Update HISTORY.md and version to 7.8.3

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

.circleci/config.yml

@@ -120,7 +120,7 @@ commands:
     steps:
       - run:
           name: "Test low-variance benchmarks"
-          command: ./tools/benchmark_ci.py --db_dir /tmp/rocksdb-benchmark-datadir --output_dir /tmp/benchmark-results --num_keys 50000
+          command: ./tools/benchmark_ci.py --db_dir /tmp/rocksdb-benchmark-datadir --output_dir /tmp/benchmark-results --num_keys 10000000
           environment:
             LD_LIBRARY_PATH: /usr/local/lib
             # How long to run parts of the test(s)
@@ -239,7 +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:

.github/workflows/sanity_check.yml

@@ -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

.gitignore

@@ -56,6 +56,7 @@ io_tracer_parser
 .DS_Store
 .vs
 .vscode
+.clangd
 
 java/out
 java/target

CMakeLists.txt

@@ -1065,28 +1065,34 @@ set(ROCKSDB_PLUGIN_BUILTINS "")
 message(STATUS "ROCKSDB PLUGINS TO BUILD ${ROCKSDB_PLUGINS}")
 foreach(PLUGIN IN LISTS PLUGINS)
   set(PLUGIN_ROOT "${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()
 

HISTORY.md

@@ -1,42 +1,58 @@
 # Rocksdb Change Log
-## 7.7.8 (11/27/2022)
-### Bug Fixes
+## 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.
-* Fixed a regression in iterator where range tombstones after `iterate_upper_bound` is processed.
 
-## 7.7.7 (11/15/2022)
 ### 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.
 
-## 7.7.6 (11/03/2022)
+## 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.
 
-## 7.7.5 (10/28/2022)
-### Bug Fixes
-* Fixed an iterator performance regression for delete range users when scanning through a consecutive sequence of range tombstones (#10877).
+## 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.
 
-## 7.7.4 (10/28/2022)
-### Bug Fixes
-* Fixed a case of calling malloc_usable_size on result of operator new[].
-
-## 7.7.3 (10/11/2022)
-### Bug Fixes
-* Fixed a memory safety bug in experimental HyperClockCache (#10768)
-
-## 7.7.2 (10/05/2022)
 ### 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).
 
-## 7.7.1 (09/26/2022)
-### Bug Fixes
-* 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).
+### 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
@@ -49,10 +65,6 @@
 * 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
-* 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.
-* 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.
 
 ### 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

Makefile

@@ -509,7 +509,7 @@ ifneq ($(strip $(FOLLY_PATH)),)
 
 	# 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 += $(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 $(FOLLY_PATH)/lib/libfolly.a $(FMT_LIB_PATH)/libfmt.a $(GLOG_LIB_PATH)/libglog.so $(GFLAGS_PATH)/lib/libgflags.so.2.2 -ldl
+	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
@@ -658,7 +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) $(DEV_HEADER_DIRS) -type f -name '*.h' | grep -E -v 'port/|plugin/|lua/|range_tree/')
 else
 	DEV_HEADERS :=
 endif
@@ -795,9 +795,9 @@ TOOLS_LIBRARY=$(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
@@ -1832,7 +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)

buckifier/buckify_rocksdb.py

@@ -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()

buckifier/targets_builder.py

@@ -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

buckifier/targets_cfg.py

@@ -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})
 """

buckifier/util.py

@@ -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:

build_tools/amalgamate.py

@@ -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()

build_tools/benchmark_log_tool.py

@@ -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())

build_tools/error_filter.py

@@ -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())

build_tools/gnu_parallel

@@ -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;

build_tools/ps_with_stack

@@ -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;

build_tools/ubuntu20_image/Dockerfile

@@ -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

cache/cache_reservation_manager_test.cc

@@ -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();
 }

cache/cache_test.cc

@@ -1023,21 +1023,21 @@ TEST_P(CacheTest, DefaultShardBits) {
       (GetParam() == kHyperClock ? 32U * 1024U : 512U) * 1024U;
 
   std::shared_ptr<Cache> cache = NewCache(32U * min_shard_size);
-  ShardedCache* sc = dynamic_cast<ShardedCache*>(cache.get());
+  ShardedCacheBase* sc = dynamic_cast<ShardedCacheBase*>(cache.get());
   ASSERT_EQ(5, sc->GetNumShardBits());
 
   cache = NewCache(min_shard_size / 1000U * 999U);
-  sc = dynamic_cast<ShardedCache*>(cache.get());
+  sc = dynamic_cast<ShardedCacheBase*>(cache.get());
   ASSERT_EQ(0, sc->GetNumShardBits());
 
   cache = NewCache(3U * 1024U * 1024U * 1024U);
-  sc = dynamic_cast<ShardedCache*>(cache.get());
+  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<ShardedCache*>(cache.get());
+    sc = dynamic_cast<ShardedCacheBase*>(cache.get());
     // current maximum of 6
     ASSERT_EQ(6, sc->GetNumShardBits());
   }

cache/clock_cache.cc

@@ -12,6 +12,7 @@
 #include <cassert>
 #include <functional>
 
+#include "cache/cache_key.h"
 #include "monitoring/perf_context_imp.h"
 #include "monitoring/statistics.h"
 #include "port/lang.h"
@@ -29,16 +30,22 @@ inline uint64_t GetRefcount(uint64_t meta) {
          ClockHandle::kCounterMask;
 }
 
+void ClockHandleBasicData::FreeData() const {
+  if (deleter) {
+    UniqueId64x2 unhashed;
+    (*deleter)(ClockCacheShard::ReverseHash(hashed_key, &unhashed), value);
+  }
+}
+
 static_assert(sizeof(ClockHandle) == 64U,
               "Expecting size / alignment with common cache line size");
 
 ClockHandleTable::ClockHandleTable(int hash_bits, bool initial_charge_metadata)
     : length_bits_(hash_bits),
-      length_bits_mask_(Lower32of64((uint64_t{1} << length_bits_) - 1)),
-      occupancy_limit_(static_cast<uint32_t>((uint64_t{1} << length_bits_) *
-                                             kStrictLoadFactor)),
+      length_bits_mask_((size_t{1} << length_bits_) - 1),
+      occupancy_limit_(static_cast<size_t>((uint64_t{1} << length_bits_) *
+                                           kStrictLoadFactor)),
       array_(new ClockHandle[size_t{1} << length_bits_]) {
-  assert(hash_bits <= 32);  // FIXME: ensure no overlap with sharding bits
   if (initial_charge_metadata) {
     usage_ += size_t{GetTableSize()} * sizeof(ClockHandle);
   }
@@ -47,7 +54,7 @@ ClockHandleTable::ClockHandleTable(int hash_bits, bool initial_charge_metadata)
 ClockHandleTable::~ClockHandleTable() {
   // Assumes there are no references or active operations on any slot/element
   // in the table.
-  for (uint32_t i = 0; i < GetTableSize(); i++) {
+  for (size_t i = 0; i < GetTableSize(); i++) {
     ClockHandle& h = array_[i];
     switch (h.meta >> ClockHandle::kStateShift) {
       case ClockHandle::kStateEmpty:
@@ -58,7 +65,7 @@ ClockHandleTable::~ClockHandleTable() {
         assert(GetRefcount(h.meta) == 0);
         h.FreeData();
 #ifndef NDEBUG
-        Rollback(h.hash, &h);
+        Rollback(h.hashed_key, &h);
         usage_.fetch_sub(h.total_charge, std::memory_order_relaxed);
         occupancy_.fetch_sub(1U, std::memory_order_relaxed);
 #endif
@@ -71,7 +78,7 @@ ClockHandleTable::~ClockHandleTable() {
   }
 
 #ifndef NDEBUG
-  for (uint32_t i = 0; i < GetTableSize(); i++) {
+  for (size_t i = 0; i < GetTableSize(); i++) {
     assert(array_[i].displacements.load() == 0);
   }
 #endif
@@ -154,12 +161,12 @@ inline void CorrectNearOverflow(uint64_t old_meta,
   }
 }
 
-Status ClockHandleTable::Insert(const ClockHandleMoreData& proto,
+Status ClockHandleTable::Insert(const ClockHandleBasicData& proto,
                                 ClockHandle** handle, Cache::Priority priority,
                                 size_t capacity, bool strict_capacity_limit) {
   // Do we have the available occupancy? Optimistically assume we do
   // and deal with it if we don't.
-  uint32_t old_occupancy = occupancy_.fetch_add(1, std::memory_order_acquire);
+  size_t old_occupancy = occupancy_.fetch_add(1, std::memory_order_acquire);
   auto revert_occupancy_fn = [&]() {
     occupancy_.fetch_sub(1, std::memory_order_relaxed);
   };
@@ -198,7 +205,7 @@ Status ClockHandleTable::Insert(const ClockHandleMoreData& proto,
     }
     if (request_evict_charge > 0) {
       size_t evicted_charge = 0;
-      uint32_t evicted_count = 0;
+      size_t evicted_count = 0;
       Evict(request_evict_charge, &evicted_charge, &evicted_count);
       occupancy_.fetch_sub(evicted_count, std::memory_order_release);
       if (LIKELY(evicted_charge > need_evict_charge)) {
@@ -263,7 +270,7 @@ Status ClockHandleTable::Insert(const ClockHandleMoreData& proto,
       need_evict_charge = 1;
     }
     size_t evicted_charge = 0;
-    uint32_t evicted_count = 0;
+    size_t evicted_count = 0;
     if (need_evict_charge > 0) {
       Evict(need_evict_charge, &evicted_charge, &evicted_count);
       // Deal with potential occupancy deficit
@@ -323,9 +330,9 @@ Status ClockHandleTable::Insert(const ClockHandleMoreData& proto,
     }
     assert(initial_countdown > 0);
 
-    uint32_t probe = 0;
+    size_t probe = 0;
     ClockHandle* e = FindSlot(
-        proto.hash,
+        proto.hashed_key,
         [&](ClockHandle* h) {
           // Optimistically transition the slot from "empty" to
           // "under construction" (no effect on other states)
@@ -338,7 +345,7 @@ Status ClockHandleTable::Insert(const ClockHandleMoreData& proto,
           if (old_state == ClockHandle::kStateEmpty) {
             // We've started inserting into an available slot, and taken
             // ownership Save data fields
-            ClockHandleMoreData* h_alias = h;
+            ClockHandleBasicData* h_alias = h;
             *h_alias = proto;
 
             // Transition from "under construction" state to "visible" state
@@ -375,7 +382,7 @@ Status ClockHandleTable::Insert(const ClockHandleMoreData& proto,
           if ((old_meta >> ClockHandle::kStateShift) ==
               ClockHandle::kStateVisible) {
             // Acquired a read reference
-            if (h->key == proto.key) {
+            if (h->hashed_key == proto.hashed_key) {
               // Match. Release in a way that boosts the clock state
               old_meta = h->meta.fetch_add(
                   ClockHandle::kReleaseIncrement * initial_countdown,
@@ -431,7 +438,7 @@ Status ClockHandleTable::Insert(const ClockHandleMoreData& proto,
       return Status::OK();
     }
     // Roll back table insertion
-    Rollback(proto.hash, e);
+    Rollback(proto.hashed_key, e);
     revert_occupancy_fn();
     // Maybe fall back on detached insert
     if (handle == nullptr) {
@@ -446,7 +453,7 @@ Status ClockHandleTable::Insert(const ClockHandleMoreData& proto,
   assert(use_detached_insert);
 
   ClockHandle* h = new ClockHandle();
-  ClockHandleMoreData* h_alias = h;
+  ClockHandleBasicData* h_alias = h;
   *h_alias = proto;
   h->detached = true;
   // Single reference (detached entries only created if returning a refed
@@ -467,10 +474,10 @@ Status ClockHandleTable::Insert(const ClockHandleMoreData& proto,
   return Status::OkOverwritten();
 }
 
-ClockHandle* ClockHandleTable::Lookup(const CacheKeyBytes& key, uint32_t hash) {
-  uint32_t probe = 0;
+ClockHandle* ClockHandleTable::Lookup(const UniqueId64x2& hashed_key) {
+  size_t probe = 0;
   ClockHandle* e = FindSlot(
-      hash,
+      hashed_key,
       [&](ClockHandle* h) {
         // Mostly branch-free version (similar performance)
         /*
@@ -501,7 +508,7 @@ ClockHandle* ClockHandleTable::Lookup(const CacheKeyBytes& key, uint32_t hash) {
         if ((old_meta >> ClockHandle::kStateShift) ==
             ClockHandle::kStateVisible) {
           // Acquired a read reference
-          if (h->key == key) {
+          if (h->hashed_key == hashed_key) {
             // Match
             return true;
           } else {
@@ -596,7 +603,7 @@ bool ClockHandleTable::Release(ClockHandle* h, bool useful,
       delete h;
       detached_usage_.fetch_sub(total_charge, std::memory_order_relaxed);
     } else {
-      uint32_t hash = h->hash;
+      UniqueId64x2 hashed_key = h->hashed_key;
 #ifndef NDEBUG
       // Mark slot as empty, with assertion
       old_meta = h->meta.exchange(0, std::memory_order_release);
@@ -607,7 +614,7 @@ bool ClockHandleTable::Release(ClockHandle* h, bool useful,
       h->meta.store(0, std::memory_order_release);
 #endif
       occupancy_.fetch_sub(1U, std::memory_order_release);
-      Rollback(hash, h);
+      Rollback(hashed_key, h);
     }
     usage_.fetch_sub(total_charge, std::memory_order_relaxed);
     assert(usage_.load(std::memory_order_relaxed) < SIZE_MAX / 2);
@@ -654,10 +661,10 @@ void ClockHandleTable::TEST_ReleaseN(ClockHandle* h, size_t n) {
   }
 }
 
-void ClockHandleTable::Erase(const CacheKeyBytes& key, uint32_t hash) {
-  uint32_t probe = 0;
+void ClockHandleTable::Erase(const UniqueId64x2& hashed_key) {
+  size_t probe = 0;
   (void)FindSlot(
-      hash,
+      hashed_key,
       [&](ClockHandle* h) {
         // Could be multiple entries in rare cases. Erase them all.
         // Optimistically increment acquire counter
@@ -667,7 +674,7 @@ void ClockHandleTable::Erase(const CacheKeyBytes& key, uint32_t hash) {
         if ((old_meta >> ClockHandle::kStateShift) ==
             ClockHandle::kStateVisible) {
           // Acquired a read reference
-          if (h->key == key) {
+          if (h->hashed_key == hashed_key) {
             // Match. Set invisible.
             old_meta =
                 h->meta.fetch_and(~(uint64_t{ClockHandle::kStateVisibleBit}
@@ -691,7 +698,7 @@ void ClockHandleTable::Erase(const CacheKeyBytes& key, uint32_t hash) {
                                  << ClockHandle::kStateShift,
                              std::memory_order_acq_rel)) {
                 // Took ownership
-                assert(hash == h->hash);
+                assert(hashed_key == h->hashed_key);
                 // TODO? Delay freeing?
                 h->FreeData();
                 usage_.fetch_sub(h->total_charge, std::memory_order_relaxed);
@@ -706,7 +713,7 @@ void ClockHandleTable::Erase(const CacheKeyBytes& key, uint32_t hash) {
                 h->meta.store(0, std::memory_order_release);
 #endif
                 occupancy_.fetch_sub(1U, std::memory_order_release);
-                Rollback(hash, h);
+                Rollback(hashed_key, h);
                 break;
               }
             }
@@ -735,14 +742,14 @@ void ClockHandleTable::Erase(const CacheKeyBytes& key, uint32_t hash) {
 }
 
 void ClockHandleTable::ConstApplyToEntriesRange(
-    std::function<void(const ClockHandle&)> func, uint32_t index_begin,
-    uint32_t index_end, bool apply_if_will_be_deleted) const {
+    std::function<void(const ClockHandle&)> func, size_t index_begin,
+    size_t index_end, bool apply_if_will_be_deleted) const {
   uint64_t check_state_mask = ClockHandle::kStateShareableBit;
   if (!apply_if_will_be_deleted) {
     check_state_mask |= ClockHandle::kStateVisibleBit;
   }
 
-  for (uint32_t i = index_begin; i < index_end; i++) {
+  for (size_t i = index_begin; i < index_end; i++) {
     ClockHandle& h = array_[i];
 
     // Note: to avoid using compare_exchange, we have to be extra careful.
@@ -776,7 +783,7 @@ void ClockHandleTable::ConstApplyToEntriesRange(
 }
 
 void ClockHandleTable::EraseUnRefEntries() {
-  for (uint32_t i = 0; i <= this->length_bits_mask_; i++) {
+  for (size_t i = 0; i <= this->length_bits_mask_; i++) {
     ClockHandle& h = array_[i];
 
     uint64_t old_meta = h.meta.load(std::memory_order_relaxed);
@@ -788,7 +795,7 @@ void ClockHandleTable::EraseUnRefEntries() {
                                            << ClockHandle::kStateShift,
                                        std::memory_order_acquire)) {
       // Took ownership
-      uint32_t hash = h.hash;
+      UniqueId64x2 hashed_key = h.hashed_key;
       h.FreeData();
       usage_.fetch_sub(h.total_charge, std::memory_order_relaxed);
 #ifndef NDEBUG
@@ -801,37 +808,29 @@ void ClockHandleTable::EraseUnRefEntries() {
       h.meta.store(0, std::memory_order_release);
 #endif
       occupancy_.fetch_sub(1U, std::memory_order_release);
-      Rollback(hash, &h);
+      Rollback(hashed_key, &h);
     }
   }
 }
 
-namespace {
-inline uint32_t Remix1(uint32_t hash) {
-  return Lower32of64((uint64_t{hash} * 0xbc9f1d35) >> 29);
-}
-
-inline uint32_t Remix2(uint32_t hash) {
-  return Lower32of64((uint64_t{hash} * 0x7a2bb9d5) >> 29);
-}
-}  // namespace
-
 ClockHandle* ClockHandleTable::FindSlot(
-    uint32_t hash, std::function<bool(ClockHandle*)> match_fn,
+    const UniqueId64x2& hashed_key, std::function<bool(ClockHandle*)> match_fn,
     std::function<bool(ClockHandle*)> abort_fn,
-    std::function<void(ClockHandle*)> update_fn, uint32_t& probe) {
+    std::function<void(ClockHandle*)> update_fn, size_t& probe) {
+  // NOTE: upper 32 bits of hashed_key[0] is used for sharding
+  //
   // We use double-hashing probing. Every probe in the sequence is a
   // pseudorandom integer, computed as a linear function of two random hashes,
   // which we call base and increment. Specifically, the i-th probe is base + i
   // * increment modulo the table size.
-  uint32_t base = ModTableSize(Remix1(hash));
+  size_t base = static_cast<size_t>(hashed_key[1]);
   // We use an odd increment, which is relatively prime with the power-of-two
   // table size. This implies that we cycle back to the first probe only
   // after probing every slot exactly once.
   // TODO: we could also reconsider linear probing, though locality benefits
   // are limited because each slot is a full cache line
-  uint32_t increment = Remix2(hash) | 1U;
-  uint32_t current = ModTableSize(base + probe * increment);
+  size_t increment = static_cast<size_t>(hashed_key[0]) | 1U;
+  size_t current = ModTableSize(base + probe * increment);
   while (probe <= length_bits_mask_) {
     ClockHandle* h = &array_[current];
     if (match_fn(h)) {
@@ -849,22 +848,23 @@ ClockHandle* ClockHandleTable::FindSlot(
   return nullptr;
 }
 
-void ClockHandleTable::Rollback(uint32_t hash, const ClockHandle* h) {
-  uint32_t current = ModTableSize(Remix1(hash));
-  uint32_t increment = Remix2(hash) | 1U;
-  for (uint32_t i = 0; &array_[current] != h; i++) {
+void ClockHandleTable::Rollback(const UniqueId64x2& hashed_key,
+                                const ClockHandle* h) {
+  size_t current = ModTableSize(hashed_key[1]);
+  size_t increment = static_cast<size_t>(hashed_key[0]) | 1U;
+  for (size_t i = 0; &array_[current] != h; i++) {
     array_[current].displacements.fetch_sub(1, std::memory_order_relaxed);
     current = ModTableSize(current + increment);
   }
 }
 
 void ClockHandleTable::Evict(size_t requested_charge, size_t* freed_charge,
-                             uint32_t* freed_count) {
+                             size_t* freed_count) {
   // precondition
   assert(requested_charge > 0);
 
   // TODO: make a tuning parameter?
-  constexpr uint32_t step_size = 4;
+  constexpr size_t step_size = 4;
 
   // First (concurrent) increment clock pointer
   uint64_t old_clock_pointer =
@@ -879,7 +879,7 @@ void ClockHandleTable::Evict(size_t requested_charge, size_t* freed_charge,
       old_clock_pointer + (ClockHandle::kMaxCountdown << length_bits_);
 
   for (;;) {
-    for (uint32_t i = 0; i < step_size; i++) {
+    for (size_t i = 0; i < step_size; i++) {
       ClockHandle& h = array_[ModTableSize(Lower32of64(old_clock_pointer + i))];
       uint64_t meta = h.meta.load(std::memory_order_relaxed);
 
