+#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.
-  assert(!FilesRangeOverlapWithCompaction(input_files, output_level));
+  int start_level = Compaction::kInvalidLevel;
+  for (const auto& in : input_files) {
+    // input_files should already be sorted by level
+    if (!in.empty()) {
+      start_level = in.level;
+      break;
+    }
+  }
+  assert(output_level == 0 ||
+         !FilesRangeOverlapWithCompaction(
+             input_files, output_level,
+             Compaction::EvaluatePenultimateLevel(vstorage, ioptions_,
+                                                  start_level, output_level)));
+#endif /* !NDEBUG */

  CompressionType compression_type;
  if (compact_options.compression == kDisableCompressionOption) {
@@ -505,8 +514,8 @@ bool CompactionPicker::SetupOtherInputs(
  if (!output_level_inputs->empty()) {
    const uint64_t limit = mutable_cf_options.max_compaction_bytes;
    const uint64_t output_level_inputs_size =
-        TotalCompensatedFileSize(output_level_inputs->files);
-    const uint64_t inputs_size = TotalCompensatedFileSize(inputs->files);
+        TotalFileSize(output_level_inputs->files);
+    const uint64_t inputs_size = TotalFileSize(inputs->files);
    bool expand_inputs = false;

    CompactionInputFiles expanded_inputs;
@@ -525,13 +534,13 @@ bool CompactionPicker::SetupOtherInputs(
                                     &expanded_inputs.files, base_index,
                                     nullptr);
    }
-    uint64_t expanded_inputs_size =
-        TotalCompensatedFileSize(expanded_inputs.files);
+    uint64_t expanded_inputs_size = TotalFileSize(expanded_inputs.files);
    if (!ExpandInputsToCleanCut(cf_name, vstorage, &expanded_inputs)) {
      try_overlapping_inputs = false;
    }
    if (try_overlapping_inputs && expanded_inputs.size() > inputs->size() &&
-        output_level_inputs_size + expanded_inputs_size < limit &&
+        (mutable_cf_options.ignore_max_compaction_bytes_for_input ||
+         output_level_inputs_size + expanded_inputs_size < limit) &&
        !AreFilesInCompaction(expanded_inputs.files)) {
      InternalKey new_start, new_limit;
      GetRange(expanded_inputs, &new_start, &new_limit);
@@ -552,9 +561,10 @@ bool CompactionPicker::SetupOtherInputs(
      vstorage->GetCleanInputsWithinInterval(input_level, &all_start,
                                             &all_limit, &expanded_inputs.files,
                                             base_index, nullptr);
-      expanded_inputs_size = TotalCompensatedFileSize(expanded_inputs.files);
+      expanded_inputs_size = TotalFileSize(expanded_inputs.files);
      if (expanded_inputs.size() > inputs->size() &&
-          output_level_inputs_size + expanded_inputs_size < limit &&
+          (mutable_cf_options.ignore_max_compaction_bytes_for_input ||
+           output_level_inputs_size + expanded_inputs_size < limit) &&
          !AreFilesInCompaction(expanded_inputs.files)) {
        expand_inputs = true;
      }
@@ -655,7 +665,10 @@ Compaction* CompactionPicker::CompactRange(

    // 2 non-exclusive manual compactions could run at the same time producing
    // overlaping outputs in the same level.
-    if (FilesRangeOverlapWithCompaction(inputs, output_level)) {
+    if (FilesRangeOverlapWithCompaction(
+            inputs, output_level,
+            Compaction::EvaluatePenultimateLevel(vstorage, ioptions_,
+                                                 start_level, output_level))) {
      // This compaction output could potentially conflict with the output
      // of a currently running compaction, we cannot run it.
      *manual_conflict = true;
@@ -722,18 +735,18 @@ Compaction* CompactionPicker::CompactRange(
      }
      largest = &inputs[i]->largest;

-      uint64_t s = inputs[i]->compensated_file_size;
+      uint64_t input_file_size = inputs[i]->fd.GetFileSize();
      uint64_t output_level_total = 0;
      if (output_level < vstorage->num_non_empty_levels()) {
        std::vector<FileMetaData*> files;
        vstorage->GetOverlappingInputsRangeBinarySearch(
            output_level, smallest, largest, &files, hint_index, &hint_index);
        for (const auto& file : files) {
-          output_level_total += file->compensated_file_size;
+          output_level_total += file->fd.GetFileSize();
        }
      }

-      input_level_total += s;
+      input_level_total += input_file_size;

      if (input_level_total + output_level_total >= limit) {
        covering_the_whole_range = false;
@@ -834,7 +847,10 @@ Compaction* CompactionPicker::CompactRange(

  // 2 non-exclusive manual compactions could run at the same time producing
  // overlaping outputs in the same level.
-  if (FilesRangeOverlapWithCompaction(compaction_inputs, output_level)) {
+  if (FilesRangeOverlapWithCompaction(
+          compaction_inputs, output_level,
+          Compaction::EvaluatePenultimateLevel(vstorage, ioptions_, input_level,
+                                               output_level))) {
    // This compaction output could potentially conflict with the output
    // of a currently running compaction, we cannot run it.
    *manual_conflict = true;
@@ -878,21 +894,21 @@ namespace {
 // Test whether two files have overlapping key-ranges.
 bool HaveOverlappingKeyRanges(const Comparator* c, const SstFileMetaData& a,
                              const SstFileMetaData& b) {
-  if (c->Compare(a.smallestkey, b.smallestkey) >= 0) {
-    if (c->Compare(a.smallestkey, b.largestkey) <= 0) {
+  if (c->CompareWithoutTimestamp(a.smallestkey, b.smallestkey) >= 0) {
+    if (c->CompareWithoutTimestamp(a.smallestkey, b.largestkey) <= 0) {
      // b.smallestkey <= a.smallestkey <= b.largestkey
      return true;
    }
-  } else if (c->Compare(a.largestkey, b.smallestkey) >= 0) {
+  } else if (c->CompareWithoutTimestamp(a.largestkey, b.smallestkey) >= 0) {
    // a.smallestkey < b.smallestkey <= a.largestkey
    return true;
  }
-  if (c->Compare(a.largestkey, b.largestkey) <= 0) {
-    if (c->Compare(a.largestkey, b.smallestkey) >= 0) {
+  if (c->CompareWithoutTimestamp(a.largestkey, b.largestkey) <= 0) {
+    if (c->CompareWithoutTimestamp(a.largestkey, b.smallestkey) >= 0) {
      // b.smallestkey <= a.largestkey <= b.largestkey
      return true;
    }
-  } else if (c->Compare(a.smallestkey, b.largestkey) <= 0) {
+  } else if (c->CompareWithoutTimestamp(a.smallestkey, b.largestkey) <= 0) {
    // a.smallestkey <= b.largestkey < a.largestkey
    return true;
  }
@@ -931,15 +947,16 @@ Status CompactionPicker::SanitizeCompactionInputFilesForAllLevels(
    // identify the first and the last compaction input files
    // in the current level.
    for (size_t f = 0; f < current_files.size(); ++f) {
-      if (input_files->find(TableFileNameToNumber(current_files[f].name)) !=
-          input_files->end()) {
-        first_included = std::min(first_included, static_cast<int>(f));
-        last_included = std::max(last_included, static_cast<int>(f));
-        if (is_first == false) {
-          smallestkey = current_files[f].smallestkey;
-          largestkey = current_files[f].largestkey;
-          is_first = true;
-        }
+      const uint64_t file_number = TableFileNameToNumber(current_files[f].name);
+      if (input_files->find(file_number) == input_files->end()) {
+        continue;
+      }
+      first_included = std::min(first_included, static_cast<int>(f));
+      last_included = std::max(last_included, static_cast<int>(f));
+      if (is_first == false) {
+        smallestkey = current_files[f].smallestkey;
+        largestkey = current_files[f].largestkey;
+        is_first = true;
      }
    }
    if (last_included == kNotFound) {
@@ -947,21 +964,22 @@ Status CompactionPicker::SanitizeCompactionInputFilesForAllLevels(
    }

    if (l != 0) {
-      // expend the compaction input of the current level if it
+      // expand the compaction input of the current level if it
      // has overlapping key-range with other non-compaction input
      // files in the same level.
      while (first_included > 0) {
-        if (comparator->Compare(current_files[first_included - 1].largestkey,
-                                current_files[first_included].smallestkey) <
-            0) {
+        if (comparator->CompareWithoutTimestamp(
+                current_files[first_included - 1].largestkey,
+                current_files[first_included].smallestkey) < 0) {
          break;
        }
        first_included--;
      }

      while (last_included < static_cast<int>(current_files.size()) - 1) {
-        if (comparator->Compare(current_files[last_included + 1].smallestkey,
-                                current_files[last_included].largestkey) > 0) {
+        if (comparator->CompareWithoutTimestamp(
+                current_files[last_included + 1].smallestkey,
+                current_files[last_included].largestkey) > 0) {
          break;
        }
        last_included++;
@@ -983,21 +1001,22 @@ Status CompactionPicker::SanitizeCompactionInputFilesForAllLevels(
    // update smallest and largest key
    if (l == 0) {
      for (int f = first_included; f <= last_included; ++f) {
-        if (comparator->Compare(smallestkey, current_files[f].smallestkey) >
-            0) {
+        if (comparator->CompareWithoutTimestamp(
+                smallestkey, current_files[f].smallestkey) > 0) {
          smallestkey = current_files[f].smallestkey;
        }
-        if (comparator->Compare(largestkey, current_files[f].largestkey) < 0) {
+        if (comparator->CompareWithoutTimestamp(
+                largestkey, current_files[f].largestkey) < 0) {
          largestkey = current_files[f].largestkey;
        }
      }
    } else {
-      if (comparator->Compare(smallestkey,
-                              current_files[first_included].smallestkey) > 0) {
+      if (comparator->CompareWithoutTimestamp(
+              smallestkey, current_files[first_included].smallestkey) > 0) {
        smallestkey = current_files[first_included].smallestkey;
      }
-      if (comparator->Compare(largestkey,
-                              current_files[last_included].largestkey) < 0) {
+      if (comparator->CompareWithoutTimestamp(
+              largestkey, current_files[last_included].largestkey) < 0) {
        largestkey = current_files[last_included].largestkey;
      }
    }
@@ -1116,7 +1135,8 @@ void CompactionPicker::RegisterCompaction(Compaction* c) {
  }
  assert(ioptions_.compaction_style != kCompactionStyleLevel ||
         c->output_level() == 0 ||
-         !FilesRangeOverlapWithCompaction(*c->inputs(), c->output_level()));
+         !FilesRangeOverlapWithCompaction(*c->inputs(), c->output_level(),
+                                          c->GetPenultimateLevel()));
  if (c->start_level() == 0 ||
      ioptions_.compaction_style == kCompactionStyleUniversal) {
    level0_compactions_in_progress_.insert(c);
@@ -182,7 +182,8 @@ class CompactionPicker {
  // Returns true if the key range that `inputs` files cover overlap with the
  // key range of a currently running compaction.
  bool FilesRangeOverlapWithCompaction(
-      const std::vector<CompactionInputFiles>& inputs, int level) const;
+      const std::vector<CompactionInputFiles>& inputs, int level,
+      int penultimate_level) const;

  bool SetupOtherInputs(const std::string& cf_name,
                        const MutableCFOptions& mutable_cf_options,
@@ -121,20 +121,45 @@ Compaction* FIFOCompactionPicker::PickTTLCompaction(
  return c;
 }

+// The size-based compaction picker for FIFO.
+//
+// When the entire column family size exceeds max_table_files_size, FIFO will
+// try to delete the oldest sst file(s) until the resulting column family size
+// is smaller than max_table_files_size.
+//
+// This function also takes care the case where a DB is migrating from level /
+// universal compaction to FIFO compaction.  During the migration, the column
+// family will also have non-L0 files while FIFO can only create L0 files.
+// In this case, this function will first purge the sst files in the bottom-
+// most non-empty level first, and the DB will eventually converge to the
+// regular FIFO case where there're only L0 files.  Note that during the
+// migration case, the purge order will only be an approximation of "FIFO"
+// as entries inside lower-level files might sometimes be newer than some
+// entries inside upper-level files.
 Compaction* FIFOCompactionPicker::PickSizeCompaction(
    const std::string& cf_name, const MutableCFOptions& mutable_cf_options,
    const MutableDBOptions& mutable_db_options, VersionStorageInfo* vstorage,
    LogBuffer* log_buffer) {
-  const int kLevel0 = 0;
-  const std::vector<FileMetaData*>& level_files = vstorage->LevelFiles(kLevel0);
-  uint64_t total_size = GetTotalFilesSize(level_files);
+  // compute the total size and identify the last non-empty level
+  int last_level = 0;
+  uint64_t total_size = 0;
+  for (int level = 0; level < vstorage->num_levels(); ++level) {
+    auto level_size = GetTotalFilesSize(vstorage->LevelFiles(level));
+    total_size += level_size;
+    if (level_size > 0) {
+      last_level = level;
+    }
+  }
+  const std::vector<FileMetaData*>& last_level_files =
+      vstorage->LevelFiles(last_level);

-  if (total_size <=
-          mutable_cf_options.compaction_options_fifo.max_table_files_size ||
-      level_files.size() == 0) {
-    // total size not exceeded
+  if (last_level == 0 &&
+      total_size <=
+          mutable_cf_options.compaction_options_fifo.max_table_files_size) {
+    // total size not exceeded, try to find intra level 0 compaction if enabled
+    const std::vector<FileMetaData*>& level0_files = vstorage->LevelFiles(0);
    if (mutable_cf_options.compaction_options_fifo.allow_compaction &&
-        level_files.size() > 0) {
+        level0_files.size() > 0) {
      CompactionInputFiles comp_inputs;
      // try to prevent same files from being compacted multiple times, which
      // could produce large files that may never TTL-expire. Achieve this by
@@ -146,7 +171,7 @@ Compaction* FIFOCompactionPicker::PickSizeCompaction(
              static_cast<uint64_t>(mutable_cf_options.write_buffer_size),
              1.1));
      if (FindIntraL0Compaction(
-              level_files,
+              level0_files,
              mutable_cf_options
                  .level0_file_num_compaction_trigger /* min_files_to_compact */
              ,
@@ -187,27 +212,58 @@ Compaction* FIFOCompactionPicker::PickSizeCompaction(

  std::vector<CompactionInputFiles> inputs;
  inputs.emplace_back();
-  inputs[0].level = 0;
+  inputs[0].level = last_level;

-  for (auto ritr = level_files.rbegin(); ritr != level_files.rend(); ++ritr) {
-    auto f = *ritr;
-    total_size -= f->fd.file_size;
-    inputs[0].files.push_back(f);
-    char tmp_fsize[16];
-    AppendHumanBytes(f->fd.GetFileSize(), tmp_fsize, sizeof(tmp_fsize));
-    ROCKS_LOG_BUFFER(log_buffer,
-                     "[%s] FIFO compaction: picking file %" PRIu64
-                     " with size %s for deletion",
-                     cf_name.c_str(), f->fd.GetNumber(), tmp_fsize);
-    if (total_size <=
-        mutable_cf_options.compaction_options_fifo.max_table_files_size) {
-      break;
+  if (last_level == 0) {
+    // In L0, right-most files are the oldest files.
+    for (auto ritr = last_level_files.rbegin(); ritr != last_level_files.rend();
+         ++ritr) {
+      auto f = *ritr;
+      total_size -= f->fd.file_size;
+      inputs[0].files.push_back(f);
+      char tmp_fsize[16];
+      AppendHumanBytes(f->fd.GetFileSize(), tmp_fsize, sizeof(tmp_fsize));
+      ROCKS_LOG_BUFFER(log_buffer,
+                       "[%s] FIFO compaction: picking file %" PRIu64
+                       " with size %s for deletion",
+                       cf_name.c_str(), f->fd.GetNumber(), tmp_fsize);
+      if (total_size <=
+          mutable_cf_options.compaction_options_fifo.max_table_files_size) {
+        break;
+      }
+    }
+  } else {
+    // If the last level is non-L0, we actually don't know which file is
+    // logically the oldest since the file creation time only represents
+    // when this file was compacted to this level, which is independent
+    // to when the entries in this file were first inserted.
+    //
+    // As a result, we delete files from the left instead.  This means the sst
+    // file with the smallest key will be deleted first.  This design decision
+    // better serves a major type of FIFO use cases where smaller keys are
+    // associated with older data.
+    for (const auto& f : last_level_files) {
+      total_size -= f->fd.file_size;
+      inputs[0].files.push_back(f);
+      char tmp_fsize[16];
+      AppendHumanBytes(f->fd.GetFileSize(), tmp_fsize, sizeof(tmp_fsize));
+      ROCKS_LOG_BUFFER(log_buffer,
+                       "[%s] FIFO compaction: picking file %" PRIu64
+                       " with size %s for deletion",
+                       cf_name.c_str(), f->fd.GetNumber(), tmp_fsize);
+      if (total_size <=
+          mutable_cf_options.compaction_options_fifo.max_table_files_size) {
+        break;
+      }
    }
  }

  Compaction* c = new Compaction(
      vstorage, ioptions_, mutable_cf_options, mutable_db_options,
-      std::move(inputs), 0, 0, 0, 0, kNoCompression,
+      std::move(inputs), last_level,
+      /* target_file_size */ 0,
+      /* max_compaction_bytes */ 0,
+      /* output_path_id */ 0, kNoCompression,
      mutable_cf_options.compression_opts, Temperature::kUnknown,
      /* max_subcompactions */ 0, {}, /* is manual */ false,
      /* trim_ts */ "", vstorage->CompactionScore(0),
@@ -224,6 +280,13 @@ Compaction* FIFOCompactionPicker::PickCompactionToWarm(
    return nullptr;
  }

+  // PickCompactionToWarm is only triggered if there is no non-L0 files.
+  for (int level = 1; level < vstorage->num_levels(); ++level) {
+    if (GetTotalFilesSize(vstorage->LevelFiles(level)) > 0) {
+      return nullptr;
+    }
+  }
+
  const int kLevel0 = 0;
  const std::vector<FileMetaData*>& level_files = vstorage->LevelFiles(kLevel0);

@@ -327,8 +390,6 @@ Compaction* FIFOCompactionPicker::PickCompaction(
    const std::string& cf_name, const MutableCFOptions& mutable_cf_options,
    const MutableDBOptions& mutable_db_options, VersionStorageInfo* vstorage,
    LogBuffer* log_buffer, SequenceNumber /*earliest_memtable_seqno*/) {
-  assert(vstorage->num_levels() == 1);
-
  Compaction* c = nullptr;
  if (mutable_cf_options.ttl > 0) {
    c = PickTTLCompaction(cf_name, mutable_cf_options, mutable_db_options,
@@ -254,6 +254,25 @@ void LevelCompactionBuilder::SetupInitialFiles() {
  }

  // TTL Compaction
+  if (ioptions_.compaction_pri == kRoundRobin &&
+      !vstorage_->ExpiredTtlFiles().empty()) {
+    auto expired_files = vstorage_->ExpiredTtlFiles();
+    // the expired files list should already be sorted by level
+    start_level_ = expired_files.front().first;
+#ifndef NDEBUG
+    for (const auto& file : expired_files) {
+      assert(start_level_ <= file.first);
+    }
+#endif
+    if (start_level_ > 0) {
+      output_level_ = start_level_ + 1;
+      if (PickFileToCompact()) {
+        compaction_reason_ = CompactionReason::kRoundRobinTtl;
+        return;
+      }
+    }
+  }
+
  PickFileToCompact(vstorage_->ExpiredTtlFiles(), true);
  if (!start_level_inputs_.empty()) {
    compaction_reason_ = CompactionReason::kTtl;
@@ -318,7 +337,7 @@ void LevelCompactionBuilder::SetupOtherFilesWithRoundRobinExpansion() {
  // Constraint 3 (pre-calculate the ideal max bytes to compact)
  for (auto f : level_files) {
    if (!f->being_compacted) {
-      start_lvl_bytes_no_compacting += f->compensated_file_size;
+      start_lvl_bytes_no_compacting += f->fd.GetFileSize();
    }
  }
  if (start_lvl_bytes_no_compacting >
@@ -341,7 +360,7 @@ void LevelCompactionBuilder::SetupOtherFilesWithRoundRobinExpansion() {
    }
  }
  // Constraint 3
-  if (start_level_inputs_[0]->compensated_file_size >=
+  if (start_level_inputs_[0]->fd.GetFileSize() >=
      start_lvl_max_bytes_to_compact) {
    return;
  }
@@ -360,7 +379,9 @@ void LevelCompactionBuilder::SetupOtherFilesWithRoundRobinExpansion() {
    if (!compaction_picker_->ExpandInputsToCleanCut(cf_name_, vstorage_,
                                                    &tmp_start_level_inputs) ||
        compaction_picker_->FilesRangeOverlapWithCompaction(
-            {tmp_start_level_inputs}, output_level_)) {
+            {tmp_start_level_inputs}, output_level_,
+            Compaction::EvaluatePenultimateLevel(
+                vstorage_, ioptions_, start_level_, output_level_))) {
      // Constraint 1a
      tmp_start_level_inputs.clear();
      return;
@@ -368,7 +389,7 @@ void LevelCompactionBuilder::SetupOtherFilesWithRoundRobinExpansion() {

    curr_bytes_to_compact = 0;
    for (auto start_lvl_f : tmp_start_level_inputs.files) {
-      curr_bytes_to_compact += start_lvl_f->compensated_file_size;
+      curr_bytes_to_compact += start_lvl_f->fd.GetFileSize();
    }

    // Check whether any output level files are locked
@@ -385,7 +406,7 @@ void LevelCompactionBuilder::SetupOtherFilesWithRoundRobinExpansion() {

    uint64_t start_lvl_curr_bytes_to_compact = curr_bytes_to_compact;
    for (auto output_lvl_f : output_level_inputs.files) {
-      curr_bytes_to_compact += output_lvl_f->compensated_file_size;
+      curr_bytes_to_compact += output_lvl_f->fd.GetFileSize();
    }
    if (curr_bytes_to_compact > mutable_cf_options_.max_compaction_bytes) {
      // Constraint 2
@@ -434,7 +455,9 @@ bool LevelCompactionBuilder::SetupOtherInputsIfNeeded() {
      // (2) AddFile ingest a new file into the LSM tree
      // We need to disallow this from happening.
      if (compaction_picker_->FilesRangeOverlapWithCompaction(
-              compaction_inputs_, output_level_)) {
+              compaction_inputs_, output_level_,
+              Compaction::EvaluatePenultimateLevel(
+                  vstorage_, ioptions_, start_level_, output_level_))) {
        // This compaction output could potentially conflict with the output
        // of a currently running compaction, we cannot run it.
        return false;
@@ -662,9 +685,14 @@ bool LevelCompactionBuilder::TryExtendNonL0TrivialMove(int start_index) {
        break;
      }
      if (i < static_cast<int>(level_files.size()) - 1 &&
-          compaction_picker_->icmp()->user_comparator()->Compare(
-              next_file->largest.user_key(),
-              level_files[i + 1]->smallest.user_key()) == 0) {
+          compaction_picker_->icmp()
+                  ->user_comparator()
+                  ->CompareWithoutTimestamp(
+                      next_file->largest.user_key(),
+                      level_files[i + 1]->smallest.user_key()) == 0) {
+        TEST_SYNC_POINT_CALLBACK(
+            "LevelCompactionBuilder::TryExtendNonL0TrivialMove:NoCleanCut",
+            nullptr);
        // Not a clean up after adding the next file. Skip.
        break;
      }
@@ -731,7 +759,9 @@ bool LevelCompactionBuilder::PickFileToCompact() {
    if (!compaction_picker_->ExpandInputsToCleanCut(cf_name_, vstorage_,
                                                    &start_level_inputs_) ||
        compaction_picker_->FilesRangeOverlapWithCompaction(
-            {start_level_inputs_}, output_level_)) {
+            {start_level_inputs_}, output_level_,
+            Compaction::EvaluatePenultimateLevel(
+                vstorage_, ioptions_, start_level_, output_level_))) {
      // A locked (pending compaction) input-level file was pulled in due to
      // user-key overlap.
      start_level_inputs_.clear();
@@ -27,7 +27,7 @@ class CountingLogger : public Logger {
  size_t log_count;
 };

-class CompactionPickerTest : public testing::Test {
+class CompactionPickerTestBase : public testing::Test {
 public:
  const Comparator* ucmp_;
  InternalKeyComparator icmp_;
@@ -49,9 +49,10 @@ class CompactionPickerTest : public testing::Test {
  std::vector<CompactionInputFiles> input_files_;
  int compaction_level_start_;

-  CompactionPickerTest()
-      : ucmp_(BytewiseComparator()),
+  explicit CompactionPickerTestBase(const Comparator* _ucmp)
+      : ucmp_(_ucmp),
        icmp_(ucmp_),
+        options_(CreateOptions(ucmp_)),
        ioptions_(options_),
        mutable_cf_options_(options_),
        mutable_db_options_(),
@@ -71,7 +72,7 @@ class CompactionPickerTest : public testing::Test {
                                    std::numeric_limits<uint64_t>::max());
  }

-  ~CompactionPickerTest() override {}
+  ~CompactionPickerTestBase() override {}

  void NewVersionStorage(int num_levels, CompactionStyle style) {
    DeleteVersionStorage();
@@ -97,12 +98,17 @@ class CompactionPickerTest : public testing::Test {
    input_files_.clear();
  }

+  // REQUIRES: smallest and largest are c-style strings ending with '\0'
  void Add(int level, uint32_t file_number, const char* smallest,
           const char* largest, uint64_t file_size = 1, uint32_t path_id = 0,
           SequenceNumber smallest_seq = 100, SequenceNumber largest_seq = 100,
           size_t compensated_file_size = 0, bool marked_for_compact = false,
           Temperature temperature = Temperature::kUnknown,
-           uint64_t oldest_ancestor_time = kUnknownOldestAncesterTime) {
+           uint64_t oldest_ancestor_time = kUnknownOldestAncesterTime,
+           Slice ts_of_smallest = Slice(), Slice ts_of_largest = Slice()) {
+    assert(ts_of_smallest.size() == ucmp_->timestamp_size());
+    assert(ts_of_largest.size() == ucmp_->timestamp_size());
+
    VersionStorageInfo* vstorage;
    if (temp_vstorage_) {
      vstorage = temp_vstorage_.get();
@@ -110,19 +116,46 @@ class CompactionPickerTest : public testing::Test {
      vstorage = vstorage_.get();
    }
    assert(level < vstorage->num_levels());
+    char* smallest_key_buf = nullptr;
+    char* largest_key_buf = nullptr;
+
+    if (!ts_of_smallest.empty()) {
+      smallest_key_buf = new char[strlen(smallest) + ucmp_->timestamp_size()];
+      memcpy(smallest_key_buf, smallest, strlen(smallest));
+      memcpy(smallest_key_buf + strlen(smallest), ts_of_smallest.data(),
+             ucmp_->timestamp_size());
+      largest_key_buf = new char[strlen(largest) + ucmp_->timestamp_size()];
+      memcpy(largest_key_buf, largest, strlen(largest));
+      memcpy(largest_key_buf + strlen(largest), ts_of_largest.data(),
+             ucmp_->timestamp_size());
+    }
+
+    InternalKey smallest_ikey = InternalKey(
+        smallest_key_buf ? Slice(smallest_key_buf,
+                                 ucmp_->timestamp_size() + strlen(smallest))
+                         : smallest,
+        smallest_seq, kTypeValue);
+    InternalKey largest_ikey = InternalKey(
+        largest_key_buf
+            ? Slice(largest_key_buf, ucmp_->timestamp_size() + strlen(largest))
+            : largest,
+        largest_seq, kTypeValue);
+
    FileMetaData* f = new FileMetaData(
-        file_number, path_id, file_size,
-        InternalKey(smallest, smallest_seq, kTypeValue),
-        InternalKey(largest, largest_seq, kTypeValue), smallest_seq,
-        largest_seq, marked_for_compact, temperature, kInvalidBlobFileNumber,
-        kUnknownOldestAncesterTime, kUnknownFileCreationTime,
-        kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2);
+        file_number, path_id, file_size, smallest_ikey, largest_ikey,
+        smallest_seq, largest_seq, marked_for_compact, temperature,
+        kInvalidBlobFileNumber, kUnknownOldestAncesterTime,
+        kUnknownFileCreationTime, kUnknownFileChecksum,
+        kUnknownFileChecksumFuncName, kNullUniqueId64x2);
    f->compensated_file_size =
        (compensated_file_size != 0) ? compensated_file_size : file_size;
    f->oldest_ancester_time = oldest_ancestor_time;
    vstorage->AddFile(level, f);
    files_.emplace_back(f);
    file_map_.insert({file_number, {f, level}});
+
+    delete[] smallest_key_buf;
+    delete[] largest_key_buf;
  }

  void SetCompactionInputFilesLevels(int level_count, int start_level) {
@@ -155,9 +188,31 @@ class CompactionPickerTest : public testing::Test {
  }

 private:
+  Options CreateOptions(const Comparator* ucmp) const {
+    Options opts;
+    opts.comparator = ucmp;
+    return opts;
+  }
+
  std::unique_ptr<VersionStorageInfo> temp_vstorage_;
 };

+class CompactionPickerTest : public CompactionPickerTestBase {
+ public:
+  explicit CompactionPickerTest()
+      : CompactionPickerTestBase(BytewiseComparator()) {}
+
+  ~CompactionPickerTest() override {}
+};
+
+class CompactionPickerU64TsTest : public CompactionPickerTestBase {
+ public:
+  explicit CompactionPickerU64TsTest()
+      : CompactionPickerTestBase(test::BytewiseComparatorWithU64TsWrapper()) {}
+
+  ~CompactionPickerU64TsTest() override {}
+};
+
 TEST_F(CompactionPickerTest, Empty) {
  NewVersionStorage(6, kCompactionStyleLevel);
  UpdateVersionStorageInfo();
@@ -1285,6 +1340,7 @@ TEST_F(CompactionPickerTest, CompactionPriMinOverlapping4) {
  ioptions_.compaction_pri = kMinOverlappingRatio;
  mutable_cf_options_.max_bytes_for_level_base = 10000000;
  mutable_cf_options_.max_bytes_for_level_multiplier = 10;
+  mutable_cf_options_.ignore_max_compaction_bytes_for_input = false;

  // file 7 and 8 over lap with the same file, but file 8 is smaller so
  // it will be picked.
@@ -2303,6 +2359,7 @@ TEST_F(CompactionPickerTest, IsBottommostLevelTest) {
 TEST_F(CompactionPickerTest, MaxCompactionBytesHit) {
  mutable_cf_options_.max_bytes_for_level_base = 1000000u;
  mutable_cf_options_.max_compaction_bytes = 800000u;
+  mutable_cf_options_.ignore_max_compaction_bytes_for_input = false;
  ioptions_.level_compaction_dynamic_level_bytes = false;
  NewVersionStorage(6, kCompactionStyleLevel);
  // A compaction should be triggered and pick file 2 and 5.
@@ -2329,6 +2386,7 @@ TEST_F(CompactionPickerTest, MaxCompactionBytesHit) {
 TEST_F(CompactionPickerTest, MaxCompactionBytesNotHit) {
  mutable_cf_options_.max_bytes_for_level_base = 800000u;
  mutable_cf_options_.max_compaction_bytes = 1000000u;
+  mutable_cf_options_.ignore_max_compaction_bytes_for_input = false;
  ioptions_.level_compaction_dynamic_level_bytes = false;
  NewVersionStorage(6, kCompactionStyleLevel);
  // A compaction should be triggered and pick file 2 and 5.
@@ -3091,7 +3149,7 @@ TEST_F(CompactionPickerTest, UniversalMarkedL0Overlap2) {
  // should fail
  NewVersionStorage(1, kCompactionStyleUniversal);

-  // Mark file number 4 for compaction
+  // Mark file number 5 for compaction
  Add(0, 4U, "260", "300", 1 * kFileSize, 0, 260, 300);
  Add(0, 5U, "240", "290", 2 * kFileSize, 0, 201, 250, 0, true);
  Add(0, 3U, "301", "350", 4 * kFileSize, 0, 101, 150);
@@ -3317,6 +3375,107 @@ TEST_F(CompactionPickerTest, UniversalSizeAmpTierCompactionLastLevel) {
  ASSERT_EQ(compaction->input_levels(6)->num_files, 0);
 }

+TEST_F(CompactionPickerU64TsTest, Overlap) {
+  int num_levels = ioptions_.num_levels;
+  NewVersionStorage(num_levels, kCompactionStyleLevel);
+
+  constexpr int level = 0;
+  constexpr uint64_t file_number = 20ULL;
+  constexpr char smallest[4] = "500";
+  constexpr char largest[4] = "600";
+  constexpr uint64_t ts_of_smallest = 12345ULL;
+  constexpr uint64_t ts_of_largest = 56789ULL;
+
+  {
+    std::string ts1;
+    PutFixed64(&ts1, ts_of_smallest);
+    std::string ts2;
+    PutFixed64(&ts2, ts_of_largest);
+    Add(level, file_number, smallest, largest,
+        /*file_size=*/1U, /*path_id=*/0,
+        /*smallest_seq=*/100, /*largest_seq=*/100, /*compensated_file_size=*/0,
+        /*marked_for_compact=*/false, /*temperature=*/Temperature::kUnknown,
+        /*oldest_ancestor_time=*/kUnknownOldestAncesterTime, ts1, ts2);
+    UpdateVersionStorageInfo();
+  }
+
+  std::unordered_set<uint64_t> input{file_number};
+
+  std::vector<CompactionInputFiles> input_files;
+  ASSERT_OK(level_compaction_picker.GetCompactionInputsFromFileNumbers(
+      &input_files, &input, vstorage_.get(), CompactionOptions()));
+  std::unique_ptr<Compaction> comp1(level_compaction_picker.CompactFiles(
+      CompactionOptions(), input_files, level, vstorage_.get(),
+      mutable_cf_options_, mutable_db_options_, /*output_path_id=*/0));
+
+  {
+    // [600, ts=50000] to [600, ts=50000] is the range to check.
+    // ucmp->Compare(smallest_user_key, c->GetLargestUserKey()) > 0, but
+    // ucmp->CompareWithoutTimestamp(smallest_user_key,
+    //                               c->GetLargestUserKey()) == 0.
+    // Should still be considered overlapping.
+    std::string user_key_with_ts1(largest);
+    PutFixed64(&user_key_with_ts1, ts_of_largest - 1);
+    std::string user_key_with_ts2(largest);
+    PutFixed64(&user_key_with_ts2, ts_of_largest - 1);
+    ASSERT_TRUE(level_compaction_picker.RangeOverlapWithCompaction(
+        user_key_with_ts1, user_key_with_ts2, level));
+  }
+  {
+    // [500, ts=60000] to [500, ts=60000] is the range to check.
+    // ucmp->Compare(largest_user_key, c->GetSmallestUserKey()) < 0, but
+    // ucmp->CompareWithoutTimestamp(largest_user_key,
+    //                               c->GetSmallestUserKey()) == 0.
+    // Should still be considered overlapping.
+    std::string user_key_with_ts1(smallest);
+    PutFixed64(&user_key_with_ts1, ts_of_smallest + 1);
+    std::string user_key_with_ts2(smallest);
+    PutFixed64(&user_key_with_ts2, ts_of_smallest + 1);
+    ASSERT_TRUE(level_compaction_picker.RangeOverlapWithCompaction(
+        user_key_with_ts1, user_key_with_ts2, level));
+  }
+}
+
+TEST_F(CompactionPickerU64TsTest, CannotTrivialMoveUniversal) {
+  constexpr uint64_t kFileSize = 100000;
+
+  mutable_cf_options_.level0_file_num_compaction_trigger = 2;
+  mutable_cf_options_.compaction_options_universal.allow_trivial_move = true;
+  NewVersionStorage(1, kCompactionStyleUniversal);
+  UniversalCompactionPicker universal_compaction_picker(ioptions_, &icmp_);
+  UpdateVersionStorageInfo();
+  // must return false when there's no files.
+  ASSERT_FALSE(universal_compaction_picker.NeedsCompaction(vstorage_.get()));
+
+  std::string ts1;
+  PutFixed64(&ts1, 9000);
+  std::string ts2;
+  PutFixed64(&ts2, 8000);
+  std::string ts3;
+  PutFixed64(&ts3, 7000);
+  std::string ts4;
+  PutFixed64(&ts4, 6000);
+
+  NewVersionStorage(3, kCompactionStyleUniversal);
+  // A compaction should be triggered and pick file 2
+  Add(1, 1U, "150", "150", kFileSize, /*path_id=*/0, /*smallest_seq=*/100,
+      /*largest_seq=*/100, /*compensated_file_size=*/kFileSize,
+      /*marked_for_compact=*/false, Temperature::kUnknown,
+      kUnknownOldestAncesterTime, ts1, ts2);
+  Add(2, 2U, "150", "150", kFileSize, /*path_id=*/0, /*smallest_seq=*/100,
+      /*largest_seq=*/100, /*compensated_file_size=*/kFileSize,
+      /*marked_for_compact=*/false, Temperature::kUnknown,
+      kUnknownOldestAncesterTime, ts3, ts4);
+  UpdateVersionStorageInfo();
+
+  std::unique_ptr<Compaction> compaction(
+      universal_compaction_picker.PickCompaction(
+          cf_name_, mutable_cf_options_, mutable_db_options_, vstorage_.get(),
+          &log_buffer_));
+  assert(compaction);
+  ASSERT_TRUE(!compaction->is_trivial_move());
+}
+
 class PerKeyPlacementCompactionPickerTest
    : public CompactionPickerTest,
      public testing::WithParamInterface<bool> {
@@ -3365,9 +3524,11 @@ TEST_P(PerKeyPlacementCompactionPickerTest, OverlapWithNormalCompaction) {
  ASSERT_OK(level_compaction_picker.GetCompactionInputsFromFileNumbers(
      &input_files, &input_set, vstorage_.get(), comp_options));

-  ASSERT_EQ(
-      enable_per_key_placement_,
-      level_compaction_picker.FilesRangeOverlapWithCompaction(input_files, 6));
+  ASSERT_EQ(enable_per_key_placement_,
+            level_compaction_picker.FilesRangeOverlapWithCompaction(
+                input_files, 6,
+                Compaction::EvaluatePenultimateLevel(vstorage_.get(), ioptions_,
+                                                     0, 6)));
 }

 TEST_P(PerKeyPlacementCompactionPickerTest, NormalCompactionOverlap) {
@@ -3408,9 +3569,9 @@ TEST_P(PerKeyPlacementCompactionPickerTest, NormalCompactionOverlap) {
  ASSERT_OK(level_compaction_picker.GetCompactionInputsFromFileNumbers(
      &input_files, &input_set, vstorage_.get(), comp_options));

-  ASSERT_EQ(
-      enable_per_key_placement_,
-      level_compaction_picker.FilesRangeOverlapWithCompaction(input_files, 5));
+  ASSERT_EQ(enable_per_key_placement_,
+            level_compaction_picker.FilesRangeOverlapWithCompaction(
+                input_files, 5, Compaction::kInvalidLevel));
 }

 TEST_P(PerKeyPlacementCompactionPickerTest,
@@ -3453,7 +3614,9 @@ TEST_P(PerKeyPlacementCompactionPickerTest,

  ASSERT_EQ(enable_per_key_placement_,
            universal_compaction_picker.FilesRangeOverlapWithCompaction(
-                input_files, 6));
+                input_files, 6,
+                Compaction::EvaluatePenultimateLevel(vstorage_.get(), ioptions_,
+                                                     0, 6)));
 }

 TEST_P(PerKeyPlacementCompactionPickerTest, NormalCompactionOverlapUniversal) {
@@ -3497,7 +3660,294 @@ TEST_P(PerKeyPlacementCompactionPickerTest, NormalCompactionOverlapUniversal) {

  ASSERT_EQ(enable_per_key_placement_,
            universal_compaction_picker.FilesRangeOverlapWithCompaction(
-                input_files, 5));
+                input_files, 5, Compaction::kInvalidLevel));
+}
+
+TEST_P(PerKeyPlacementCompactionPickerTest, PenultimateOverlapUniversal) {
+  // This test is make sure the Tiered compaction would lock whole range of
+  // both output level and penultimate level
+  if (enable_per_key_placement_) {
+    ioptions_.preclude_last_level_data_seconds = 10000;
+  }
+
+  int num_levels = ioptions_.num_levels;
+  ioptions_.compaction_style = kCompactionStyleUniversal;
+  UniversalCompactionPicker universal_compaction_picker(ioptions_, &icmp_);
+  NewVersionStorage(num_levels, kCompactionStyleUniversal);
+
+  // L4:   [200, 220]  [230, 250]       [360, 380]
+  // L5:
+  // L6: [101,                    351]
+  Add(4, 40U, "200", "220", 60000000U);
+  Add(4, 41U, "230", "250", 60000000U);
+  Add(4, 42U, "360", "380", 60000000U);
+  Add(6, 60U, "101", "351", 60000000U);
+  UpdateVersionStorageInfo();
+
+  // the existing compaction is the 1st L4 file + L6 file
+  // then compaction of the 2nd L4 file to L5 (penultimate level) is overlapped
+  // when the tiered compaction feature is on.
+  CompactionOptions comp_options;
+  std::unordered_set<uint64_t> input_set;
+  input_set.insert(40);
+  input_set.insert(60);
+  std::vector<CompactionInputFiles> input_files;
+  ASSERT_OK(universal_compaction_picker.GetCompactionInputsFromFileNumbers(
+      &input_files, &input_set, vstorage_.get(), comp_options));
+
+  std::unique_ptr<Compaction> comp1(universal_compaction_picker.CompactFiles(
+      comp_options, input_files, 6, vstorage_.get(), mutable_cf_options_,
+      mutable_db_options_, 0));
+
+  input_set.clear();
+  input_files.clear();
+  input_set.insert(41);
+  ASSERT_OK(universal_compaction_picker.GetCompactionInputsFromFileNumbers(
+      &input_files, &input_set, vstorage_.get(), comp_options));
+
+  ASSERT_EQ(enable_per_key_placement_,
+            universal_compaction_picker.FilesRangeOverlapWithCompaction(
+                input_files, 5, Compaction::kInvalidLevel));
+
+  // compacting the 3rd L4 file is always safe:
+  input_set.clear();
+  input_files.clear();
+  input_set.insert(42);
+  ASSERT_OK(universal_compaction_picker.GetCompactionInputsFromFileNumbers(
+      &input_files, &input_set, vstorage_.get(), comp_options));
+
+  ASSERT_FALSE(universal_compaction_picker.FilesRangeOverlapWithCompaction(
+      input_files, 5, Compaction::kInvalidLevel));
+}
+
+TEST_P(PerKeyPlacementCompactionPickerTest, LastLevelOnlyOverlapUniversal) {
+  if (enable_per_key_placement_) {
+    ioptions_.preclude_last_level_data_seconds = 10000;
+  }
+
+  int num_levels = ioptions_.num_levels;
+  ioptions_.compaction_style = kCompactionStyleUniversal;
+  UniversalCompactionPicker universal_compaction_picker(ioptions_, &icmp_);
+  NewVersionStorage(num_levels, kCompactionStyleUniversal);
+
+  // L4:   [200, 220]  [230, 250]       [360, 380]
+  // L5:
+  // L6: [101,                    351]
+  Add(4, 40U, "200", "220", 60000000U);
+  Add(4, 41U, "230", "250", 60000000U);
+  Add(4, 42U, "360", "380", 60000000U);
+  Add(6, 60U, "101", "351", 60000000U);
+  UpdateVersionStorageInfo();
+
+  CompactionOptions comp_options;
+  std::unordered_set<uint64_t> input_set;
+  input_set.insert(60);
+  std::vector<CompactionInputFiles> input_files;
+  ASSERT_OK(universal_compaction_picker.GetCompactionInputsFromFileNumbers(
+      &input_files, &input_set, vstorage_.get(), comp_options));
+
+  std::unique_ptr<Compaction> comp1(universal_compaction_picker.CompactFiles(
+      comp_options, input_files, 6, vstorage_.get(), mutable_cf_options_,
+      mutable_db_options_, 0));
+
+  // cannot compact file 41 if the preclude_last_level feature is on, otherwise
+  // compact file 41 is okay.
+  input_set.clear();
+  input_files.clear();
+  input_set.insert(41);
+  ASSERT_OK(universal_compaction_picker.GetCompactionInputsFromFileNumbers(
+      &input_files, &input_set, vstorage_.get(), comp_options));
+
+  ASSERT_EQ(enable_per_key_placement_,
+            universal_compaction_picker.FilesRangeOverlapWithCompaction(
+                input_files, 5, Compaction::kInvalidLevel));
+
+  // compacting the 3rd L4 file is always safe:
+  input_set.clear();
+  input_files.clear();
+  input_set.insert(42);
+  ASSERT_OK(universal_compaction_picker.GetCompactionInputsFromFileNumbers(
+      &input_files, &input_set, vstorage_.get(), comp_options));
+
+  ASSERT_FALSE(universal_compaction_picker.FilesRangeOverlapWithCompaction(
+      input_files, 5, Compaction::kInvalidLevel));
+}
+
+TEST_P(PerKeyPlacementCompactionPickerTest,
+       LastLevelOnlyFailPenultimateUniversal) {
+  // This is to test last_level only compaction still unable to do the
+  // penultimate level compaction if there's already a file in the penultimate
+  // level.
+  // This should rarely happen in universal compaction, as the non-empty L5
+  // should be included in the compaction.
+  if (enable_per_key_placement_) {
+    ioptions_.preclude_last_level_data_seconds = 10000;
+  }
+
+  int num_levels = ioptions_.num_levels;
+  ioptions_.compaction_style = kCompactionStyleUniversal;
+  UniversalCompactionPicker universal_compaction_picker(ioptions_, &icmp_);
+  NewVersionStorage(num_levels, kCompactionStyleUniversal);
+
+  // L4:   [200, 220]
+  // L5:              [230, 250]
+  // L6: [101,                    351]
+  Add(4, 40U, "200", "220", 60000000U);
+  Add(5, 50U, "230", "250", 60000000U);
+  Add(6, 60U, "101", "351", 60000000U);
+  UpdateVersionStorageInfo();
+
+  CompactionOptions comp_options;
+  std::unordered_set<uint64_t> input_set;
+  input_set.insert(60);
+  std::vector<CompactionInputFiles> input_files;
+  ASSERT_OK(universal_compaction_picker.GetCompactionInputsFromFileNumbers(
+      &input_files, &input_set, vstorage_.get(), comp_options));
+
+  std::unique_ptr<Compaction> comp1(universal_compaction_picker.CompactFiles(
+      comp_options, input_files, 6, vstorage_.get(), mutable_cf_options_,
+      mutable_db_options_, 0));
+
+  ASSERT_TRUE(comp1);
+  ASSERT_EQ(comp1->GetPenultimateLevel(), Compaction::kInvalidLevel);
+
+  // As comp1 cannot be output to the penultimate level, compacting file 40 to
+  // L5 is always safe.
+  input_set.clear();
+  input_files.clear();
+  input_set.insert(40);
+  ASSERT_OK(universal_compaction_picker.GetCompactionInputsFromFileNumbers(
+      &input_files, &input_set, vstorage_.get(), comp_options));
+
+  ASSERT_FALSE(universal_compaction_picker.FilesRangeOverlapWithCompaction(
+      input_files, 5, Compaction::kInvalidLevel));
+
+  std::unique_ptr<Compaction> comp2(universal_compaction_picker.CompactFiles(
+      comp_options, input_files, 5, vstorage_.get(), mutable_cf_options_,
+      mutable_db_options_, 0));
+  ASSERT_TRUE(comp2);
+  ASSERT_EQ(Compaction::kInvalidLevel, comp2->GetPenultimateLevel());
+}
+
+TEST_P(PerKeyPlacementCompactionPickerTest,
+       LastLevelOnlyConflictWithOngoingUniversal) {
+  // This is to test last_level only compaction still unable to do the
+  // penultimate level compaction if there's already an ongoing compaction to
+  // the penultimate level
+  if (enable_per_key_placement_) {
+    ioptions_.preclude_last_level_data_seconds = 10000;
+  }
+
+  int num_levels = ioptions_.num_levels;
+  ioptions_.compaction_style = kCompactionStyleUniversal;
+  UniversalCompactionPicker universal_compaction_picker(ioptions_, &icmp_);
+  NewVersionStorage(num_levels, kCompactionStyleUniversal);
+
+  // L4:   [200, 220]  [230, 250]       [360, 380]
+  // L5:
+  // L6: [101,                    351]
+  Add(4, 40U, "200", "220", 60000000U);
+  Add(4, 41U, "230", "250", 60000000U);
+  Add(4, 42U, "360", "380", 60000000U);
+  Add(6, 60U, "101", "351", 60000000U);
+  UpdateVersionStorageInfo();
+
+  // create an ongoing compaction to L5 (penultimate level)
+  CompactionOptions comp_options;
+  std::unordered_set<uint64_t> input_set;
+  input_set.insert(40);
+  std::vector<CompactionInputFiles> input_files;
+  ASSERT_OK(universal_compaction_picker.GetCompactionInputsFromFileNumbers(
+      &input_files, &input_set, vstorage_.get(), comp_options));
+
+  std::unique_ptr<Compaction> comp1(universal_compaction_picker.CompactFiles(
+      comp_options, input_files, 5, vstorage_.get(), mutable_cf_options_,
+      mutable_db_options_, 0));
+
+  ASSERT_TRUE(comp1);
+  ASSERT_EQ(comp1->GetPenultimateLevel(), Compaction::kInvalidLevel);
+
+  input_set.clear();
+  input_files.clear();
+  input_set.insert(60);
+  ASSERT_OK(universal_compaction_picker.GetCompactionInputsFromFileNumbers(
+      &input_files, &input_set, vstorage_.get(), comp_options));
+
+  ASSERT_EQ(enable_per_key_placement_,
+            universal_compaction_picker.FilesRangeOverlapWithCompaction(
+                input_files, 6,
+                Compaction::EvaluatePenultimateLevel(vstorage_.get(), ioptions_,
+                                                     6, 6)));
+
+  if (!enable_per_key_placement_) {
+    std::unique_ptr<Compaction> comp2(universal_compaction_picker.CompactFiles(
+        comp_options, input_files, 6, vstorage_.get(), mutable_cf_options_,
+        mutable_db_options_, 0));
+    ASSERT_TRUE(comp2);
+    ASSERT_EQ(Compaction::kInvalidLevel, comp2->GetPenultimateLevel());
+  }
+}
+
+TEST_P(PerKeyPlacementCompactionPickerTest,
+       LastLevelOnlyNoConflictWithOngoingUniversal) {
+  // This is similar to `LastLevelOnlyConflictWithOngoingUniversal`, the only
+  // change is the ongoing compaction to L5 has no overlap with the last level
+  // compaction, so it's safe to move data from the last level to the
+  // penultimate level.
+  if (enable_per_key_placement_) {
+    ioptions_.preclude_last_level_data_seconds = 10000;
+  }
+
+  int num_levels = ioptions_.num_levels;
+  ioptions_.compaction_style = kCompactionStyleUniversal;
+  UniversalCompactionPicker universal_compaction_picker(ioptions_, &icmp_);
+  NewVersionStorage(num_levels, kCompactionStyleUniversal);
+
+  // L4:   [200, 220]  [230, 250]       [360, 380]
+  // L5:
+  // L6: [101,                    351]
+  Add(4, 40U, "200", "220", 60000000U);
+  Add(4, 41U, "230", "250", 60000000U);
+  Add(4, 42U, "360", "380", 60000000U);
+  Add(6, 60U, "101", "351", 60000000U);
+  UpdateVersionStorageInfo();
+
+  // create an ongoing compaction to L5 (penultimate level)
+  CompactionOptions comp_options;
+  std::unordered_set<uint64_t> input_set;
+  input_set.insert(42);
+  std::vector<CompactionInputFiles> input_files;
+  ASSERT_OK(universal_compaction_picker.GetCompactionInputsFromFileNumbers(
+      &input_files, &input_set, vstorage_.get(), comp_options));
+
+  std::unique_ptr<Compaction> comp1(universal_compaction_picker.CompactFiles(
+      comp_options, input_files, 5, vstorage_.get(), mutable_cf_options_,
+      mutable_db_options_, 0));
+
+  ASSERT_TRUE(comp1);
+  ASSERT_EQ(comp1->GetPenultimateLevel(), Compaction::kInvalidLevel);
+
+  input_set.clear();
+  input_files.clear();
+  input_set.insert(60);
+  ASSERT_OK(universal_compaction_picker.GetCompactionInputsFromFileNumbers(
+      &input_files, &input_set, vstorage_.get(), comp_options));
+
+  // always safe to move data up
+  ASSERT_FALSE(universal_compaction_picker.FilesRangeOverlapWithCompaction(
+      input_files, 6,
+      Compaction::EvaluatePenultimateLevel(vstorage_.get(), ioptions_, 6, 6)));
+
+  // 2 compactions can be run in parallel
+  std::unique_ptr<Compaction> comp2(universal_compaction_picker.CompactFiles(
+      comp_options, input_files, 6, vstorage_.get(), mutable_cf_options_,
+      mutable_db_options_, 0));
+  ASSERT_TRUE(comp2);
+  if (enable_per_key_placement_) {
+    ASSERT_NE(Compaction::kInvalidLevel, comp2->GetPenultimateLevel());
+  } else {
+    ASSERT_EQ(Compaction::kInvalidLevel, comp2->GetPenultimateLevel());
+  }
 }