@@ -919,11 +919,13 @@ void ClockHandleTable::Evict(size_t requested_charge, size_t* freed_charge,
               uint64_t{ClockHandle::kStateConstruction}
                   << ClockHandle::kStateShift,
               std::memory_order_acquire)) {
-        // Took ownership
-        uint32_t hash = h.hash;
+        // Took ownership.
+        // Save info about h to minimize dependences between atomic updates
+        // (e.g. fully relaxed Rollback after h released by marking empty)
+        const UniqueId64x2 h_hashed_key = h.hashed_key;
+        size_t h_total_charge = h.total_charge;
         // TODO? Delay freeing?
         h.FreeData();
-        *freed_charge += h.total_charge;
 #ifndef NDEBUG
         // Mark slot as empty, with assertion
         meta = h.meta.exchange(0, std::memory_order_release);
@@ -934,7 +936,8 @@ void ClockHandleTable::Evict(size_t requested_charge, size_t* freed_charge,
         h.meta.store(0, std::memory_order_release);
 #endif
         *freed_count += 1;
-        Rollback(hash, &h);
+        *freed_charge += h_total_charge;
+        Rollback(h_hashed_key, &h);
       }
     }
 
@@ -955,7 +958,7 @@ void ClockHandleTable::Evict(size_t requested_charge, size_t* freed_charge,
 ClockCacheShard::ClockCacheShard(
     size_t capacity, size_t estimated_value_size, bool strict_capacity_limit,
     CacheMetadataChargePolicy metadata_charge_policy)
-    : CacheShard(metadata_charge_policy),
+    : CacheShardBase(metadata_charge_policy),
       table_(
           CalcHashBits(capacity, estimated_value_size, metadata_charge_policy),
           /*initial_charge_metadata*/ metadata_charge_policy ==
@@ -971,31 +974,33 @@ void ClockCacheShard::EraseUnRefEntries() { table_.EraseUnRefEntries(); }
 void ClockCacheShard::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.
-  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_.ConstApplyToEntriesRange(
       [callback](const ClockHandle& h) {
-        callback(h.KeySlice(), h.value, h.total_charge, h.deleter);
+        UniqueId64x2 unhashed;
+        callback(ReverseHash(h.hashed_key, &unhashed), h.value, h.total_charge,
+                 h.deleter);
       },
       index_begin, index_end, false);
 }
@@ -1011,7 +1016,7 @@ int ClockCacheShard::CalcHashBits(
   uint64_t num_slots =
       static_cast<uint64_t>(capacity / average_slot_charge + 0.999999);
 
-  int hash_bits = std::min(FloorLog2((num_slots << 1) - 1), 32);
+  int hash_bits = FloorLog2((num_slots << 1) - 1);
   if (metadata_charge_policy == kFullChargeCacheMetadata) {
     // For very small estimated value sizes, it's possible to overshoot
     while (hash_bits > 0 &&
@@ -1033,17 +1038,16 @@ void ClockCacheShard::SetStrictCapacityLimit(bool strict_capacity_limit) {
   // next Insert will take care of any necessary evictions
 }
 
-Status ClockCacheShard::Insert(const Slice& key, uint32_t hash, void* value,
-                               size_t charge, Cache::DeleterFn deleter,
-                               Cache::Handle** handle,
+Status ClockCacheShard::Insert(const Slice& key, const UniqueId64x2& hashed_key,
+                               void* value, size_t charge,
+                               Cache::DeleterFn deleter, ClockHandle** handle,
                                Cache::Priority priority) {
   if (UNLIKELY(key.size() != kCacheKeySize)) {
     return Status::NotSupported("ClockCache only supports key size " +
                                 std::to_string(kCacheKeySize) + "B");
   }
-  ClockHandleMoreData proto;
-  proto.key = *reinterpret_cast<const CacheKeyBytes*>(key.data());
-  proto.hash = hash;
+  ClockHandleBasicData proto;
+  proto.hashed_key = hashed_key;
   proto.value = value;
   proto.deleter = deleter;
   proto.total_charge = charge;
@@ -1054,49 +1058,47 @@ Status ClockCacheShard::Insert(const Slice& key, uint32_t hash, void* value,
   return s;
 }
 
-Cache::Handle* ClockCacheShard::Lookup(const Slice& key, uint32_t hash) {
+ClockHandle* ClockCacheShard::Lookup(const Slice& key,
+                                     const UniqueId64x2& hashed_key) {
   if (UNLIKELY(key.size() != kCacheKeySize)) {
     return nullptr;
   }
-  auto key_bytes = reinterpret_cast<const CacheKeyBytes*>(key.data());
-  return reinterpret_cast<Cache::Handle*>(table_.Lookup(*key_bytes, hash));
+  return table_.Lookup(hashed_key);
 }
 
-bool ClockCacheShard::Ref(Cache::Handle* h) {
+bool ClockCacheShard::Ref(ClockHandle* h) {
   if (h == nullptr) {
     return false;
   }
-  table_.Ref(*reinterpret_cast<ClockHandle*>(h));
+  table_.Ref(*h);
   return true;
 }
 
-bool ClockCacheShard::Release(Cache::Handle* handle, bool useful,
+bool ClockCacheShard::Release(ClockHandle* handle, bool useful,
                               bool erase_if_last_ref) {
   if (handle == nullptr) {
     return false;
   }
-  return table_.Release(reinterpret_cast<ClockHandle*>(handle), useful,
-                        erase_if_last_ref);
+  return table_.Release(handle, useful, erase_if_last_ref);
 }
 
-void ClockCacheShard::TEST_RefN(Cache::Handle* h, size_t n) {
-  table_.TEST_RefN(*reinterpret_cast<ClockHandle*>(h), n);
+void ClockCacheShard::TEST_RefN(ClockHandle* h, size_t n) {
+  table_.TEST_RefN(*h, n);
 }
 
-void ClockCacheShard::TEST_ReleaseN(Cache::Handle* h, size_t n) {
-  table_.TEST_ReleaseN(reinterpret_cast<ClockHandle*>(h), n);
+void ClockCacheShard::TEST_ReleaseN(ClockHandle* h, size_t n) {
+  table_.TEST_ReleaseN(h, n);
 }
 
-bool ClockCacheShard::Release(Cache::Handle* handle, bool erase_if_last_ref) {
+bool ClockCacheShard::Release(ClockHandle* handle, bool erase_if_last_ref) {
   return Release(handle, /*useful=*/true, erase_if_last_ref);
 }
 
-void ClockCacheShard::Erase(const Slice& key, uint32_t hash) {
+void ClockCacheShard::Erase(const Slice& key, const UniqueId64x2& hashed_key) {
   if (UNLIKELY(key.size() != kCacheKeySize)) {
     return;
   }
-  auto key_bytes = reinterpret_cast<const CacheKeyBytes*>(key.data());
-  table_.Erase(*key_bytes, hash);
+  table_.Erase(hashed_key);
 }
 
 size_t ClockCacheShard::GetUsage() const { return table_.GetUsage(); }
@@ -1140,39 +1142,19 @@ size_t ClockCacheShard::GetTableAddressCount() const {
 HyperClockCache::HyperClockCache(
     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),
-      num_shards_(1 << num_shard_bits) {
+    CacheMetadataChargePolicy metadata_charge_policy,
+    std::shared_ptr<MemoryAllocator> memory_allocator)
+    : ShardedCache(capacity, num_shard_bits, strict_capacity_limit,
+                   std::move(memory_allocator)) {
   assert(estimated_value_size > 0 ||
          metadata_charge_policy != kDontChargeCacheMetadata);
   // TODO: should not need to go through two levels of pointer indirection to
   // get to table entries
-  shards_ = reinterpret_cast<ClockCacheShard*>(
-      port::cacheline_aligned_alloc(sizeof(ClockCacheShard) * num_shards_));
-  size_t per_shard = (capacity + (num_shards_ - 1)) / num_shards_;
-  for (int i = 0; i < num_shards_; i++) {
-    new (&shards_[i])
-        ClockCacheShard(per_shard, estimated_value_size, strict_capacity_limit,
-                        metadata_charge_policy);
-  }
-}
-
-HyperClockCache::~HyperClockCache() {
-  if (shards_ != nullptr) {
-    assert(num_shards_ > 0);
-    for (int i = 0; i < num_shards_; i++) {
-      shards_[i].~ClockCacheShard();
-    }
-    port::cacheline_aligned_free(shards_);
-  }
-}
-
-CacheShard* HyperClockCache::GetShard(uint32_t shard) {
-  return reinterpret_cast<CacheShard*>(&shards_[shard]);
-}
-
-const CacheShard* HyperClockCache::GetShard(uint32_t shard) const {
-  return reinterpret_cast<CacheShard*>(&shards_[shard]);
+  size_t per_shard = GetPerShardCapacity();
+  InitShards([=](ClockCacheShard* cs) {
+    new (cs) ClockCacheShard(per_shard, estimated_value_size,
+                             strict_capacity_limit, metadata_charge_policy);
+  });
 }
 
 void* HyperClockCache::Value(Handle* handle) {
@@ -1188,18 +1170,6 @@ Cache::DeleterFn HyperClockCache::GetDeleter(Handle* handle) const {
   return h->deleter;
 }
 
-uint32_t HyperClockCache::GetHash(Handle* handle) const {
-  return reinterpret_cast<const ClockHandle*>(handle)->hash;
-}
-
-void HyperClockCache::DisownData() {
-  // Leak data only if that won't generate an ASAN/valgrind warning.
-  if (!kMustFreeHeapAllocations) {
-    shards_ = nullptr;
-    num_shards_ = 0;
-  }
-}
-
 }  // namespace hyper_clock_cache
 
 // DEPRECATED (see public API)
@@ -1225,7 +1195,7 @@ std::shared_ptr<Cache> HyperClockCacheOptions::MakeSharedCache() const {
   }
   return std::make_shared<hyper_clock_cache::HyperClockCache>(
       capacity, estimated_entry_charge, my_num_shard_bits,
-      strict_capacity_limit, metadata_charge_policy);
+      strict_capacity_limit, metadata_charge_policy, memory_allocator);
 }
 
 }  // namespace ROCKSDB_NAMESPACE

cache/clock_cache.h

@@ -303,30 +303,24 @@ constexpr double kLoadFactor = 0.7;
 // strict upper bound on the load factor.
 constexpr double kStrictLoadFactor = 0.84;
 
-using CacheKeyBytes = std::array<char, kCacheKeySize>;
-
 struct ClockHandleBasicData {
   void* value = nullptr;
   Cache::DeleterFn deleter = nullptr;
-  CacheKeyBytes key = {};
+  // A lossless, reversible hash of the fixed-size (16 byte) cache key. This
+  // eliminates the need to store a hash separately.
+  UniqueId64x2 hashed_key = kNullUniqueId64x2;
   size_t total_charge = 0;
 
-  Slice KeySlice() const { return Slice(key.data(), kCacheKeySize); }
+  // Calls deleter (if non-null) on cache key and value
+  void FreeData() const;
 
-  void FreeData() const {
-    if (deleter) {
-      (*deleter)(KeySlice(), value);
-    }
-  }
-};
-
-struct ClockHandleMoreData : public ClockHandleBasicData {
-  uint32_t hash = 0;
+  // Required by concept HandleImpl
+  const UniqueId64x2& GetHash() const { return hashed_key; }
 };
 
 // Target size to be exactly a common cache line size (see static_assert in
 // clock_cache.cc)
-struct ALIGN_AS(64U) ClockHandle : public ClockHandleMoreData {
+struct ALIGN_AS(64U) ClockHandle : public ClockHandleBasicData {
   // Constants for handling the atomic `meta` word, which tracks most of the
   // state of the handle. The meta word looks like this:
   // low bits                                                     high bits
@@ -391,31 +385,31 @@ class ClockHandleTable {
   explicit ClockHandleTable(int hash_bits, bool initial_charge_metadata);
   ~ClockHandleTable();
 
-  Status Insert(const ClockHandleMoreData& proto, ClockHandle** handle,
+  Status Insert(const ClockHandleBasicData& proto, ClockHandle** handle,
                 Cache::Priority priority, size_t capacity,
                 bool strict_capacity_limit);
 
-  ClockHandle* Lookup(const CacheKeyBytes& key, uint32_t hash);
+  ClockHandle* Lookup(const UniqueId64x2& hashed_key);
 
   bool Release(ClockHandle* handle, bool useful, bool erase_if_last_ref);
 
   void Ref(ClockHandle& handle);
 
-  void Erase(const CacheKeyBytes& key, uint32_t hash);
+  void Erase(const UniqueId64x2& hashed_key);
 
   void ConstApplyToEntriesRange(std::function<void(const ClockHandle&)> func,
-                                uint32_t index_begin, uint32_t index_end,
+                                size_t index_begin, size_t index_end,
                                 bool apply_if_will_be_deleted) const;
 
   void EraseUnRefEntries();
 
-  uint32_t GetTableSize() const { return uint32_t{1} << length_bits_; }
+  size_t GetTableSize() const { return size_t{1} << length_bits_; }
 
   int GetLengthBits() const { return length_bits_; }
 
-  uint32_t GetOccupancyLimit() const { return occupancy_limit_; }
+  size_t GetOccupancyLimit() const { return occupancy_limit_; }
 
-  uint32_t GetOccupancy() const {
+  size_t GetOccupancy() const {
     return occupancy_.load(std::memory_order_relaxed);
   }
 
@@ -431,13 +425,15 @@ class ClockHandleTable {
 
  private:  // functions
   // Returns x mod 2^{length_bits_}.
-  uint32_t ModTableSize(uint32_t x) { return x & length_bits_mask_; }
+  inline size_t ModTableSize(uint64_t x) {
+    return static_cast<size_t>(x) & length_bits_mask_;
+  }
 
   // Runs the clock eviction algorithm trying to reclaim at least
   // requested_charge. Returns how much is evicted, which could be less
   // if it appears impossible to evict the requested amount without blocking.
   void Evict(size_t requested_charge, size_t* freed_charge,
-             uint32_t* freed_count);
+             size_t* freed_count);
 
   // Returns the first slot in the probe sequence, starting from the given
   // probe number, with a handle e such that match(e) is true. At every
@@ -450,15 +446,15 @@ class ClockHandleTable {
   // value of probe is one more than the last non-aborting probe during the
   // call. This is so that that the variable can be used to keep track of
   // progress across consecutive calls to FindSlot.
-  inline ClockHandle* FindSlot(uint32_t hash,
+  inline ClockHandle* FindSlot(const UniqueId64x2& hashed_key,
                                std::function<bool(ClockHandle*)> match,
                                std::function<bool(ClockHandle*)> stop,
                                std::function<void(ClockHandle*)> update,
-                               uint32_t& probe);
+                               size_t& probe);
 
   // Re-decrement all displacements in probe path starting from beginning
   // until (not including) the given handle
-  void Rollback(uint32_t hash, const ClockHandle* h);
+  void Rollback(const UniqueId64x2& hashed_key, const ClockHandle* h);
 
  private:  // data
   // Number of hash bits used for table index.
@@ -466,10 +462,10 @@ class ClockHandleTable {
   const int length_bits_;
 
   // For faster computation of ModTableSize.
-  const uint32_t length_bits_mask_;
+  const size_t length_bits_mask_;
 
   // Maximum number of elements the user can store in the table.
-  const uint32_t occupancy_limit_;
+  const size_t occupancy_limit_;
 
   // Array of slots comprising the hash table.
   const std::unique_ptr<ClockHandle[]> array_;
@@ -484,7 +480,7 @@ class ClockHandleTable {
 
   ALIGN_AS(CACHE_LINE_SIZE)
   // Number of elements in the table.
-  std::atomic<uint32_t> occupancy_{};
+  std::atomic<size_t> occupancy_{};
 
   // Memory usage by entries tracked by the cache (including detached)
   std::atomic<size_t> usage_{};
@@ -494,78 +490,107 @@ class ClockHandleTable {
 };  // class ClockHandleTable
 
 // A single shard of sharded cache.
-class ALIGN_AS(CACHE_LINE_SIZE) ClockCacheShard final : public CacheShard {
+class ALIGN_AS(CACHE_LINE_SIZE) ClockCacheShard final : public CacheShardBase {
  public:
   ClockCacheShard(size_t capacity, size_t estimated_value_size,
                   bool strict_capacity_limit,
                   CacheMetadataChargePolicy metadata_charge_policy);
-  ~ClockCacheShard() override = default;
 
-  // TODO: document limitations
-  void SetCapacity(size_t capacity) override;
+  // For CacheShard concept
+  using HandleImpl = ClockHandle;
+  // Hash is lossless hash of 128-bit key
+  using HashVal = UniqueId64x2;
+  using HashCref = const HashVal&;
+  static inline uint32_t HashPieceForSharding(HashCref hash) {
+    return Upper32of64(hash[0]);
+  }
+  static inline HashVal ComputeHash(const Slice& key) {
+    assert(key.size() == kCacheKeySize);
+    HashVal in;
+    HashVal out;
+    // NOTE: endian dependence
+    // TODO: use GetUnaligned?
+    std::memcpy(&in, key.data(), kCacheKeySize);
+    BijectiveHash2x64(in[1], in[0], &out[1], &out[0]);
+    return out;
+  }
 
-  void SetStrictCapacityLimit(bool strict_capacity_limit) override;
+  // For reconstructing key from hashed_key. Requires the caller to provide
+  // backing storage for the Slice in `unhashed`
+  static inline Slice ReverseHash(const UniqueId64x2& hashed,
+                                  UniqueId64x2* unhashed) {
+    BijectiveUnhash2x64(hashed[1], hashed[0], &(*unhashed)[1], &(*unhashed)[0]);
+    // NOTE: endian dependence
+    return Slice(reinterpret_cast<const char*>(unhashed), kCacheKeySize);
+  }
 
-  Status Insert(const Slice& key, uint32_t hash, void* value, size_t charge,
-                Cache::DeleterFn deleter, Cache::Handle** handle,
-                Cache::Priority priority) override;
+  // Although capacity is dynamically changeable, the number of table slots is
+  // not, so growing capacity substantially could lead to hitting occupancy
+  // limit.
+  void SetCapacity(size_t capacity);
 
-  Cache::Handle* Lookup(const Slice& key, uint32_t hash) override;
+  void SetStrictCapacityLimit(bool strict_capacity_limit);
 
-  bool Release(Cache::Handle* handle, bool useful,
-               bool erase_if_last_ref) override;
+  Status Insert(const Slice& key, const UniqueId64x2& hashed_key, void* value,
+                size_t charge, Cache::DeleterFn deleter, ClockHandle** handle,
+                Cache::Priority priority);
 
-  bool Release(Cache::Handle* handle, bool erase_if_last_ref = false) override;
+  ClockHandle* Lookup(const Slice& key, const UniqueId64x2& hashed_key);
 
-  bool Ref(Cache::Handle* handle) override;
+  bool Release(ClockHandle* handle, bool useful, bool erase_if_last_ref);
 
-  void Erase(const Slice& key, uint32_t hash) override;
+  bool Release(ClockHandle* handle, bool erase_if_last_ref = false);
 
-  size_t GetUsage() const override;
+  bool Ref(ClockHandle* handle);
 
-  size_t GetPinnedUsage() const override;
+  void Erase(const Slice& key, const UniqueId64x2& hashed_key);
 
-  size_t GetOccupancyCount() const override;
+  size_t GetUsage() const;
 
-  size_t GetTableAddressCount() const override;
+  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;
+  void EraseUnRefEntries();
 
-  std::string GetPrintableOptions() const override { return std::string{}; }
+  std::string GetPrintableOptions() const { return std::string{}; }
 
   // SecondaryCache not yet supported
-  Status Insert(const Slice& key, uint32_t hash, void* value,
+  Status Insert(const Slice& key, const UniqueId64x2& hashed_key, void* value,
                 const Cache::CacheItemHelper* helper, size_t charge,
-                Cache::Handle** handle, Cache::Priority priority) override {
-    return Insert(key, hash, value, charge, helper->del_cb, handle, priority);
+                ClockHandle** handle, Cache::Priority priority) {
+    return Insert(key, hashed_key, 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 {
-    return Lookup(key, hash);
+  ClockHandle* Lookup(const Slice& key, const UniqueId64x2& hashed_key,
+                      const Cache::CacheItemHelper* /*helper*/,
+                      const Cache::CreateCallback& /*create_cb*/,
+                      Cache::Priority /*priority*/, bool /*wait*/,
+                      Statistics* /*stats*/) {
+    return Lookup(key, hashed_key);
   }
 
-  bool IsReady(Cache::Handle* /*handle*/) override { return true; }
+  bool IsReady(ClockHandle* /*handle*/) { return true; }
 
-  void Wait(Cache::Handle* /*handle*/) override {}
+  void Wait(ClockHandle* /*handle*/) {}
 
   // Acquire/release N references
-  void TEST_RefN(Cache::Handle* handle, size_t n);
-  void TEST_ReleaseN(Cache::Handle* handle, size_t n);
+  void TEST_RefN(ClockHandle* handle, size_t n);
+  void TEST_ReleaseN(ClockHandle* handle, size_t n);
 
  private:  // functions
   friend class ClockCache;
   friend class ClockCacheTest;
 
-  ClockHandle* DetachedInsert(const ClockHandleMoreData& h);
+  ClockHandle* DetachedInsert(const ClockHandleBasicData& h);
 
   // Returns the number of bits used to hash an element in the hash
   // table.
@@ -586,35 +611,20 @@ class HyperClockCache
 #ifdef NDEBUG
     final
 #endif
-    : public ShardedCache {
+    : public ShardedCache<ClockCacheShard> {
  public:
   HyperClockCache(size_t capacity, size_t estimated_value_size,
                   int num_shard_bits, bool strict_capacity_limit,
-                  CacheMetadataChargePolicy metadata_charge_policy =
-                      kDontChargeCacheMetadata);
-
-  ~HyperClockCache() override;
+                  CacheMetadataChargePolicy metadata_charge_policy,
+                  std::shared_ptr<MemoryAllocator> memory_allocator);
 
   const char* Name() const override { return "HyperClockCache"; }
 
-  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:
-  ClockCacheShard* shards_ = nullptr;
-
-  int num_shards_;
 };  // class HyperClockCache
 
 }  // namespace hyper_clock_cache

cache/compressed_secondary_cache.cc

@@ -178,6 +178,19 @@ Status CompressedSecondaryCache::Insert(const Slice& key, void* value,
 
 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);
@@ -194,9 +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()};

cache/compressed_secondary_cache.h

@@ -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;
@@ -78,7 +79,7 @@ class CompressedSecondaryCache : public SecondaryCache {
       CompressionType compression_type = CompressionType::kLZ4Compression,
       uint32_t compress_format_version = 2,
       bool enable_custom_split_merge = false);
-  virtual ~CompressedSecondaryCache() override;
+  ~CompressedSecondaryCache() override;
 
   const char* Name() const override { return "CompressedSecondaryCache"; }
 
@@ -95,6 +96,10 @@ class CompressedSecondaryCache : public SecondaryCache {
 
   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:
@@ -116,7 +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
@@ -128,6 +133,7 @@ class CompressedSecondaryCache : public SecondaryCache {
   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

cache/compressed_secondary_cache_test.cc

@@ -5,28 +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 {
@@ -34,10 +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_;
@@ -50,7 +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);
@@ -315,12 +308,12 @@ class CompressedSecondaryCacheTest : public testing::Test {
     get_perf_context()->Reset();
     Random rnd(301);
     std::string str1 = rnd.RandomString(1001);
-    TestItem* item1_1 = new TestItem(str1.data(), str1.length());
+    auto item1_1 = new TestItem(str1.data(), str1.length());
     ASSERT_OK(cache->Insert(
         "k1", item1_1, &CompressedSecondaryCacheTest::helper_, str1.length()));
 
     std::string str2 = rnd.RandomString(1012);
-    TestItem* item2_1 = new TestItem(str2.data(), str2.length());
+    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(
@@ -330,7 +323,7 @@ class CompressedSecondaryCacheTest : public testing::Test {
     ASSERT_EQ(get_perf_context()->compressed_sec_cache_compressed_bytes, 0);
 
     std::string str3 = rnd.RandomString(1024);
-    TestItem* item3_1 = new TestItem(str3.data(), str3.length());
+    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(
@@ -339,14 +332,14 @@ class CompressedSecondaryCacheTest : public testing::Test {
 
     // After this Insert, primary cache contains k1 and secondary cache contains
     // k1's dummy item, k2's dummy item, and k3's dummy item.
-    TestItem* item1_2 = new TestItem(str1.data(), str1.length());
+    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.
-    TestItem* item2_2 = new TestItem(str2.data(), str2.length());
+    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);
@@ -362,7 +355,7 @@ class CompressedSecondaryCacheTest : public testing::Test {
 
     // After this Insert, primary cache contains k3 and secondary cache contains
     // k1's item and k2's item.
-    TestItem* item3_2 = new TestItem(str3.data(), str3.length());
+    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);
@@ -381,7 +374,7 @@ class CompressedSecondaryCacheTest : public testing::Test {
                            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_2->Buf(), item3_2->Size()), 0);
     cache->Release(handle);
@@ -399,18 +392,20 @@ class CompressedSecondaryCacheTest : public testing::Test {
                            stats.get());
     ASSERT_NE(handle, nullptr);
     ASSERT_EQ(get_perf_context()->block_cache_standalone_handle_count, 1);
-    TestItem* val1_1 = static_cast<TestItem*>(cache->Value(handle));
+    auto val1_1 = static_cast<TestItem*>(cache->Value(handle));
     ASSERT_NE(val1_1, nullptr);
     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.
@@ -421,6 +416,54 @@ class CompressedSecondaryCacheTest : public testing::Test {
     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();
   }
@@ -452,8 +495,7 @@ class CompressedSecondaryCacheTest : public testing::Test {
 
     Random rnd(301);
     std::string str1 = rnd.RandomString(1001);
-    auto item1 =
-        std::unique_ptr<TestItem>(new TestItem(str1.data(), str1.length()));
+    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_,
@@ -500,13 +542,13 @@ class CompressedSecondaryCacheTest : public testing::Test {
 
     Random rnd(301);
     std::string str1 = rnd.RandomString(1001);
-    TestItem* item1 = new TestItem(str1.data(), str1.length());
+    auto item1 = new TestItem(str1.data(), str1.length());
     ASSERT_OK(cache->Insert("k1", item1,
                             &CompressedSecondaryCacheTest::helper_fail_,
                             str1.length()));
 
     std::string str2 = rnd.RandomString(1002);
-    TestItem* item2 = new TestItem(str2.data(), str2.length());
+    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_,
@@ -559,12 +601,12 @@ class CompressedSecondaryCacheTest : public testing::Test {
 
     Random rnd(301);
     std::string str1 = rnd.RandomString(1001);
-    TestItem* item1 = new TestItem(str1.data(), str1.length());
+    auto item1 = new TestItem(str1.data(), str1.length());
     ASSERT_OK(cache->Insert("k1", item1, &CompressedSecondaryCacheTest::helper_,
                             str1.length()));
 
     std::string str2 = rnd.RandomString(1002);
-    TestItem* item2 = new TestItem(str2.data(), str2.length());
+    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()));
@@ -617,13 +659,13 @@ class CompressedSecondaryCacheTest : public testing::Test {
 
     Random rnd(301);
     std::string str1 = rnd.RandomString(1001);
-    TestItem* item1_1 = new TestItem(str1.data(), str1.length());
+    auto item1_1 = new TestItem(str1.data(), str1.length());
     ASSERT_OK(cache->Insert(
         "k1", item1_1, &CompressedSecondaryCacheTest::helper_, str1.length()));
 
     std::string str2 = rnd.RandomString(1002);
     std::string str2_clone{str2};
-    TestItem* item2 = new TestItem(str2.data(), str2.length());
+    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_,
@@ -631,11 +673,11 @@ class CompressedSecondaryCacheTest : public testing::Test {
 
     // After this Insert, primary cache contains k1 and secondary cache contains
     // k1's dummy item and k2's dummy item.
-    TestItem* item1_2 = new TestItem(str1.data(), str1.length());
+    auto item1_2 = new TestItem(str1.data(), str1.length());
     ASSERT_OK(cache->Insert(
         "k1", item1_2, &CompressedSecondaryCacheTest::helper_, str1.length()));
 
-    TestItem* item2_2 = new TestItem(str2.data(), str2.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(
@@ -957,6 +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();
 }

cache/fast_lru_cache.cc

@@ -173,7 +173,7 @@ 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)
-    : CacheShard(metadata_charge_policy),
+    : CacheShardBase(metadata_charge_policy),
       capacity_(capacity),
       strict_capacity_limit_(strict_capacity_limit),
       table_(
@@ -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;
   {
@@ -535,41 +532,18 @@ size_t LRUCacheShard::GetTableAddressCount() const {
   return table_.GetTableSize();
 }
 
-std::string LRUCacheShard::GetPrintableOptions() const { return std::string{}; }
-
 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) {
@@ -577,12 +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 {
@@ -590,18 +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(

cache/fast_lru_cache.h

@@ -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
@@ -337,48 +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 GetOccupancyCount() const override;
-  size_t GetTableAddressCount() 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;
@@ -446,25 +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
 

cache/lru_cache.cc

@@ -22,6 +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_] {}),
@@ -35,7 +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) {
@@ -110,12 +113,13 @@ 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)
-    : CacheShard(metadata_charge_policy),
+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),
@@ -162,27 +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(
@@ -361,7 +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;
@@ -411,7 +415,7 @@ Status LRUCacheShard::InsertItem(LRUHandle* e, Cache::Handle** handle,
         if (!e->HasRefs()) {
           e->Ref();
         }
-        *handle = reinterpret_cast<Cache::Handle*>(e);
+        *handle = e;
       }
     }
   }
@@ -425,16 +429,20 @@ void LRUCacheShard::Promote(LRUHandle* e) {
   SecondaryCacheResultHandle* secondary_handle = e->sec_handle;
 
   assert(secondary_handle->IsReady());
-  e->SetIncomplete(false);
-  e->SetInCache(false);
+  // 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;
 
   if (e->value) {
+    e->CalcTotalCharge(value_size, metadata_charge_policy_);
     Status s;
-    if (secondary_cache_ && secondary_cache_->SupportForceErase() &&
-        e->IsStandalone()) {
+    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;
@@ -464,15 +472,16 @@ void LRUCacheShard::Promote(LRUHandle* e) {
 
       // 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, /*value=*/nullptr, 0,
+      s = Insert(e->key(), e->hash, kDummyValueMarker, /*charge=*/0,
                  /*deleter=*/nullptr, /*helper=*/nullptr, /*handle=*/nullptr,
                  priority);
     } else {
       e->SetInCache(true);
-      e->SetIsStandalone(false);
-      Cache::Handle* handle = reinterpret_cast<Cache::Handle*>(e);
+      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
@@ -489,23 +498,19 @@ void LRUCacheShard::Promote(LRUHandle* e) {
       // 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);
-    e->SetIsStandalone(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};
   {
@@ -513,21 +518,22 @@ Cache::Handle* LRUCacheShard::Lookup(
     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);
-      }
-      e->Ref();
-      e->SetHit();
-
-      // 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.
-      if (!e->value && secondary_cache_ &&
-          secondary_cache_->SupportForceErase()) {
+      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();
       }
     }
   }
@@ -541,21 +547,13 @@ Cache::Handle* LRUCacheShard::Lookup(
   // 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.
-  //
-  // Only support synchronous for now.
-  // TODO: Support asynchronous lookup in secondary cache
-  if ((!e || found_dummy_entry) && secondary_cache_ && helper &&
-      helper->saveto_cb) {
+  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
     // a deleter would not be required is for dummy entries inserted for
     // accounting purposes, which we won't demote to the secondary cache
     // anyway.
     assert(create_cb && helper->del_cb);
-    // Release the dummy handle.
-    if (e) {
-      Release(reinterpret_cast<Cache::Handle*>(e), true /*erase_if_last_ref*/);
-    }
     bool is_in_sec_cache{false};
     std::unique_ptr<SecondaryCacheResultHandle> secondary_handle =
         secondary_cache_->Lookup(key, create_cb, wait, found_dummy_entry,
@@ -563,7 +561,8 @@ Cache::Handle* LRUCacheShard::Lookup(
     if (secondary_handle != nullptr) {
       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();
@@ -577,10 +576,8 @@ Cache::Handle* LRUCacheShard::Lookup(
       e->total_charge = 0;
       e->Ref();
       e->SetIsInSecondaryCache(is_in_sec_cache);
-
-      if (secondary_cache_->SupportForceErase() && !found_dummy_entry) {
-        e->SetIsStandalone(true);
-      }
+      e->SetIsStandalone(secondary_cache_->SupportForceErase() &&
+                         !found_dummy_entry);
 
       if (wait) {
         Promote(e);
@@ -598,24 +595,27 @@ Cache::Handle* LRUCacheShard::Lookup(
       } 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 {
-      e = nullptr;
+      // 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;
 }
@@ -634,12 +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();
@@ -657,12 +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;
     }
@@ -679,7 +678,7 @@ 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.
@@ -687,14 +686,17 @@ Status LRUCacheShard::Insert(const Slice& key, uint32_t hash, void* value,
       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();
@@ -735,9 +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_);
@@ -768,7 +768,7 @@ size_t LRUCacheShard::GetTableAddressCount() const {
   return size_t{1} << table_.GetLengthBits();
 }
 
-std::string LRUCacheShard::GetPrintableOptions() const {
+void LRUCacheShard::AppendPrintableOptions(std::string& str) const {
   const int kBufferSize = 200;
   char buffer[kBufferSize];
   {
@@ -778,7 +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,
@@ -787,51 +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 {
@@ -843,32 +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) {
@@ -894,22 +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
@@ -929,7 +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) {

cache/lru_cache.h

@@ -38,21 +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)
-// 4. The handle is never inserted into the LRUCache (both hash table and LRU
-//    list) and it doesn't experience the above three states.
-//    The entry can be freed when refs becomes 0.
-//    (refs >= 1 && in_cache == false && IS_STANDALONE == true)
-// All newly created LRUHandles are in state 1 or 4. 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;
@@ -60,7 +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.
@@ -77,45 +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),
-    // Whether this entry is not inserted into the cache (both hash table and
-    // LRU list).
-    IS_STANDALONE = (1 << 9),
+    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++; }
 
@@ -129,104 +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 IsStandalone() const { return flags & IS_STANDALONE; }
+  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) {
-      flags |= IS_STANDALONE;
+      im_flags |= IM_IS_STANDALONE;
     } else {
-      flags &= ~IS_STANDALONE;
+      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()) {
@@ -281,9 +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();
 
@@ -292,8 +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;
@@ -332,23 +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);
@@ -357,58 +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;
-  virtual size_t GetOccupancyCount() const override;
-  virtual size_t GetTableAddressCount() 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);
 
@@ -422,17 +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
@@ -519,14 +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,
@@ -534,27 +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_;
 };
 

cache/lru_cache_test.cc

@@ -67,7 +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;
@@ -529,22 +529,27 @@ class ClockCacheTest : public testing::Test {
                                  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 InsertWithLen(char key, size_t len) {
-    return Insert(std::string(len, key));
+    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, bool useful = true) {
-    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, useful, /*erase_if_last_ref=*/false);
       return true;
@@ -552,44 +557,29 @@ class ClockCacheTest : public testing::Test {
     return false;
   }
 
+  bool Lookup(const UniqueId64x2& hashed_key, bool useful = true) {
+    return Lookup(TestKey(hashed_key), hashed_key, useful);
+  }
+
   bool Lookup(char key, bool useful = true) {
-    return Lookup(std::string(kCacheKeySize, key), useful);
+    return Lookup(TestHashedKey(key), useful);
   }
 
-  void Erase(const std::string& key) { shard_->Erase(key, 0 /*hash*/); }
-
-#if 0  // FIXME
-  size_t CalcEstimatedHandleChargeWrapper(
-      size_t estimated_value_size,
-      CacheMetadataChargePolicy metadata_charge_policy) {
-    return ClockCacheShard::CalcEstimatedHandleCharge(estimated_value_size,
-                                                      metadata_charge_policy);
+  void Erase(char key) {
+    UniqueId64x2 hashed_key = TestHashedKey(key);
+    shard_->Erase(TestKey(hashed_key), hashed_key);
   }
 
-  int CalcHashBitsWrapper(size_t capacity, size_t estimated_value_size,
-                          CacheMetadataChargePolicy metadata_charge_policy) {
-    return ClockCacheShard::CalcHashBits(capacity, estimated_value_size,
-                                         metadata_charge_policy);
+  static inline Slice TestKey(const UniqueId64x2& hashed_key) {
+    return Slice(reinterpret_cast<const char*>(&hashed_key), 16U);
   }
 
-  // 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);
+  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};
   }
 
-  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);
-    }
-  }
-#endif
-
   ClockCacheShard* shard_ = nullptr;
 };
 
@@ -607,10 +597,10 @@ TEST_F(ClockCacheTest, Misc) {
 
   // Some of this is motivated by code coverage
   std::string wrong_size_key(15, 'x');
-  EXPECT_FALSE(Lookup(wrong_size_key));
+  EXPECT_FALSE(Lookup(wrong_size_key, TestHashedKey('x')));
   EXPECT_FALSE(shard_->Ref(nullptr));
   EXPECT_FALSE(shard_->Release(nullptr));
-  shard_->Erase(wrong_size_key, /*hash*/ 42);  // no-op
+  shard_->Erase(wrong_size_key, TestHashedKey('x'));  // no-op
 }
 
 TEST_F(ClockCacheTest, Limits) {
@@ -622,11 +612,11 @@ TEST_F(ClockCacheTest, Limits) {
     // Also tests switching between strict limit and not
     shard_->SetStrictCapacityLimit(strict_capacity_limit);
 
-    std::string key(16, 'x');
+    UniqueId64x2 hkey = TestHashedKey('x');
 
     // Single entry charge beyond capacity
     {
-      Status s = shard_->Insert(key, 0 /*hash*/, nullptr /*value*/,
+      Status s = shard_->Insert(TestKey(hkey), hkey, nullptr /*value*/,
                                 5 /*charge*/, nullptr /*deleter*/,
                                 nullptr /*handle*/, Cache::Priority::LOW);
       if (strict_capacity_limit) {
@@ -638,9 +628,10 @@ TEST_F(ClockCacheTest, Limits) {
 
     // Single entry fills capacity
     {
-      Cache::Handle* h;
-      ASSERT_OK(shard_->Insert(key, 0 /*hash*/, nullptr /*value*/, 3 /*charge*/,
-                               nullptr /*deleter*/, &h, Cache::Priority::LOW));
+      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) {
@@ -657,11 +648,11 @@ TEST_F(ClockCacheTest, Limits) {
     // entries) to exceed occupancy limit.
     {
       size_t n = shard_->GetTableAddressCount() + 1;
-      std::unique_ptr<Cache::Handle* []> ha { new Cache::Handle* [n] {} };
+      std::unique_ptr<ClockHandle* []> ha { new ClockHandle* [n] {} };
       Status s;
       for (size_t i = 0; i < n && s.ok(); ++i) {
-        EncodeFixed64(&key[0], i);
-        s = shard_->Insert(key, 0 /*hash*/, nullptr /*value*/, 0 /*charge*/,
+        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);
@@ -807,12 +798,11 @@ void IncrementIntDeleter(const Slice& /*key*/, void* value) {
 // Testing calls to CorrectNearOverflow in Release
 TEST_F(ClockCacheTest, ClockCounterOverflowTest) {
   NewShard(6, /*strict_capacity_limit*/ false);
-  Cache::Handle* h;
+  ClockHandle* h;
   int deleted = 0;
-  std::string my_key(kCacheKeySize, 'x');
-  uint32_t my_hash = 42;
-  ASSERT_OK(shard_->Insert(my_key, my_hash, &deleted, 1, IncrementIntDeleter,
-                           &h, Cache::Priority::HIGH));
+  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);
@@ -822,7 +812,7 @@ TEST_F(ClockCacheTest, ClockCounterOverflowTest) {
     shard_->TEST_ReleaseN(h, 1234567);
   }
   // Mark it invisible (to reach a different CorrectNearOverflow() in Release)
-  shard_->Erase(my_key, my_hash);
+  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) {
@@ -844,63 +834,65 @@ TEST_F(ClockCacheTest, ClockCounterOverflowTest) {
 TEST_F(ClockCacheTest, CollidingInsertEraseTest) {
   NewShard(6, /*strict_capacity_limit*/ false);
   int deleted = 0;
-  std::string key1(kCacheKeySize, 'x');
-  std::string key2(kCacheKeySize, 'y');
-  std::string key3(kCacheKeySize, 'z');
-  uint32_t my_hash = 42;
-  Cache::Handle* h1;
-  ASSERT_OK(shard_->Insert(key1, my_hash, &deleted, 1, IncrementIntDeleter, &h1,
+  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));
-  Cache::Handle* h2;
-  ASSERT_OK(shard_->Insert(key2, my_hash, &deleted, 1, IncrementIntDeleter, &h2,
+  ClockHandle* h2;
+  ASSERT_OK(shard_->Insert(key2, hkey2, &deleted, 1, IncrementIntDeleter, &h2,
                            Cache::Priority::HIGH));
-  Cache::Handle* h3;
-  ASSERT_OK(shard_->Insert(key3, my_hash, &deleted, 1, IncrementIntDeleter, &h3,
+  ClockHandle* h3;
+  ASSERT_OK(shard_->Insert(key3, hkey3, &deleted, 1, IncrementIntDeleter, &h3,
                            Cache::Priority::HIGH));
 
   // Can repeatedly lookup+release despite the hash collision
-  Cache::Handle* tmp_h;
+  ClockHandle* tmp_h;
   for (bool erase_if_last_ref : {true, false}) {  // but not last ref
-    tmp_h = shard_->Lookup(key1, my_hash);
+    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, my_hash);
+    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, my_hash);
+    tmp_h = shard_->Lookup(key3, hkey3);
     ASSERT_EQ(h3, tmp_h);
     ASSERT_FALSE(shard_->Release(tmp_h, erase_if_last_ref));
   }
 
   // Make h1 invisible
-  shard_->Erase(key1, my_hash);
+  shard_->Erase(key1, hkey1);
   // Redundant erase
-  shard_->Erase(key1, my_hash);
+  shard_->Erase(key1, hkey1);
 
   // All still alive
   ASSERT_EQ(deleted, 0);
 
   // Invisible to Lookup
-  tmp_h = shard_->Lookup(key1, my_hash);
+  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, my_hash);
+    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, my_hash);
+    tmp_h = shard_->Lookup(key3, hkey3);
     ASSERT_EQ(h3, tmp_h);
     ASSERT_FALSE(shard_->Release(tmp_h, erase_if_last_ref));
   }
 