 INSTANTIATE_TEST_CASE_P(PerKeyPlacementCompactionPickerTest,
@@ -3508,6 +3958,7 @@ INSTANTIATE_TEST_CASE_P(PerKeyPlacementCompactionPickerTest,
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -164,8 +164,8 @@ struct SmallestKeyHeapComparator {
  explicit SmallestKeyHeapComparator(const Comparator* ucmp) { ucmp_ = ucmp; }

  bool operator()(InputFileInfo i1, InputFileInfo i2) const {
-    return (ucmp_->Compare(i1.f->smallest.user_key(),
-                           i2.f->smallest.user_key()) > 0);
+    return (ucmp_->CompareWithoutTimestamp(i1.f->smallest.user_key(),
+                                           i2.f->smallest.user_key()) > 0);
  }

 private:
@@ -249,13 +249,13 @@ bool UniversalCompactionBuilder::IsInputFilesNonOverlapping(Compaction* c) {
      prev = curr;
      first_iter = 0;
    } else {
-      if (comparator->Compare(prev.f->largest.user_key(),
-                              curr.f->smallest.user_key()) >= 0) {
+      if (comparator->CompareWithoutTimestamp(
+              prev.f->largest.user_key(), curr.f->smallest.user_key()) >= 0) {
        // found overlapping files, return false
        return false;
      }
-      assert(comparator->Compare(curr.f->largest.user_key(),
-                                 prev.f->largest.user_key()) > 0);
+      assert(comparator->CompareWithoutTimestamp(
+                 curr.f->largest.user_key(), prev.f->largest.user_key()) > 0);
      prev = curr;
    }

@@ -743,6 +743,13 @@ Compaction* UniversalCompactionBuilder::PickCompactionToReduceSortedRuns(
    grandparents = vstorage_->LevelFiles(sorted_runs_[first_index_after].level);
  }

+  if (output_level != 0 &&
+      picker_->FilesRangeOverlapWithCompaction(
+          inputs, output_level,
+          Compaction::EvaluatePenultimateLevel(vstorage_, ioptions_,
+                                               start_level, output_level))) {
+    return nullptr;
+  }
  CompactionReason compaction_reason;
  if (max_number_of_files_to_compact == UINT_MAX) {
    compaction_reason = CompactionReason::kUniversalSizeRatio;
@@ -755,9 +762,9 @@ Compaction* UniversalCompactionBuilder::PickCompactionToReduceSortedRuns(
                                            kCompactionStyleUniversal),
                        GetMaxOverlappingBytes(), path_id,
                        GetCompressionType(vstorage_, mutable_cf_options_,
-                                           start_level, 1, enable_compression),
+                                           output_level, 1, enable_compression),
                        GetCompressionOptions(mutable_cf_options_, vstorage_,
-                                              start_level, enable_compression),
+                                              output_level, enable_compression),
                        Temperature::kUnknown,
                        /* max_subcompactions */ 0, grandparents,
                        /* is manual */ false, /* trim_ts */ "", score_,
@@ -1081,6 +1088,24 @@ Compaction* UniversalCompactionBuilder::PickIncrementalForReduceSizeAmp(
  inputs.push_back(second_last_level_inputs);
  inputs.push_back(bottom_level_inputs);

+  int start_level = Compaction::kInvalidLevel;
+  for (const auto& in : inputs) {
+    if (!in.empty()) {
+      // inputs should already be sorted by level
+      start_level = in.level;
+      break;
+    }
+  }
+
+  // intra L0 compactions outputs could have overlap
+  if (output_level != 0 &&
+      picker_->FilesRangeOverlapWithCompaction(
+          inputs, output_level,
+          Compaction::EvaluatePenultimateLevel(vstorage_, ioptions_,
+                                               start_level, output_level))) {
+    return nullptr;
+  }
+
  // TODO support multi paths?
  uint32_t path_id = 0;
  return new Compaction(
@@ -1210,7 +1235,10 @@ Compaction* UniversalCompactionBuilder::PickDeleteTriggeredCompaction() {
      if (!output_level_inputs.empty()) {
        inputs.push_back(output_level_inputs);
      }
-      if (picker_->FilesRangeOverlapWithCompaction(inputs, output_level)) {
+      if (picker_->FilesRangeOverlapWithCompaction(
+              inputs, output_level,
+              Compaction::EvaluatePenultimateLevel(
+                  vstorage_, ioptions_, start_level, output_level))) {
        return nullptr;
      }

@@ -1312,6 +1340,15 @@ Compaction* UniversalCompactionBuilder::PickCompactionWithSortedRunRange(
    output_level = sorted_runs_[end_index + 1].level - 1;
  }

+  // intra L0 compactions outputs could have overlap
+  if (output_level != 0 &&
+      picker_->FilesRangeOverlapWithCompaction(
+          inputs, output_level,
+          Compaction::EvaluatePenultimateLevel(vstorage_, ioptions_,
+                                               start_level, output_level))) {
+    return nullptr;
+  }
+
  // We never check size for
  // compaction_options_universal.compression_size_percent,
  // because we always compact all the files, so always compress.
@@ -289,16 +289,6 @@ TEST_F(CompactionServiceTest, BasicCompactions) {
        auto s = static_cast<Status*>(status);
        *s = Status::Aborted("MyTestCompactionService failed to compact!");
      });
-
-  // tracking success unique id verification
-  std::atomic_int verify_passed{0};
-  SyncPoint::GetInstance()->SetCallBack(
-      "Version::VerifySstUniqueIds::Passed", [&](void* arg) {
-        // override job status
-        auto id = static_cast<UniqueId64x2*>(arg);
-        assert(*id != kNullUniqueId64x2);
-        verify_passed++;
-      });
  SyncPoint::GetInstance()->EnableProcessing();

  Status s;
@@ -324,11 +314,18 @@ TEST_F(CompactionServiceTest, BasicCompactions) {
  }
  ASSERT_TRUE(s.IsAborted());

-  // Test verification
-  ASSERT_EQ(verify_passed, 0);
-  options.verify_sst_unique_id_in_manifest = true;
+  // Test re-open and successful unique id verification
+  std::atomic_int verify_passed{0};
+  SyncPoint::GetInstance()->SetCallBack(
+      "BlockBasedTable::Open::PassedVerifyUniqueId", [&](void* arg) {
+        // override job status
+        auto id = static_cast<UniqueId64x2*>(arg);
+        assert(*id != kNullUniqueId64x2);
+        verify_passed++;
+      });
  Reopen(options);
  ASSERT_GT(verify_passed, 0);
+  Close();
 }

 TEST_F(CompactionServiceTest, ManualCompaction) {
@@ -23,46 +23,6 @@ void SubcompactionState::AggregateCompactionStats(
  }
 }

-void SubcompactionState::FillFilesToCutForTtl() {
-  if (compaction->immutable_options()->compaction_style !=
-          CompactionStyle::kCompactionStyleLevel ||
-      compaction->immutable_options()->compaction_pri !=
-          CompactionPri::kMinOverlappingRatio ||
-      compaction->mutable_cf_options()->ttl == 0 ||
-      compaction->num_input_levels() < 2 || compaction->bottommost_level()) {
-    return;
-  }
-
-  // We define new file with the oldest ancestor time to be younger than 1/4
-  // TTL, and an old one to be older than 1/2 TTL time.
-  int64_t temp_current_time;
-  auto get_time_status = compaction->immutable_options()->clock->GetCurrentTime(
-      &temp_current_time);
-  if (!get_time_status.ok()) {
-    return;
-  }
-  auto current_time = static_cast<uint64_t>(temp_current_time);
-  if (current_time < compaction->mutable_cf_options()->ttl) {
-    return;
-  }
-  uint64_t old_age_thres =
-      current_time - compaction->mutable_cf_options()->ttl / 2;
-
-  const std::vector<FileMetaData*>& olevel =
-      *(compaction->inputs(compaction->num_input_levels() - 1));
-  for (FileMetaData* file : olevel) {
-    // Worth filtering out by start and end?
-    uint64_t oldest_ancester_time = file->TryGetOldestAncesterTime();
-    // We put old files if they are not too small to prevent a flood
-    // of small files.
-    if (oldest_ancester_time < old_age_thres &&
-        file->fd.GetFileSize() >
-            compaction->mutable_cf_options()->target_file_size_base / 2) {
-      files_to_cut_for_ttl_.push_back(file);
-    }
-  }
-}
-
 OutputIterator SubcompactionState::GetOutputs() const {
  return OutputIterator(penultimate_level_outputs_.outputs_,
                        compaction_outputs_.outputs_);
@@ -128,83 +88,6 @@ Slice SubcompactionState::LargestUserKey() const {
  }
 }

-bool SubcompactionState::ShouldStopBefore(const Slice& internal_key) {
-  uint64_t curr_file_size = Current().GetCurrentOutputFileSize();
-  const InternalKeyComparator* icmp =
-      &compaction->column_family_data()->internal_comparator();
-
-  // Invalid local_output_split_key indicates that we do not need to split
-  if (local_output_split_key_ != nullptr && !is_split_) {
-    // Split occurs when the next key is larger than/equal to the cursor
-    if (icmp->Compare(internal_key, local_output_split_key_->Encode()) >= 0) {
-      is_split_ = true;
-      return true;
-    }
-  }
-
-  const std::vector<FileMetaData*>& grandparents = compaction->grandparents();
-  bool grandparant_file_switched = false;
-  // Scan to find the earliest grandparent file that contains key.
-  while (grandparent_index_ < grandparents.size() &&
-         icmp->Compare(internal_key,
-                       grandparents[grandparent_index_]->largest.Encode()) >
-             0) {
-    if (seen_key_) {
-      overlapped_bytes_ += grandparents[grandparent_index_]->fd.GetFileSize();
-      grandparant_file_switched = true;
-    }
-    assert(grandparent_index_ + 1 >= grandparents.size() ||
-           icmp->Compare(
-               grandparents[grandparent_index_]->largest.Encode(),
-               grandparents[grandparent_index_ + 1]->smallest.Encode()) <= 0);
-    grandparent_index_++;
-  }
-  seen_key_ = true;
-
-  if (grandparant_file_switched &&
-      overlapped_bytes_ + curr_file_size > compaction->max_compaction_bytes()) {
-    // Too much overlap for current output; start new output
-    overlapped_bytes_ = 0;
-    return true;
-  }
-
-  if (!files_to_cut_for_ttl_.empty()) {
-    if (cur_files_to_cut_for_ttl_ != -1) {
-      // Previous key is inside the range of a file
-      if (icmp->Compare(internal_key,
-                        files_to_cut_for_ttl_[cur_files_to_cut_for_ttl_]
-                            ->largest.Encode()) > 0) {
-        next_files_to_cut_for_ttl_ = cur_files_to_cut_for_ttl_ + 1;
-        cur_files_to_cut_for_ttl_ = -1;
-        return true;
-      }
-    } else {
-      // Look for the key position
-      while (next_files_to_cut_for_ttl_ <
-             static_cast<int>(files_to_cut_for_ttl_.size())) {
-        if (icmp->Compare(internal_key,
-                          files_to_cut_for_ttl_[next_files_to_cut_for_ttl_]
-                              ->smallest.Encode()) >= 0) {
-          if (icmp->Compare(internal_key,
-                            files_to_cut_for_ttl_[next_files_to_cut_for_ttl_]
-                                ->largest.Encode()) <= 0) {
-            // With in the current file
-            cur_files_to_cut_for_ttl_ = next_files_to_cut_for_ttl_;
-            return true;
-          }
-          // Beyond the current file
-          next_files_to_cut_for_ttl_++;
-        } else {
-          // Still fall into the gap
-          break;
-        }
-      }
-    }
-  }
-
-  return false;
-}
-
 Status SubcompactionState::AddToOutput(
    const CompactionIterator& iter,
    const CompactionFileOpenFunc& open_file_func,
@@ -128,21 +128,12 @@ class SubcompactionState {
        compaction_outputs_(c, /*is_penultimate_level=*/false),
        penultimate_level_outputs_(c, /*is_penultimate_level=*/true) {
    assert(compaction != nullptr);
-    const InternalKeyComparator* icmp =
-        &compaction->column_family_data()->internal_comparator();
-    const InternalKey* output_split_key = compaction->GetOutputSplitKey();
-    // Invalid output_split_key indicates that we do not need to split
-    if (output_split_key != nullptr) {
-      // We may only split the output when the cursor is in the range. Split
-      if ((!end.has_value() ||
-           icmp->user_comparator()->Compare(
-               ExtractUserKey(output_split_key->Encode()), end.value()) < 0) &&
-          (!start.has_value() || icmp->user_comparator()->Compare(
-                                     ExtractUserKey(output_split_key->Encode()),
-                                     start.value()) > 0)) {
-        local_output_split_key_ = output_split_key;
-      }
-    }
+    // Set output split key (used for RoundRobin feature) only for normal
+    // compaction_outputs, output to penultimate_level feature doesn't support
+    // RoundRobin feature (and may never going to be supported, because for
+    // RoundRobin, the data time is mostly naturally sorted, no need to have
+    // per-key placement with output_to_penultimate_level).
+    compaction_outputs_.SetOutputSlitKey(start, end);
  }

  SubcompactionState(SubcompactionState&& state) noexcept
@@ -155,12 +146,6 @@ class SubcompactionState {
            state.notify_on_subcompaction_completion),
        compaction_job_stats(std::move(state.compaction_job_stats)),
        sub_job_id(state.sub_job_id),
-        files_to_cut_for_ttl_(std::move(state.files_to_cut_for_ttl_)),
-        cur_files_to_cut_for_ttl_(state.cur_files_to_cut_for_ttl_),
-        next_files_to_cut_for_ttl_(state.next_files_to_cut_for_ttl_),
-        grandparent_index_(state.grandparent_index_),
-        overlapped_bytes_(state.overlapped_bytes_),
-        seen_key_(state.seen_key_),
        compaction_outputs_(std::move(state.compaction_outputs_)),
        penultimate_level_outputs_(std::move(state.penultimate_level_outputs_)),
        is_current_penultimate_level_(state.is_current_penultimate_level_),
@@ -175,12 +160,6 @@ class SubcompactionState {
           penultimate_level_outputs_.HasRangeDel();
  }

-  void FillFilesToCutForTtl();
-
-  // Returns true iff we should stop building the current output
-  // before processing "internal_key".
-  bool ShouldStopBefore(const Slice& internal_key);
-
  bool IsCurrentPenultimateLevel() const {
    return is_current_penultimate_level_;
  }
@@ -224,28 +203,6 @@ class SubcompactionState {
  }

 private:
-  // Some identified files with old oldest ancester time and the range should be
-  // isolated out so that the output file(s) in that range can be merged down
-  // for TTL and clear the timestamps for the range.
-  std::vector<FileMetaData*> files_to_cut_for_ttl_;
-  int cur_files_to_cut_for_ttl_ = -1;
-  int next_files_to_cut_for_ttl_ = 0;
-
-  // An index that used to speed up ShouldStopBefore().
-  size_t grandparent_index_ = 0;
-  // The number of bytes overlapping between the current output and
-  // grandparent files used in ShouldStopBefore().
-  uint64_t overlapped_bytes_ = 0;
-  // A flag determines whether the key has been seen in ShouldStopBefore()
-  bool seen_key_ = false;
-
-  // A flag determines if this subcompaction has been split by the cursor
-  bool is_split_ = false;
-
-  // We also maintain the output split key for each subcompaction to avoid
-  // repetitive comparison in ShouldStopBefore()
-  const InternalKey* local_output_split_key_ = nullptr;
-
  // State kept for output being generated
  CompactionOutputs compaction_outputs_;
  CompactionOutputs penultimate_level_outputs_;
@@ -10,7 +10,10 @@

 #include "db/db_test_util.h"
 #include "port/stack_trace.h"
+#include "rocksdb/iostats_context.h"
 #include "rocksdb/listener.h"
+#include "rocksdb/utilities/debug.h"
+#include "test_util/mock_time_env.h"

 namespace ROCKSDB_NAMESPACE {

@@ -414,38 +417,20 @@ TEST_P(TieredCompactionTest, RangeBasedTieredStorageUniversal) {
  // No data is moved from cold tier to hot tier because no input files from L5
  // or higher, it's not safe to move data to output_to_penultimate_level level.
  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
-  ASSERT_EQ("0,0,0,0,0,0,1", FilesPerLevel());
+  ASSERT_EQ("0,0,0,0,0,1", FilesPerLevel());

-  // Add 2 keys in higher level, but in separated files, the keys within that
-  // range should be moved up to the penultimate level
+  // Add 2 keys in higher level, but in separated files, all keys can be moved
+  // up if it's hot
  ASSERT_OK(Put(Key(0), "value" + std::to_string(0)));
  ASSERT_OK(Flush());
  ASSERT_OK(Put(Key(50), "value" + std::to_string(0)));
  ASSERT_OK(Flush());
-  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
-  ASSERT_EQ("0,0,0,0,0,1,1", FilesPerLevel());
-
-  // Add an SST with a key range cover all the data we want to move from the
-  // last level to the penultimate level
-  ASSERT_OK(Put(Key(0), "value" + std::to_string(0)));
-  ASSERT_OK(Put(Key(99), "value" + std::to_string(0)));
-  ASSERT_OK(Flush());
-
-  ResetAllStats(expect_stats, expect_pl_stats);
-
  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
  ASSERT_EQ("0,0,0,0,0,1", FilesPerLevel());
+
  ASSERT_GT(GetSstSizeHelper(Temperature::kUnknown), 0);
  ASSERT_EQ(GetSstSizeHelper(Temperature::kCold), 0);

-  last_stats.Add(kBasicCompStats);
-  last_stats.ResetCompactionReason(CompactionReason::kManualCompaction);
-  last_stats.bytes_read_output_level = kHasValue;
-  last_stats.num_input_files_in_output_level = kHasValue;
-  expect_pl_stats.Add(kBasicPerKeyPlacementCompStats);
-  expect_pl_stats.ResetCompactionReason(CompactionReason::kManualCompaction);
-  VerifyCompactionStats(expect_stats, expect_pl_stats);
-
  // change to only 1 key cold, to test compaction could stop even it matches
  // size amp compaction threshold
  {
@@ -1215,6 +1200,817 @@ TEST_P(TieredCompactionTest, RangeBasedTieredStorageLevel) {
 INSTANTIATE_TEST_CASE_P(TieredCompactionTest, TieredCompactionTest,
                        testing::Bool());