   // Also Insert with invisible entry there
-  ASSERT_OK(shard_->Insert(key1, my_hash, &deleted, 1, IncrementIntDeleter,
+  ASSERT_OK(shard_->Insert(key1, hkey1, &deleted, 1, IncrementIntDeleter,
                            nullptr, Cache::Priority::HIGH));
-  tmp_h = shard_->Lookup(key1, my_hash);
+  tmp_h = shard_->Lookup(key1, hkey1);
   // Found but distinct handle
   ASSERT_NE(nullptr, tmp_h);
   ASSERT_NE(h1, tmp_h);
@@ -918,11 +910,11 @@ TEST_F(ClockCacheTest, CollidingInsertEraseTest) {
 
   // Can still find h2, h3
   for (bool erase_if_last_ref : {true, false}) {  // but not last ref
-    tmp_h = shard_->Lookup(key2, my_hash);
+    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, my_hash);
+    tmp_h = shard_->Lookup(key3, hkey3);
     ASSERT_EQ(h3, tmp_h);
     ASSERT_FALSE(shard_->Release(tmp_h, erase_if_last_ref));
   }
@@ -934,7 +926,7 @@ TEST_F(ClockCacheTest, CollidingInsertEraseTest) {
   ASSERT_EQ(deleted, 0);
 
   // Can still find it
-  tmp_h = shard_->Lookup(key2, my_hash);
+  tmp_h = shard_->Lookup(key2, hkey2);
   ASSERT_EQ(h2, tmp_h);
 
   // Release last ref on h2, with erase
@@ -942,12 +934,12 @@ TEST_F(ClockCacheTest, CollidingInsertEraseTest) {
 
   // h2 deleted
   ASSERT_EQ(deleted--, 1);
-  tmp_h = shard_->Lookup(key2, my_hash);
+  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, my_hash);
+    tmp_h = shard_->Lookup(key3, hkey3);
     ASSERT_EQ(h3, tmp_h);
     ASSERT_FALSE(shard_->Release(tmp_h, erase_if_last_ref));
   }
@@ -959,11 +951,11 @@ TEST_F(ClockCacheTest, CollidingInsertEraseTest) {
   ASSERT_EQ(deleted, 0);
 
   // Explicit erase
-  shard_->Erase(key3, my_hash);
+  shard_->Erase(key3, hkey3);
 
   // h3 deleted
   ASSERT_EQ(deleted--, 1);
-  tmp_h = shard_->Lookup(key3, my_hash);
+  tmp_h = shard_->Lookup(key3, hkey3);
   ASSERT_EQ(nullptr, tmp_h);
 }
 
@@ -1371,9 +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 =
@@ -2703,6 +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();
 }

cache/sharded_cache.cc

@@ -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);
@@ -210,7 +75,7 @@ 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;
 }
 
@@ -226,25 +91,10 @@ int GetDefaultCacheShardBits(size_t capacity, size_t min_shard_size) {
   return num_shard_bits;
 }
 
-int ShardedCache::GetNumShardBits() const { return BitsSetToOne(shard_mask_); }
-
-uint32_t ShardedCache::GetNumShards() const { return shard_mask_ + 1; }
-
-size_t ShardedCache::GetOccupancyCount() const {
-  size_t oc = 0;
-  uint32_t num_shards = GetNumShards();
-  for (uint32_t s = 0; s < num_shards; s++) {
-    oc += GetShard(s)->GetOccupancyCount();
-  }
-  return oc;
-}
-size_t ShardedCache::GetTableAddressCount() const {
-  size_t tac = 0;
-  uint32_t num_shards = GetNumShards();
-  for (uint32_t s = 0; s < num_shards; s++) {
-    tac += GetShard(s)->GetTableAddressCount();
-  }
-  return tac;
+int ShardedCacheBase::GetNumShardBits() const {
+  return BitsSetToOne(shard_mask_);
 }
 
+uint32_t ShardedCacheBase::GetNumShards() const { return shard_mask_ + 1; }
+
 }  // namespace ROCKSDB_NAMESPACE

cache/sharded_cache.h

@@ -10,122 +10,309 @@
 #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:
-  explicit CacheShard(CacheMetadataChargePolicy metadata_charge_policy)
+  explicit CacheShardBase(CacheMetadataChargePolicy metadata_charge_policy)
       : metadata_charge_policy_(metadata_charge_policy) {}
-  virtual ~CacheShard() = default;
 
   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;
-  virtual size_t GetOccupancyCount() const = 0;
-  virtual size_t GetTableAddressCount() 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 ""; }
+      size_t average_entries_per_lock, size_t* state) = 0;
+  void EraseUnRefEntries() = 0;
+  */
 
  protected:
   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 size_t GetOccupancyCount() const override;
-  virtual size_t GetTableAddressCount() 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;
 
+  uint64_t NewId() override;
+
+  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_;
+
+  // Dynamic configuration parameters, guarded by config_mutex_
+  bool strict_capacity_limit_;
+  size_t capacity_;
+  mutable port::Mutex config_mutex_;
+};
+
+// 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 uint32_t Shard(uint32_t hash) { return hash & shard_mask_; }
+  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:
-  const uint32_t shard_mask_;
-  mutable port::Mutex capacity_mutex_;
-  size_t capacity_;
-  bool strict_capacity_limit_;
-  std::atomic<uint64_t> last_id_;
+  CacheShard* const shards_;
+  bool destroy_shards_in_dtor_;
 };
 
 // 512KB is traditional minimum shard size.

coverage/parse_gcov_output.py

@@ -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()

db/blob/blob_counting_iterator_test.cc

@@ -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();
 }

db/blob/blob_file_addition_test.cc

@@ -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();
 }

db/blob/blob_file_builder_test.cc

@@ -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();
 }

db/blob/blob_file_cache_test.cc

@@ -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();
 }

db/blob/blob_file_garbage_test.cc

@@ -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();
 }

db/blob/blob_file_reader_test.cc

@@ -1018,6 +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();
 }

db/blob/blob_garbage_meter_test.cc

@@ -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();
 }

db/blob/blob_source_test.cc

@@ -1304,12 +1304,20 @@ TEST_F(BlobSecondaryCacheTest, GetBlobsFromSecondaryCache) {
                                                blob_offsets[1]));
 
       // 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() can only
-      // get a dummy handle.
+      // 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);
-      // handl1 is a dummy handle.
-      ASSERT_EQ(blob_cache->Value(handle1), nullptr);
+      ASSERT_NE(blob_cache->Value(handle1), nullptr);
       blob_cache->Release(handle1);
     }
   }
@@ -1610,6 +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();
 }

db/builder.cc

@@ -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) {

db/c.cc

@@ -6468,6 +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;

db/column_family.cc

@@ -291,7 +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();

db/compact_files_test.cc

@@ -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();
 }

db/compaction/clipping_iterator_test.cc

@@ -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();
 }

db/compaction/compaction.cc

@@ -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;
   }
 

db/compaction/compaction.h

@@ -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

db/compaction/compaction_iterator.cc

@@ -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_,
@@ -1092,7 +1106,7 @@ void CompactionIterator::DecideOutputLevel() {
 
   // 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;
   }
@@ -1154,7 +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(

db/compaction/compaction_iterator.h

@@ -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(); }
 

db/compaction/compaction_iterator_test.cc

@@ -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();
 }

db/compaction/compaction_job.cc

@@ -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,
@@ -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);
+      }
     }
   }
 }
@@ -1035,7 +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.
@@ -1213,8 +1254,8 @@ 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
@@ -1237,6 +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.
@@ -1455,7 +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");
@@ -1661,7 +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) {
@@ -1950,7 +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();
@@ -1971,6 +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_);
+      }
+    }
   }
 }
 

db/compaction/compaction_job.h

@@ -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_`.

db/compaction/compaction_job_test.cc

@@ -228,6 +228,9 @@ class CompactionJobTestBase : public testing::Test {
         test::CreateEnvFromSystem(ConfigOptions(), &base_env, &env_guard_));
     env_ = base_env;
     fs_ = env_->GetFileSystem();
+    // set default for the tests
+    mutable_cf_options_.target_file_size_base = 1024 * 1024;
+    mutable_cf_options_.max_compaction_bytes = 10 * 1024 * 1024;
   }
 
   void SetUp() override {
@@ -415,8 +418,9 @@ class CompactionJobTestBase : public testing::Test {
 
     auto cfd = versions_->GetColumnFamilySet()->GetDefault();
     if (table_type_ == TableTypeForTest::kMockTable) {
-      assert(expected_results.size() == 1);
-      mock_table_factory_->AssertLatestFile(expected_results[0]);
+      ASSERT_EQ(compaction_job_stats_.num_output_files,
+                expected_results.size());
+      mock_table_factory_->AssertLatestFiles(expected_results);
     } else {
       assert(table_type_ == TableTypeForTest::kBlockBasedTable);
     }
@@ -426,7 +430,8 @@ class CompactionJobTestBase : public testing::Test {
     ASSERT_EQ(expected_output_file_num, output_files.size());
 
     if (table_type_ == TableTypeForTest::kMockTable) {
-      assert(output_files.size() == 1);
+      assert(output_files.size() ==
+             static_cast<size_t>(expected_output_file_num));
       const FileMetaData* const output_file = output_files[0];
       ASSERT_EQ(output_file->oldest_blob_file_number,
                 expected_oldest_blob_file_numbers[0]);
@@ -620,12 +625,22 @@ class CompactionJobTestBase : public testing::Test {
       num_input_files += level_files.size();
     }
 
+    std::vector<FileMetaData*> grandparents;
+    // it should actually be the next non-empty level
+    const int kGrandparentsLevel = output_level + 1;
+    if (kGrandparentsLevel < cf_options_.num_levels) {
+      grandparents =
+          cfd_->current()->storage_info()->LevelFiles(kGrandparentsLevel);
+    }
+
     Compaction compaction(
         cfd->current()->storage_info(), *cfd->ioptions(),
         *cfd->GetLatestMutableCFOptions(), mutable_db_options_,
-        compaction_input_files, output_level, 1024 * 1024, 10 * 1024 * 1024, 0,
-        kNoCompression, cfd->GetLatestMutableCFOptions()->compression_opts,
-        Temperature::kUnknown, max_subcompactions, {}, true);
+        compaction_input_files, output_level,
+        mutable_cf_options_.target_file_size_base,
+        mutable_cf_options_.max_compaction_bytes, 0, kNoCompression,
+        cfd->GetLatestMutableCFOptions()->compression_opts,
+        Temperature::kUnknown, max_subcompactions, grandparents, true);
     compaction.SetInputVersion(cfd->current());
 
     assert(db_options_.info_log);
@@ -1721,6 +1736,392 @@ TEST_F(CompactionJobTest, ResultSerialization) {
   }
 }
 
+class CompactionJobDynamicFileSizeTest
+    : public CompactionJobTestBase,
+      public ::testing::WithParamInterface<bool> {
+ public:
+  CompactionJobDynamicFileSizeTest()
+      : CompactionJobTestBase(
+            test::PerThreadDBPath("compaction_job_dynamic_file_size_test"),
+            BytewiseComparator(), [](uint64_t /*ts*/) { return ""; },
+            /*test_io_priority=*/false, TableTypeForTest::kMockTable) {}
+};
+
+TEST_P(CompactionJobDynamicFileSizeTest, CutForMaxCompactionBytes) {
+  // dynamic_file_size option should have no impact on cutting for max
+  // compaction bytes.
+  bool enable_dyanmic_file_size = GetParam();
+  cf_options_.level_compaction_dynamic_file_size = enable_dyanmic_file_size;
+
+  NewDB();
+  mutable_cf_options_.target_file_size_base = 80;
+  mutable_cf_options_.max_compaction_bytes = 21;
+
+  auto file1 = mock::MakeMockFile({
+      {KeyStr("c", 5U, kTypeValue), "val2"},
+      {KeyStr("n", 6U, kTypeValue), "val3"},
+  });
+  AddMockFile(file1);
+
+  auto file2 = mock::MakeMockFile({{KeyStr("h", 3U, kTypeValue), "val"},
+                                   {KeyStr("j", 4U, kTypeValue), "val"}});
+  AddMockFile(file2, 1);
+
+  // Create three L2 files, each size 10.
+  // max_compaction_bytes 21 means the compaction output in L1 will
+  // be cut to at least two files.
+  auto file3 = mock::MakeMockFile({{KeyStr("b", 1U, kTypeValue), "val"},
+                                   {KeyStr("c", 1U, kTypeValue), "val"},
+                                   {KeyStr("c1", 1U, kTypeValue), "val"},
+                                   {KeyStr("c2", 1U, kTypeValue), "val"},
+                                   {KeyStr("c3", 1U, kTypeValue), "val"},
+                                   {KeyStr("c4", 1U, kTypeValue), "val"},
+                                   {KeyStr("d", 1U, kTypeValue), "val"},
+                                   {KeyStr("e", 2U, kTypeValue), "val"}});
+  AddMockFile(file3, 2);
+
+  auto file4 = mock::MakeMockFile({{KeyStr("h", 1U, kTypeValue), "val"},
+                                   {KeyStr("i", 1U, kTypeValue), "val"},
+                                   {KeyStr("i1", 1U, kTypeValue), "val"},
+                                   {KeyStr("i2", 1U, kTypeValue), "val"},
+                                   {KeyStr("i3", 1U, kTypeValue), "val"},
+                                   {KeyStr("i4", 1U, kTypeValue), "val"},
+                                   {KeyStr("j", 1U, kTypeValue), "val"},
+                                   {KeyStr("k", 2U, kTypeValue), "val"}});
+  AddMockFile(file4, 2);
+
+  auto file5 = mock::MakeMockFile({{KeyStr("l", 1U, kTypeValue), "val"},
+                                   {KeyStr("m", 1U, kTypeValue), "val"},
+                                   {KeyStr("m1", 1U, kTypeValue), "val"},
+                                   {KeyStr("m2", 1U, kTypeValue), "val"},
+                                   {KeyStr("m3", 1U, kTypeValue), "val"},
+                                   {KeyStr("m4", 1U, kTypeValue), "val"},
+                                   {KeyStr("n", 1U, kTypeValue), "val"},
+                                   {KeyStr("o", 2U, kTypeValue), "val"}});
+  AddMockFile(file5, 2);
+
+  // The expected output should be:
+  //  L1:   [c,   h,  j]        [n]
+  //  L2: [b ... e] [h ... k] [l ... o]
+  // It's better to have "j" in the first file, because anyway it's overlapping
+  // with the second file on L2.
+  // (Note: before this PR, it was cut at "h" because it's using the internal
+  // comparator which think L1 "h" with seqno 3 is smaller than L2 "h" with
+  // seqno 1, but actually they're overlapped with the compaction picker).
+
+  auto expected_file1 =
+      mock::MakeMockFile({{KeyStr("c", 5U, kTypeValue), "val2"},
+                          {KeyStr("h", 3U, kTypeValue), "val"},
+                          {KeyStr("j", 4U, kTypeValue), "val"}});
+  auto expected_file2 =
+      mock::MakeMockFile({{KeyStr("n", 6U, kTypeValue), "val3"}});
+
+  SetLastSequence(6U);
+
+  const std::vector<int> input_levels = {0, 1};
+  auto lvl0_files = cfd_->current()->storage_info()->LevelFiles(0);
+  auto lvl1_files = cfd_->current()->storage_info()->LevelFiles(1);
+
+  RunCompaction({lvl0_files, lvl1_files}, input_levels,
+                {expected_file1, expected_file2});
+}
+
+TEST_P(CompactionJobDynamicFileSizeTest, CutToSkipGrandparentFile) {
+  bool enable_dyanmic_file_size = GetParam();
+  cf_options_.level_compaction_dynamic_file_size = enable_dyanmic_file_size;
+
+  NewDB();
+  // Make sure the grandparent level file size (10) qualifies skipping.
+  // Currently, it has to be > 1/8 of target file size.
+  mutable_cf_options_.target_file_size_base = 70;
+
+  auto file1 = mock::MakeMockFile({
+      {KeyStr("a", 5U, kTypeValue), "val2"},
+      {KeyStr("z", 6U, kTypeValue), "val3"},
+  });
+  AddMockFile(file1);
+
+  auto file2 = mock::MakeMockFile({{KeyStr("c", 3U, kTypeValue), "val"},
+                                   {KeyStr("x", 4U, kTypeValue), "val"}});
+  AddMockFile(file2, 1);
+
+  auto file3 = mock::MakeMockFile({{KeyStr("b", 1U, kTypeValue), "val"},
+                                   {KeyStr("d", 2U, kTypeValue), "val"}});
+  AddMockFile(file3, 2);
+
+  auto file4 = mock::MakeMockFile({{KeyStr("h", 1U, kTypeValue), "val"},
+                                   {KeyStr("i", 2U, kTypeValue), "val"}});
+  AddMockFile(file4, 2);
+
+  auto file5 = mock::MakeMockFile({{KeyStr("v", 1U, kTypeValue), "val"},
+                                   {KeyStr("y", 2U, kTypeValue), "val"}});
+  AddMockFile(file5, 2);
+
+  auto expected_file1 =
+      mock::MakeMockFile({{KeyStr("a", 5U, kTypeValue), "val2"},
+                          {KeyStr("c", 3U, kTypeValue), "val"}});
+  auto expected_file2 =
+      mock::MakeMockFile({{KeyStr("x", 4U, kTypeValue), "val"},
+                          {KeyStr("z", 6U, kTypeValue), "val3"}});
+
+  auto expected_file_disable_dynamic_file_size =
+      mock::MakeMockFile({{KeyStr("a", 5U, kTypeValue), "val2"},
+                          {KeyStr("c", 3U, kTypeValue), "val"},
+                          {KeyStr("x", 4U, kTypeValue), "val"},
+                          {KeyStr("z", 6U, kTypeValue), "val3"}});
+
+  SetLastSequence(6U);
+  const std::vector<int> input_levels = {0, 1};
+  auto lvl0_files = cfd_->current()->storage_info()->LevelFiles(0);
+  auto lvl1_files = cfd_->current()->storage_info()->LevelFiles(1);
+  if (enable_dyanmic_file_size) {
+    RunCompaction({lvl0_files, lvl1_files}, input_levels,
+                  {expected_file1, expected_file2});
+  } else {
+    RunCompaction({lvl0_files, lvl1_files}, input_levels,
+                  {expected_file_disable_dynamic_file_size});
+  }
+}
+
+TEST_P(CompactionJobDynamicFileSizeTest, CutToAlignGrandparentBoundary) {
+  bool enable_dyanmic_file_size = GetParam();
+  cf_options_.level_compaction_dynamic_file_size = enable_dyanmic_file_size;
+  NewDB();
+
+  // MockTable has 1 byte per entry by default and each file is 10 bytes.
+  // When the file size is smaller than 100, it won't cut file earlier to align
+  // with its grandparent boundary.
+  const size_t kKeyValueSize = 10000;
+  mock_table_factory_->SetKeyValueSize(kKeyValueSize);
+
+  mutable_cf_options_.target_file_size_base = 10 * kKeyValueSize;
+
+  mock::KVVector file1;
+  char ch = 'd';
+  // Add value from d -> o
+  for (char i = 0; i < 12; i++) {
+    file1.emplace_back(KeyStr(std::string(1, ch + i), i + 10, kTypeValue),
+                       "val" + std::to_string(i));
+  }
+
+  AddMockFile(file1);
+
+  auto file2 = mock::MakeMockFile({{KeyStr("e", 3U, kTypeValue), "val"},
+                                   {KeyStr("s", 4U, kTypeValue), "val"}});
+  AddMockFile(file2, 1);
+
+  // the 1st grandparent file should be skipped
+  auto file3 = mock::MakeMockFile({{KeyStr("a", 1U, kTypeValue), "val"},
+                                   {KeyStr("b", 2U, kTypeValue), "val"}});
+  AddMockFile(file3, 2);
+
+  auto file4 = mock::MakeMockFile({{KeyStr("c", 1U, kTypeValue), "val"},
+                                   {KeyStr("e", 2U, kTypeValue), "val"}});
+  AddMockFile(file4, 2);
+
+  auto file5 = mock::MakeMockFile({{KeyStr("h", 1U, kTypeValue), "val"},
+                                   {KeyStr("j", 2U, kTypeValue), "val"}});
+  AddMockFile(file5, 2);
+
+  auto file6 = mock::MakeMockFile({{KeyStr("k", 1U, kTypeValue), "val"},
+                                   {KeyStr("n", 2U, kTypeValue), "val"}});
+  AddMockFile(file6, 2);
+
+  auto file7 = mock::MakeMockFile({{KeyStr("q", 1U, kTypeValue), "val"},
+                                   {KeyStr("t", 2U, kTypeValue), "val"}});
+  AddMockFile(file7, 2);
+
+  // The expected outputs are:
+  //  L1:         [d,e,f,g,h,i,j] [k,l,m,n,o,s]
+  //  L2: [a, b] [c,  e]   [h, j] [k, n]  [q, t]
+  // The first output cut earlier at "j", so it could be aligned with L2 files.
+  // If dynamic_file_size is not enabled, it will be cut based on the
+  // target_file_size
+  mock::KVVector expected_file1;
+  for (char i = 0; i < 7; i++) {
+    expected_file1.emplace_back(
+        KeyStr(std::string(1, ch + i), i + 10, kTypeValue),
+        "val" + std::to_string(i));
+  }
+
+  mock::KVVector expected_file2;
+  for (char i = 7; i < 12; i++) {
+    expected_file2.emplace_back(
+        KeyStr(std::string(1, ch + i), i + 10, kTypeValue),
+        "val" + std::to_string(i));
+  }
+  expected_file2.emplace_back(KeyStr("s", 4U, kTypeValue), "val");
+
+  mock::KVVector expected_file_disable_dynamic_file_size1;
+  for (char i = 0; i < 10; i++) {
+    expected_file_disable_dynamic_file_size1.emplace_back(
+        KeyStr(std::string(1, ch + i), i + 10, kTypeValue),
+        "val" + std::to_string(i));
+  }
+
+  mock::KVVector expected_file_disable_dynamic_file_size2;
+  for (char i = 10; i < 12; i++) {
+    expected_file_disable_dynamic_file_size2.emplace_back(
+        KeyStr(std::string(1, ch + i), i + 10, kTypeValue),
+        "val" + std::to_string(i));
+  }
+
+  expected_file_disable_dynamic_file_size2.emplace_back(
+      KeyStr("s", 4U, kTypeValue), "val");
+
+  SetLastSequence(22U);
+  const std::vector<int> input_levels = {0, 1};
+  auto lvl0_files = cfd_->current()->storage_info()->LevelFiles(0);
+  auto lvl1_files = cfd_->current()->storage_info()->LevelFiles(1);
+  if (enable_dyanmic_file_size) {
+    RunCompaction({lvl0_files, lvl1_files}, input_levels,
+                  {expected_file1, expected_file2});
+  } else {
+    RunCompaction({lvl0_files, lvl1_files}, input_levels,
+                  {expected_file_disable_dynamic_file_size1,
+                   expected_file_disable_dynamic_file_size2});
+  }
+}
+
+TEST_P(CompactionJobDynamicFileSizeTest, CutToAlignGrandparentBoundarySameKey) {
+  bool enable_dyanmic_file_size = GetParam();
+  cf_options_.level_compaction_dynamic_file_size = enable_dyanmic_file_size;
+  NewDB();
+
+  // MockTable has 1 byte per entry by default and each file is 10 bytes.
+  // When the file size is smaller than 100, it won't cut file earlier to align
+  // with its grandparent boundary.
+  const size_t kKeyValueSize = 10000;
+  mock_table_factory_->SetKeyValueSize(kKeyValueSize);
+
+  mutable_cf_options_.target_file_size_base = 10 * kKeyValueSize;
+
+  mock::KVVector file1;
+  for (int i = 0; i < 7; i++) {
+    file1.emplace_back(KeyStr("a", 100 - i, kTypeValue),
+                       "val" + std::to_string(100 - i));
+  }
+  file1.emplace_back(KeyStr("b", 90, kTypeValue), "valb");
+
+  AddMockFile(file1);
+
+  auto file2 = mock::MakeMockFile({{KeyStr("a", 93U, kTypeValue), "val93"},
+                                   {KeyStr("b", 90U, kTypeValue), "valb"}});
+  AddMockFile(file2, 1);
+
+  auto file3 = mock::MakeMockFile({{KeyStr("a", 89U, kTypeValue), "val"},
+                                   {KeyStr("a", 88U, kTypeValue), "val"}});
+  AddMockFile(file3, 2);
+
+  auto file4 = mock::MakeMockFile({{KeyStr("a", 87U, kTypeValue), "val"},
+                                   {KeyStr("a", 86U, kTypeValue), "val"}});
+  AddMockFile(file4, 2);
+
+  auto file5 = mock::MakeMockFile({{KeyStr("b", 85U, kTypeValue), "val"},
+                                   {KeyStr("b", 84U, kTypeValue), "val"}});
+  AddMockFile(file5, 2);
+
+  mock::KVVector expected_file1;
+  mock::KVVector expected_file_disable_dynamic_file_size;
+
+  for (int i = 0; i < 8; i++) {
+    expected_file1.emplace_back(KeyStr("a", 100 - i, kTypeValue),
+                                "val" + std::to_string(100 - i));
+    expected_file_disable_dynamic_file_size.emplace_back(
+        KeyStr("a", 100 - i, kTypeValue), "val" + std::to_string(100 - i));
+  }
+
+  // make sure `b` is cut in a separated file (so internally it's not using
+  // internal comparator, which will think the "b:90" (seqno 90) here is smaller
+  // than "b:85" on L2.)
+  auto expected_file2 =
+      mock::MakeMockFile({{KeyStr("b", 90U, kTypeValue), "valb"}});
+
+  expected_file_disable_dynamic_file_size.emplace_back(
+      KeyStr("b", 90U, kTypeValue), "valb");
+
+  SetLastSequence(122U);
+  const std::vector<int> input_levels = {0, 1};
+  auto lvl0_files = cfd_->current()->storage_info()->LevelFiles(0);
+  auto lvl1_files = cfd_->current()->storage_info()->LevelFiles(1);
+
+  // Just keep all the history
+  std::vector<SequenceNumber> snapshots;
+  for (int i = 80; i <= 100; i++) {
+    snapshots.emplace_back(i);
+  }
+  if (enable_dyanmic_file_size) {
+    RunCompaction({lvl0_files, lvl1_files}, input_levels,
+                  {expected_file1, expected_file2}, snapshots);
+  } else {
+    RunCompaction({lvl0_files, lvl1_files}, input_levels,
+                  {expected_file_disable_dynamic_file_size}, snapshots);
+  }
+}
+
+TEST_P(CompactionJobDynamicFileSizeTest, CutForMaxCompactionBytesSameKey) {
+  // dynamic_file_size option should have no impact on cutting for max
+  // compaction bytes.
+  bool enable_dyanmic_file_size = GetParam();
+  cf_options_.level_compaction_dynamic_file_size = enable_dyanmic_file_size;
+
+  NewDB();
+  mutable_cf_options_.target_file_size_base = 80;
+  mutable_cf_options_.max_compaction_bytes = 20;
+
+  auto file1 = mock::MakeMockFile({{KeyStr("a", 104U, kTypeValue), "val1"},
+                                   {KeyStr("b", 103U, kTypeValue), "val"}});
+  AddMockFile(file1);
+
+  auto file2 = mock::MakeMockFile({{KeyStr("a", 102U, kTypeValue), "val2"},
+                                   {KeyStr("c", 101U, kTypeValue), "val"}});
+  AddMockFile(file2, 1);
+
+  for (int i = 0; i < 10; i++) {
+    auto file =
+        mock::MakeMockFile({{KeyStr("a", 100 - (i * 2), kTypeValue), "val"},
+                            {KeyStr("a", 99 - (i * 2), kTypeValue), "val"}});
+    AddMockFile(file, 2);
+  }
+
+  for (int i = 0; i < 10; i++) {
+    auto file =
+        mock::MakeMockFile({{KeyStr("b", 80 - (i * 2), kTypeValue), "val"},
+                            {KeyStr("b", 79 - (i * 2), kTypeValue), "val"}});
+    AddMockFile(file, 2);
+  }
+
+  auto file5 = mock::MakeMockFile({{KeyStr("c", 60U, kTypeValue), "valc"},
+                                   {KeyStr("c", 59U, kTypeValue), "valc"}});
+
+  // "a" has 10 overlapped grandparent files (each size 10), which is far
+  // exceeded the `max_compaction_bytes`, but make sure 2 "a" are not separated,
+  // as splitting them won't help reducing the compaction size.
+  // also make sure "b" and "c" are cut separately.
+  mock::KVVector expected_file1 =
+      mock::MakeMockFile({{KeyStr("a", 104U, kTypeValue), "val1"},
+                          {KeyStr("a", 102U, kTypeValue), "val2"}});
+  mock::KVVector expected_file2 =
+      mock::MakeMockFile({{KeyStr("b", 103U, kTypeValue), "val"}});
+  mock::KVVector expected_file3 =
+      mock::MakeMockFile({{KeyStr("c", 101U, kTypeValue), "val"}});
+
+  SetLastSequence(122U);
+  const std::vector<int> input_levels = {0, 1};
+  auto lvl0_files = cfd_->current()->storage_info()->LevelFiles(0);
+  auto lvl1_files = cfd_->current()->storage_info()->LevelFiles(1);
+
+  // Just keep all the history
+  std::vector<SequenceNumber> snapshots;
+  for (int i = 80; i <= 105; i++) {
+    snapshots.emplace_back(i);
+  }
+  RunCompaction({lvl0_files, lvl1_files}, input_levels,
+                {expected_file1, expected_file2, expected_file3}, snapshots);
+}
+
+INSTANTIATE_TEST_CASE_P(CompactionJobDynamicFileSizeTest,
+                        CompactionJobDynamicFileSizeTest, testing::Bool());
 
 class CompactionJobTimestampTest : public CompactionJobTestBase {
  public:
@@ -2033,6 +2434,7 @@ TEST_F(CompactionJobIOPriorityTest, GetRateLimiterPriority) {
 }  // namespace ROCKSDB_NAMESPACE
 
 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
   ::testing::InitGoogleTest(&argc, argv);
   RegisterCustomObjects(argc, argv);
   return RUN_ALL_TESTS();

db/compaction/compaction_outputs.cc

@@ -76,14 +76,131 @@ 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 (HasBuilder() && partitioner_ &&
-      partitioner_->ShouldPartition(
-          PartitionerRequest(last_key_for_partitioner_, c_iter.user_key(),
-                             current_output_file_size_)) == kRequired) {
+  if (partitioner_ && partitioner_->ShouldPartition(PartitionerRequest(
+                          last_key_for_partitioner_, c_iter.user_key(),
+                          current_output_file_size_)) == kRequired) {
     return true;
   }
 
@@ -92,12 +209,11 @@ bool CompactionOutputs::ShouldStopBefore(const CompactionIterator& c_iter) {
     return false;
   }
 
-  // reach the target file size
+  // reach the max file size
   if (current_output_file_size_ >= compaction_->max_output_file_size()) {
     return true;
   }
 
-  const Slice& internal_key = c_iter.key();
   const InternalKeyComparator* icmp =
       &compaction_->column_family_data()->internal_comparator();
 
@@ -111,32 +227,67 @@ bool CompactionOutputs::ShouldStopBefore(const CompactionIterator& c_iter) {
     }
   }
 
-  // Update grandparent information
-  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;
+  // 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;
     }
-    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_ + current_output_file_size_ >
-          compaction_->max_compaction_bytes()) {
-    // Too much overlap for current output; start new output
-    overlapped_bytes_ = 0;
-    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
@@ -189,6 +340,10 @@ Status CompactionOutputs::AddToOutput(
     if (!s.ok()) {
       return s;
     }
+    // reset grandparent information
+    grandparent_boundary_switched_num_ = 0;
+    grandparent_overlapped_bytes_ =
+        GetCurrentKeyGrandparentOverlappedBytes(key);
   }
 
   // Open output file if necessary
@@ -199,10 +354,9 @@ Status CompactionOutputs::AddToOutput(
     }
   }
 
-  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;
   }
@@ -232,10 +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();
@@ -249,7 +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
@@ -269,21 +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;
   }
@@ -292,8 +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
@@ -307,7 +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
@@ -320,7 +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.
@@ -331,12 +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.
@@ -356,13 +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.
@@ -376,9 +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;

db/compaction/compaction_outputs.h

@@ -168,11 +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);
+                      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 {
@@ -216,9 +221,16 @@ 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);
+
+  // 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.
@@ -311,12 +323,21 @@ class CompactionOutputs {
   // 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 overlapped_bytes_ = 0;
+  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

db/compaction/compaction_picker.cc

@@ -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;
@@ -307,13 +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;
     }
   }
@@ -322,17 +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;
+      }
     }
   }
 
@@ -356,11 +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) {
@@ -507,8 +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;
@@ -527,13 +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);
@@ -554,9 +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;
       }
@@ -657,7 +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;
@@ -724,18 +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;
@@ -836,7 +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;
@@ -1121,7 +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);

db/compaction/compaction_picker.h

@@ -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,

db/compaction/compaction_picker_fifo.cc

@@ -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,

db/compaction/compaction_picker_level.cc

@@ -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;
@@ -736,7 +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();

db/compaction/compaction_picker_test.cc

@@ -1340,6 +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.
@@ -2358,6 +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.
@@ -2384,6 +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.
@@ -3146,7 +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);
@@ -3521,9 +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) {
@@ -3564,9 +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,
@@ -3609,7 +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) {
@@ -3653,7 +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,
@@ -3664,6 +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();
 }

db/compaction/compaction_picker_universal.cc

@@ -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;
@@ -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.

db/compaction/compaction_service_test.cc

@@ -325,6 +325,7 @@ TEST_F(CompactionServiceTest, BasicCompactions) {
       });
   Reopen(options);
   ASSERT_GT(verify_passed, 0);
+  Close();
 }
 
 TEST_F(CompactionServiceTest, ManualCompaction) {

db/compaction/tiered_compaction_test.cc

@@ -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

db/comparator_db_test.cc

@@ -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();
 }

db/cuckoo_table_db_test.cc

@@ -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 {

db/db_block_cache_test.cc

@@ -170,13 +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
 };
@@ -1500,23 +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();

db/db_bloom_filter_test.cc

@@ -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"));

db/db_compaction_test.cc

@@ -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;

db/db_flush_test.cc

@@ -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());
 

db/db_impl/db_impl.cc

@@ -847,11 +847,13 @@ Status DBImpl::RegisterRecordSeqnoTimeWorker() {
     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()) {
@@ -863,8 +865,11 @@ 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;
@@ -1438,11 +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) {
@@ -1453,6 +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;
@@ -1551,7 +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());
@@ -1559,7 +1576,7 @@ Status DBImpl::LockWAL() {
     // future writes
     WriteStatusCheck(status);
   }
-  return std::move(status);
+  return static_cast<Status>(status);
 }
 
 Status DBImpl::UnlockWAL() {
@@ -3092,7 +3109,8 @@ Status DBImpl::CreateColumnFamilyImpl(const ColumnFamilyOptions& cf_options,
     }
   }  // InstrumentedMutexLock l(&mutex_)
 
-  if (cf_options.preclude_last_level_data_seconds > 0) {
+  if (cf_options.preserve_internal_time_seconds > 0 ||
+      cf_options.preclude_last_level_data_seconds > 0) {
     s = RegisterRecordSeqnoTimeWorker();
   }
   sv_context.Clean();
@@ -3183,7 +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();
   }
 
@@ -4395,10 +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";
@@ -4580,8 +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);
@@ -4621,8 +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 ||
@@ -4644,7 +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);
     }
 
@@ -4652,7 +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) {
@@ -4793,7 +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;
   }
@@ -5251,7 +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) {
@@ -5262,7 +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;
@@ -5742,8 +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() {

db/db_impl/db_impl.h

@@ -222,6 +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,
@@ -423,7 +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;
@@ -565,9 +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;
 
@@ -1877,12 +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,

db/db_impl/db_impl_compaction_flush.cc

@@ -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);
 

db/db_impl/db_impl_debug.cc

@@ -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;

db/db_impl/db_impl_files.cc

@@ -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(

db/db_impl/db_impl_open.cc

@@ -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)");
@@ -570,7 +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);
@@ -678,7 +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()) {
@@ -1929,18 +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;
       }
@@ -2022,9 +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;
@@ -2060,10 +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());

db/db_impl/db_impl_secondary.cc

@@ -100,7 +100,11 @@ Status DBImplSecondary::FindNewLogNumbers(std::vector<uint64_t>* logs) {
   assert(logs != nullptr);
   std::vector<std::string> filenames;
   Status s;
-  s = env_->GetChildren(immutable_db_options_.GetWalDir(), &filenames);
+  IOOptions io_opts;
+  io_opts.do_not_recurse = true;
+  s = immutable_db_options_.fs->GetChildren(immutable_db_options_.GetWalDir(),
+                                            io_opts, &filenames,
+                                            /*IODebugContext*=*/nullptr);
   if (s.IsNotFound()) {
     return Status::InvalidArgument("Failed to open wal_dir",
                                    immutable_db_options_.GetWalDir());

db/db_impl/db_impl_write.cc

@@ -111,6 +111,17 @@ Status DBImpl::DeleteRange(const WriteOptions& write_options,
   return DB::DeleteRange(write_options, column_family, begin_key, end_key);
 }
 
+Status DBImpl::DeleteRange(const WriteOptions& write_options,
+                           ColumnFamilyHandle* column_family,
+                           const Slice& begin_key, const Slice& end_key,
+                           const Slice& ts) {
+  const Status s = FailIfTsMismatchCf(column_family, ts, /*ts_for_read=*/false);
+  if (!s.ok()) {
+    return s;
+  }
+  return DB::DeleteRange(write_options, column_family, begin_key, end_key, ts);
+}
+
 void DBImpl::SetRecoverableStatePreReleaseCallback(
     PreReleaseCallback* callback) {
   recoverable_state_pre_release_callback_.reset(callback);
@@ -1749,6 +1760,7 @@ Status DBImpl::DelayWrite(uint64_t num_bytes,
                  &time_delayed);
     uint64_t delay =
         write_controller_.GetDelay(immutable_db_options_.clock, num_bytes);
+    TEST_SYNC_POINT("DBImpl::DelayWrite:Start");
     if (delay > 0) {
       if (write_options.no_slowdown) {
         return Status::Incomplete("Write stall");
@@ -1758,8 +1770,8 @@ Status DBImpl::DelayWrite(uint64_t num_bytes,
       // Notify write_thread_ about the stall so it can setup a barrier and
       // fail any pending writers with no_slowdown
       write_thread_.BeginWriteStall();
-      TEST_SYNC_POINT("DBImpl::DelayWrite:BeginWriteStallDone");
       mutex_.Unlock();
+      TEST_SYNC_POINT("DBImpl::DelayWrite:BeginWriteStallDone");
       // We will delay the write until we have slept for `delay` microseconds
       // or we don't need a delay anymore. We check for cancellation every 1ms
       // (slightly longer because WriteController minimum delay is 1ms, in
@@ -1784,7 +1796,8 @@ Status DBImpl::DelayWrite(uint64_t num_bytes,
     // might wait here indefinitely as the background compaction may never
     // finish successfully, resulting in the stall condition lasting
     // indefinitely
-    while (error_handler_.GetBGError().ok() && write_controller_.IsStopped()) {
+    while (error_handler_.GetBGError().ok() && write_controller_.IsStopped() &&
+           !shutting_down_.load(std::memory_order_relaxed)) {
       if (write_options.no_slowdown) {
         return Status::Incomplete("Write stall");
       }
@@ -1810,9 +1823,13 @@ Status DBImpl::DelayWrite(uint64_t num_bytes,
   // proceed
   Status s;
   if (write_controller_.IsStopped()) {
-    // If writes are still stopped, it means we bailed due to a background
-    // error
-    s = Status::Incomplete(error_handler_.GetBGError().ToString());
+    if (!shutting_down_.load(std::memory_order_relaxed)) {
+      // If writes are still stopped and db not shutdown, it means we bailed
+      // due to a background error
+      s = Status::Incomplete(error_handler_.GetBGError().ToString());
+    } else {
+      s = Status::ShutdownInProgress("stalled writes");
+    }
   }
   if (error_handler_.IsDBStopped()) {
     s = error_handler_.GetBGError();
@@ -2361,6 +2378,24 @@ Status DB::DeleteRange(const WriteOptions& opt,
   return Write(opt, &batch);
 }
 
+Status DB::DeleteRange(const WriteOptions& opt,
+                       ColumnFamilyHandle* column_family,
+                       const Slice& begin_key, const Slice& end_key,
+                       const Slice& ts) {
+  ColumnFamilyHandle* default_cf = DefaultColumnFamily();
+  assert(default_cf);
+  const Comparator* const default_cf_ucmp = default_cf->GetComparator();
+  assert(default_cf_ucmp);
+  WriteBatch batch(0 /* reserved_bytes */, 0 /* max_bytes */,
+                   opt.protection_bytes_per_key,
+                   default_cf_ucmp->timestamp_size());
+  Status s = batch.DeleteRange(column_family, begin_key, end_key, ts);
+  if (!s.ok()) {
+    return s;
+  }
+  return Write(opt, &batch);
+}
+
 Status DB::Merge(const WriteOptions& opt, ColumnFamilyHandle* column_family,
                  const Slice& key, const Slice& value) {
   WriteBatch batch(0 /* reserved_bytes */, 0 /* max_bytes */,

db/db_iter.cc

@@ -90,7 +90,8 @@ DBIter::DBIter(Env* _env, const ReadOptions& read_options,
     iter_.iter()->SetPinnedItersMgr(&pinned_iters_mgr_);
   }
   status_.PermitUncheckedError();
-  assert(timestamp_size_ == user_comparator_.timestamp_size());
+  assert(timestamp_size_ ==
+         user_comparator_.user_comparator()->timestamp_size());
 }
 
 Status DBIter::GetProperty(std::string prop_name, std::string* prop) {

db/db_iter_stress_test.cc

@@ -652,6 +652,7 @@ TEST_F(DBIteratorStressTest, StressTest) {
 }  // namespace ROCKSDB_NAMESPACE
 
 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
   ::testing::InitGoogleTest(&argc, argv);
   ParseCommandLineFlags(&argc, &argv, true);
   return RUN_ALL_TESTS();

db/db_iter_test.cc

@@ -3187,6 +3187,7 @@ TEST_F(DBIteratorTest, ReverseToForwardWithDisappearingKeys) {
 }  // namespace ROCKSDB_NAMESPACE
 