+class PrecludeLastLevelTest : public DBTestBase {
+ public:
+  PrecludeLastLevelTest()
+      : DBTestBase("preclude_last_level_test", /*env_do_fsync=*/false) {
+    mock_clock_ = std::make_shared<MockSystemClock>(env_->GetSystemClock());
+    mock_env_ = std::make_unique<CompositeEnvWrapper>(env_, mock_clock_);
+  }
+
+ protected:
+  std::unique_ptr<Env> mock_env_;
+  std::shared_ptr<MockSystemClock> mock_clock_;
+
+  void SetUp() override {
+    mock_clock_->InstallTimedWaitFixCallback();
+    SyncPoint::GetInstance()->SetCallBack(
+        "DBImpl::StartPeriodicTaskScheduler:Init", [&](void* arg) {
+          auto periodic_task_scheduler_ptr =
+              reinterpret_cast<PeriodicTaskScheduler*>(arg);
+          periodic_task_scheduler_ptr->TEST_OverrideTimer(mock_clock_.get());
+        });
+    mock_clock_->SetCurrentTime(0);
+  }
+};
+
+TEST_F(PrecludeLastLevelTest, MigrationFromPreserveTimeManualCompaction) {
+  const int kNumTrigger = 4;
+  const int kNumLevels = 7;
+  const int kNumKeys = 100;
+  const int kKeyPerSec = 10;
+
+  Options options = CurrentOptions();
+  options.compaction_style = kCompactionStyleUniversal;
+  options.preserve_internal_time_seconds = 10000;
+  options.env = mock_env_.get();
+  options.level0_file_num_compaction_trigger = kNumTrigger;
+  options.num_levels = kNumLevels;
+  DestroyAndReopen(options);
+
+  // pass some time first, otherwise the first a few keys write time are going
+  // to be zero, and internally zero has special meaning: kUnknownSeqnoTime
+  dbfull()->TEST_WaitForPeridicTaskRun(
+      [&] { mock_clock_->MockSleepForSeconds(static_cast<int>(kKeyPerSec)); });
+
+  int sst_num = 0;
+  // Write files that are overlap and enough to trigger compaction
+  for (; sst_num < kNumTrigger; sst_num++) {
+    for (int i = 0; i < kNumKeys; i++) {
+      ASSERT_OK(Put(Key(sst_num * (kNumKeys - 1) + i), "value"));
+      dbfull()->TEST_WaitForPeridicTaskRun([&] {
+        mock_clock_->MockSleepForSeconds(static_cast<int>(kKeyPerSec));
+      });
+    }
+    ASSERT_OK(Flush());
+  }
+  ASSERT_OK(dbfull()->WaitForCompact(true));
+
+  // all data is pushed to the last level
+  ASSERT_EQ("0,0,0,0,0,0,1", FilesPerLevel());
+
+  // enable preclude feature
+  options.preclude_last_level_data_seconds = 10000;
+  options.last_level_temperature = Temperature::kCold;
+  Reopen(options);
+
+  // all data is hot, even they're in the last level
+  ASSERT_EQ(GetSstSizeHelper(Temperature::kCold), 0);
+  ASSERT_GT(GetSstSizeHelper(Temperature::kUnknown), 0);
+
+  // Generate a sstable and trigger manual compaction
+  ASSERT_OK(Put(Key(10), "value"));
+  ASSERT_OK(Flush());
+
+  CompactRangeOptions cro;
+  cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
+  ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
+
+  // all data is moved up to the penultimate level
+  ASSERT_EQ("0,0,0,0,0,1", FilesPerLevel());
+  ASSERT_EQ(GetSstSizeHelper(Temperature::kCold), 0);
+  ASSERT_GT(GetSstSizeHelper(Temperature::kUnknown), 0);
+
+  // close explicitly, because the env is local variable which will be released
+  // first.
+  Close();
+}
+
+TEST_F(PrecludeLastLevelTest, MigrationFromPreserveTimeAutoCompaction) {
+  const int kNumTrigger = 4;
+  const int kNumLevels = 7;
+  const int kNumKeys = 100;
+  const int kKeyPerSec = 10;
+
+  Options options = CurrentOptions();
+  options.compaction_style = kCompactionStyleUniversal;
+  options.preserve_internal_time_seconds = 10000;
+  options.env = mock_env_.get();
+  options.level0_file_num_compaction_trigger = kNumTrigger;
+  options.num_levels = kNumLevels;
+  DestroyAndReopen(options);
+
+  // pass some time first, otherwise the first a few keys write time are going
+  // to be zero, and internally zero has special meaning: kUnknownSeqnoTime
+  dbfull()->TEST_WaitForPeridicTaskRun(
+      [&] { mock_clock_->MockSleepForSeconds(static_cast<int>(kKeyPerSec)); });
+
+  int sst_num = 0;
+  // Write files that are overlap and enough to trigger compaction
+  for (; sst_num < kNumTrigger; sst_num++) {
+    for (int i = 0; i < kNumKeys; i++) {
+      ASSERT_OK(Put(Key(sst_num * (kNumKeys - 1) + i), "value"));
+      dbfull()->TEST_WaitForPeridicTaskRun([&] {
+        mock_clock_->MockSleepForSeconds(static_cast<int>(kKeyPerSec));
+      });
+    }
+    ASSERT_OK(Flush());
+  }
+  ASSERT_OK(dbfull()->WaitForCompact(true));
+
+  // all data is pushed to the last level
+  ASSERT_EQ("0,0,0,0,0,0,1", FilesPerLevel());
+
+  // enable preclude feature
+  options.preclude_last_level_data_seconds = 10000;
+  options.last_level_temperature = Temperature::kCold;
+  // make sure it won't trigger Size Amp compaction, unlike normal Size Amp
+  // compaction which is typically a last level compaction, when tiered Storage
+  // ("preclude_last_level") is enabled, size amp won't include the last level.
+  // As the last level would be in cold tier and the size would not be a
+  // problem, which also avoid frequent hot to cold storage compaction.
+  options.compaction_options_universal.max_size_amplification_percent = 400;
+  Reopen(options);
+
+  // all data is hot, even they're in the last level
+  ASSERT_EQ(GetSstSizeHelper(Temperature::kCold), 0);
+  ASSERT_GT(GetSstSizeHelper(Temperature::kUnknown), 0);
+
+  // Write more data, but still all hot until the 10th SST, as:
+  // write a key every 10 seconds, 100 keys per SST, each SST takes 1000 seconds
+  // The preclude_last_level_data_seconds is 10k
+  Random rnd(301);
+  for (; sst_num < kNumTrigger * 2 - 1; sst_num++) {
+    for (int i = 0; i < kNumKeys; i++) {
+      // the value needs to be big enough to trigger full compaction
+      ASSERT_OK(Put(Key(sst_num * (kNumKeys - 1) + i), rnd.RandomString(100)));
+      dbfull()->TEST_WaitForPeridicTaskRun([&] {
+        mock_clock_->MockSleepForSeconds(static_cast<int>(kKeyPerSec));
+      });
+    }
+    ASSERT_OK(Flush());
+    ASSERT_OK(dbfull()->WaitForCompact(true));
+  }
+
+  // all data is moved up to the penultimate level
+  ASSERT_EQ("0,0,0,0,0,1", FilesPerLevel());
+  ASSERT_EQ(GetSstSizeHelper(Temperature::kCold), 0);
+  ASSERT_GT(GetSstSizeHelper(Temperature::kUnknown), 0);
+
+  // close explicitly, because the env is local variable which will be released
+  // first.
+  Close();
+}
+
+TEST_F(PrecludeLastLevelTest, MigrationFromPreserveTimePartial) {
+  const int kNumTrigger = 4;
+  const int kNumLevels = 7;
+  const int kNumKeys = 100;
+  const int kKeyPerSec = 10;
+
+  Options options = CurrentOptions();
+  options.compaction_style = kCompactionStyleUniversal;
+  options.preserve_internal_time_seconds = 2000;
+  options.env = mock_env_.get();
+  options.level0_file_num_compaction_trigger = kNumTrigger;
+  options.num_levels = kNumLevels;
+  DestroyAndReopen(options);
+
+  // pass some time first, otherwise the first a few keys write time are going
+  // to be zero, and internally zero has special meaning: kUnknownSeqnoTime
+  dbfull()->TEST_WaitForPeridicTaskRun(
+      [&] { mock_clock_->MockSleepForSeconds(static_cast<int>(kKeyPerSec)); });
+
+  int sst_num = 0;
+  // Write files that are overlap and enough to trigger compaction
+  for (; sst_num < kNumTrigger; sst_num++) {
+    for (int i = 0; i < kNumKeys; i++) {
+      ASSERT_OK(Put(Key(sst_num * (kNumKeys - 1) + i), "value"));
+      dbfull()->TEST_WaitForPeridicTaskRun([&] {
+        mock_clock_->MockSleepForSeconds(static_cast<int>(kKeyPerSec));
+      });
+    }
+    ASSERT_OK(Flush());
+  }
+  ASSERT_OK(dbfull()->WaitForCompact(true));
+
+  // all data is pushed to the last level
+  ASSERT_EQ("0,0,0,0,0,0,1", FilesPerLevel());
+
+  std::vector<KeyVersion> key_versions;
+  ASSERT_OK(GetAllKeyVersions(db_, Slice(), Slice(),
+                              std::numeric_limits<size_t>::max(),
+                              &key_versions));
+
+  // make sure there're more than 300 keys and first 100 keys are having seqno
+  // zeroed out, the last 100 key seqno not zeroed out
+  ASSERT_GT(key_versions.size(), 300);
+  for (int i = 0; i < 100; i++) {
+    ASSERT_EQ(key_versions[i].sequence, 0);
+  }
+  auto rit = key_versions.rbegin();
+  for (int i = 0; i < 100; i++) {
+    ASSERT_GT(rit->sequence, 0);
+    rit++;
+  }
+
+  // enable preclude feature
+  options.preclude_last_level_data_seconds = 2000;
+  options.last_level_temperature = Temperature::kCold;
+  Reopen(options);
+
+  // Generate a sstable and trigger manual compaction
+  ASSERT_OK(Put(Key(10), "value"));
+  ASSERT_OK(Flush());
+
+  CompactRangeOptions cro;
+  cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
+  ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
+
+  // some data are moved up, some are not
+  ASSERT_EQ("0,0,0,0,0,1,1", FilesPerLevel());
+  ASSERT_GT(GetSstSizeHelper(Temperature::kCold), 0);
+  ASSERT_GT(GetSstSizeHelper(Temperature::kUnknown), 0);
+
+  Close();
+}
+
+TEST_F(PrecludeLastLevelTest, SmallPrecludeTime) {
+  const int kNumTrigger = 4;
+  const int kNumLevels = 7;
+  const int kNumKeys = 100;
+
+  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;
+  DestroyAndReopen(options);
+
+  Random rnd(301);
+
+  dbfull()->TEST_WaitForPeridicTaskRun([&] {
+    mock_clock_->MockSleepForSeconds(static_cast<int>(rnd.Uniform(10) + 1));
+  });
+
+  for (int i = 0; i < kNumKeys; i++) {
+    ASSERT_OK(Put(Key(i), rnd.RandomString(100)));
+    dbfull()->TEST_WaitForPeridicTaskRun([&] {
+      mock_clock_->MockSleepForSeconds(static_cast<int>(rnd.Uniform(2)));
+    });
+  }
+  ASSERT_OK(Flush());
+
+  TablePropertiesCollection tables_props;
+  ASSERT_OK(dbfull()->GetPropertiesOfAllTables(&tables_props));
+  ASSERT_EQ(tables_props.size(), 1);
+  ASSERT_FALSE(tables_props.begin()->second->seqno_to_time_mapping.empty());
+  SeqnoToTimeMapping tp_mapping;
+  ASSERT_OK(
+      tp_mapping.Add(tables_props.begin()->second->seqno_to_time_mapping));
+  ASSERT_OK(tp_mapping.Sort());
+  ASSERT_FALSE(tp_mapping.Empty());
+  auto seqs = tp_mapping.TEST_GetInternalMapping();
+  ASSERT_FALSE(seqs.empty());
+
+  // Wait more than preclude_last_level time, then make sure all the data is
+  // compacted to the last level even there's no write (no seqno -> time
+  // information was flushed to any SST).
+  mock_clock_->MockSleepForSeconds(100);
+
+  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  ASSERT_EQ("0,0,0,0,0,0,1", FilesPerLevel());
+  ASSERT_EQ(GetSstSizeHelper(Temperature::kUnknown), 0);
+  ASSERT_GT(GetSstSizeHelper(Temperature::kCold), 0);
+
+  Close();
+}
+
+TEST_F(PrecludeLastLevelTest, LastLevelOnlyCompactionPartial) {
+  const int kNumTrigger = 4;
+  const int kNumLevels = 7;
+  const int kNumKeys = 100;
+  const int kKeyPerSec = 10;
+
+  Options options = CurrentOptions();
+  options.compaction_style = kCompactionStyleUniversal;
+  options.preserve_internal_time_seconds = 2000;
+  options.env = mock_env_.get();
+  options.level0_file_num_compaction_trigger = kNumTrigger;
+  options.num_levels = kNumLevels;
+  DestroyAndReopen(options);
+
+  // pass some time first, otherwise the first a few keys write time are going
+  // to be zero, and internally zero has special meaning: kUnknownSeqnoTime
+  dbfull()->TEST_WaitForPeridicTaskRun(
+      [&] { mock_clock_->MockSleepForSeconds(static_cast<int>(kKeyPerSec)); });
+
+  int sst_num = 0;
+  // Write files that are overlap and enough to trigger compaction
+  for (; sst_num < kNumTrigger; sst_num++) {
+    for (int i = 0; i < kNumKeys; i++) {
+      ASSERT_OK(Put(Key(sst_num * (kNumKeys - 1) + i), "value"));
+      dbfull()->TEST_WaitForPeridicTaskRun([&] {
+        mock_clock_->MockSleepForSeconds(static_cast<int>(kKeyPerSec));
+      });
+    }
+    ASSERT_OK(Flush());
+  }
+  ASSERT_OK(dbfull()->WaitForCompact(true));
+
+  // all data is pushed to the last level
+  ASSERT_EQ("0,0,0,0,0,0,1", FilesPerLevel());
+
+  // enable preclude feature
+  options.preclude_last_level_data_seconds = 2000;
+  options.last_level_temperature = Temperature::kCold;
+  Reopen(options);
+
+  CompactRangeOptions cro;
+  cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
+  ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
+
+  // some data are moved up, some are not
+  ASSERT_EQ("0,0,0,0,0,1,1", FilesPerLevel());
+  ASSERT_GT(GetSstSizeHelper(Temperature::kCold), 0);
+  ASSERT_GT(GetSstSizeHelper(Temperature::kUnknown), 0);
+
+  std::vector<KeyVersion> key_versions;
+  ASSERT_OK(GetAllKeyVersions(db_, Slice(), Slice(),
+                              std::numeric_limits<size_t>::max(),
+                              &key_versions));
+
+  // make sure there're more than 300 keys and first 100 keys are having seqno
+  // zeroed out, the last 100 key seqno not zeroed out
+  ASSERT_GT(key_versions.size(), 300);
+  for (int i = 0; i < 100; i++) {
+    ASSERT_EQ(key_versions[i].sequence, 0);
+  }
+  auto rit = key_versions.rbegin();
+  for (int i = 0; i < 100; i++) {
+    ASSERT_GT(rit->sequence, 0);
+    rit++;
+  }
+
+  Close();
+}
+
+class PrecludeLastLevelTestWithParms
+    : public PrecludeLastLevelTest,
+      public testing::WithParamInterface<bool> {
+ public:
+  PrecludeLastLevelTestWithParms() : PrecludeLastLevelTest() {}
+};
+
+TEST_P(PrecludeLastLevelTestWithParms, LastLevelOnlyCompactionNoPreclude) {
+  const int kNumTrigger = 4;
+  const int kNumLevels = 7;
+  const int kNumKeys = 100;
+  const int kKeyPerSec = 10;
+
+  bool enable_preclude_last_level = GetParam();
+
+  Options options = CurrentOptions();
+  options.compaction_style = kCompactionStyleUniversal;
+  options.preserve_internal_time_seconds = 2000;
+  options.env = mock_env_.get();
+  options.level0_file_num_compaction_trigger = kNumTrigger;
+  options.num_levels = kNumLevels;
+  DestroyAndReopen(options);
+
+  // pass some time first, otherwise the first a few keys write time are going
+  // to be zero, and internally zero has special meaning: kUnknownSeqnoTime
+  dbfull()->TEST_WaitForPeridicTaskRun(
+      [&] { mock_clock_->MockSleepForSeconds(static_cast<int>(kKeyPerSec)); });
+
+  Random rnd(301);
+  int sst_num = 0;
+  // Write files that are overlap and enough to trigger compaction
+  for (; sst_num < kNumTrigger; sst_num++) {
+    for (int i = 0; i < kNumKeys; i++) {
+      ASSERT_OK(Put(Key(sst_num * (kNumKeys - 1) + i), rnd.RandomString(100)));
+      dbfull()->TEST_WaitForPeridicTaskRun([&] {
+        mock_clock_->MockSleepForSeconds(static_cast<int>(kKeyPerSec));
+      });
+    }
+    ASSERT_OK(Flush());
+  }
+  ASSERT_OK(dbfull()->WaitForCompact(true));
+
+  // all data is pushed to the last level
+  ASSERT_EQ("0,0,0,0,0,0,1", FilesPerLevel());
+
+  std::atomic_bool is_manual_compaction_running = false;
+  std::atomic_bool verified_compaction_order = false;
+
+  // Make sure the manual compaction is in progress and try to trigger a
+  // SizeRatio compaction by flushing 4 files to L0. The compaction will try to
+  // compact 4 files at L0 to L5 (the last empty level).
+  // If the preclude_last_feature is enabled, the auto triggered compaction
+  // cannot be picked. Otherwise, the auto triggered compaction can run in
+  // parallel with the last level compaction.
+  // L0: [a] [b] [c] [d]
+  // L5:     (locked if preclude_last_level is enabled)
+  // L6: [z] (locked: manual compaction in progress)
+  // TODO: in this case, L0 files should just be compacted to L4, so the 2
+  //  compactions won't be overlapped.
+  SyncPoint::GetInstance()->SetCallBack(
+      "CompactionJob::ProcessKeyValueCompaction()::Processing", [&](void* arg) {
+        auto compaction = static_cast<Compaction*>(arg);
+        if (compaction->is_manual_compaction()) {
+          is_manual_compaction_running = true;
+          TEST_SYNC_POINT(
+              "PrecludeLastLevelTest::LastLevelOnlyCompactionConflit:"
+              "ManualCompaction1");
+          TEST_SYNC_POINT(
+              "PrecludeLastLevelTest::LastLevelOnlyCompactionConflit:"
+              "ManualCompaction2");
+          is_manual_compaction_running = false;
+        }
+      });
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "UniversalCompactionBuilder::PickCompaction:Return", [&](void* arg) {
+        auto compaction = static_cast<Compaction*>(arg);
+        if (enable_preclude_last_level && is_manual_compaction_running) {
+          ASSERT_TRUE(compaction == nullptr);
+          verified_compaction_order = true;
+        } else {
+          ASSERT_TRUE(compaction != nullptr);
+          verified_compaction_order = true;
+        }
+        if (!compaction || !compaction->is_manual_compaction()) {
+          TEST_SYNC_POINT(
+              "PrecludeLastLevelTest::LastLevelOnlyCompactionConflit:"
+              "AutoCompactionPicked");
+        }
+      });
+
+  SyncPoint::GetInstance()->LoadDependency({
+      {"PrecludeLastLevelTest::LastLevelOnlyCompactionConflit:"
+       "ManualCompaction1",
+       "PrecludeLastLevelTest::LastLevelOnlyCompactionConflit:StartWrite"},
+      {"PrecludeLastLevelTest::LastLevelOnlyCompactionConflit:"
+       "AutoCompactionPicked",
+       "PrecludeLastLevelTest::LastLevelOnlyCompactionConflit:"
+       "ManualCompaction2"},
+  });
+
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  // only enable if the Parameter is true
+  if (enable_preclude_last_level) {
+    options.preclude_last_level_data_seconds = 2000;
+  }
+  options.max_background_jobs = 8;
+  options.last_level_temperature = Temperature::kCold;
+  Reopen(options);
+
+  auto manual_compaction_thread = port::Thread([this]() {
+    CompactRangeOptions cro;
+    cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
+    cro.exclusive_manual_compaction = false;
+    ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
+  });
+
+  TEST_SYNC_POINT(
+      "PrecludeLastLevelTest::LastLevelOnlyCompactionConflit:StartWrite");
+  auto stop_token =
+      dbfull()->TEST_write_controler().GetCompactionPressureToken();
+
+  for (; sst_num < kNumTrigger * 2; sst_num++) {
+    for (int i = 0; i < kNumKeys; i++) {
+      // the value needs to be big enough to trigger full compaction
+      ASSERT_OK(Put(Key(sst_num * (kNumKeys - 1) + i), "value"));
+      dbfull()->TEST_WaitForPeridicTaskRun([&] {
+        mock_clock_->MockSleepForSeconds(static_cast<int>(kKeyPerSec));
+      });
+    }
+    ASSERT_OK(Flush());
+  }
+
+  manual_compaction_thread.join();
+
+  ASSERT_OK(dbfull()->WaitForCompact(true));
+
+  if (enable_preclude_last_level) {
+    ASSERT_NE("0,0,0,0,0,1,1", FilesPerLevel());
+  } else {
+    ASSERT_EQ("0,0,0,0,0,1,1", FilesPerLevel());
+  }
+  ASSERT_TRUE(verified_compaction_order);
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+  stop_token.reset();
+
+  Close();
+}
+
+INSTANTIATE_TEST_CASE_P(PrecludeLastLevelTestWithParms,
+                        PrecludeLastLevelTestWithParms, testing::Bool());
+
+// partition the SST into 3 ranges [0, 19] [20, 39] [40, ...]
+class ThreeRangesPartitioner : public SstPartitioner {
+ public:
+  const char* Name() const override { return "SingleKeySstPartitioner"; }
+
+  PartitionerResult ShouldPartition(
+      const PartitionerRequest& request) override {
+    if ((cmp->CompareWithoutTimestamp(*request.current_user_key,
+                                      DBTestBase::Key(20)) >= 0 &&
+         cmp->CompareWithoutTimestamp(*request.prev_user_key,
+                                      DBTestBase::Key(20)) < 0) ||
+        (cmp->CompareWithoutTimestamp(*request.current_user_key,
+                                      DBTestBase::Key(40)) >= 0 &&
+         cmp->CompareWithoutTimestamp(*request.prev_user_key,
+                                      DBTestBase::Key(40)) < 0)) {
+      return kRequired;
+    } else {
+      return kNotRequired;
+    }
+  }
+
+  bool CanDoTrivialMove(const Slice& /*smallest_user_key*/,
+                        const Slice& /*largest_user_key*/) override {
+    return false;
+  }
+
+  const Comparator* cmp = BytewiseComparator();
+};
+
+class ThreeRangesPartitionerFactory : public SstPartitionerFactory {
+ public:
+  static const char* kClassName() {
+    return "TombstoneTestSstPartitionerFactory";
+  }
+  const char* Name() const override { return kClassName(); }
+
+  std::unique_ptr<SstPartitioner> CreatePartitioner(
+      const SstPartitioner::Context& /* context */) const override {
+    return std::unique_ptr<SstPartitioner>(new ThreeRangesPartitioner());
+  }
+};
+
+TEST_F(PrecludeLastLevelTest, PartialPenultimateLevelCompaction) {
+  const int kNumTrigger = 4;
+  const int kNumLevels = 7;
+  const int kKeyPerSec = 10;
+
+  Options options = CurrentOptions();
+  options.compaction_style = kCompactionStyleUniversal;
+  options.env = mock_env_.get();
+  options.level0_file_num_compaction_trigger = kNumTrigger;
+  options.preserve_internal_time_seconds = 10000;
+  options.num_levels = kNumLevels;
+  DestroyAndReopen(options);
+
+  // pass some time first, otherwise the first a few keys write time are going
+  // to be zero, and internally zero has special meaning: kUnknownSeqnoTime
+  dbfull()->TEST_WaitForPeridicTaskRun(
+      [&] { mock_clock_->MockSleepForSeconds(static_cast<int>(10)); });
+
+  Random rnd(301);
+
+  for (int i = 0; i < 300; i++) {
+    ASSERT_OK(Put(Key(i), rnd.RandomString(100)));
+    dbfull()->TEST_WaitForPeridicTaskRun(
+        [&] { mock_clock_->MockSleepForSeconds(kKeyPerSec); });
+  }
+  ASSERT_OK(Flush());
+  CompactRangeOptions cro;
+  cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
+
+  ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
+
+  // make sure all data is compacted to the last level
+  ASSERT_EQ("0,0,0,0,0,0,1", FilesPerLevel());
+
+  // Create 3 L5 files
+  auto factory = std::make_shared<ThreeRangesPartitionerFactory>();
+  options.sst_partitioner_factory = factory;
+
+  Reopen(options);
+
+  for (int i = 0; i < kNumTrigger - 1; i++) {
+    for (int j = 0; j < 100; j++) {
+      ASSERT_OK(Put(Key(i * 100 + j), rnd.RandomString(10)));
+    }
+    ASSERT_OK(Flush());
+  }
+
+  ASSERT_OK(dbfull()->WaitForCompact(true));
+
+  // L5: [0,19] [20,39] [40,299]
+  // L6: [0,                299]
+  ASSERT_EQ("0,0,0,0,0,3,1", FilesPerLevel());
+
+  // enable tiered storage feature
+  options.preclude_last_level_data_seconds = 10000;
+  options.last_level_temperature = Temperature::kCold;
+  options.statistics = CreateDBStatistics();
+  Reopen(options);
+
+  ColumnFamilyMetaData meta;
+  db_->GetColumnFamilyMetaData(&meta);
+  ASSERT_EQ(meta.levels[5].files.size(), 3);
+  ASSERT_EQ(meta.levels[6].files.size(), 1);
+  ASSERT_EQ(meta.levels[6].files[0].smallestkey, Key(0));
+  ASSERT_EQ(meta.levels[6].files[0].largestkey, Key(299));
+
+  std::string file_path = meta.levels[5].files[1].db_path;
+  std::vector<std::string> files;
+  // pick 3rd file @L5 + file@L6 for compaction
+  files.push_back(file_path + "/" + meta.levels[5].files[2].name);
+  files.push_back(file_path + "/" + meta.levels[6].files[0].name);
+  ASSERT_OK(db_->CompactFiles(CompactionOptions(), files, 6));
+
+  // The compaction only moved partial of the hot data to hot tier, range[0,39]
+  // is unsafe to move up, otherwise, they will be overlapped with the existing
+  // files@L5.
+  // The output should be:
+  //  L5: [0,19] [20,39] [40,299]    <-- Temperature::kUnknown
+  //  L6: [0,19] [20,39]             <-- Temperature::kCold
+  // L6 file is split because of the customized partitioner
+  ASSERT_EQ("0,0,0,0,0,3,2", FilesPerLevel());
+
+  // even all the data is hot, but not all data are moved to the hot tier
+  ASSERT_GT(GetSstSizeHelper(Temperature::kUnknown), 0);
+  ASSERT_GT(GetSstSizeHelper(Temperature::kCold), 0);
+
+  db_->GetColumnFamilyMetaData(&meta);
+  ASSERT_EQ(meta.levels[5].files.size(), 3);
+  ASSERT_EQ(meta.levels[6].files.size(), 2);
+  for (const auto& file : meta.levels[5].files) {
+    ASSERT_EQ(file.temperature, Temperature::kUnknown);
+  }
+  for (const auto& file : meta.levels[6].files) {
+    ASSERT_EQ(file.temperature, Temperature::kCold);
+  }
+  ASSERT_EQ(meta.levels[6].files[0].smallestkey, Key(0));
+  ASSERT_EQ(meta.levels[6].files[0].largestkey, Key(19));
+  ASSERT_EQ(meta.levels[6].files[1].smallestkey, Key(20));
+  ASSERT_EQ(meta.levels[6].files[1].largestkey, Key(39));
+
+  Close();
+}
+
+struct TestPropertiesCollector : public TablePropertiesCollector {
+  Status AddUserKey(const Slice& key, const Slice& /*value*/,
+                    EntryType /*type*/, SequenceNumber /*seq*/,
+                    uint64_t /*file_size*/) override {
+    if (cmp->Compare(key, DBTestBase::Key(100)) == 0) {
+      has_key_100 = true;
+    }
+    if (cmp->Compare(key, DBTestBase::Key(200)) == 0) {
+      has_key_200 = true;
+    }
+
+    return Status::OK();
+  }
+
+  const char* Name() const override { return "TestTablePropertiesCollector"; }
+
+  UserCollectedProperties GetReadableProperties() const override {
+    UserCollectedProperties ret;
+    return ret;
+  }
+
+  Status Finish(UserCollectedProperties* /*properties*/) override {
+    // The LSM tree would be like:
+    // L5: [0,19] [20,39] [40,299]
+    // L6: [0,                299]
+    // the 3rd file @L5 has both 100 and 200, which will be marked for
+    // compaction
+    // Also avoid marking flushed SST for compaction, which won't have both 100
+    // and 200
+    if (has_key_100 && has_key_200) {
+      need_compact_ = true;
+    } else {
+      need_compact_ = false;
+    }
+    has_key_100 = false;
+    has_key_200 = false;
+    return Status::OK();
+  }
+
+  bool NeedCompact() const override { return need_compact_; }
+
+  const Comparator* cmp = BytewiseComparator();
+
+ private:
+  bool has_key_100 = false;
+  bool has_key_200 = false;
+
+  bool need_compact_ = false;
+};
+
+class TestPropertiesCollectorFactory : public TablePropertiesCollectorFactory {
+ public:
+  TablePropertiesCollector* CreateTablePropertiesCollector(
+      TablePropertiesCollectorFactory::Context /*context*/) override {
+    return new TestPropertiesCollector;
+  }
+  const char* Name() const override { return "TestTablePropertiesCollector"; }
+};
+
+TEST_F(PrecludeLastLevelTest, PartialPenultimateLevelCompactionWithRangeDel) {
+  const int kNumTrigger = 4;
+  const int kNumLevels = 7;
+  const int kKeyPerSec = 10;
+
+  Options options = CurrentOptions();
+  options.compaction_style = kCompactionStyleUniversal;
+  options.env = mock_env_.get();
+  options.level0_file_num_compaction_trigger = kNumTrigger;
+  options.preserve_internal_time_seconds = 10000;
+  options.num_levels = kNumLevels;
+  // set a small max_compaction_bytes to avoid input level expansion
+  options.max_compaction_bytes = 30000;
+  options.ignore_max_compaction_bytes_for_input = false;
+  DestroyAndReopen(options);
+
+  // pass some time first, otherwise the first a few keys write time are going
+  // to be zero, and internally zero has special meaning: kUnknownSeqnoTime
+  dbfull()->TEST_WaitForPeridicTaskRun(
+      [&] { mock_clock_->MockSleepForSeconds(static_cast<int>(10)); });
+
+  Random rnd(301);
+
+  for (int i = 0; i < 300; i++) {
+    ASSERT_OK(Put(Key(i), rnd.RandomString(100)));
+    dbfull()->TEST_WaitForPeridicTaskRun(
+        [&] { mock_clock_->MockSleepForSeconds(kKeyPerSec); });
+  }
+  ASSERT_OK(Flush());
+  CompactRangeOptions cro;
+  cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
+
+  ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
+
+  // make sure all data is compacted to the last level
+  ASSERT_EQ("0,0,0,0,0,0,1", FilesPerLevel());
+
+  // Create 3 L5 files
+  auto factory = std::make_shared<ThreeRangesPartitionerFactory>();
+  options.sst_partitioner_factory = factory;
+
+  // the user defined properties_collector will mark the 3rd file for compaction
+  auto collector_factory = std::make_shared<TestPropertiesCollectorFactory>();
+  options.table_properties_collector_factories.resize(1);
+  options.table_properties_collector_factories[0] = collector_factory;
+  // enable tiered storage feature
+  options.preclude_last_level_data_seconds = 10000;
+  options.last_level_temperature = Temperature::kCold;
+  Reopen(options);
+
+  for (int i = 0; i < kNumTrigger - 2; i++) {
+    for (int j = 0; j < 100; j++) {
+      ASSERT_OK(Put(Key(i * 100 + j), rnd.RandomString(10)));
+    }
+    ASSERT_OK(Flush());
+  }
+
+  // make sure there is one and only one compaction supports per-key placement
+  // but has the penultimate level output disabled.
+  std::atomic_int per_key_comp_num = 0;
+  SyncPoint::GetInstance()->SetCallBack(
+      "UniversalCompactionBuilder::PickCompaction:Return", [&](void* arg) {
+        auto compaction = static_cast<Compaction*>(arg);
+        if (compaction->SupportsPerKeyPlacement()) {
+          ASSERT_EQ(compaction->GetPenultimateOutputRangeType(),
+                    Compaction::PenultimateOutputRangeType::kDisabled);
+          per_key_comp_num++;
+        }
+      });
+
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  for (int j = 0; j < 100; j++) {
+    ASSERT_OK(Put(Key(200 + j), rnd.RandomString(10)));
+  }
+  ASSERT_OK(db_->DeleteRange(WriteOptions(), db_->DefaultColumnFamily(),
+                             Key(32), Key(40)));
+  ASSERT_OK(Flush());
+
+  // Before the per-key placement compaction, the LSM tress should be like:
+  // L5: [0,19] [20,40] [40,299]
+  // L6: [0,                299]
+  // The 2nd file @L5 has the largest key 40 because of range del
+
+  ASSERT_OK(dbfull()->WaitForCompact(true));
+
+  ASSERT_EQ(per_key_comp_num, 1);
+
+  // the compaction won't move any data to the penultimate level
+  ASSERT_EQ("0,0,0,0,0,2,3", FilesPerLevel());
+
+  Close();
+}
+
 #endif  // !defined(ROCKSDB_LITE)