 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
   ::testing::InitGoogleTest(&argc, argv);
   return RUN_ALL_TESTS();
 }

db/db_kv_checksum_test.cc

@@ -879,6 +879,7 @@ TEST_P(DbMemtableKVChecksumTest, FlushWithCorruptAfterMemtableInsert) {
 }  // namespace ROCKSDB_NAMESPACE
 
 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
   ::testing::InitGoogleTest(&argc, argv);
   return RUN_ALL_TESTS();
 }

db/db_logical_block_size_cache_test.cc

@@ -515,6 +515,7 @@ TEST_F(DBLogicalBlockSizeCacheTest, MultiDBWithSamePaths) {
 #endif  // OS_LINUX
 
 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
   ::testing::InitGoogleTest(&argc, argv);
   return RUN_ALL_TESTS();
 }

db/db_properties_test.cc

@@ -1225,7 +1225,7 @@ class BlockCountingTablePropertiesCollector : public TablePropertiesCollector {
     return Status::OK();
   }
 
-  void BlockAdd(uint64_t /* block_raw_bytes */,
+  void BlockAdd(uint64_t /* block_uncomp_bytes */,
                 uint64_t block_compressed_bytes_fast,
                 uint64_t block_compressed_bytes_slow) override {
     if (block_compressed_bytes_fast > 0 || block_compressed_bytes_slow > 0) {

db/db_range_del_test.cc

@@ -13,6 +13,7 @@
 
 namespace ROCKSDB_NAMESPACE {
 
+// TODO(cbi): parameterize the test to cover user-defined timestamp cases
 class DBRangeDelTest : public DBTestBase {
  public:
   DBRangeDelTest() : DBTestBase("db_range_del_test", /*env_do_fsync=*/false) {}
@@ -1029,7 +1030,11 @@ TEST_F(DBRangeDelTest, CompactionTreatsSplitInputLevelDeletionAtomically) {
   options.level0_file_num_compaction_trigger = kNumFilesPerLevel;
   options.memtable_factory.reset(
       test::NewSpecialSkipListFactory(2 /* num_entries_flush */));
-  options.target_file_size_base = kValueBytes;
+  // max file size could be 2x of target file size, so set it to half of that
+  options.target_file_size_base = kValueBytes / 2;
+  // disable dynamic_file_size, as it will cut L1 files into more files (than
+  // kNumFilesPerLevel).
+  options.level_compaction_dynamic_file_size = false;
   options.max_compaction_bytes = 1500;
   // i == 0: CompactFiles
   // i == 1: CompactRange
@@ -1100,6 +1105,12 @@ TEST_F(DBRangeDelTest, RangeTombstoneEndKeyAsSstableUpperBound) {
       test::NewSpecialSkipListFactory(2 /* num_entries_flush */));
   options.target_file_size_base = kValueBytes;
   options.disable_auto_compactions = true;
+  // disable it for now, otherwise the L1 files are going be cut before data 1:
+  // L1: [0]   [1,4]
+  // L2: [0,0]
+  // because the grandparent file is between [0]->[1] and it's size is more than
+  // 1/8 of target size (4k).
+  options.level_compaction_dynamic_file_size = false;
 
   DestroyAndReopen(options);
 
@@ -1723,9 +1734,8 @@ TEST_F(DBRangeDelTest, OverlappedKeys) {
 
   ASSERT_OK(dbfull()->TEST_CompactRange(1, nullptr, nullptr, nullptr,
                                         true /* disallow_trivial_move */));
-  ASSERT_EQ(
-      3, NumTableFilesAtLevel(
-             2));  // L1->L2 compaction size is limited to max_compaction_bytes
+  // L1->L2 compaction size is limited to max_compaction_bytes
+  ASSERT_EQ(3, NumTableFilesAtLevel(2));
   ASSERT_EQ(0, NumTableFilesAtLevel(1));
 }
 
@@ -1974,6 +1984,40 @@ TEST_F(DBRangeDelTest, TombstoneFromCurrentLevel) {
   delete iter;
 }
 
+class TombstoneTestSstPartitioner : public SstPartitioner {
+ public:
+  const char* Name() const override { return "SingleKeySstPartitioner"; }
+
+  PartitionerResult ShouldPartition(
+      const PartitionerRequest& request) override {
+    if (cmp->Compare(*request.current_user_key, DBTestBase::Key(5)) == 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 TombstoneTestSstPartitionerFactory : 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 TombstoneTestSstPartitioner());
+  }
+};
+
 TEST_F(DBRangeDelTest, TombstoneAcrossFileBoundary) {
   // Verify that a range tombstone across file boundary covers keys from older
   // levels. Test set up:
@@ -1987,11 +2031,17 @@ TEST_F(DBRangeDelTest, TombstoneAcrossFileBoundary) {
   options.target_file_size_base = 2 * 1024;
   options.max_compaction_bytes = 2 * 1024;
 
+  // Make sure L1 files are split before "5"
+  auto factory = std::make_shared<TombstoneTestSstPartitionerFactory>();
+  options.sst_partitioner_factory = factory;
+
   DestroyAndReopen(options);
 
   Random rnd(301);
   // L2
-  ASSERT_OK(db_->Put(WriteOptions(), Key(5), rnd.RandomString(1 << 10)));
+  // the file should be smaller than max_compaction_bytes, otherwise the file
+  // will be cut before 7.
+  ASSERT_OK(db_->Put(WriteOptions(), Key(5), rnd.RandomString(1 << 9)));
   ASSERT_OK(db_->Flush(FlushOptions()));
   MoveFilesToLevel(2);
   ASSERT_EQ(1, NumTableFilesAtLevel(2));

db/db_tailing_iter_test.cc

@@ -18,15 +18,22 @@
 
 namespace ROCKSDB_NAMESPACE {
 
-class DBTestTailingIterator : public DBTestBase {
+class DBTestTailingIterator : public DBTestBase,
+                              public ::testing::WithParamInterface<bool> {
  public:
   DBTestTailingIterator()
       : DBTestBase("db_tailing_iterator_test", /*env_do_fsync=*/true) {}
 };
 
-TEST_F(DBTestTailingIterator, TailingIteratorSingle) {
+INSTANTIATE_TEST_CASE_P(DBTestTailingIterator, DBTestTailingIterator,
+                        ::testing::Bool());
+
+TEST_P(DBTestTailingIterator, TailingIteratorSingle) {
   ReadOptions read_options;
   read_options.tailing = true;
+  if (GetParam()) {
+    read_options.async_io = true;
+  }
 
   std::unique_ptr<Iterator> iter(db_->NewIterator(read_options));
   iter->SeekToFirst();
@@ -43,11 +50,13 @@ TEST_F(DBTestTailingIterator, TailingIteratorSingle) {
   ASSERT_TRUE(!iter->Valid());
 }
 
-TEST_F(DBTestTailingIterator, TailingIteratorKeepAdding) {
+TEST_P(DBTestTailingIterator, TailingIteratorKeepAdding) {
   CreateAndReopenWithCF({"pikachu"}, CurrentOptions());
   ReadOptions read_options;
   read_options.tailing = true;
-
+  if (GetParam()) {
+    read_options.async_io = true;
+  }
   std::unique_ptr<Iterator> iter(db_->NewIterator(read_options, handles_[1]));
   ASSERT_OK(iter->status());
   std::string value(1024, 'a');
@@ -66,11 +75,13 @@ TEST_F(DBTestTailingIterator, TailingIteratorKeepAdding) {
   }
 }
 
-TEST_F(DBTestTailingIterator, TailingIteratorSeekToNext) {
+TEST_P(DBTestTailingIterator, TailingIteratorSeekToNext) {
   CreateAndReopenWithCF({"pikachu"}, CurrentOptions());
   ReadOptions read_options;
   read_options.tailing = true;
-
+  if (GetParam()) {
+    read_options.async_io = true;
+  }
   std::unique_ptr<Iterator> iter(db_->NewIterator(read_options, handles_[1]));
   ASSERT_OK(iter->status());
   std::unique_ptr<Iterator> itern(db_->NewIterator(read_options, handles_[1]));
@@ -125,7 +136,7 @@ TEST_F(DBTestTailingIterator, TailingIteratorSeekToNext) {
   }
 }
 
-TEST_F(DBTestTailingIterator, TailingIteratorTrimSeekToNext) {
+TEST_P(DBTestTailingIterator, TailingIteratorTrimSeekToNext) {
   const uint64_t k150KB = 150 * 1024;
   Options options;
   options.write_buffer_size = k150KB;
@@ -135,6 +146,9 @@ TEST_F(DBTestTailingIterator, TailingIteratorTrimSeekToNext) {
   CreateAndReopenWithCF({"pikachu"}, options);
   ReadOptions read_options;
   read_options.tailing = true;
+  if (GetParam()) {
+    read_options.async_io = true;
+  }
   int num_iters, deleted_iters;
 
   char bufe[32];
@@ -265,10 +279,13 @@ TEST_F(DBTestTailingIterator, TailingIteratorTrimSeekToNext) {
   }
 }
 
-TEST_F(DBTestTailingIterator, TailingIteratorDeletes) {
+TEST_P(DBTestTailingIterator, TailingIteratorDeletes) {
   CreateAndReopenWithCF({"pikachu"}, CurrentOptions());
   ReadOptions read_options;
   read_options.tailing = true;
+  if (GetParam()) {
+    read_options.async_io = true;
+  }
 
   std::unique_ptr<Iterator> iter(db_->NewIterator(read_options, handles_[1]));
   ASSERT_OK(iter->status());
@@ -300,15 +317,18 @@ TEST_F(DBTestTailingIterator, TailingIteratorDeletes) {
 
   // make sure we can read all new records using the existing iterator
   int count = 0;
-  for (; iter->Valid(); iter->Next(), ++count) ;
+  for (; iter->Valid(); iter->Next(), ++count)
+    ;
 
   ASSERT_EQ(count, num_records);
 }
 
-TEST_F(DBTestTailingIterator, TailingIteratorPrefixSeek) {
+TEST_P(DBTestTailingIterator, TailingIteratorPrefixSeek) {
   ReadOptions read_options;
   read_options.tailing = true;
-
+  if (GetParam()) {
+    read_options.async_io = true;
+  }
   Options options = CurrentOptions();
   options.create_if_missing = true;
   options.disable_auto_compactions = true;
@@ -338,10 +358,13 @@ TEST_F(DBTestTailingIterator, TailingIteratorPrefixSeek) {
   ASSERT_TRUE(!iter->Valid());
 }
 
-TEST_F(DBTestTailingIterator, TailingIteratorIncomplete) {
+TEST_P(DBTestTailingIterator, TailingIteratorIncomplete) {
   CreateAndReopenWithCF({"pikachu"}, CurrentOptions());
   ReadOptions read_options;
   read_options.tailing = true;
+  if (GetParam()) {
+    read_options.async_io = true;
+  }
   read_options.read_tier = kBlockCacheTier;
 
   std::string key("key");
@@ -361,7 +384,7 @@ TEST_F(DBTestTailingIterator, TailingIteratorIncomplete) {
   ASSERT_TRUE(iter->Valid() || iter->status().IsIncomplete());
 }
 
-TEST_F(DBTestTailingIterator, TailingIteratorSeekToSame) {
+TEST_P(DBTestTailingIterator, TailingIteratorSeekToSame) {
   Options options = CurrentOptions();
   options.compaction_style = kCompactionStyleUniversal;
   options.write_buffer_size = 1000;
@@ -369,7 +392,9 @@ TEST_F(DBTestTailingIterator, TailingIteratorSeekToSame) {
 
   ReadOptions read_options;
   read_options.tailing = true;
-
+  if (GetParam()) {
+    read_options.async_io = true;
+  }
   const int NROWS = 10000;
   // Write rows with keys 00000, 00002, 00004 etc.
   for (int i = 0; i < NROWS; ++i) {
@@ -400,14 +425,16 @@ TEST_F(DBTestTailingIterator, TailingIteratorSeekToSame) {
 // Sets iterate_upper_bound and verifies that ForwardIterator doesn't call
 // Seek() on immutable iterators when target key is >= prev_key and all
 // iterators, including the memtable iterator, are over the upper bound.
-TEST_F(DBTestTailingIterator, TailingIteratorUpperBound) {
+TEST_P(DBTestTailingIterator, TailingIteratorUpperBound) {
   CreateAndReopenWithCF({"pikachu"}, CurrentOptions());
 
   const Slice upper_bound("20", 3);
   ReadOptions read_options;
   read_options.tailing = true;
   read_options.iterate_upper_bound = &upper_bound;
-
+  if (GetParam()) {
+    read_options.async_io = true;
+  }
   ASSERT_OK(Put(1, "11", "11"));
   ASSERT_OK(Put(1, "12", "12"));
   ASSERT_OK(Put(1, "22", "22"));
@@ -439,10 +466,14 @@ TEST_F(DBTestTailingIterator, TailingIteratorUpperBound) {
 
   ASSERT_FALSE(it->Valid());
   ASSERT_OK(it->status());
-  ASSERT_EQ(0, immutable_seeks);
+  if (GetParam()) {
+    ASSERT_EQ(1, immutable_seeks);
+  } else {
+    ASSERT_EQ(0, immutable_seeks);
+  }
 }
 
-TEST_F(DBTestTailingIterator, TailingIteratorGap) {
+TEST_P(DBTestTailingIterator, TailingIteratorGap) {
   // level 1:            [20, 25]  [35, 40]
   // level 2:  [10 - 15]                    [45 - 50]
   // level 3:            [20,    30,    40]
@@ -455,7 +486,9 @@ TEST_F(DBTestTailingIterator, TailingIteratorGap) {
 
   ReadOptions read_options;
   read_options.tailing = true;
-
+  if (GetParam()) {
+    read_options.async_io = true;
+  }
   ASSERT_OK(Put(1, "20", "20"));
   ASSERT_OK(Put(1, "30", "30"));
   ASSERT_OK(Put(1, "40", "40"));
@@ -497,9 +530,12 @@ TEST_F(DBTestTailingIterator, TailingIteratorGap) {
   ASSERT_OK(it->status());
 }
 
-TEST_F(DBTestTailingIterator, SeekWithUpperBoundBug) {
+TEST_P(DBTestTailingIterator, SeekWithUpperBoundBug) {
   ReadOptions read_options;
   read_options.tailing = true;
+  if (GetParam()) {
+    read_options.async_io = true;
+  }
   const Slice upper_bound("cc", 3);
   read_options.iterate_upper_bound = &upper_bound;
 
@@ -520,9 +556,12 @@ TEST_F(DBTestTailingIterator, SeekWithUpperBoundBug) {
   ASSERT_EQ(iter->key().ToString(), "aa");
 }
 
-TEST_F(DBTestTailingIterator, SeekToFirstWithUpperBoundBug) {
+TEST_P(DBTestTailingIterator, SeekToFirstWithUpperBoundBug) {
   ReadOptions read_options;
   read_options.tailing = true;
+  if (GetParam()) {
+    read_options.async_io = true;
+  }
   const Slice upper_bound("cc", 3);
   read_options.iterate_upper_bound = &upper_bound;
 

db/db_test.cc

@@ -483,6 +483,135 @@ TEST_F(DBTest, LevelLimitReopen) {
 }
 #endif  // ROCKSDB_LITE
 
+#ifndef ROCKSDB_LITE
+TEST_F(DBTest, LevelReopenWithFIFO) {
+  const int kLevelCount = 4;
+  const int kKeyCount = 5;
+  const int kTotalSstFileCount = kLevelCount * kKeyCount;
+  const int kCF = 1;
+
+  Options options = CurrentOptions();
+  // Config level0_file_num_compaction_trigger to prevent L0 files being
+  // automatically compacted while we are constructing a LSM tree structure
+  // to test multi-level FIFO compaction.
+  options.level0_file_num_compaction_trigger = kKeyCount + 1;
+  CreateAndReopenWithCF({"pikachu"}, options);
+
+  // The expected number of files per level after each file creation.
+  const std::string expected_files_per_level[kLevelCount][kKeyCount] = {
+      {"0,0,0,1", "0,0,0,2", "0,0,0,3", "0,0,0,4", "0,0,0,5"},
+      {"0,0,1,5", "0,0,2,5", "0,0,3,5", "0,0,4,5", "0,0,5,5"},
+      {"0,1,5,5", "0,2,5,5", "0,3,5,5", "0,4,5,5", "0,5,5,5"},
+      {"1,5,5,5", "2,5,5,5", "3,5,5,5", "4,5,5,5", "5,5,5,5"},
+  };
+
+  const std::string expected_entries[kKeyCount][kLevelCount + 1] = {
+      {"[ ]", "[ a3 ]", "[ a2, a3 ]", "[ a1, a2, a3 ]", "[ a0, a1, a2, a3 ]"},
+      {"[ ]", "[ b3 ]", "[ b2, b3 ]", "[ b1, b2, b3 ]", "[ b0, b1, b2, b3 ]"},
+      {"[ ]", "[ c3 ]", "[ c2, c3 ]", "[ c1, c2, c3 ]", "[ c0, c1, c2, c3 ]"},
+      {"[ ]", "[ d3 ]", "[ d2, d3 ]", "[ d1, d2, d3 ]", "[ d0, d1, d2, d3 ]"},
+      {"[ ]", "[ e3 ]", "[ e2, e3 ]", "[ e1, e2, e3 ]", "[ e0, e1, e2, e3 ]"},
+  };
+
+  // The loop below creates the following LSM tree where each (k, v) pair
+  // represents a file that contains that entry.  When a file is created,
+  // the db is reopend with FIFO compaction and verified the LSM tree
+  // structure is still the same.
+  //
+  // The resulting LSM tree will contain 5 different keys.  Each key as
+  // 4 different versions, located in different level.
+  //
+  // L0:  (e, e0) (d, d0) (c, c0) (b, b0) (a, a0)
+  // L1:  (a, a1) (b, b1) (c, c1) (d, d1) (e, e1)
+  // L2:  (a, a2) (b, b2) (c, c2) (d, d2) (e, e2)
+  // L3:  (a, a3) (b, b3) (c, c3) (d, d3) (e, e3)
+  for (int l = 0; l < kLevelCount; ++l) {
+    int level = kLevelCount - 1 - l;
+    for (int p = 0; p < kKeyCount; ++p) {
+      std::string put_key = std::string(1, char('a' + p));
+      ASSERT_OK(Put(kCF, put_key, put_key + std::to_string(level)));
+      ASSERT_OK(Flush(kCF));
+      ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
+      for (int g = 0; g < kKeyCount; ++g) {
+        int entry_count = (p >= g) ? l + 1 : l;
+        std::string get_key = std::string(1, char('a' + g));
+        CheckAllEntriesWithFifoReopen(expected_entries[g][entry_count], get_key,
+                                      kCF, {"pikachu"}, options);
+      }
+      if (level != 0) {
+        MoveFilesToLevel(level, kCF);
+        for (int g = 0; g < kKeyCount; ++g) {
+          int entry_count = (p >= g) ? l + 1 : l;
+          std::string get_key = std::string(1, char('a' + g));
+          CheckAllEntriesWithFifoReopen(expected_entries[g][entry_count],
+                                        get_key, kCF, {"pikachu"}, options);
+        }
+      }
+      ASSERT_EQ(expected_files_per_level[l][p], FilesPerLevel(kCF));
+    }
+  }
+
+  // The expected number of sst files in each level after each FIFO compaction
+  // that deletes the oldest sst file.
+  const std::string expected_files_per_level_after_fifo[] = {
+      "5,5,5,4", "5,5,5,3", "5,5,5,2", "5,5,5,1", "5,5,5", "5,5,4", "5,5,3",
+      "5,5,2",   "5,5,1",   "5,5",     "5,4",     "5,3",   "5,2",   "5,1",
+      "5",       "4",       "3",       "2",       "1",     "",
+  };
+
+  // The expected value entries of each key after each FIFO compaction.
+  // This verifies whether FIFO removes the file with the smallest key in non-L0
+  // files first then the oldest files in L0.
+  const std::string expected_entries_after_fifo[kKeyCount][kLevelCount + 1] = {
+      {"[ a0, a1, a2, a3 ]", "[ a0, a1, a2 ]", "[ a0, a1 ]", "[ a0 ]", "[ ]"},
+      {"[ b0, b1, b2, b3 ]", "[ b0, b1, b2 ]", "[ b0, b1 ]", "[ b0 ]", "[ ]"},
+      {"[ c0, c1, c2, c3 ]", "[ c0, c1, c2 ]", "[ c0, c1 ]", "[ c0 ]", "[ ]"},
+      {"[ d0, d1, d2, d3 ]", "[ d0, d1, d2 ]", "[ d0, d1 ]", "[ d0 ]", "[ ]"},
+      {"[ e0, e1, e2, e3 ]", "[ e0, e1, e2 ]", "[ e0, e1 ]", "[ e0 ]", "[ ]"},
+  };
+
+  // In the 2nd phase, we reopen the DB with FIFO compaction.  In each reopen,
+  // we config max_table_files_size so that FIFO will remove exactly one file
+  // at a time upon compaction, and we will use it to verify whether the sst
+  // files are deleted in the correct order.
+  for (int i = 0; i < kTotalSstFileCount; ++i) {
+    uint64_t total_sst_files_size = 0;
+    ASSERT_TRUE(dbfull()->GetIntProperty(
+        handles_[1], "rocksdb.total-sst-files-size", &total_sst_files_size));
+    ASSERT_TRUE(total_sst_files_size > 0);
+
+    Options fifo_options(options);
+    fifo_options.compaction_style = kCompactionStyleFIFO;
+    options.create_if_missing = false;
+    fifo_options.max_open_files = -1;
+    fifo_options.disable_auto_compactions = false;
+    // Config max_table_files_size to be total_sst_files_size - 1 so that
+    // FIFO will delete one file.
+    fifo_options.compaction_options_fifo.max_table_files_size =
+        total_sst_files_size - 1;
+    ASSERT_OK(
+        TryReopenWithColumnFamilies({"default", "pikachu"}, fifo_options));
+    // For FIFO to pick a compaction
+    ASSERT_OK(dbfull()->TEST_CompactRange(0, nullptr, nullptr, handles_[1]));
+    ASSERT_OK(dbfull()->TEST_WaitForCompact(false));
+    for (int g = 0; g < kKeyCount; ++g) {
+      std::string get_key = std::string(1, char('a' + g));
+      int status_index = i / kKeyCount;
+      if ((i % kKeyCount) >= g) {
+        // If true, then it means the sst file containing the get_key in the
+        // current level has already been deleted, so we need to move the
+        // status_index for checking the expected value.
+        status_index++;
+      }
+      CheckAllEntriesWithFifoReopen(
+          expected_entries_after_fifo[g][status_index], get_key, kCF,
+          {"pikachu"}, options);
+    }
+    ASSERT_EQ(expected_files_per_level_after_fifo[i], FilesPerLevel(kCF));
+  }
+}
+#endif  // !ROCKSDB_LITE
+
 TEST_F(DBTest, PutSingleDeleteGet) {
   do {
     CreateAndReopenWithCF({"pikachu"}, CurrentOptions());
@@ -668,7 +797,7 @@ TEST_F(DBTest, SingleDeletePutFlush) {
     ASSERT_OK(Flush(1));
 