 }  // namespace ROCKSDB_NAMESPACE
@@ -673,6 +673,7 @@ TEST_P(ComparatorDBTest, SeparatorSuccessorRandomizeTest) {
 }  // namespace ROCKSDB_NAMESPACE

 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
  ::testing::InitGoogleTest(&argc, argv);
  return RUN_ALL_TESTS();
 }
@@ -345,6 +345,7 @@ TEST_F(CuckooTableDBTest, AdaptiveTable) {

 int main(int argc, char** argv) {
  if (ROCKSDB_NAMESPACE::port::kLittleEndian) {
+    ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
    ::testing::InitGoogleTest(&argc, argv);
    return RUN_ALL_TESTS();
  } else {
@@ -253,6 +253,7 @@ TEST_F(DBBasicTest, CompactedDB) {
  ASSERT_OK(Put("bbb", DummyString(kFileSize / 2, 'b')));
  ASSERT_OK(Put("eee", DummyString(kFileSize / 2, 'e')));
  ASSERT_OK(Flush());
+  ASSERT_OK(Put("something_not_flushed", "x"));
  Close();

  ASSERT_OK(ReadOnlyReopen(options));
@@ -260,6 +261,20 @@ TEST_F(DBBasicTest, CompactedDB) {
  s = Put("new", "value");
  ASSERT_EQ(s.ToString(),
            "Not implemented: Not supported operation in read only mode.");
+
+  // TODO: validate that other write ops return NotImplemented
+  // (DBImplReadOnly is missing some overrides)
+
+  // Ensure no deadlock on flush triggered by another API function
+  // (Old deadlock bug depends on something_not_flushed above.)
+  std::vector<std::string> files;
+  uint64_t manifest_file_size;
+  ASSERT_OK(db_->GetLiveFiles(files, &manifest_file_size, /*flush*/ true));
+  LiveFilesStorageInfoOptions lfsi_opts;
+  lfsi_opts.wal_size_for_flush = 0;  // always
+  std::vector<LiveFileStorageInfo> files2;
+  ASSERT_OK(db_->GetLiveFilesStorageInfo(lfsi_opts, &files2));
+
  Close();

  // Full compaction
@@ -290,6 +305,13 @@ TEST_F(DBBasicTest, CompactedDB) {
  ASSERT_EQ(DummyString(kFileSize / 2, 'j'), Get("jjj"));
  ASSERT_EQ("NOT_FOUND", Get("kkk"));

+  // TODO: validate that other write ops return NotImplemented
+  // (CompactedDB is missing some overrides)
+
+  // Ensure no deadlock on flush triggered by another API function
+  ASSERT_OK(db_->GetLiveFiles(files, &manifest_file_size, /*flush*/ true));
+  ASSERT_OK(db_->GetLiveFilesStorageInfo(lfsi_opts, &files2));
+
  // MultiGet
  std::vector<std::string> values;
  std::vector<Status> status_list = dbfull()->MultiGet(
@@ -2014,7 +2036,7 @@ TEST_P(DBMultiGetTestWithParam, MultiGetBatchedValueSize) {

 TEST_P(DBMultiGetTestWithParam, MultiGetBatchedValueSizeMultiLevelMerge) {
  if (std::get<1>(GetParam())) {
-    ROCKSDB_GTEST_SKIP("This test needs to be fixed for async IO");
+    ROCKSDB_GTEST_BYPASS("This test needs to be fixed for async IO");
    return;
  }
  // Skip for unbatched MultiGet
@@ -2239,7 +2261,9 @@ TEST_P(DBMultiGetAsyncIOTest, GetFromL0) {
    ASSERT_EQ(multiget_io_batch_size.max, 3);
    ASSERT_EQ(statistics()->getTickerCount(MULTIGET_COROUTINE_COUNT), 3);
  } else {
-    ASSERT_EQ(multiget_io_batch_size.count, 0);
+    // Without Async IO, MultiGet will call MultiRead 3 times, once for each
+    // L0 file
+    ASSERT_EQ(multiget_io_batch_size.count, 3);
  }
 }

@@ -2356,8 +2380,9 @@ TEST_P(DBMultiGetAsyncIOTest, GetFromL1AndL2) {
  statistics()->histogramData(MULTIGET_IO_BATCH_SIZE, &multiget_io_batch_size);

  // There are 2 keys in L1 in twp separate files, and 1 in L2. With
-  // async IO, all three lookups will happen in parallel
-  ASSERT_EQ(multiget_io_batch_size.count, 1);
+  // optimize_multiget_for_io, all three lookups will happen in parallel.
+  // Otherwise, the L2 lookup will happen after L1.
+  ASSERT_EQ(multiget_io_batch_size.count, GetParam() ? 1 : 2);
  ASSERT_EQ(multiget_io_batch_size.max, GetParam() ? 3 : 2);
 }

@@ -13,7 +13,6 @@

 #include "cache/cache_entry_roles.h"
 #include "cache/cache_key.h"
-#include "cache/clock_cache.h"
 #include "cache/fast_lru_cache.h"
 #include "cache/lru_cache.h"
 #include "db/column_family.h"
@@ -171,13 +170,16 @@ class DBBlockCacheTest : public DBTestBase {
 #ifndef ROCKSDB_LITE
  const std::array<size_t, kNumCacheEntryRoles> GetCacheEntryRoleCountsBg() {
    // Verify in cache entry role stats
-    ColumnFamilyHandleImpl* cfh =
-        static_cast<ColumnFamilyHandleImpl*>(dbfull()->DefaultColumnFamily());
-    InternalStats* internal_stats_ptr = cfh->cfd()->internal_stats();
-    InternalStats::CacheEntryRoleStats stats;
-    internal_stats_ptr->TEST_GetCacheEntryRoleStats(&stats,
-                                                    /*foreground=*/false);
-    return stats.entry_counts;
+    std::array<size_t, kNumCacheEntryRoles> cache_entry_role_counts;
+    std::map<std::string, std::string> values;
+    EXPECT_TRUE(db_->GetMapProperty(DB::Properties::kFastBlockCacheEntryStats,
+                                    &values));
+    for (size_t i = 0; i < kNumCacheEntryRoles; ++i) {
+      auto role = static_cast<CacheEntryRole>(i);
+      cache_entry_role_counts[i] =
+          ParseSizeT(values[BlockCacheEntryStatsMapKeys::EntryCount(role)]);
+    }
+    return cache_entry_role_counts;
  }
 #endif  // ROCKSDB_LITE
 };
@@ -678,8 +680,8 @@ INSTANTIATE_TEST_CASE_P(DBBlockCacheTest1, DBBlockCacheTest1,
 TEST_P(DBBlockCacheTest1, WarmCacheWithBlocksDuringFlush) {
  Options options = CurrentOptions();
  options.create_if_missing = true;
+  options.disable_auto_compactions = true;
  options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
-  options.max_compaction_bytes = 2000;

  BlockBasedTableOptions table_options;
  table_options.block_cache = NewLRUCache(1 << 25, 0, false);
@@ -737,8 +739,10 @@ TEST_P(DBBlockCacheTest1, WarmCacheWithBlocksDuringFlush) {
  }

  // Verify compaction not counted
-  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), /*begin=*/nullptr,
-                              /*end=*/nullptr));
+  CompactRangeOptions cro;
+  // Ensure files are rewritten, not just trivially moved.
+  cro.bottommost_level_compaction = BottommostLevelCompaction::kForceOptimized;
+  ASSERT_OK(db_->CompactRange(cro, /*begin=*/nullptr, /*end=*/nullptr));
  EXPECT_EQ(kNumBlocks,
            options.statistics->getTickerCount(BLOCK_CACHE_DATA_ADD));
  // Index and filter blocks are automatically warmed when the new table file
@@ -936,9 +940,11 @@ TEST_F(DBBlockCacheTest, AddRedundantStats) {
  int iterations_tested = 0;
  for (std::shared_ptr<Cache> base_cache :
       {NewLRUCache(capacity, num_shard_bits),
-        ExperimentalNewClockCache(
-            capacity, 1 /*estimated_value_size*/, num_shard_bits,
-            false /*strict_capacity_limit*/, kDefaultCacheMetadataChargePolicy),
+        HyperClockCacheOptions(
+            capacity,
+            BlockBasedTableOptions().block_size /*estimated_value_size*/,
+            num_shard_bits)
+            .MakeSharedCache(),
        NewFastLRUCache(capacity, 1 /*estimated_value_size*/, num_shard_bits,
                        false /*strict_capacity_limit*/,
                        kDefaultCacheMetadataChargePolicy)}) {
@@ -1296,10 +1302,10 @@ TEST_F(DBBlockCacheTest, CacheEntryRoleStats) {
  for (bool partition : {false, true}) {
    for (std::shared_ptr<Cache> cache :
         {NewLRUCache(capacity),
-          ExperimentalNewClockCache(capacity, 1 /*estimated_value_size*/,
-                                    -1 /*num_shard_bits*/,
-                                    false /*strict_capacity_limit*/,
-                                    kDefaultCacheMetadataChargePolicy)}) {
+          HyperClockCacheOptions(
+              capacity,
+              BlockBasedTableOptions().block_size /*estimated_value_size*/)
+              .MakeSharedCache()}) {
      if (!cache) {
        // Skip clock cache when not supported
        continue;
@@ -1497,23 +1503,38 @@ TEST_F(DBBlockCacheTest, CacheEntryRoleStats) {
      dbfull()->DumpStats();
      ASSERT_EQ(scan_count, 1);

-      env_->MockSleepForSeconds(10000);
+      env_->MockSleepForSeconds(60);
+      ASSERT_TRUE(db_->GetMapProperty(DB::Properties::kFastBlockCacheEntryStats,
+                                      &values));
+      ASSERT_EQ(scan_count, 1);
      ASSERT_TRUE(
          db_->GetMapProperty(DB::Properties::kBlockCacheEntryStats, &values));
      ASSERT_EQ(scan_count, 2);

      env_->MockSleepForSeconds(10000);
-      std::string value_str;
-      ASSERT_TRUE(
-          db_->GetProperty(DB::Properties::kBlockCacheEntryStats, &value_str));
+      ASSERT_TRUE(db_->GetMapProperty(DB::Properties::kFastBlockCacheEntryStats,
+                                      &values));
      ASSERT_EQ(scan_count, 3);

+      env_->MockSleepForSeconds(60);
+      std::string value_str;
+      ASSERT_TRUE(db_->GetProperty(DB::Properties::kFastBlockCacheEntryStats,
+                                   &value_str));
+      ASSERT_EQ(scan_count, 3);
+      ASSERT_TRUE(
+          db_->GetProperty(DB::Properties::kBlockCacheEntryStats, &value_str));
+      ASSERT_EQ(scan_count, 4);
+
      env_->MockSleepForSeconds(10000);
+      ASSERT_TRUE(db_->GetProperty(DB::Properties::kFastBlockCacheEntryStats,
+                                   &value_str));
+      ASSERT_EQ(scan_count, 5);
+
      ASSERT_TRUE(db_->GetProperty(DB::Properties::kCFStats, &value_str));
      // To match historical speed, querying this property no longer triggers
      // a scan, even if results are old. But periodic dump stats should keep
      // things reasonably updated.
-      ASSERT_EQ(scan_count, /*unchanged*/ 3);
+      ASSERT_EQ(scan_count, /*unchanged*/ 5);

      SyncPoint::GetInstance()->DisableProcessing();
      SyncPoint::GetInstance()->ClearAllCallBacks();
@@ -1568,6 +1589,10 @@ TEST_P(DBBlockCacheKeyTest, StableCacheKeys) {
  options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
  options.env = test_env.get();

+  // Corrupting the table properties corrupts the unique id.
+  // Ignore the unique id recorded in the manifest.
+  options.verify_sst_unique_id_in_manifest = false;
+
  BlockBasedTableOptions table_options;

  int key_count = 0;
@@ -1706,7 +1706,7 @@ TEST_F(DBBloomFilterTest, ContextCustomFilterPolicy) {
    ASSERT_OK(Put(1, Key(maxKey + 55555), Key(maxKey + 55555)));
    Flush(1);
    EXPECT_EQ(policy->DumpTestReport(),
-              fifo ? "cf=abe,s=kCompactionStyleFIFO,n=1,l=0,b=0,r=kFlush\n"
+              fifo ? "cf=abe,s=kCompactionStyleFIFO,n=7,l=0,b=0,r=kFlush\n"
                   : "cf=bob,s=kCompactionStyleLevel,n=7,l=0,b=0,r=kFlush\n");

    for (int i = maxKey / 2; i < maxKey; i++) {
@@ -1714,7 +1714,7 @@ TEST_F(DBBloomFilterTest, ContextCustomFilterPolicy) {
    }
    Flush(1);
    EXPECT_EQ(policy->DumpTestReport(),
-              fifo ? "cf=abe,s=kCompactionStyleFIFO,n=1,l=0,b=0,r=kFlush\n"
+              fifo ? "cf=abe,s=kCompactionStyleFIFO,n=7,l=0,b=0,r=kFlush\n"
                   : "cf=bob,s=kCompactionStyleLevel,n=7,l=0,b=0,r=kFlush\n");

    // Check that they can be found
@@ -1738,7 +1738,7 @@ TEST_F(DBBloomFilterTest, ContextCustomFilterPolicy) {
      EXPECT_LE(useful_count, maxKey * 2 * (fifo ? 0.9995 : 0.98));
    }

-    if (!fifo) {  // FIFO only has L0
+    if (!fifo) {  // FIFO doesn't fully support CompactRange
      // Full compaction
      ASSERT_OK(db_->CompactRange(CompactRangeOptions(), handles_[1], nullptr,
                                  nullptr));
@@ -2496,7 +2496,7 @@ TEST_F(DBBloomFilterTest, OptimizeFiltersForHits) {
  for (int i = 0; i < numkeys; i += 2) {
    keys.push_back(i);
  }
-  RandomShuffle(std::begin(keys), std::end(keys));
+  RandomShuffle(std::begin(keys), std::end(keys), /*seed*/ 42);
  int num_inserted = 0;
  for (int key : keys) {
    ASSERT_OK(Put(1, Key(key), "val"));
@@ -328,11 +328,9 @@ TEST_F(DBTestCompactionFilter, CompactionFilter) {
  Arena arena;
  {
    InternalKeyComparator icmp(options.comparator);
-    ReadRangeDelAggregator range_del_agg(&icmp,
-                                         kMaxSequenceNumber /* upper_bound */);
    ReadOptions read_options;
    ScopedArenaIterator iter(dbfull()->NewInternalIterator(
-        read_options, &arena, &range_del_agg, kMaxSequenceNumber, handles_[1]));
+        read_options, &arena, kMaxSequenceNumber, handles_[1]));
    iter->SeekToFirst();
    ASSERT_OK(iter->status());
    while (iter->Valid()) {
@@ -422,11 +420,9 @@ TEST_F(DBTestCompactionFilter, CompactionFilter) {
  count = 0;
  {
    InternalKeyComparator icmp(options.comparator);
-    ReadRangeDelAggregator range_del_agg(&icmp,
-                                         kMaxSequenceNumber /* upper_bound */);
    ReadOptions read_options;
    ScopedArenaIterator iter(dbfull()->NewInternalIterator(
-        read_options, &arena, &range_del_agg, kMaxSequenceNumber, handles_[1]));
+        read_options, &arena, kMaxSequenceNumber, handles_[1]));
    iter->SeekToFirst();
    ASSERT_OK(iter->status());
    while (iter->Valid()) {
@@ -701,11 +697,9 @@ TEST_F(DBTestCompactionFilter, CompactionFilterContextManual) {
    int total = 0;
    Arena arena;
    InternalKeyComparator icmp(options.comparator);
-    ReadRangeDelAggregator range_del_agg(&icmp,
-                                         kMaxSequenceNumber /* snapshots */);
    ReadOptions read_options;
-    ScopedArenaIterator iter(dbfull()->NewInternalIterator(
-        read_options, &arena, &range_del_agg, kMaxSequenceNumber));
+    ScopedArenaIterator iter(dbfull()->NewInternalIterator(read_options, &arena,
+                                                           kMaxSequenceNumber));
    iter->SeekToFirst();
    ASSERT_OK(iter->status());
    while (iter->Valid()) {
@@ -3824,6 +3824,242 @@ TEST_F(DBCompactionTest, NoCompactBottomLevelFilesWithDeletions) {
  }
 }