     ASSERT_EQ("[ ]", AllEntriesFor("a", 1));
-    // Skip FIFO and universal compaction beccaus they do not apply to the test
+    // Skip FIFO and universal compaction because they do not apply to the test
     // case. Skip MergePut because single delete does not get removed when it
     // encounters a merge.
   } while (ChangeOptions(kSkipFIFOCompaction | kSkipUniversalCompaction |
@@ -7211,6 +7340,46 @@ TEST_F(DBTest, MemoryUsageWithMaxWriteBufferSizeToMaintain) {
   }
 }
 
+TEST_F(DBTest, ShuttingDownNotBlockStalledWrites) {
+  Options options = CurrentOptions();
+  options.disable_auto_compactions = true;
+  Reopen(options);
+  Random rnd(403);
+
+  for (int i = 0; i < 20; i++) {
+    ASSERT_OK(Put("key_" + std::to_string(i), rnd.RandomString(10)));
+    ASSERT_OK(Flush());
+  }
+  ASSERT_EQ(GetSstFileCount(dbname_), 20);
+
+  // We need !disable_auto_compactions for writes to stall but also want to
+  // delay compaction so stalled writes unblocked due to kShutdownInProgress. BG
+  // compaction will first wait for the sync point
+  // DBTest::ShuttingDownNotBlockStalledWrites. Then it waits extra 2 sec to
+  // allow CancelAllBackgroundWork() to set shutting_down_.
+  SyncPoint::GetInstance()->SetCallBack(
+      "BackgroundCallCompaction:0",
+      [&](void* /* arg */) { env_->SleepForMicroseconds(2 * 1000 * 1000); });
+  SyncPoint::GetInstance()->LoadDependency(
+      {{"DBImpl::DelayWrite:Wait", "DBTest::ShuttingDownNotBlockStalledWrites"},
+       {"DBTest::ShuttingDownNotBlockStalledWrites",
+        "BackgroundCallCompaction:0"}});
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  options.level0_stop_writes_trigger = 20;
+  options.disable_auto_compactions = false;
+  Reopen(options);
+
+  std::thread thd([&]() {
+    Status s = Put("key_" + std::to_string(101), "101");
+    ASSERT_EQ(s.code(), Status::kShutdownInProgress);
+  });
+
+  TEST_SYNC_POINT("DBTest::ShuttingDownNotBlockStalledWrites");
+  CancelAllBackgroundWork(db_, true);
+
+  thd.join();
+}
 #endif
 
 }  // namespace ROCKSDB_NAMESPACE

db/db_test2.cc

@@ -1796,15 +1796,14 @@ TEST_P(CompressionFailuresTest, CompressionFailures) {
         });
   } else if (compression_failure_type_ == kTestDecompressionFail) {
     ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
-        "UncompressBlockContentsForCompressionType:TamperWithReturnValue",
-        [](void* arg) {
+        "UncompressBlockData:TamperWithReturnValue", [](void* arg) {
           Status* ret = static_cast<Status*>(arg);
           ASSERT_OK(*ret);
           *ret = Status::Corruption("kTestDecompressionFail");
         });
   } else if (compression_failure_type_ == kTestDecompressionCorruption) {
     ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
-        "UncompressBlockContentsForCompressionType:"
+        "UncompressBlockData:"
         "TamperWithDecompressionOutput",
         [](void* arg) {
           BlockContents* contents = static_cast<BlockContents*>(arg);
@@ -1872,7 +1871,7 @@ TEST_P(CompressionFailuresTest, CompressionFailures) {
               "Could not decompress: kTestDecompressionFail");
   } else if (compression_failure_type_ == kTestDecompressionCorruption) {
     ASSERT_EQ(std::string(s.getState()),
-              "Decompressed block did not match raw block");
+              "Decompressed block did not match pre-compression block");
   }
 }
 
@@ -2380,9 +2379,15 @@ TEST_F(DBTest2, MaxCompactionBytesTest) {
   }
   ASSERT_OK(dbfull()->TEST_WaitForCompact());
 
-  // Output files to L1 are cut to three pieces, according to
-  // options.max_compaction_bytes
-  ASSERT_EQ("0,3,8", FilesPerLevel(0));
+  // Output files to L1 are cut to 4 pieces, according to
+  // options.max_compaction_bytes (300K)
+  // There are 8 files on L2 (grandparents level), each one is 100K. The first
+  // file overlaps with a, b which max_compaction_bytes is less than 300K, the
+  // second one overlaps with d, e, which is also less than 300K. Including any
+  // extra grandparent file will make the future compaction larger than 300K.
+  // L1: [  1  ] [  2 ]  [  3  ] [ 4 ]
+  // L2: [a] [b] [c] [d] [e] [f] [g] [h]
+  ASSERT_EQ("0,4,8", FilesPerLevel(0));
 }
 
 static void UniqueIdCallback(void* arg) {
@@ -6918,7 +6923,7 @@ TEST_F(DBTest2, LastLevelTemperatureUniversal) {
   ASSERT_EQ(size, 0);
   ASSERT_EQ(iostats->file_io_stats_by_temperature.hot_file_read_count, 0);
   ASSERT_EQ(iostats->file_io_stats_by_temperature.warm_file_read_count, 0);
-  ASSERT_EQ(iostats->file_io_stats_by_temperature.hot_file_read_count, 0);
+  ASSERT_EQ(iostats->file_io_stats_by_temperature.cold_file_read_count, 0);
   ASSERT_EQ(options.statistics->getTickerCount(HOT_FILE_READ_BYTES), 0);
   ASSERT_EQ(options.statistics->getTickerCount(WARM_FILE_READ_BYTES), 0);
   ASSERT_EQ(options.statistics->getTickerCount(COLD_FILE_READ_BYTES), 0);
@@ -7501,6 +7506,80 @@ TEST_F(DBTest2, SstUniqueIdVerifyMultiCFs) {
   ASSERT_TRUE(s.IsCorruption());
 }
 
+TEST_F(DBTest2, BestEffortsRecoveryWithSstUniqueIdVerification) {
+  const auto tamper_with_uniq_id = [&](void* arg) {
+    auto props = static_cast<TableProperties*>(arg);
+    assert(props);
+    // update table property session_id to a different one
+    props->db_session_id = DBImpl::GenerateDbSessionId(nullptr);
+  };
+
+  const auto assert_db = [&](size_t expected_count,
+                             const std::string& expected_v) {
+    std::unique_ptr<Iterator> it(db_->NewIterator(ReadOptions()));
+    size_t cnt = 0;
+    for (it->SeekToFirst(); it->Valid(); it->Next(), ++cnt) {
+      ASSERT_EQ(std::to_string(cnt), it->key());
+      ASSERT_EQ(expected_v, it->value());
+    }
+    ASSERT_EQ(expected_count, cnt);
+  };
+
+  const int num_l0_compaction_trigger = 8;
+  const int num_l0 = num_l0_compaction_trigger - 1;
+  Options options = CurrentOptions();
+  options.level0_file_num_compaction_trigger = num_l0_compaction_trigger;
+
+  for (int k = 0; k < num_l0; ++k) {
+    // Allow mismatch for now
+    options.verify_sst_unique_id_in_manifest = false;
+
+    DestroyAndReopen(options);
+
+    constexpr size_t num_keys_per_file = 10;
+    for (int i = 0; i < num_l0; ++i) {
+      for (size_t j = 0; j < num_keys_per_file; ++j) {
+        ASSERT_OK(Put(std::to_string(j), "v" + std::to_string(i)));
+      }
+      if (i == k) {
+        SyncPoint::GetInstance()->DisableProcessing();
+        SyncPoint::GetInstance()->SetCallBack(
+            "PropertyBlockBuilder::AddTableProperty:Start",
+            tamper_with_uniq_id);
+        SyncPoint::GetInstance()->EnableProcessing();
+      }
+      ASSERT_OK(Flush());
+    }
+
+    options.verify_sst_unique_id_in_manifest = true;
+    Status s = TryReopen(options);
+    ASSERT_TRUE(s.IsCorruption());
+
+    options.best_efforts_recovery = true;
+    Reopen(options);
+    assert_db(k == 0 ? 0 : num_keys_per_file, "v" + std::to_string(k - 1));
+
+    // Reopen with regular recovery
+    options.best_efforts_recovery = false;
+    Reopen(options);
+    assert_db(k == 0 ? 0 : num_keys_per_file, "v" + std::to_string(k - 1));
+
+    SyncPoint::GetInstance()->DisableProcessing();
+    SyncPoint::GetInstance()->ClearAllCallBacks();
+
+    for (size_t i = 0; i < num_keys_per_file; ++i) {
+      ASSERT_OK(Put(std::to_string(i), "v"));
+    }
+    ASSERT_OK(Flush());
+    Reopen(options);
+    {
+      for (size_t i = 0; i < num_keys_per_file; ++i) {
+        ASSERT_EQ("v", Get(std::to_string(i)));
+      }
+    }
+  }
+}
+
 #ifndef ROCKSDB_LITE
 TEST_F(DBTest2, GetLatestSeqAndTsForKey) {
   Destroy(last_options_);

db/db_test_util.cc

@@ -487,8 +487,10 @@ Options DBTestBase::GetOptions(
     case kInfiniteMaxOpenFiles:
       options.max_open_files = -1;
       break;
-    case kXXH3Checksum: {
-      table_options.checksum = kXXH3;
+    case kCRC32cChecksum: {
+      // Old default was CRC32c, but XXH3 (new default) is faster on common
+      // hardware
+      table_options.checksum = kCRC32c;
       // Thrown in here for basic coverage:
       options.DisableExtraChecks();
       break;
@@ -963,6 +965,25 @@ std::string DBTestBase::Contents(int cf) {
   return result;
 }
 
+void DBTestBase::CheckAllEntriesWithFifoReopen(
+    const std::string& expected_value, const Slice& user_key, int cf,
+    const std::vector<std::string>& cfs, const Options& options) {
+  ASSERT_EQ(AllEntriesFor(user_key, cf), expected_value);
+
+  std::vector<std::string> cfs_plus_default = cfs;
+  cfs_plus_default.insert(cfs_plus_default.begin(), kDefaultColumnFamilyName);
+
+  Options fifo_options(options);
+  fifo_options.compaction_style = kCompactionStyleFIFO;
+  fifo_options.max_open_files = -1;
+  fifo_options.disable_auto_compactions = true;
+  ASSERT_OK(TryReopenWithColumnFamilies(cfs_plus_default, fifo_options));
+  ASSERT_EQ(AllEntriesFor(user_key, cf), expected_value);
+
+  ASSERT_OK(TryReopenWithColumnFamilies(cfs_plus_default, options));
+  ASSERT_EQ(AllEntriesFor(user_key, cf), expected_value);
+}
+
 std::string DBTestBase::AllEntriesFor(const Slice& user_key, int cf) {
   Arena arena;
   auto options = CurrentOptions();

db/db_test_util.h

@@ -714,6 +714,16 @@ class FileTemperatureTestFS : public FileSystemWrapper {
       MutexLock lock(&mu_);
       requested_sst_file_temperatures_.emplace_back(number, opts.temperature);
       if (s.ok()) {
+        if (opts.temperature != Temperature::kUnknown) {
+          // Be extra picky and don't open if a wrong non-unknown temperature is
+          // provided
+          auto e = current_sst_file_temperatures_.find(number);
+          if (e != current_sst_file_temperatures_.end() &&
+              e->second != opts.temperature) {
+            result->reset();
+            return IOStatus::PathNotFound("Temperature mismatch on " + fname);
+          }
+        }
         *result = WrapWithTemperature<FSSequentialFileOwnerWrapper>(
             number, std::move(*result));
       }
@@ -733,6 +743,16 @@ class FileTemperatureTestFS : public FileSystemWrapper {
       MutexLock lock(&mu_);
       requested_sst_file_temperatures_.emplace_back(number, opts.temperature);
       if (s.ok()) {
+        if (opts.temperature != Temperature::kUnknown) {
+          // Be extra picky and don't open if a wrong non-unknown temperature is
+          // provided
+          auto e = current_sst_file_temperatures_.find(number);
+          if (e != current_sst_file_temperatures_.end() &&
+              e->second != opts.temperature) {
+            result->reset();
+            return IOStatus::PathNotFound("Temperature mismatch on " + fname);
+          }
+        }
         *result = WrapWithTemperature<FSRandomAccessFileOwnerWrapper>(
             number, std::move(*result));
       }
@@ -1025,7 +1045,7 @@ class DBTestBase : public testing::Test {
     kUniversalCompactionMultiLevel = 20,
     kCompressedBlockCache = 21,
     kInfiniteMaxOpenFiles = 22,
-    kXXH3Checksum = 23,
+    kCRC32cChecksum = 23,
     kFIFOCompaction = 24,
     kOptimizeFiltersForHits = 25,
     kRowCache = 26,
@@ -1219,6 +1239,15 @@ class DBTestBase : public testing::Test {
 
   std::string AllEntriesFor(const Slice& user_key, int cf = 0);
 
+  // Similar to AllEntriesFor but this function also covers reopen with fifo.
+  // Note that test cases with snapshots or entries in memtable should simply
+  // use AllEntriesFor instead as snapshots and entries in memtable will
+  // survive after db reopen.
+  void CheckAllEntriesWithFifoReopen(const std::string& expected_value,
+                                     const Slice& user_key, int cf,
+                                     const std::vector<std::string>& cfs,
+                                     const Options& options);
+
 #ifndef ROCKSDB_LITE
   int NumSortedRuns(int cf = 0);
 

db/db_with_timestamp_basic_test.cc

@@ -56,7 +56,7 @@ TEST_F(DBBasicTestWithTimestamp, SanityChecks) {
       db_->SingleDelete(WriteOptions(), "key", dummy_ts).IsInvalidArgument());
   ASSERT_TRUE(db_->DeleteRange(WriteOptions(), db_->DefaultColumnFamily(),
                                "begin_key", "end_key", dummy_ts)
-                  .IsNotSupported());
+                  .IsInvalidArgument());
 
   // Perform non-timestamp operations on "data" cf.
   ASSERT_TRUE(
@@ -85,6 +85,11 @@ TEST_F(DBBasicTestWithTimestamp, SanityChecks) {
     ASSERT_OK(wb.SingleDelete(handle, "key"));
     ASSERT_TRUE(db_->Write(WriteOptions(), &wb).IsInvalidArgument());
   }
+  {
+    WriteBatch wb;
+    ASSERT_OK(wb.DeleteRange(handle, "begin_key", "end_key"));
+    ASSERT_TRUE(db_->Write(WriteOptions(), &wb).IsInvalidArgument());
+  }
 
   // Perform timestamp operations with timestamps of incorrect size.
   const std::string wrong_ts(sizeof(uint32_t), '\0');
@@ -98,7 +103,7 @@ TEST_F(DBBasicTestWithTimestamp, SanityChecks) {
                   .IsInvalidArgument());
   ASSERT_TRUE(
       db_->DeleteRange(WriteOptions(), handle, "begin_key", "end_key", wrong_ts)
-          .IsNotSupported());
+          .IsInvalidArgument());
 
   delete handle;
 }
@@ -215,6 +220,10 @@ TEST_F(DBBasicTestWithTimestamp, GcPreserveLatestVersionBelowFullHistoryLow) {
   ts_str = Timestamp(4, 0);
   ASSERT_OK(db_->Put(wopts, "k1", ts_str, "v5"));
 
+  ts_str = Timestamp(5, 0);
+  ASSERT_OK(
+      db_->DeleteRange(wopts, db_->DefaultColumnFamily(), "k0", "k9", ts_str));
+
   ts_str = Timestamp(3, 0);
   Slice ts = ts_str;
   CompactRangeOptions cro;
@@ -234,6 +243,13 @@ TEST_F(DBBasicTestWithTimestamp, GcPreserveLatestVersionBelowFullHistoryLow) {
   ASSERT_TRUE(db_->Get(ropts, "k3", &value, &key_ts).IsNotFound());
   ASSERT_EQ(Timestamp(2, 0), key_ts);
 
+  ts_str = Timestamp(5, 0);
+  ts = ts_str;
+  ropts.timestamp = &ts;
+  ASSERT_TRUE(db_->Get(ropts, "k2", &value, &key_ts).IsNotFound());
+  ASSERT_EQ(Timestamp(5, 0), key_ts);
+  ASSERT_TRUE(db_->Get(ropts, "k2", &value).IsNotFound());
+
   Close();
 }
 
@@ -590,6 +606,19 @@ TEST_F(DBBasicTestWithTimestamp, TrimHistoryTest) {
   check_value_by_ts(db_, "k1", Timestamp(7, 0), Status::OK(), "v2",
                     Timestamp(4, 0));
   Close();
+
+  Reopen(options);
+  ASSERT_OK(db_->DeleteRange(WriteOptions(), db_->DefaultColumnFamily(), "k1",
+                             "k3", Timestamp(7, 0)));
+  check_value_by_ts(db_, "k1", Timestamp(8, 0), Status::NotFound(), "",
+                    Timestamp(7, 0));
+  Close();
+  // Trim data whose timestamp > Timestamp(6, 0), read(k1, ts(8)) <- v2
+  ASSERT_OK(DB::OpenAndTrimHistory(db_options, dbname_, column_families,
+                                   &handles_, &db_, Timestamp(6, 0)));
+  check_value_by_ts(db_, "k1", Timestamp(8, 0), Status::OK(), "v2",
+                    Timestamp(4, 0));
+  Close();
 }
 
 TEST_F(DBBasicTestWithTimestamp, OpenAndTrimHistoryInvalidOptionTest) {
@@ -2014,7 +2043,7 @@ constexpr int DataVisibilityTest::kTestDataSize;
 //                               seq'=11
 //                               write finishes
 //         GetImpl(ts,seq)
-// It is OK to return <k, t1, s1> if ts>=t1 AND seq>=s1. If ts>=1t1 but seq<s1,
+// It is OK to return <k, t1, s1> if ts>=t1 AND seq>=s1. If ts>=t1 but seq<s1,
 // the key should not be returned.
 TEST_F(DataVisibilityTest, PointLookupWithoutSnapshot1) {
   Options options = CurrentOptions();
@@ -3249,6 +3278,418 @@ TEST_F(UpdateFullHistoryTsLowTest, ConcurrentUpdate) {
   Close();
 }
 