+TEST_F(DBCompactionTest, RoundRobinTtlCompactionNormal) {
+  Options options = CurrentOptions();
+  options.compression = kNoCompression;
+  options.level0_file_num_compaction_trigger = 20;
+  options.ttl = 24 * 60 * 60;  // 24 hours
+  options.compaction_pri = kRoundRobin;
+  env_->now_cpu_count_.store(0);
+  env_->SetMockSleep();
+  options.env = env_;
+
+  // add a small second for each wait time, to make sure the file is expired
+  int small_seconds = 1;
+
+  std::atomic_int ttl_compactions{0};
+  std::atomic_int round_robin_ttl_compactions{0};
+  std::atomic_int other_compactions{0};
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "LevelCompactionPicker::PickCompaction:Return", [&](void* arg) {
+        Compaction* compaction = reinterpret_cast<Compaction*>(arg);
+        auto compaction_reason = compaction->compaction_reason();
+        if (compaction_reason == CompactionReason::kTtl) {
+          ttl_compactions++;
+        } else if (compaction_reason == CompactionReason::kRoundRobinTtl) {
+          round_robin_ttl_compactions++;
+        } else {
+          other_compactions++;
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  DestroyAndReopen(options);
+
+  // Setup the files from lower level to up level, each file is 1 hour's older
+  // than the next one.
+  // create 10 files on the last level (L6)
+  for (int i = 0; i < 10; i++) {
+    for (int j = 0; j < 100; j++) {
+      ASSERT_OK(Put(Key(i * 100 + j), "value" + std::to_string(i * 100 + j)));
+    }
+    ASSERT_OK(Flush());
+    env_->MockSleepForSeconds(60 * 60);  // generate 1 file per hour
+  }
+  MoveFilesToLevel(6);
+
+  // create 5 files on L5
+  for (int i = 0; i < 5; i++) {
+    for (int j = 0; j < 200; j++) {
+      ASSERT_OK(Put(Key(i * 200 + j), "value" + std::to_string(i * 200 + j)));
+    }
+    ASSERT_OK(Flush());
+    env_->MockSleepForSeconds(60 * 60);
+  }
+  MoveFilesToLevel(5);
+
+  // create 3 files on L4
+  for (int i = 0; i < 3; i++) {
+    for (int j = 0; j < 300; j++) {
+      ASSERT_OK(Put(Key(i * 300 + j), "value" + std::to_string(i * 300 + j)));
+    }
+    ASSERT_OK(Flush());
+    env_->MockSleepForSeconds(60 * 60);
+  }
+  MoveFilesToLevel(4);
+
+  // The LSM tree should be like:
+  // L4: [0,            299], [300,      599], [600,     899]
+  // L5: [0,  199]      [200,         399]...............[800, 999]
+  // L6: [0,99][100,199][200,299][300,399]...............[800,899][900,999]
+  ASSERT_EQ("0,0,0,0,3,5,10", FilesPerLevel());
+
+  // make sure the first L5 file is expired
+  env_->MockSleepForSeconds(16 * 60 * 60 + small_seconds++);
+
+  // trigger TTL compaction
+  ASSERT_OK(Put(Key(4), "value" + std::to_string(1)));
+  ASSERT_OK(Put(Key(5), "value" + std::to_string(1)));
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  // verify there's a RoundRobin TTL compaction
+  ASSERT_EQ(1, round_robin_ttl_compactions);
+  round_robin_ttl_compactions = 0;
+
+  // expire 2 more files
+  env_->MockSleepForSeconds(2 * 60 * 60 + small_seconds++);
+  // trigger TTL compaction
+  ASSERT_OK(Put(Key(4), "value" + std::to_string(2)));
+  ASSERT_OK(Put(Key(5), "value" + std::to_string(2)));
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  ASSERT_EQ(2, round_robin_ttl_compactions);
+  round_robin_ttl_compactions = 0;
+
+  // expire 4 more files, 2 out of 3 files on L4 are expired
+  env_->MockSleepForSeconds(4 * 60 * 60 + small_seconds++);
+  // trigger TTL compaction
+  ASSERT_OK(Put(Key(6), "value" + std::to_string(3)));
+  ASSERT_OK(Put(Key(7), "value" + std::to_string(3)));
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  ASSERT_EQ(1, NumTableFilesAtLevel(4));
+  ASSERT_EQ(0, NumTableFilesAtLevel(5));
+
+  ASSERT_GT(round_robin_ttl_compactions, 0);
+  round_robin_ttl_compactions = 0;
+
+  // make the first L0 file expired, which triggers a normal TTL compaction
+  // instead of roundrobin TTL compaction, it will also include an extra file
+  // from L0 because of overlap
+  ASSERT_EQ(0, ttl_compactions);
+  env_->MockSleepForSeconds(19 * 60 * 60 + small_seconds++);
+
+  // trigger TTL compaction
+  ASSERT_OK(Put(Key(6), "value" + std::to_string(4)));
+  ASSERT_OK(Put(Key(7), "value" + std::to_string(4)));
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  // L0 -> L1 compaction is normal TTL compaction, L1 -> next levels compactions
+  // are RoundRobin TTL compaction.
+  ASSERT_GT(ttl_compactions, 0);
+  ttl_compactions = 0;
+  ASSERT_GT(round_robin_ttl_compactions, 0);
+  round_robin_ttl_compactions = 0;
+
+  // All files are expired, so only the last level has data
+  env_->MockSleepForSeconds(24 * 60 * 60);
+  // trigger TTL compaction
+  ASSERT_OK(Put(Key(6), "value" + std::to_string(4)));
+  ASSERT_OK(Put(Key(7), "value" + std::to_string(4)));
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+  ASSERT_EQ("0,0,0,0,0,0,2", FilesPerLevel());
+
+  ASSERT_GT(ttl_compactions, 0);
+  ttl_compactions = 0;
+  ASSERT_GT(round_robin_ttl_compactions, 0);
+  round_robin_ttl_compactions = 0;
+
+  ASSERT_EQ(0, other_compactions);
+}
+
+TEST_F(DBCompactionTest, RoundRobinTtlCompactionUnsortedTime) {
+  // This is to test the case that the RoundRobin compaction cursor not pointing
+  // to the oldest file, RoundRobin compaction should still compact the file
+  // after cursor until all expired files are compacted.
+  Options options = CurrentOptions();
+  options.compression = kNoCompression;
+  options.level0_file_num_compaction_trigger = 20;
+  options.ttl = 24 * 60 * 60;  // 24 hours
+  options.compaction_pri = kRoundRobin;
+  env_->now_cpu_count_.store(0);
+  env_->SetMockSleep();
+  options.env = env_;
+
+  std::atomic_int ttl_compactions{0};
+  std::atomic_int round_robin_ttl_compactions{0};
+  std::atomic_int other_compactions{0};
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "LevelCompactionPicker::PickCompaction:Return", [&](void* arg) {
+        Compaction* compaction = reinterpret_cast<Compaction*>(arg);
+        auto compaction_reason = compaction->compaction_reason();
+        if (compaction_reason == CompactionReason::kTtl) {
+          ttl_compactions++;
+        } else if (compaction_reason == CompactionReason::kRoundRobinTtl) {
+          round_robin_ttl_compactions++;
+        } else {
+          other_compactions++;
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  DestroyAndReopen(options);
+
+  // create 10 files on the last level (L6)
+  for (int i = 0; i < 10; i++) {
+    for (int j = 0; j < 100; j++) {
+      ASSERT_OK(Put(Key(i * 100 + j), "value" + std::to_string(i * 100 + j)));
+    }
+    ASSERT_OK(Flush());
+    env_->MockSleepForSeconds(60 * 60);  // generate 1 file per hour
+  }
+  MoveFilesToLevel(6);
+
+  // create 5 files on L5
+  for (int i = 0; i < 5; i++) {
+    for (int j = 0; j < 200; j++) {
+      ASSERT_OK(Put(Key(i * 200 + j), "value" + std::to_string(i * 200 + j)));
+    }
+    ASSERT_OK(Flush());
+    env_->MockSleepForSeconds(60 * 60);  // 1 hour
+  }
+  MoveFilesToLevel(5);
+
+  // The LSM tree should be like:
+  // L5: [0,  199]      [200,         399] [400,599] [600,799] [800, 999]
+  // L6: [0,99][100,199][200,299][300,399]....................[800,899][900,999]
+  ASSERT_EQ("0,0,0,0,0,5,10", FilesPerLevel());
+
+  // point the compaction cursor to the 4th file on L5
+  VersionSet* const versions = dbfull()->GetVersionSet();
+  assert(versions);
+  ColumnFamilyData* const cfd = versions->GetColumnFamilySet()->GetDefault();
+  ASSERT_NE(cfd, nullptr);
+  Version* const current = cfd->current();
+  ASSERT_NE(current, nullptr);
+  VersionStorageInfo* storage_info = current->storage_info();
+  ASSERT_NE(storage_info, nullptr);
+  const InternalKey split_cursor = InternalKey(Key(600), 100000, kTypeValue);
+  storage_info->AddCursorForOneLevel(5, split_cursor);
+
+  // make the first file on L5 expired, there should be 3 TTL compactions:
+  // 4th one, 5th one, then 1st one.
+  env_->MockSleepForSeconds(19 * 60 * 60 + 1);
+  // trigger TTL compaction
+  ASSERT_OK(Put(Key(6), "value" + std::to_string(4)));
+  ASSERT_OK(Put(Key(7), "value" + std::to_string(4)));
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+  ASSERT_EQ(2, NumTableFilesAtLevel(5));
+
+  ASSERT_EQ(3, round_robin_ttl_compactions);
+  ASSERT_EQ(0, ttl_compactions);
+  ASSERT_EQ(0, other_compactions);
+}
+
 TEST_F(DBCompactionTest, LevelCompactExpiredTtlFiles) {
  const int kNumKeysPerFile = 32;
  const int kNumLevelFiles = 2;
@@ -7958,6 +8194,42 @@ TEST_F(DBCompactionTest, BottomPriCompactionCountsTowardConcurrencyLimit) {
  compact_range_thread.join();
 }

+TEST_F(DBCompactionTest, BottommostFileCompactionAllowIngestBehind) {
+  // allow_ingest_behind prevents seqnum zeroing, and could cause
+  // compaction loop with reason kBottommostFiles.
+  Options options = CurrentOptions();
+  options.env = env_;
+  options.compaction_style = kCompactionStyleLevel;
+  options.allow_ingest_behind = true;
+  options.comparator = BytewiseComparator();
+  DestroyAndReopen(options);
+
+  WriteOptions write_opts;
+  ASSERT_OK(db_->Put(write_opts, "infinite", "compaction loop"));
+  ASSERT_OK(db_->Put(write_opts, "infinite", "loop"));
+
+  ASSERT_OK(Flush());
+  MoveFilesToLevel(1);
+  ASSERT_OK(db_->Put(write_opts, "bumpseqnum", ""));
+  ASSERT_OK(Flush());
+  auto snapshot = db_->GetSnapshot();
+  // Bump up oldest_snapshot_seqnum_ in VersionStorageInfo.
+  db_->ReleaseSnapshot(snapshot);
+  bool compacted = false;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "LevelCompactionPicker::PickCompaction:Return", [&](void* /* arg */) {
+        // There should not be a compaction.
+        compacted = true;
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+  // Wait for compaction to be scheduled.
+  env_->SleepForMicroseconds(2000000);
+  ASSERT_FALSE(compacted);
+  // The following assert can be used to check for compaction loop:
+  // it used to wait forever before the fix.
+  // ASSERT_OK(dbfull()->TEST_WaitForCompact(true /* wait_unscheduled */));
+}
+
 #endif  // !defined(ROCKSDB_LITE)

 }  // namespace ROCKSDB_NAMESPACE
@@ -3032,6 +3032,44 @@ TEST_P(DBAtomicFlushTest, BgThreadNoWaitAfterManifestError) {
  SyncPoint::GetInstance()->ClearAllCallBacks();
 }

+TEST_P(DBAtomicFlushTest, NoWaitWhenWritesStopped) {
+  Options options = GetDefaultOptions();
+  options.create_if_missing = true;
+  options.atomic_flush = GetParam();
+  options.max_write_buffer_number = 2;
+  options.memtable_factory.reset(test::NewSpecialSkipListFactory(1));
+
+  Reopen(options);
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->LoadDependency(
+      {{"DBImpl::DelayWrite:Start",
+        "DBAtomicFlushTest::NoWaitWhenWritesStopped:0"}});
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(dbfull()->PauseBackgroundWork());
+  for (int i = 0; i < options.max_write_buffer_number; ++i) {
+    ASSERT_OK(Put("k" + std::to_string(i), "v" + std::to_string(i)));
+  }
+  std::thread stalled_writer([&]() { ASSERT_OK(Put("k", "v")); });
+
+  TEST_SYNC_POINT("DBAtomicFlushTest::NoWaitWhenWritesStopped:0");
+
+  {
+    FlushOptions flush_opts;
+    flush_opts.wait = false;
+    flush_opts.allow_write_stall = true;
+    ASSERT_TRUE(db_->Flush(flush_opts).IsTryAgain());
+  }
+
+  ASSERT_OK(dbfull()->ContinueBackgroundWork());
+  ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
+
+  stalled_writer.join();
+
+  SyncPoint::GetInstance()->DisableProcessing();
+}
+
 INSTANTIATE_TEST_CASE_P(DBFlushDirectIOTest, DBFlushDirectIOTest,
                        testing::Bool());

@@ -242,7 +242,7 @@ Status CompactedDBImpl::Open(const Options& options,
  std::unique_ptr<CompactedDBImpl> db(new CompactedDBImpl(db_options, dbname));
  Status s = db->Init(options);
  if (s.ok()) {
-    s = db->StartPeriodicWorkScheduler();
+    s = db->StartPeriodicTaskScheduler();
  }
  if (s.ok()) {
    ROCKS_LOG_INFO(db->immutable_db_options_.info_log,
@@ -11,6 +11,7 @@

 namespace ROCKSDB_NAMESPACE {

+// TODO: Share common structure with DBImplSecondary and DBImplReadOnly
 class CompactedDBImpl : public DBImpl {
 public:
  CompactedDBImpl(const DBOptions& options, const std::string& dbname);
@@ -127,6 +128,17 @@ class CompactedDBImpl : public DBImpl {
    return Status::NotSupported("Not supported in compacted db mode.");
  }

+  // FIXME: some missing overrides for more "write" functions
+  // Share with DBImplReadOnly?
+
+ protected:
+#ifndef ROCKSDB_LITE
+  Status FlushForGetLiveFiles() override {
+    // No-op for read-only DB
+    return Status::OK();
+  }
+#endif  // !ROCKSDB_LITE
+
 private:
  friend class DB;
  inline size_t FindFile(const Slice& key);
@@ -45,7 +45,7 @@
 #include "db/memtable_list.h"
 #include "db/merge_context.h"
 #include "db/merge_helper.h"
-#include "db/periodic_work_scheduler.h"
+#include "db/periodic_task_scheduler.h"
 #include "db/range_tombstone_fragmenter.h"
 #include "db/table_cache.h"
 #include "db/table_properties_collector.h"
@@ -165,8 +165,13 @@ DBImpl::DBImpl(const DBOptions& options, const std::string& dbname,
      fs_(immutable_db_options_.fs, io_tracer_),
      mutable_db_options_(initial_db_options_),
      stats_(immutable_db_options_.stats),
+#ifdef COERCE_CONTEXT_SWITCH
+      mutex_(stats_, immutable_db_options_.clock, DB_MUTEX_WAIT_MICROS, &bg_cv_,
+             immutable_db_options_.use_adaptive_mutex),
+#else   // COERCE_CONTEXT_SWITCH
      mutex_(stats_, immutable_db_options_.clock, DB_MUTEX_WAIT_MICROS,
             immutable_db_options_.use_adaptive_mutex),
+#endif  // COERCE_CONTEXT_SWITCH
      default_cf_handle_(nullptr),
      error_handler_(this, immutable_db_options_, &mutex_),
      event_logger_(immutable_db_options_.info_log.get()),
@@ -217,7 +222,7 @@ DBImpl::DBImpl(const DBOptions& options, const std::string& dbname,
      refitting_level_(false),
      opened_successfully_(false),
 #ifndef ROCKSDB_LITE
-      periodic_work_scheduler_(nullptr),
+      periodic_task_scheduler_(),
 #endif  // ROCKSDB_LITE
      two_write_queues_(options.two_write_queues),
      manual_wal_flush_(options.manual_wal_flush),
@@ -260,6 +265,18 @@ DBImpl::DBImpl(const DBOptions& options, const std::string& dbname,
  SetDbSessionId();
  assert(!db_session_id_.empty());

+#ifndef ROCKSDB_LITE
+  periodic_task_functions_.emplace(PeriodicTaskType::kDumpStats,
+                                   [this]() { this->DumpStats(); });
+  periodic_task_functions_.emplace(PeriodicTaskType::kPersistStats,
+                                   [this]() { this->PersistStats(); });
+  periodic_task_functions_.emplace(PeriodicTaskType::kFlushInfoLog,
+                                   [this]() { this->FlushInfoLog(); });
+  periodic_task_functions_.emplace(
+      PeriodicTaskType::kRecordSeqnoTime,
+      [this]() { this->RecordSeqnoToTimeMapping(); });
+#endif  // ROCKSDB_LITE
+
  versions_.reset(new VersionSet(dbname_, &immutable_db_options_, file_options_,
                                 table_cache_.get(), write_buffer_manager_,
                                 &write_controller_, &block_cache_tracer_,
@@ -480,9 +497,15 @@ void DBImpl::CancelAllBackgroundWork(bool wait) {
                 "Shutdown: canceling all background work");

 #ifndef ROCKSDB_LITE
-  if (periodic_work_scheduler_ != nullptr) {
-    periodic_work_scheduler_->Unregister(this);
-    periodic_work_scheduler_->UnregisterRecordSeqnoTimeWorker(this);
+  for (uint8_t task_type = 0;
+       task_type < static_cast<uint8_t>(PeriodicTaskType::kMax); task_type++) {
+    Status s = periodic_task_scheduler_.Unregister(
+        static_cast<PeriodicTaskType>(task_type));
+    if (!s.ok()) {
+      ROCKS_LOG_WARN(immutable_db_options_.info_log,
+                     "Failed to unregister periodic task %d, status: %s",
+                     task_type, s.ToString().c_str());
+    }
  }
 #endif  // !ROCKSDB_LITE

@@ -767,30 +790,50 @@ void DBImpl::PrintStatistics() {
  }
 }

-Status DBImpl::StartPeriodicWorkScheduler() {
+Status DBImpl::StartPeriodicTaskScheduler() {
 #ifndef ROCKSDB_LITE

 #ifndef NDEBUG
  // It only used by test to disable scheduler
  bool disable_scheduler = false;
  TEST_SYNC_POINT_CALLBACK(
-      "DBImpl::StartPeriodicWorkScheduler:DisableScheduler",
+      "DBImpl::StartPeriodicTaskScheduler:DisableScheduler",
      &disable_scheduler);
  if (disable_scheduler) {
    return Status::OK();
  }
-#endif  // !NDEBUG

  {
    InstrumentedMutexLock l(&mutex_);
-    periodic_work_scheduler_ = PeriodicWorkScheduler::Default();
-    TEST_SYNC_POINT_CALLBACK("DBImpl::StartPeriodicWorkScheduler:Init",
-                             &periodic_work_scheduler_);
+    TEST_SYNC_POINT_CALLBACK("DBImpl::StartPeriodicTaskScheduler:Init",
+                             &periodic_task_scheduler_);
  }

-  return periodic_work_scheduler_->Register(
-      this, mutable_db_options_.stats_dump_period_sec,
-      mutable_db_options_.stats_persist_period_sec);
+#endif  // !NDEBUG
+  if (mutable_db_options_.stats_dump_period_sec > 0) {
+    Status s = periodic_task_scheduler_.Register(
+        PeriodicTaskType::kDumpStats,
+        periodic_task_functions_.at(PeriodicTaskType::kDumpStats),
+        mutable_db_options_.stats_dump_period_sec);
+    if (!s.ok()) {
+      return s;
+    }
+  }
+  if (mutable_db_options_.stats_persist_period_sec > 0) {
+    Status s = periodic_task_scheduler_.Register(
+        PeriodicTaskType::kPersistStats,
+        periodic_task_functions_.at(PeriodicTaskType::kPersistStats),
+        mutable_db_options_.stats_persist_period_sec);
+    if (!s.ok()) {
+      return s;
+    }
+  }
+
+  Status s = periodic_task_scheduler_.Register(
+      PeriodicTaskType::kFlushInfoLog,
+      periodic_task_functions_.at(PeriodicTaskType::kFlushInfoLog));
+
+  return s;
 #else
  return Status::OK();
 #endif  // !ROCKSDB_LITE
@@ -798,20 +841,19 @@ Status DBImpl::StartPeriodicWorkScheduler() {

 Status DBImpl::RegisterRecordSeqnoTimeWorker() {
 #ifndef ROCKSDB_LITE
-  if (!periodic_work_scheduler_) {
-    return Status::OK();
-  }
  uint64_t min_time_duration = std::numeric_limits<uint64_t>::max();
  uint64_t max_time_duration = std::numeric_limits<uint64_t>::min();
  {
    InstrumentedMutexLock l(&mutex_);

    for (auto cfd : *versions_->GetColumnFamilySet()) {
-      uint64_t preclude_last_option =
-          cfd->ioptions()->preclude_last_level_data_seconds;
-      if (!cfd->IsDropped() && preclude_last_option > 0) {
-        min_time_duration = std::min(preclude_last_option, min_time_duration);
-        max_time_duration = std::max(preclude_last_option, max_time_duration);
+      // preserve time is the max of 2 options.
+      uint64_t preserve_time_duration =
+          std::max(cfd->ioptions()->preserve_internal_time_seconds,
+                   cfd->ioptions()->preclude_last_level_data_seconds);
+      if (!cfd->IsDropped() && preserve_time_duration > 0) {
+        min_time_duration = std::min(preserve_time_duration, min_time_duration);
+        max_time_duration = std::max(preserve_time_duration, max_time_duration);
      }
    }
    if (min_time_duration == std::numeric_limits<uint64_t>::max()) {
@@ -823,31 +865,21 @@ Status DBImpl::RegisterRecordSeqnoTimeWorker() {

  uint64_t seqno_time_cadence = 0;
  if (min_time_duration != std::numeric_limits<uint64_t>::max()) {
+    // round up to 1 when the time_duration is smaller than
+    // kMaxSeqnoTimePairsPerCF
    seqno_time_cadence =
-        min_time_duration / SeqnoToTimeMapping::kMaxSeqnoTimePairsPerCF;
+        (min_time_duration + SeqnoToTimeMapping::kMaxSeqnoTimePairsPerCF - 1) /
+        SeqnoToTimeMapping::kMaxSeqnoTimePairsPerCF;
  }

  Status s;
-  if (seqno_time_cadence != record_seqno_time_cadence_) {
-    if (seqno_time_cadence == 0) {
-      periodic_work_scheduler_->UnregisterRecordSeqnoTimeWorker(this);
-    } else {
-      s = periodic_work_scheduler_->RegisterRecordSeqnoTimeWorker(
-          this, seqno_time_cadence);
-    }
-
-    if (s.ok()) {
-      record_seqno_time_cadence_ = seqno_time_cadence;
-    }
-
-    if (s.IsNotSupported()) {
-      // TODO: Fix the timer cannot cancel and re-add the same task
-      ROCKS_LOG_WARN(
-          immutable_db_options_.info_log,
-          "Updating seqno to time worker cadence is not supported yet, to make "
-          "the change effective, please reopen the DB instance.");
-      s = Status::OK();
-    }
+  if (seqno_time_cadence == 0) {
+    s = periodic_task_scheduler_.Unregister(PeriodicTaskType::kRecordSeqnoTime);
+  } else {
+    s = periodic_task_scheduler_.Register(
+        PeriodicTaskType::kRecordSeqnoTime,
+        periodic_task_functions_.at(PeriodicTaskType::kRecordSeqnoTime),
+        seqno_time_cadence);
  }

  return s;
@@ -1087,6 +1119,10 @@ void DBImpl::DumpStats() {
  PrintStatistics();
 }

+// Periodically flush info log out of application buffer at a low frequency.
+// This improves debuggability in case of RocksDB hanging since it ensures the
+// log messages leading up to the hang will eventually become visible in the
+// log.
 void DBImpl::FlushInfoLog() {
  if (shutdown_initiated_) {
    return;
@@ -1279,22 +1315,36 @@ Status DBImpl::SetDBOptions(
        MaybeScheduleFlushOrCompaction();
      }

-      if (new_options.stats_dump_period_sec !=
-              mutable_db_options_.stats_dump_period_sec ||
-          new_options.stats_persist_period_sec !=
-              mutable_db_options_.stats_persist_period_sec) {
-        mutex_.Unlock();
-        periodic_work_scheduler_->Unregister(this);
-        s = periodic_work_scheduler_->Register(
-            this, new_options.stats_dump_period_sec,
-            new_options.stats_persist_period_sec);
-        mutex_.Lock();
+      mutex_.Unlock();
+      if (new_options.stats_dump_period_sec == 0) {
+        s = periodic_task_scheduler_.Unregister(PeriodicTaskType::kDumpStats);
+      } else {
+        s = periodic_task_scheduler_.Register(
+            PeriodicTaskType::kDumpStats,
+            periodic_task_functions_.at(PeriodicTaskType::kDumpStats),
+            new_options.stats_dump_period_sec);
      }
      if (new_options.max_total_wal_size !=
          mutable_db_options_.max_total_wal_size) {
        max_total_wal_size_.store(new_options.max_total_wal_size,
                                  std::memory_order_release);
      }
+      if (s.ok()) {
+        if (new_options.stats_persist_period_sec == 0) {
+          s = periodic_task_scheduler_.Unregister(
+              PeriodicTaskType::kPersistStats);
+        } else {
+          s = periodic_task_scheduler_.Register(
+              PeriodicTaskType::kPersistStats,
+              periodic_task_functions_.at(PeriodicTaskType::kPersistStats),
+              new_options.stats_persist_period_sec);
+        }
+      }
+      mutex_.Lock();
+      if (!s.ok()) {
+        return s;
+      }
+
      write_controller_.set_max_delayed_write_rate(
          new_options.delayed_write_rate);
      table_cache_.get()->SetCapacity(new_options.max_open_files == -1
@@ -1393,11 +1443,11 @@ Status DBImpl::FlushWAL(bool sync) {
      // future writes
      IOStatusCheck(io_s);
      // whether sync or not, we should abort the rest of function upon error
-      return std::move(io_s);
+      return static_cast<Status>(io_s);
    }
    if (!sync) {
      ROCKS_LOG_DEBUG(immutable_db_options_.info_log, "FlushWAL sync=false");
-      return std::move(io_s);
+      return static_cast<Status>(io_s);
    }
  }
  if (!sync) {
@@ -1408,6 +1458,18 @@ Status DBImpl::FlushWAL(bool sync) {
  return SyncWAL();
 }

+bool DBImpl::WALBufferIsEmpty(bool lock) {
+  if (lock) {
+    log_write_mutex_.Lock();
+  }
+  log::Writer* cur_log_writer = logs_.back().writer;
+  auto res = cur_log_writer->BufferIsEmpty();
+  if (lock) {
+    log_write_mutex_.Unlock();
+  }
+  return res;
+}
+
 Status DBImpl::SyncWAL() {
  TEST_SYNC_POINT("DBImpl::SyncWAL:Begin");
  autovector<log::Writer*, 1> logs_to_sync;
@@ -1494,8 +1556,8 @@ Status DBImpl::SyncWAL() {
 Status DBImpl::ApplyWALToManifest(VersionEdit* synced_wals) {
  // not empty, write to MANIFEST.
  mutex_.AssertHeld();
-  Status status =
-      versions_->LogAndApplyToDefaultColumnFamily(synced_wals, &mutex_);
+  Status status = versions_->LogAndApplyToDefaultColumnFamily(
+      synced_wals, &mutex_, directories_.GetDbDir());
  if (!status.ok() && versions_->io_status().IsIOError()) {
    status = error_handler_.SetBGError(versions_->io_status(),
                                       BackgroundErrorReason::kManifestWrite);
@@ -1506,7 +1568,7 @@ Status DBImpl::ApplyWALToManifest(VersionEdit* synced_wals) {
 Status DBImpl::LockWAL() {
  log_write_mutex_.Lock();
  auto cur_log_writer = logs_.back().writer;
-  auto status = cur_log_writer->WriteBuffer();
+  IOStatus status = cur_log_writer->WriteBuffer();
  if (!status.ok()) {
    ROCKS_LOG_ERROR(immutable_db_options_.info_log, "WAL flush error %s",
                    status.ToString().c_str());
@@ -1514,7 +1576,7 @@ Status DBImpl::LockWAL() {
    // future writes
    WriteStatusCheck(status);
  }
-  return std::move(status);
+  return static_cast<Status>(status);
 }

 Status DBImpl::UnlockWAL() {
@@ -1532,20 +1594,28 @@ void DBImpl::MarkLogsSynced(uint64_t up_to, bool synced_dir,
    auto& wal = *it;
    assert(wal.IsSyncing());

-    if (logs_.size() > 1) {
+    if (wal.number < logs_.back().number) {
+      // Inactive WAL
      if (immutable_db_options_.track_and_verify_wals_in_manifest &&
          wal.GetPreSyncSize() > 0) {
        synced_wals->AddWal(wal.number, WalMetadata(wal.GetPreSyncSize()));
      }
-      logs_to_free_.push_back(wal.ReleaseWriter());
-      it = logs_.erase(it);
+      if (wal.GetPreSyncSize() == wal.writer->file()->GetFlushedSize()) {
+        // Fully synced
+        logs_to_free_.push_back(wal.ReleaseWriter());
+        it = logs_.erase(it);
+      } else {
+        assert(wal.GetPreSyncSize() < wal.writer->file()->GetFlushedSize());
+        wal.FinishSync();
+        ++it;
+      }
    } else {
+      assert(wal.number == logs_.back().number);
+      // Active WAL
      wal.FinishSync();
      ++it;
    }
  }
-  assert(logs_.empty() || logs_[0].number > up_to ||
-         (logs_.size() == 1 && !logs_[0].IsSyncing()));
  log_sync_cv_.SignalAll();
 }

@@ -1593,7 +1663,6 @@ Status DBImpl::GetFullHistoryTsLow(ColumnFamilyHandle* column_family,

 InternalIterator* DBImpl::NewInternalIterator(const ReadOptions& read_options,
                                              Arena* arena,
-                                              RangeDelAggregator* range_del_agg,
                                              SequenceNumber sequence,
                                              ColumnFamilyHandle* column_family,
                                              bool allow_unprepared_value) {
@@ -1608,8 +1677,8 @@ InternalIterator* DBImpl::NewInternalIterator(const ReadOptions& read_options,
  mutex_.Lock();
  SuperVersion* super_version = cfd->GetSuperVersion()->Ref();
  mutex_.Unlock();
-  return NewInternalIterator(read_options, cfd, super_version, arena,
-                             range_del_agg, sequence, allow_unprepared_value);
+  return NewInternalIterator(read_options, cfd, super_version, arena, sequence,
+                             allow_unprepared_value);
 }

 void DBImpl::SchedulePurge() {
@@ -1740,48 +1809,54 @@ static void CleanupGetMergeOperandsState(void* arg1, void* /*arg2*/) {

 }  // namespace

-InternalIterator* DBImpl::NewInternalIterator(const ReadOptions& read_options,
-                                              ColumnFamilyData* cfd,
-                                              SuperVersion* super_version,
-                                              Arena* arena,
-                                              RangeDelAggregator* range_del_agg,
-                                              SequenceNumber sequence,
-                                              bool allow_unprepared_value) {
+InternalIterator* DBImpl::NewInternalIterator(
+    const ReadOptions& read_options, ColumnFamilyData* cfd,
+    SuperVersion* super_version, Arena* arena, SequenceNumber sequence,
+    bool allow_unprepared_value, ArenaWrappedDBIter* db_iter) {
  InternalIterator* internal_iter;
  assert(arena != nullptr);
-  assert(range_del_agg != nullptr);
  // Need to create internal iterator from the arena.
  MergeIteratorBuilder merge_iter_builder(
      &cfd->internal_comparator(), arena,
      !read_options.total_order_seek &&
-          super_version->mutable_cf_options.prefix_extractor != nullptr);
-  // Collect iterator for mutable mem
-  merge_iter_builder.AddIterator(
-      super_version->mem->NewIterator(read_options, arena));
-  std::unique_ptr<FragmentedRangeTombstoneIterator> range_del_iter;
+          super_version->mutable_cf_options.prefix_extractor != nullptr,
+      read_options.iterate_upper_bound);
+  // Collect iterator for mutable memtable
+  auto mem_iter = super_version->mem->NewIterator(read_options, arena);
  Status s;
  if (!read_options.ignore_range_deletions) {
-    range_del_iter.reset(super_version->mem->NewRangeTombstoneIterator(
-        read_options, sequence, false /* immutable_memtable */));
-    range_del_agg->AddTombstones(std::move(range_del_iter));
+    TruncatedRangeDelIterator* mem_tombstone_iter = nullptr;
+    auto range_del_iter = super_version->mem->NewRangeTombstoneIterator(
+        read_options, sequence, false /* immutable_memtable */);
+    if (range_del_iter == nullptr || range_del_iter->empty()) {
+      delete range_del_iter;
+    } else {
+      mem_tombstone_iter = new TruncatedRangeDelIterator(
+          std::unique_ptr<FragmentedRangeTombstoneIterator>(range_del_iter),
+          &cfd->ioptions()->internal_comparator, nullptr /* smallest */,
+          nullptr /* largest */);
+    }
+    merge_iter_builder.AddPointAndTombstoneIterator(mem_iter,
+                                                    mem_tombstone_iter);
+  } else {
+    merge_iter_builder.AddIterator(mem_iter);
  }
+
  // Collect all needed child iterators for immutable memtables
  if (s.ok()) {
-    super_version->imm->AddIterators(read_options, &merge_iter_builder);
-    if (!read_options.ignore_range_deletions) {
-      s = super_version->imm->AddRangeTombstoneIterators(read_options, arena,
-                                                         range_del_agg);
-    }
+    super_version->imm->AddIterators(read_options, &merge_iter_builder,
+                                     !read_options.ignore_range_deletions);
  }
  TEST_SYNC_POINT_CALLBACK("DBImpl::NewInternalIterator:StatusCallback", &s);
  if (s.ok()) {
    // Collect iterators for files in L0 - Ln
    if (read_options.read_tier != kMemtableTier) {
      super_version->current->AddIterators(read_options, file_options_,
-                                           &merge_iter_builder, range_del_agg,
+                                           &merge_iter_builder,
                                           allow_unprepared_value);
    }
-    internal_iter = merge_iter_builder.Finish();
+    internal_iter = merge_iter_builder.Finish(
+        read_options.ignore_range_deletions ? nullptr : db_iter);
    SuperVersionHandle* cleanup = new SuperVersionHandle(
        this, &mutex_, super_version,
        read_options.background_purge_on_iterator_cleanup ||
@@ -3034,13 +3109,9 @@ Status DBImpl::CreateColumnFamilyImpl(const ColumnFamilyOptions& cf_options,
    }
  }  // InstrumentedMutexLock l(&mutex_)

-  if (cf_options.preclude_last_level_data_seconds > 0) {
-    // TODO(zjay): Fix the timer issue and re-enable this.
-    ROCKS_LOG_ERROR(
-        immutable_db_options_.info_log,
-        "Creating column family with `preclude_last_level_data_seconds` needs "
-        "to restart DB to take effect");
-    //    s = RegisterRecordSeqnoTimeWorker();
+  if (cf_options.preserve_internal_time_seconds > 0 ||
+      cf_options.preclude_last_level_data_seconds > 0) {
+    s = RegisterRecordSeqnoTimeWorker();
  }
  sv_context.Clean();
  // this is outside the mutex
@@ -3106,7 +3177,7 @@ Status DBImpl::DropColumnFamilyImpl(ColumnFamilyHandle* column_family) {
      WriteThread::Writer w;
      write_thread_.EnterUnbatched(&w, &mutex_);
      s = versions_->LogAndApply(cfd, *cfd->GetLatestMutableCFOptions(), &edit,
-                                 &mutex_);
+                                 &mutex_, directories_.GetDbDir());
      write_thread_.ExitUnbatched(&w);
    }
    if (s.ok()) {
@@ -3130,7 +3201,8 @@ Status DBImpl::DropColumnFamilyImpl(ColumnFamilyHandle* column_family) {
    bg_cv_.SignalAll();
  }

-  if (cfd->ioptions()->preclude_last_level_data_seconds > 0) {
+  if (cfd->ioptions()->preserve_internal_time_seconds > 0 ||
+      cfd->ioptions()->preclude_last_level_data_seconds > 0) {
    s = RegisterRecordSeqnoTimeWorker();
  }

@@ -3311,9 +3383,8 @@ ArenaWrappedDBIter* DBImpl::NewIteratorImpl(const ReadOptions& read_options,
      read_options.snapshot != nullptr ? false : allow_refresh);

  InternalIterator* internal_iter = NewInternalIterator(
-      db_iter->GetReadOptions(), cfd, sv, db_iter->GetArena(),
-      db_iter->GetRangeDelAggregator(), snapshot,
-      /* allow_unprepared_value */ true);
+      db_iter->GetReadOptions(), cfd, sv, db_iter->GetArena(), snapshot,
+      /* allow_unprepared_value */ true, db_iter);
  db_iter->SetIterUnderDBIter(internal_iter);

  return db_iter;
@@ -3611,14 +3682,16 @@ void DBImpl::ReleaseSnapshot(const Snapshot* s) {
    if (oldest_snapshot > bottommost_files_mark_threshold_) {
      CfdList cf_scheduled;
      for (auto* cfd : *versions_->GetColumnFamilySet()) {
-        cfd->current()->storage_info()->UpdateOldestSnapshot(oldest_snapshot);
-        if (!cfd->current()
-                 ->storage_info()
-                 ->BottommostFilesMarkedForCompaction()
-                 .empty()) {
-          SchedulePendingCompaction(cfd);
-          MaybeScheduleFlushOrCompaction();
-          cf_scheduled.push_back(cfd);
+        if (!cfd->ioptions()->allow_ingest_behind) {
+          cfd->current()->storage_info()->UpdateOldestSnapshot(oldest_snapshot);
+          if (!cfd->current()
+                   ->storage_info()
+                   ->BottommostFilesMarkedForCompaction()
+                   .empty()) {
+            SchedulePendingCompaction(cfd);
+            MaybeScheduleFlushOrCompaction();
+            cf_scheduled.push_back(cfd);
+          }
        }
      }

@@ -3627,7 +3700,8 @@ void DBImpl::ReleaseSnapshot(const Snapshot* s) {
      // mutex might be unlocked during the loop, making the result inaccurate.
      SequenceNumber new_bottommost_files_mark_threshold = kMaxSequenceNumber;
      for (auto* cfd : *versions_->GetColumnFamilySet()) {
-        if (CfdListContains(cf_scheduled, cfd)) {
+        if (CfdListContains(cf_scheduled, cfd) ||
+            cfd->ioptions()->allow_ingest_behind) {
          continue;
        }
        new_bottommost_files_mark_threshold = std::min(
@@ -4340,10 +4414,14 @@ Status DBImpl::CheckConsistency() {
      }
      files_by_directory[md.db_path].push_back(fname);
    }
+
+    IOOptions io_opts;
+    io_opts.do_not_recurse = true;
    for (const auto& dir_files : files_by_directory) {
      std::string directory = dir_files.first;
      std::vector<std::string> existing_files;
-      Status s = env_->GetChildren(directory, &existing_files);
+      Status s = fs_->GetChildren(directory, io_opts, &existing_files,
+                                  /*IODebugContext*=*/nullptr);
      if (!s.ok()) {
        corruption_messages +=
            "Can't list files in " + directory + ": " + s.ToString() + "\n";
@@ -4525,8 +4603,12 @@ Status DestroyDB(const std::string& dbname, const Options& options,
  // Reset the logger because it holds a handle to the
  // log file and prevents cleanup and directory removal
  soptions.info_log.reset();
+  IOOptions io_opts;
  // Ignore error in case directory does not exist
-  env->GetChildren(dbname, &filenames).PermitUncheckedError();
+  soptions.fs
+      ->GetChildren(dbname, io_opts, &filenames,
+                    /*IODebugContext*=*/nullptr)
+      .PermitUncheckedError();

  FileLock* lock;
  const std::string lockname = LockFileName(dbname);
@@ -4566,8 +4648,12 @@ Status DestroyDB(const std::string& dbname, const Options& options,
        paths.insert(cf_path.path);
      }
    }
+
    for (const auto& path : paths) {
-      if (env->GetChildren(path, &filenames).ok()) {
+      if (soptions.fs
+              ->GetChildren(path, io_opts, &filenames,
+                            /*IODebugContext*=*/nullptr)
+              .ok()) {
        for (const auto& fname : filenames) {
          if (ParseFileName(fname, &number, &type) &&
              (type == kTableFile ||
@@ -4589,7 +4675,11 @@ Status DestroyDB(const std::string& dbname, const Options& options,
    std::string archivedir = ArchivalDirectory(dbname);
    bool wal_dir_exists = false;
    if (!soptions.IsWalDirSameAsDBPath(dbname)) {
-      wal_dir_exists = env->GetChildren(soptions.wal_dir, &walDirFiles).ok();
+      wal_dir_exists =
+          soptions.fs
+              ->GetChildren(soptions.wal_dir, io_opts, &walDirFiles,
+                            /*IODebugContext*=*/nullptr)
+              .ok();
      archivedir = ArchivalDirectory(soptions.wal_dir);
    }

@@ -4597,7 +4687,10 @@ Status DestroyDB(const std::string& dbname, const Options& options,
    // processed and removed before those otherwise we have issues
    // removing them
    std::vector<std::string> archiveFiles;
-    if (env->GetChildren(archivedir, &archiveFiles).ok()) {
+    if (soptions.fs
+            ->GetChildren(archivedir, io_opts, &archiveFiles,
+                          /*IODebugContext*=*/nullptr)
+            .ok()) {
      // Delete archival files.
      for (const auto& file : archiveFiles) {
        if (ParseFileName(file, &number, &type) && type == kWalFile) {
@@ -4738,7 +4831,10 @@ Status DBImpl::DeleteObsoleteOptionsFiles() {
  // to the oldest.
  std::map<uint64_t, std::string> options_filenames;
  Status s;
-  s = GetEnv()->GetChildren(GetName(), &filenames);
+  IOOptions io_opts;
+  io_opts.do_not_recurse = true;
+  s = fs_->GetChildren(GetName(), io_opts, &filenames,
+                       /*IODebugContext*=*/nullptr);
  if (!s.ok()) {
    return s;
  }
@@ -5196,7 +5292,7 @@ Status DBImpl::IngestExternalFiles(
        mutex_.Unlock();
        status = AtomicFlushMemTables(cfds_to_flush, flush_opts,
                                      FlushReason::kExternalFileIngestion,
-                                      true /* writes_stopped */);
+                                      true /* entered_write_thread */);
        mutex_.Lock();
      } else {
        for (size_t i = 0; i != num_cfs; ++i) {
@@ -5207,7 +5303,7 @@ Status DBImpl::IngestExternalFiles(
                    ->cfd();
            status = FlushMemTable(cfd, flush_opts,
                                   FlushReason::kExternalFileIngestion,
-                                   true /* writes_stopped */);
+                                   true /* entered_write_thread */);
            mutex_.Lock();
            if (!status.ok()) {
              break;
@@ -5687,8 +5783,24 @@ Status DBImpl::NewDefaultReplayer(
 Status DBImpl::StartBlockCacheTrace(
    const TraceOptions& trace_options,
    std::unique_ptr<TraceWriter>&& trace_writer) {
-  return block_cache_tracer_.StartTrace(immutable_db_options_.clock,
-                                        trace_options, std::move(trace_writer));
+  BlockCacheTraceOptions block_trace_opts;
+  block_trace_opts.sampling_frequency = trace_options.sampling_frequency;
+
+  BlockCacheTraceWriterOptions trace_writer_opt;
+  trace_writer_opt.max_trace_file_size = trace_options.max_trace_file_size;
+
+  std::unique_ptr<BlockCacheTraceWriter> block_cache_trace_writer =
+      NewBlockCacheTraceWriter(env_->GetSystemClock().get(), trace_writer_opt,
+                               std::move(trace_writer));
+
+  return block_cache_tracer_.StartTrace(block_trace_opts,
+                                        std::move(block_cache_trace_writer));
+}
+
+Status DBImpl::StartBlockCacheTrace(
+    const BlockCacheTraceOptions& trace_options,
+    std::unique_ptr<BlockCacheTraceWriter>&& trace_writer) {
+  return block_cache_tracer_.StartTrace(trace_options, std::move(trace_writer));
 }

 Status DBImpl::EndBlockCacheTrace() {
@@ -32,6 +32,7 @@
 #include "db/log_writer.h"
 #include "db/logs_with_prep_tracker.h"
 #include "db/memtable_list.h"
+#include "db/periodic_task_scheduler.h"
 #include "db/post_memtable_callback.h"
 #include "db/pre_release_callback.h"
 #include "db/range_del_aggregator.h"
@@ -75,10 +76,6 @@ class ArenaWrappedDBIter;
 class InMemoryStatsHistoryIterator;
 class MemTable;
 class PersistentStatsHistoryIterator;
-class PeriodicWorkScheduler;
-#ifndef NDEBUG
-class PeriodicWorkTestScheduler;
-#endif  // !NDEBUG
 class TableCache;
 class TaskLimiterToken;
 class Version;
@@ -225,6 +222,9 @@ class DBImpl : public DB {
  Status DeleteRange(const WriteOptions& options,
                     ColumnFamilyHandle* column_family, const Slice& begin_key,
                     const Slice& end_key) override;
+  Status DeleteRange(const WriteOptions& options,
+                     ColumnFamilyHandle* column_family, const Slice& begin_key,
+                     const Slice& end_key, const Slice& ts) override;

  using DB::Write;
  virtual Status Write(const WriteOptions& options,
@@ -426,7 +426,7 @@ class DBImpl : public DB {
      const FlushOptions& options,
      const std::vector<ColumnFamilyHandle*>& column_families) override;
  virtual Status FlushWAL(bool sync) override;
-  bool TEST_WALBufferIsEmpty(bool lock = true);
+  bool WALBufferIsEmpty(bool lock = true);
  virtual Status SyncWAL() override;
  virtual Status LockWAL() override;
  virtual Status UnlockWAL() override;
@@ -568,9 +568,13 @@ class DBImpl : public DB {

  using DB::StartBlockCacheTrace;
  Status StartBlockCacheTrace(
-      const TraceOptions& options,
+      const TraceOptions& trace_options,
      std::unique_ptr<TraceWriter>&& trace_writer) override;

+  Status StartBlockCacheTrace(
+      const BlockCacheTraceOptions& options,
+      std::unique_ptr<BlockCacheTraceWriter>&& trace_writer) override;
+
  using DB::EndBlockCacheTrace;
  Status EndBlockCacheTrace() override;

@@ -742,13 +746,29 @@ class DBImpl : public DB {
  // the value and so will require PrepareValue() to be called before value();
  // allow_unprepared_value = false is convenient when this optimization is not
  // useful, e.g. when reading the whole column family.
+  //
+  // read_options.ignore_range_deletions determines whether range tombstones are
+  // processed in the returned interator internally, i.e., whether range
+  // tombstone covered keys are in this iterator's output.
  // @param read_options Must outlive the returned iterator.
  InternalIterator* NewInternalIterator(
-      const ReadOptions& read_options, Arena* arena,
-      RangeDelAggregator* range_del_agg, SequenceNumber sequence,
+      const ReadOptions& read_options, Arena* arena, SequenceNumber sequence,
      ColumnFamilyHandle* column_family = nullptr,
      bool allow_unprepared_value = false);

+  // Note: to support DB iterator refresh, memtable range tombstones in the
+  // underlying merging iterator needs to be refreshed. If db_iter is not
+  // nullptr, db_iter->SetMemtableRangetombstoneIter() is called with the
+  // memtable range tombstone iterator used by the underlying merging iterator.
+  // This range tombstone iterator can be refreshed later by db_iter.
+  // @param read_options Must outlive the returned iterator.
+  InternalIterator* NewInternalIterator(const ReadOptions& read_options,
+                                        ColumnFamilyData* cfd,
+                                        SuperVersion* super_version,
+                                        Arena* arena, SequenceNumber sequence,
+                                        bool allow_unprepared_value,
+                                        ArenaWrappedDBIter* db_iter = nullptr);
+
  LogsWithPrepTracker* logs_with_prep_tracker() {
    return &logs_with_prep_tracker_;
  }
@@ -871,15 +891,6 @@ class DBImpl : public DB {

  const WriteController& write_controller() { return write_controller_; }

-  // @param read_options Must outlive the returned iterator.
-  InternalIterator* NewInternalIterator(const ReadOptions& read_options,
-                                        ColumnFamilyData* cfd,
-                                        SuperVersion* super_version,
-                                        Arena* arena,
-                                        RangeDelAggregator* range_del_agg,
-                                        SequenceNumber sequence,
-                                        bool allow_unprepared_value);
-
  // hollow transactions shell used for recovery.
  // these will then be passed to TransactionDB so that
  // locks can be reacquired before writing can resume.
@@ -1147,7 +1158,7 @@ class DBImpl : public DB {
  int TEST_BGCompactionsAllowed() const;
  int TEST_BGFlushesAllowed() const;
  size_t TEST_GetWalPreallocateBlockSize(uint64_t write_buffer_size) const;
-  void TEST_WaitForPeridicWorkerRun(std::function<void()> callback) const;
+  void TEST_WaitForPeridicTaskRun(std::function<void()> callback) const;
  SeqnoToTimeMapping TEST_GetSeqnoToTimeMapping() const;
  size_t TEST_EstimateInMemoryStatsHistorySize() const;

@@ -1162,7 +1173,7 @@ class DBImpl : public DB {
  }

 #ifndef ROCKSDB_LITE
-  PeriodicWorkTestScheduler* TEST_GetPeriodicWorkScheduler() const;
+  const PeriodicTaskScheduler& TEST_GetPeriodicTaskScheduler() const;
 #endif  // !ROCKSDB_LITE

 #endif  // NDEBUG
@@ -1390,7 +1401,7 @@ class DBImpl : public DB {
  void NotifyOnExternalFileIngested(
      ColumnFamilyData* cfd, const ExternalSstFileIngestionJob& ingestion_job);

-  Status FlushForGetLiveFiles();
+  virtual Status FlushForGetLiveFiles();
 #endif  // !ROCKSDB_LITE

  void NewThreadStatusCfInfo(ColumnFamilyData* cfd) const;
@@ -1873,12 +1884,13 @@ class DBImpl : public DB {

  // Force current memtable contents to be flushed.
  Status FlushMemTable(ColumnFamilyData* cfd, const FlushOptions& options,
-                       FlushReason flush_reason, bool writes_stopped = false);
+                       FlushReason flush_reason,
+                       bool entered_write_thread = false);

  Status AtomicFlushMemTables(
      const autovector<ColumnFamilyData*>& column_family_datas,
      const FlushOptions& options, FlushReason flush_reason,
-      bool writes_stopped = false);
+      bool entered_write_thread = false);

  // Wait until flushing this column family won't stall writes
  Status WaitUntilFlushWouldNotStallWrites(ColumnFamilyData* cfd,
@@ -2069,7 +2081,7 @@ class DBImpl : public DB {
                              LogBuffer* log_buffer);

  // Schedule background tasks
-  Status StartPeriodicWorkScheduler();
+  Status StartPeriodicTaskScheduler();

  Status RegisterRecordSeqnoTimeWorker();

@@ -2175,11 +2187,6 @@ class DBImpl : public DB {
  IOStatus CreateWAL(uint64_t log_file_num, uint64_t recycle_log_number,
                     size_t preallocate_block_size, log::Writer** new_log);

-  // Verify SST file unique id between Manifest and table properties to make
-  // sure they're the same. Currently only used during DB open when
-  // `verify_sst_unique_id_in_manifest = true`.
-  Status VerifySstUniqueIdInManifest();
-
  // Validate self-consistency of DB options
  static Status ValidateOptions(const DBOptions& db_options);
  // Validate self-consistency of DB options and its consistency with cf options
@@ -2611,14 +2618,11 @@ class DBImpl : public DB {

 #ifndef ROCKSDB_LITE
  // Scheduler to run DumpStats(), PersistStats(), and FlushInfoLog().
-  // Currently, it always use a global instance from
-  // PeriodicWorkScheduler::Default(). Only in unittest, it can be overrided by
-  // PeriodicWorkTestScheduler.
-  PeriodicWorkScheduler* periodic_work_scheduler_;
+  // Currently, internally it has a global timer instance for running the tasks.
+  PeriodicTaskScheduler periodic_task_scheduler_;

-  // Current cadence of the periodic worker for recording sequence number to
-  // time.
-  uint64_t record_seqno_time_cadence_ = 0;
+  // It contains the implementations for each periodic task.
+  std::map<PeriodicTaskType, const PeriodicTaskFunc> periodic_task_functions_;
 #endif

  // When set, we use a separate queue for writes that don't write to memtable.
@@ -988,7 +988,7 @@ Status DBImpl::IncreaseFullHistoryTsLowImpl(ColumnFamilyData* cfd,
  }

  Status s = versions_->LogAndApply(cfd, *cfd->GetLatestMutableCFOptions(),
-                                    &edit, &mutex_);
+                                    &edit, &mutex_, directories_.GetDbDir());
  if (!s.ok()) {
    return s;
  }
@@ -1055,10 +1055,10 @@ Status DBImpl::CompactRangeInternal(const CompactRangeOptions& options,
      SelectColumnFamiliesForAtomicFlush(&cfds);
      mutex_.Unlock();
      s = AtomicFlushMemTables(cfds, fo, FlushReason::kManualCompaction,
-                               false /* writes_stopped */);
+                               false /* entered_write_thread */);
    } else {
      s = FlushMemTable(cfd, fo, FlushReason::kManualCompaction,
-                        false /* writes_stopped*/);
+                        false /* entered_write_thread */);
    }
    if (!s.ok()) {
      LogFlush(immutable_db_options_.info_log);
@@ -2016,9 +2016,16 @@ void DBImpl::GenerateFlushRequest(const autovector<ColumnFamilyData*>& cfds,

 Status DBImpl::FlushMemTable(ColumnFamilyData* cfd,
                             const FlushOptions& flush_options,
-                             FlushReason flush_reason, bool writes_stopped) {
+                             FlushReason flush_reason,
+                             bool entered_write_thread) {
  // This method should not be called if atomic_flush is true.
  assert(!immutable_db_options_.atomic_flush);
+  if (!flush_options.wait && write_controller_.IsStopped()) {
+    std::ostringstream oss;
+    oss << "Writes have been stopped, thus unable to perform manual flush. "
+           "Please try again later after writes are resumed";
+    return Status::TryAgain(oss.str());
+  }
  Status s;
  if (!flush_options.allow_write_stall) {
    bool flush_needed = true;
@@ -2029,6 +2036,7 @@ Status DBImpl::FlushMemTable(ColumnFamilyData* cfd,
    }
  }

+  const bool needs_to_join_write_thread = !entered_write_thread;
  autovector<FlushRequest> flush_reqs;
  autovector<uint64_t> memtable_ids_to_wait;
  {
@@ -2037,7 +2045,7 @@ Status DBImpl::FlushMemTable(ColumnFamilyData* cfd,

    WriteThread::Writer w;
    WriteThread::Writer nonmem_w;
-    if (!writes_stopped) {
+    if (needs_to_join_write_thread) {
      write_thread_.EnterUnbatched(&w, &mutex_);
      if (two_write_queues_) {
        nonmem_write_thread_.EnterUnbatched(&nonmem_w, &mutex_);
@@ -2120,7 +2128,7 @@ Status DBImpl::FlushMemTable(ColumnFamilyData* cfd,
      MaybeScheduleFlushOrCompaction();
    }

-    if (!writes_stopped) {
+    if (needs_to_join_write_thread) {
      write_thread_.ExitUnbatched(&w);
      if (two_write_queues_) {
        nonmem_write_thread_.ExitUnbatched(&nonmem_w);
@@ -2156,7 +2164,14 @@ Status DBImpl::FlushMemTable(ColumnFamilyData* cfd,
 Status DBImpl::AtomicFlushMemTables(
    const autovector<ColumnFamilyData*>& column_family_datas,
    const FlushOptions& flush_options, FlushReason flush_reason,
-    bool writes_stopped) {
+    bool entered_write_thread) {
+  assert(immutable_db_options_.atomic_flush);
+  if (!flush_options.wait && write_controller_.IsStopped()) {
+    std::ostringstream oss;
+    oss << "Writes have been stopped, thus unable to perform manual flush. "
+           "Please try again later after writes are resumed";
+    return Status::TryAgain(oss.str());
+  }
  Status s;
  if (!flush_options.allow_write_stall) {
    int num_cfs_to_flush = 0;
@@ -2173,6 +2188,7 @@ Status DBImpl::AtomicFlushMemTables(
      return s;
    }
  }
+  const bool needs_to_join_write_thread = !entered_write_thread;
  FlushRequest flush_req;
  autovector<ColumnFamilyData*> cfds;
  {
@@ -2181,7 +2197,7 @@ Status DBImpl::AtomicFlushMemTables(

    WriteThread::Writer w;
    WriteThread::Writer nonmem_w;
-    if (!writes_stopped) {
+    if (needs_to_join_write_thread) {
      write_thread_.EnterUnbatched(&w, &mutex_);
      if (two_write_queues_) {
        nonmem_write_thread_.EnterUnbatched(&nonmem_w, &mutex_);
@@ -2229,7 +2245,7 @@ Status DBImpl::AtomicFlushMemTables(
      MaybeScheduleFlushOrCompaction();
    }

-    if (!writes_stopped) {
+    if (needs_to_join_write_thread) {
      write_thread_.ExitUnbatched(&w);
      if (two_write_queues_) {
        nonmem_write_thread_.ExitUnbatched(&nonmem_w);
@@ -3263,7 +3279,6 @@ Status DBImpl::BackgroundCompaction(bool* made_progress,
    TEST_SYNC_POINT_CALLBACK("DBImpl::BackgroundCompaction:BeforeCompaction",
                             c->column_family_data());
    assert(c->num_input_files(1) == 0);
-    assert(c->level() == 0);
    assert(c->column_family_data()->ioptions()->compaction_style ==
           kCompactionStyleFIFO);

@@ -3756,9 +3771,11 @@ void DBImpl::InstallSuperVersionAndScheduleWork(
  // triggered soon anyway.
  bottommost_files_mark_threshold_ = kMaxSequenceNumber;
  for (auto* my_cfd : *versions_->GetColumnFamilySet()) {
-    bottommost_files_mark_threshold_ = std::min(
-        bottommost_files_mark_threshold_,
-        my_cfd->current()->storage_info()->bottommost_files_mark_threshold());
+    if (!my_cfd->ioptions()->allow_ingest_behind) {
+      bottommost_files_mark_threshold_ = std::min(
+          bottommost_files_mark_threshold_,
+          my_cfd->current()->storage_info()->bottommost_files_mark_threshold());
+    }
  }

  // Whenever we install new SuperVersion, we might need to issue new flushes or
@@ -12,7 +12,7 @@
 #include "db/column_family.h"
 #include "db/db_impl/db_impl.h"
 #include "db/error_handler.h"
-#include "db/periodic_work_scheduler.h"
+#include "db/periodic_task_scheduler.h"
 #include "monitoring/thread_status_updater.h"
 #include "util/cast_util.h"

@@ -31,18 +31,6 @@ Status DBImpl::TEST_SwitchWAL() {
  return s;
 }

-bool DBImpl::TEST_WALBufferIsEmpty(bool lock) {
-  if (lock) {
-    log_write_mutex_.Lock();
-  }
-  log::Writer* cur_log_writer = logs_.back().writer;
-  auto res = cur_log_writer->TEST_BufferIsEmpty();
-  if (lock) {
-    log_write_mutex_.Unlock();
-  }
-  return res;
-}
-
 uint64_t DBImpl::TEST_MaxNextLevelOverlappingBytes(
    ColumnFamilyHandle* column_family) {
  ColumnFamilyData* cfd;
@@ -302,16 +290,12 @@ size_t DBImpl::TEST_GetWalPreallocateBlockSize(
 }

 #ifndef ROCKSDB_LITE
-void DBImpl::TEST_WaitForPeridicWorkerRun(
-    std::function<void()> callback) const {
-  if (periodic_work_scheduler_ != nullptr) {
-    static_cast<PeriodicWorkTestScheduler*>(periodic_work_scheduler_)
-        ->TEST_WaitForRun(callback);
-  }
+void DBImpl::TEST_WaitForPeridicTaskRun(std::function<void()> callback) const {
+  periodic_task_scheduler_.TEST_WaitForRun(callback);
 }

-PeriodicWorkTestScheduler* DBImpl::TEST_GetPeriodicWorkScheduler() const {
-  return static_cast<PeriodicWorkTestScheduler*>(periodic_work_scheduler_);
+const PeriodicTaskScheduler& DBImpl::TEST_GetPeriodicTaskScheduler() const {
+  return periodic_task_scheduler_;
 }

 SeqnoToTimeMapping DBImpl::TEST_GetSeqnoToTimeMapping() const {
@@ -193,11 +193,14 @@ void DBImpl::FindObsoleteFiles(JobContext* job_context, bool force,
      }
    }

+    IOOptions io_opts;
+    io_opts.do_not_recurse = true;
    for (auto& path : paths) {
      // set of all files in the directory. We'll exclude files that are still
      // alive in the subsequent processings.
      std::vector<std::string> files;
-      Status s = env_->GetChildren(path, &files);
+      Status s = immutable_db_options_.fs->GetChildren(
+          path, io_opts, &files, /*IODebugContext*=*/nullptr);
      s.PermitUncheckedError();  // TODO: What should we do on error?
      for (const std::string& file : files) {
        uint64_t number;
@@ -223,7 +226,9 @@ void DBImpl::FindObsoleteFiles(JobContext* job_context, bool force,
    // Add log files in wal_dir
    if (!immutable_db_options_.IsWalDirSameAsDBPath(dbname_)) {
      std::vector<std::string> log_files;
-      Status s = env_->GetChildren(immutable_db_options_.wal_dir, &log_files);
+      Status s = immutable_db_options_.fs->GetChildren(
+          immutable_db_options_.wal_dir, io_opts, &log_files,
+          /*IODebugContext*=*/nullptr);
      s.PermitUncheckedError();  // TODO: What should we do on error?
      for (const std::string& log_file : log_files) {
        job_context->full_scan_candidate_files.emplace_back(
@@ -235,8 +240,9 @@ void DBImpl::FindObsoleteFiles(JobContext* job_context, bool force,
    if (!immutable_db_options_.db_log_dir.empty() &&
        immutable_db_options_.db_log_dir != dbname_) {
      std::vector<std::string> info_log_files;
-      Status s =
-          env_->GetChildren(immutable_db_options_.db_log_dir, &info_log_files);
+      Status s = immutable_db_options_.fs->GetChildren(
+          immutable_db_options_.db_log_dir, io_opts, &info_log_files,
+          /*IODebugContext*=*/nullptr);
      s.PermitUncheckedError();  // TODO: What should we do on error?
      for (std::string& log_file : info_log_files) {
        job_context->full_scan_candidate_files.emplace_back(
@@ -11,7 +11,7 @@
 #include "db/builder.h"
 #include "db/db_impl/db_impl.h"
 #include "db/error_handler.h"
-#include "db/periodic_work_scheduler.h"
+#include "db/periodic_task_scheduler.h"
 #include "env/composite_env_wrapper.h"
 #include "file/filename.h"
 #include "file/read_write_util.h"
@@ -163,8 +163,11 @@ DBOptions SanitizeOptions(const std::string& dbname, const DBOptions& src,
    // DeleteScheduler::CleanupDirectory on the same dir later, it will be
    // safe
    std::vector<std::string> filenames;
+    IOOptions io_opts;
+    io_opts.do_not_recurse = true;
    auto wal_dir = immutable_db_options.GetWalDir();
-    Status s = result.env->GetChildren(wal_dir, &filenames);
+    Status s = immutable_db_options.fs->GetChildren(
+        wal_dir, io_opts, &filenames, /*IODebugContext*=*/nullptr);
    s.PermitUncheckedError();  //**TODO: What to do on error?
    for (std::string& filename : filenames) {
      if (filename.find(".log.trash", filename.length() -
@@ -432,7 +435,10 @@ Status DBImpl::Recover(
      s = env_->FileExists(current_fname);
    } else {
      s = Status::NotFound();
-      Status io_s = env_->GetChildren(dbname_, &files_in_dbname);
+      IOOptions io_opts;
+      io_opts.do_not_recurse = true;
+      Status io_s = immutable_db_options_.fs->GetChildren(
+          dbname_, io_opts, &files_in_dbname, /*IODebugContext*=*/nullptr);
      if (!io_s.ok()) {
        s = io_s;
        files_in_dbname.clear();
@@ -498,7 +504,10 @@ Status DBImpl::Recover(
    }
  } else if (immutable_db_options_.best_efforts_recovery) {
    assert(files_in_dbname.empty());
-    Status s = env_->GetChildren(dbname_, &files_in_dbname);
+    IOOptions io_opts;
+    io_opts.do_not_recurse = true;
+    Status s = immutable_db_options_.fs->GetChildren(
+        dbname_, io_opts, &files_in_dbname, /*IODebugContext*=*/nullptr);
    if (s.IsNotFound()) {
      return Status::InvalidArgument(dbname_,
                                     "does not exist (open for read only)");
@@ -525,12 +534,6 @@ Status DBImpl::Recover(
  if (!s.ok()) {
    return s;
  }
-  if (immutable_db_options_.verify_sst_unique_id_in_manifest) {
-    s = VerifySstUniqueIdInManifest();
-    if (!s.ok()) {
-      return s;
-    }
-  }
  s = SetupDBId(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) {
@@ -576,7 +579,10 @@ Status DBImpl::Recover(
    // produced by an older version of rocksdb.
    auto wal_dir = immutable_db_options_.GetWalDir();
    if (!immutable_db_options_.best_efforts_recovery) {
-      s = env_->GetChildren(wal_dir, &files_in_wal_dir);
+      IOOptions io_opts;
+      io_opts.do_not_recurse = true;
+      s = immutable_db_options_.fs->GetChildren(
+          wal_dir, io_opts, &files_in_wal_dir, /*IODebugContext*=*/nullptr);
    }
    if (s.IsNotFound()) {
      return Status::InvalidArgument("wal_dir not found", wal_dir);
@@ -684,7 +690,10 @@ Status DBImpl::Recover(
      } else if (normalized_dbname == normalized_wal_dir) {
        filenames = std::move(files_in_wal_dir);
      } else {
-        s = env_->GetChildren(GetName(), &filenames);
+        IOOptions io_opts;
+        io_opts.do_not_recurse = true;
+        s = immutable_db_options_.fs->GetChildren(
+            GetName(), io_opts, &filenames, /*IODebugContext*=*/nullptr);
      }
    }
    if (s.ok()) {
@@ -708,31 +717,6 @@ Status DBImpl::Recover(
  return s;
 }

-Status DBImpl::VerifySstUniqueIdInManifest() {
-  mutex_.AssertHeld();
-  ROCKS_LOG_INFO(
-      immutable_db_options_.info_log,
-      "Verifying SST unique id between MANIFEST and SST file table properties");
-  Status status;
-  for (auto cfd : *versions_->GetColumnFamilySet()) {
-    if (!cfd->IsDropped()) {
-      auto version = cfd->current();
-      version->Ref();
-      mutex_.Unlock();
-      status = version->VerifySstUniqueIds();
-      mutex_.Lock();
-      version->Unref();
-      if (!status.ok()) {
-        ROCKS_LOG_WARN(immutable_db_options_.info_log,
-                       "SST unique id mismatch in column family \"%s\": %s",
-                       cfd->GetName().c_str(), status.ToString().c_str());
-        return status;
-      }
-    }
-  }
-  return status;
-}
-
 Status DBImpl::PersistentStatsProcessFormatVersion() {
  mutex_.AssertHeld();
  Status s;
@@ -1960,18 +1944,6 @@ Status DBImpl::Open(const DBOptions& db_options, const std::string& dbname,

  if (s.ok()) {
    for (auto cfd : *impl->versions_->GetColumnFamilySet()) {
-      if (cfd->ioptions()->compaction_style == kCompactionStyleFIFO) {
-        auto* vstorage = cfd->current()->storage_info();
-        for (int i = 1; i < vstorage->num_levels(); ++i) {
-          int num_files = vstorage->NumLevelFiles(i);
-          if (num_files > 0) {
-            s = Status::InvalidArgument(
-                "Not all files are at level 0. Cannot "
-                "open with FIFO compaction style.");
-            break;
-          }
-        }
-      }
      if (!cfd->mem()->IsSnapshotSupported()) {
        impl->is_snapshot_supported_ = false;
      }
@@ -2053,9 +2025,13 @@ Status DBImpl::Open(const DBOptions& db_options, const std::string& dbname,
    // 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_.env->GetChildren(path, &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;
@@ -2091,10 +2067,16 @@ Status DBImpl::Open(const DBOptions& db_options, const std::string& dbname,
    ROCKS_LOG_HEADER(impl->immutable_db_options_.info_log, "DB pointer %p",
                     impl);
    LogFlush(impl->immutable_db_options_.info_log);
-    assert(impl->TEST_WALBufferIsEmpty());
-    // If the assert above fails then we need to FlushWAL before returning
-    // control back to the user.
-    if (!persist_options_status.ok()) {
+    if (!impl->WALBufferIsEmpty()) {
+      s = impl->FlushWAL(false);
+      if (s.ok()) {
+        // Sync is needed otherwise WAL buffered data might get lost after a
+        // power reset.
+        log::Writer* log_writer = impl->logs_.back().writer;
+        s = log_writer->file()->Sync(impl->immutable_db_options_.use_fsync);
+      }
+    }
+    if (s.ok() && !persist_options_status.ok()) {
      s = Status::IOError(
          "DB::Open() failed --- Unable to persist Options file",
          persist_options_status.ToString());
@@ -2105,7 +2087,7 @@ Status DBImpl::Open(const DBOptions& db_options, const std::string& dbname,
                   "DB::Open() failed: %s", s.ToString().c_str());
  }
  if (s.ok()) {
-    s = impl->StartPeriodicWorkScheduler();
+    s = impl->StartPeriodicTaskScheduler();
  }

  if (s.ok()) {