+TEST_F(DBBasicTestWithTimestamp,
+       GCPreserveRangeTombstoneWhenNoOrSmallFullHistoryLow) {
+  Options options = CurrentOptions();
+  options.env = env_;
+  options.create_if_missing = true;
+  const size_t kTimestampSize = Timestamp(0, 0).size();
+  TestComparator test_cmp(kTimestampSize);
+  options.comparator = &test_cmp;
+  DestroyAndReopen(options);
+
+  std::string ts_str = Timestamp(1, 0);
+  WriteOptions wopts;
+  ASSERT_OK(db_->Put(wopts, "k1", ts_str, "v1"));
+  ASSERT_OK(db_->Put(wopts, "k2", ts_str, "v2"));
+  ASSERT_OK(db_->Put(wopts, "k3", ts_str, "v3"));
+  ts_str = Timestamp(2, 0);
+  ASSERT_OK(
+      db_->DeleteRange(wopts, db_->DefaultColumnFamily(), "k1", "k3", ts_str));
+
+  ts_str = Timestamp(3, 0);
+  Slice ts = ts_str;
+  ReadOptions ropts;
+  ropts.timestamp = &ts;
+  CompactRangeOptions cro;
+  cro.full_history_ts_low = nullptr;
+  std::string value, key_ts;
+  Status s;
+  auto verify = [&] {
+    s = db_->Get(ropts, "k1", &value);
+    ASSERT_TRUE(s.IsNotFound());
+
+    s = db_->Get(ropts, "k2", &value, &key_ts);
+    ASSERT_TRUE(s.IsNotFound());
+    ASSERT_EQ(key_ts, Timestamp(2, 0));
+
+    ASSERT_OK(db_->Get(ropts, "k3", &value, &key_ts));
+    ASSERT_EQ(value, "v3");
+    ASSERT_EQ(Timestamp(1, 0), key_ts);
+
+    size_t batch_size = 3;
+    std::vector<std::string> key_strs = {"k1", "k2", "k3"};
+    std::vector<Slice> keys{key_strs.begin(), key_strs.end()};
+    std::vector<PinnableSlice> values(batch_size);
+    std::vector<Status> statuses(batch_size);
+    db_->MultiGet(ropts, db_->DefaultColumnFamily(), batch_size, keys.data(),
+                  values.data(), statuses.data(), true /* sorted_input */);
+    ASSERT_TRUE(statuses[0].IsNotFound());
+    ASSERT_TRUE(statuses[1].IsNotFound());
+    ASSERT_OK(statuses[2]);
+    ;
+    ASSERT_EQ(values[2], "v3");
+  };
+  verify();
+  ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
+  verify();
+  std::string lb = Timestamp(0, 0);
+  Slice lb_slice = lb;
+  cro.full_history_ts_low = &lb_slice;
+  ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
+  verify();
+  Close();
+}
+
+TEST_F(DBBasicTestWithTimestamp,
+       GCRangeTombstonesAndCoveredKeysRespectingTslow) {
+  Options options = CurrentOptions();
+  options.env = env_;
+  options.create_if_missing = true;
+  BlockBasedTableOptions bbto;
+  bbto.filter_policy.reset(NewBloomFilterPolicy(10, false));
+  bbto.cache_index_and_filter_blocks = true;
+  bbto.whole_key_filtering = true;
+  options.table_factory.reset(NewBlockBasedTableFactory(bbto));
+  const size_t kTimestampSize = Timestamp(0, 0).size();
+  TestComparator test_cmp(kTimestampSize);
+  options.comparator = &test_cmp;
+  options.num_levels = 2;
+  DestroyAndReopen(options);
+
+  WriteOptions wopts;
+  ASSERT_OK(db_->Put(wopts, "k1", Timestamp(1, 0), "v1"));
+  ASSERT_OK(db_->Delete(wopts, "k2", Timestamp(2, 0)));
+  ASSERT_OK(db_->DeleteRange(wopts, db_->DefaultColumnFamily(), "k1", "k3",
+                             Timestamp(3, 0)));
+  ASSERT_OK(db_->Put(wopts, "k3", Timestamp(4, 0), "v3"));
+
+  ReadOptions ropts;
+  std::string read_ts = Timestamp(5, 0);
+  Slice read_ts_slice = read_ts;
+  ropts.timestamp = &read_ts_slice;
+  size_t batch_size = 3;
+  std::vector<std::string> key_strs = {"k1", "k2", "k3"};
+  std::vector<Slice> keys = {key_strs.begin(), key_strs.end()};
+  std::vector<PinnableSlice> values(batch_size);
+  std::vector<Status> statuses(batch_size);
+  std::vector<std::string> timestamps(batch_size);
+  db_->MultiGet(ropts, db_->DefaultColumnFamily(), batch_size, keys.data(),
+                values.data(), timestamps.data(), statuses.data(),
+                true /* sorted_input */);
+  ASSERT_TRUE(statuses[0].IsNotFound());
+  ASSERT_EQ(timestamps[0], Timestamp(3, 0));
+  ASSERT_TRUE(statuses[1].IsNotFound());
+  // DeleteRange has a higher timestamp than Delete for "k2"
+  ASSERT_EQ(timestamps[1], Timestamp(3, 0));
+  ASSERT_OK(statuses[2]);
+  ASSERT_EQ(values[2], "v3");
+  ASSERT_EQ(timestamps[2], Timestamp(4, 0));
+
+  CompactRangeOptions cro;
+  // Range tombstone has timestamp >= full_history_ts_low, covered keys
+  // are not dropped.
+  std::string compaction_ts_str = Timestamp(2, 0);
+  Slice compaction_ts = compaction_ts_str;
+  cro.full_history_ts_low = &compaction_ts;
+  cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
+  ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
+  ropts.timestamp = &compaction_ts;
+  std::string value, ts;
+  ASSERT_OK(db_->Get(ropts, "k1", &value, &ts));
+  ASSERT_EQ(value, "v1");
+  // timestamp is below full_history_ts_low, zeroed out as the key goes into
+  // bottommost level
+  ASSERT_EQ(ts, Timestamp(0, 0));
+  ASSERT_TRUE(db_->Get(ropts, "k2", &value, &ts).IsNotFound());
+  ASSERT_EQ(ts, Timestamp(2, 0));
+
+  compaction_ts_str = Timestamp(4, 0);
+  compaction_ts = compaction_ts_str;
+  cro.full_history_ts_low = &compaction_ts;
+  ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
+  ropts.timestamp = &read_ts_slice;
+  // k1, k2 and the range tombstone should be dropped
+  // k3 should still exist
+  db_->MultiGet(ropts, db_->DefaultColumnFamily(), batch_size, keys.data(),
+                values.data(), timestamps.data(), statuses.data(),
+                true /* sorted_input */);
+  ASSERT_TRUE(statuses[0].IsNotFound());
+  ASSERT_TRUE(timestamps[0].empty());
+  ASSERT_TRUE(statuses[1].IsNotFound());
+  ASSERT_TRUE(timestamps[1].empty());
+  ASSERT_OK(statuses[2]);
+  ASSERT_EQ(values[2], "v3");
+  ASSERT_EQ(timestamps[2], Timestamp(4, 0));
+
+  Close();
+}
+
+TEST_P(DBBasicTestWithTimestampTableOptions, DeleteRangeBaiscReadAndIterate) {
+  const int kNum = 200, kRangeBegin = 50, kRangeEnd = 150, kNumPerFile = 25;
+  Options options = CurrentOptions();
+  options.prefix_extractor.reset(NewFixedPrefixTransform(3));
+  options.compression = kNoCompression;
+  BlockBasedTableOptions bbto;
+  bbto.index_type = GetParam();
+  bbto.block_size = 100;
+  options.table_factory.reset(NewBlockBasedTableFactory(bbto));
+  options.env = env_;
+  options.create_if_missing = true;
+  const size_t kTimestampSize = Timestamp(0, 0).size();
+  TestComparator test_cmp(kTimestampSize);
+  options.comparator = &test_cmp;
+  options.memtable_factory.reset(test::NewSpecialSkipListFactory(kNumPerFile));
+  DestroyAndReopen(options);
+
+  // Write half of the keys before the tombstone and half after the tombstone.
+  // Only covered keys (i.e., within the range and older than the tombstone)
+  // should be deleted.
+  for (int i = 0; i < kNum; ++i) {
+    if (i == kNum / 2) {
+      ASSERT_OK(db_->DeleteRange(WriteOptions(), db_->DefaultColumnFamily(),
+                                 Key1(kRangeBegin), Key1(kRangeEnd),
+                                 Timestamp(i, 0)));
+    }
+    ASSERT_OK(db_->Put(WriteOptions(), Key1(i), Timestamp(i, 0),
+                       "val" + std::to_string(i)));
+    if (i == kNum - kNumPerFile) {
+      ASSERT_OK(Flush());
+    }
+  }
+
+  ReadOptions read_opts;
+  read_opts.total_order_seek = true;
+  std::string read_ts = Timestamp(kNum, 0);
+  Slice read_ts_slice = read_ts;
+  read_opts.timestamp = &read_ts_slice;
+  {
+    std::unique_ptr<Iterator> iter(db_->NewIterator(read_opts));
+    ASSERT_OK(iter->status());
+
+    int expected = 0;
+    for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
+      ASSERT_EQ(Key1(expected), iter->key());
+      if (expected == kRangeBegin - 1) {
+        expected = kNum / 2;
+      } else {
+        ++expected;
+      }
+    }
+    ASSERT_EQ(kNum, expected);
+
+    expected = kNum / 2;
+    for (iter->Seek(Key1(kNum / 2)); iter->Valid(); iter->Next()) {
+      ASSERT_EQ(Key1(expected), iter->key());
+      ++expected;
+    }
+    ASSERT_EQ(kNum, expected);
+
+    expected = kRangeBegin - 1;
+    for (iter->SeekForPrev(Key1(kNum / 2 - 1)); iter->Valid(); iter->Prev()) {
+      ASSERT_EQ(Key1(expected), iter->key());
+      --expected;
+    }
+    ASSERT_EQ(-1, expected);
+
+    read_ts = Timestamp(0, 0);
+    read_ts_slice = read_ts;
+    read_opts.timestamp = &read_ts_slice;
+    iter.reset(db_->NewIterator(read_opts));
+    iter->SeekToFirst();
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ(iter->key(), Key1(0));
+    iter->Next();
+    ASSERT_FALSE(iter->Valid());
+    ASSERT_OK(iter->status());
+  }
+
+  read_ts = Timestamp(kNum, 0);
+  read_ts_slice = read_ts;
+  read_opts.timestamp = &read_ts_slice;
+  std::string value, timestamp;
+  Status s;
+  for (int i = 0; i < kNum; ++i) {
+    s = db_->Get(read_opts, Key1(i), &value, &timestamp);
+    if (i >= kRangeBegin && i < kNum / 2) {
+      ASSERT_TRUE(s.IsNotFound());
+      ASSERT_EQ(timestamp, Timestamp(kNum / 2, 0));
+    } else {
+      ASSERT_OK(s);
+      ASSERT_EQ(value, "val" + std::to_string(i));
+      ASSERT_EQ(timestamp, Timestamp(i, 0));
+    }
+  }
+
+  size_t batch_size = kNum;
+  std::vector<std::string> key_strs(batch_size);
+  std::vector<Slice> keys(batch_size);
+  std::vector<PinnableSlice> values(batch_size);
+  std::vector<Status> statuses(batch_size);
+  std::vector<std::string> timestamps(batch_size);
+  for (int i = 0; i < kNum; ++i) {
+    key_strs[i] = Key1(i);
+    keys[i] = key_strs[i];
+  }
+  db_->MultiGet(read_opts, db_->DefaultColumnFamily(), batch_size, keys.data(),
+                values.data(), timestamps.data(), statuses.data(),
+                true /* sorted_input */);
+  for (int i = 0; i < kNum; ++i) {
+    if (i >= kRangeBegin && i < kNum / 2) {
+      ASSERT_TRUE(statuses[i].IsNotFound());
+      ASSERT_EQ(timestamps[i], Timestamp(kNum / 2, 0));
+    } else {
+      ASSERT_OK(statuses[i]);
+      ASSERT_EQ(values[i], "val" + std::to_string(i));
+      ASSERT_EQ(timestamps[i], Timestamp(i, 0));
+    }
+  }
+  Close();
+}
+
+TEST_F(DBBasicTestWithTimestamp, DeleteRangeGetIteratorWithSnapshot) {
+  // 4 keys 0, 1, 2, 3 at timestamps 0, 1, 2, 3 respectively.
+  // A range tombstone [1, 3) at timestamp 1 and has a sequence number between
+  // key 1 and 2.
+  Options options = CurrentOptions();
+  const size_t kTimestampSize = Timestamp(0, 0).size();
+  TestComparator test_cmp(kTimestampSize);
+  options.comparator = &test_cmp;
+  DestroyAndReopen(options);
+  WriteOptions write_opts;
+  std::string put_ts = Timestamp(0, 0);
+  const int kNum = 4, kNumPerFile = 1, kRangeBegin = 1, kRangeEnd = 3;
+  options.memtable_factory.reset(test::NewSpecialSkipListFactory(kNumPerFile));
+  const Snapshot* before_tombstone = nullptr;
+  const Snapshot* after_tombstone = nullptr;
+  for (int i = 0; i < kNum; ++i) {
+    ASSERT_OK(db_->Put(WriteOptions(), Key1(i), Timestamp(i, 0),
+                       "val" + std::to_string(i)));
+    if (i == kRangeBegin) {
+      before_tombstone = db_->GetSnapshot();
+      ASSERT_OK(db_->DeleteRange(WriteOptions(), db_->DefaultColumnFamily(),
+                                 Key1(kRangeBegin), Key1(kRangeEnd),
+                                 Timestamp(kRangeBegin, 0)));
+    }
+    if (i == kNum / 2) {
+      ASSERT_OK(Flush());
+    }
+  }
+  assert(before_tombstone);
+  after_tombstone = db_->GetSnapshot();
+  // snapshot and ts before tombstone
+  std::string read_ts_str = Timestamp(kRangeBegin - 1, 0);  // (0, 0)
+  Slice read_ts = read_ts_str;
+  ReadOptions read_opts;
+  read_opts.timestamp = &read_ts;
+  read_opts.snapshot = before_tombstone;
+  std::vector<Status> expected_status = {
+      Status::OK(), Status::NotFound(), Status::NotFound(), Status::NotFound()};
+  std::vector<std::string> expected_values(kNum);
+  expected_values[0] = "val" + std::to_string(0);
+  std::vector<std::string> expected_timestamps(kNum);
+  expected_timestamps[0] = Timestamp(0, 0);
+
+  size_t batch_size = kNum;
+  std::vector<std::string> key_strs(batch_size);
+  std::vector<Slice> keys(batch_size);
+  std::vector<PinnableSlice> values(batch_size);
+  std::vector<Status> statuses(batch_size);
+  std::vector<std::string> timestamps(batch_size);
+  for (int i = 0; i < kNum; ++i) {
+    key_strs[i] = Key1(i);
+    keys[i] = key_strs[i];
+  }
+
+  auto verify = [&] {
+    db_->MultiGet(read_opts, db_->DefaultColumnFamily(), batch_size,
+                  keys.data(), values.data(), timestamps.data(),
+                  statuses.data(), true /* sorted_input */);
+    std::string value, timestamp;
+    Status s;
+    for (int i = 0; i < kNum; ++i) {
+      s = db_->Get(read_opts, Key1(i), &value, &timestamp);
+      ASSERT_EQ(s, expected_status[i]);
+      ASSERT_EQ(statuses[i], expected_status[i]);
+      if (s.ok()) {
+        ASSERT_EQ(value, expected_values[i]);
+        ASSERT_EQ(values[i], expected_values[i]);
+      }
+      if (!timestamp.empty()) {
+        ASSERT_EQ(timestamp, expected_timestamps[i]);
+        ASSERT_EQ(timestamps[i], expected_timestamps[i]);
+      } else {
+        ASSERT_TRUE(timestamps[i].empty());
+      }
+    }
+    std::unique_ptr<Iterator> iter(db_->NewIterator(read_opts));
+    std::unique_ptr<Iterator> iter_for_seek(db_->NewIterator(read_opts));
+    iter->SeekToFirst();
+    for (int i = 0; i < kNum; ++i) {
+      if (expected_status[i].ok()) {
+        auto verify_iter = [&](Iterator* iter_ptr) {
+          ASSERT_TRUE(iter_ptr->Valid());
+          ASSERT_EQ(iter_ptr->key(), keys[i]);
+          ASSERT_EQ(iter_ptr->value(), expected_values[i]);
+          ASSERT_EQ(iter_ptr->timestamp(), expected_timestamps[i]);
+        };
+        verify_iter(iter.get());
+        iter->Next();
+
+        iter_for_seek->Seek(keys[i]);
+        verify_iter(iter_for_seek.get());
+
+        iter_for_seek->SeekForPrev(keys[i]);
+        verify_iter(iter_for_seek.get());
+      }
+    }
+    ASSERT_FALSE(iter->Valid());
+    ASSERT_OK(iter->status());
+  };
+
+  verify();
+
+  // snapshot before tombstone and ts after tombstone
+  read_ts_str = Timestamp(kNum, 0);  // (4, 0)
+  read_ts = read_ts_str;
+  read_opts.timestamp = &read_ts;
+  read_opts.snapshot = before_tombstone;
+  expected_status[1] = Status::OK();
+  expected_timestamps[1] = Timestamp(1, 0);
+  expected_values[1] = "val" + std::to_string(1);
+  verify();
+
+  // snapshot after tombstone and ts before tombstone
+  read_ts_str = Timestamp(kRangeBegin - 1, 0);  // (0, 0)
+  read_ts = read_ts_str;
+  read_opts.timestamp = &read_ts;
+  read_opts.snapshot = after_tombstone;
+  expected_status[1] = Status::NotFound();
+  expected_timestamps[1].clear();
+  expected_values[1].clear();
+  verify();
+
+  // snapshot and ts after tombstone
+  read_ts_str = Timestamp(kNum, 0);  // (4, 0)
+  read_ts = read_ts_str;
+  read_opts.timestamp = &read_ts;
+  read_opts.snapshot = after_tombstone;
+  for (int i = 0; i < kNum; ++i) {
+    if (i == kRangeBegin) {
+      expected_status[i] = Status::NotFound();
+      expected_values[i].clear();
+    } else {
+      expected_status[i] = Status::OK();
+      expected_values[i] = "val" + std::to_string(i);
+    }
+    expected_timestamps[i] = Timestamp(i, 0);
+  }
+  verify();
+
+  db_->ReleaseSnapshot(before_tombstone);
+  db_->ReleaseSnapshot(after_tombstone);
+  Close();
+}
 }  // namespace ROCKSDB_NAMESPACE
 
 int main(int argc, char** argv) {

db/db_write_test.cc

@@ -453,15 +453,15 @@ TEST_P(DBWriteTest, ManualWalFlushInEffect) {
   Reopen(options);
   // try the 1st WAL created during open
   ASSERT_TRUE(Put("key" + std::to_string(0), "value").ok());
-  ASSERT_TRUE(options.manual_wal_flush != dbfull()->TEST_WALBufferIsEmpty());
+  ASSERT_TRUE(options.manual_wal_flush != dbfull()->WALBufferIsEmpty());
   ASSERT_TRUE(dbfull()->FlushWAL(false).ok());
-  ASSERT_TRUE(dbfull()->TEST_WALBufferIsEmpty());
+  ASSERT_TRUE(dbfull()->WALBufferIsEmpty());
   // try the 2nd wal created during SwitchWAL
   ASSERT_OK(dbfull()->TEST_SwitchWAL());
   ASSERT_TRUE(Put("key" + std::to_string(0), "value").ok());
-  ASSERT_TRUE(options.manual_wal_flush != dbfull()->TEST_WALBufferIsEmpty());
+  ASSERT_TRUE(options.manual_wal_flush != dbfull()->WALBufferIsEmpty());
   ASSERT_TRUE(dbfull()->FlushWAL(false).ok());
-  ASSERT_TRUE(dbfull()->TEST_WALBufferIsEmpty());
+  ASSERT_TRUE(dbfull()->WALBufferIsEmpty());
 }
 
 TEST_P(DBWriteTest, UnflushedPutRaceWithTrackedWalSync) {
@@ -609,16 +609,16 @@ TEST_P(DBWriteTest, LockWalInEffect) {
   Reopen(options);
   // try the 1st WAL created during open
   ASSERT_OK(Put("key" + std::to_string(0), "value"));
-  ASSERT_TRUE(options.manual_wal_flush != dbfull()->TEST_WALBufferIsEmpty());
+  ASSERT_TRUE(options.manual_wal_flush != dbfull()->WALBufferIsEmpty());
   ASSERT_OK(dbfull()->LockWAL());
-  ASSERT_TRUE(dbfull()->TEST_WALBufferIsEmpty(false));
+  ASSERT_TRUE(dbfull()->WALBufferIsEmpty(false));
   ASSERT_OK(dbfull()->UnlockWAL());
   // try the 2nd wal created during SwitchWAL
   ASSERT_OK(dbfull()->TEST_SwitchWAL());
   ASSERT_OK(Put("key" + std::to_string(0), "value"));
-  ASSERT_TRUE(options.manual_wal_flush != dbfull()->TEST_WALBufferIsEmpty());
+  ASSERT_TRUE(options.manual_wal_flush != dbfull()->WALBufferIsEmpty());
   ASSERT_OK(dbfull()->LockWAL());
-  ASSERT_TRUE(dbfull()->TEST_WALBufferIsEmpty(false));
+  ASSERT_TRUE(dbfull()->WALBufferIsEmpty(false));
   ASSERT_OK(dbfull()->UnlockWAL());
 }
 

db/dbformat.cc

@@ -88,6 +88,19 @@ void AppendKeyWithMaxTimestamp(std::string* result, const Slice& key,
   result->append(kTsMax.data(), ts_sz);
 }
 
+void AppendUserKeyWithMaxTimestamp(std::string* result, const Slice& key,
+                                   size_t ts_sz) {
+  assert(ts_sz > 0);
+  result->append(key.data(), key.size() - ts_sz);
+
+  static constexpr char kTsMax[] = "\xff\xff\xff\xff\xff\xff\xff\xff\xff";
+  if (ts_sz < strlen(kTsMax)) {
+    result->append(kTsMax, ts_sz);
+  } else {
+    result->append(std::string(ts_sz, '\xff'));
+  }
+}
+
 std::string ParsedInternalKey::DebugString(bool log_err_key, bool hex) const {
   std::string result = "'";
   if (log_err_key) {

db/dbformat.h

@@ -182,6 +182,11 @@ extern void AppendKeyWithMinTimestamp(std::string* result, const Slice& key,
 extern void AppendKeyWithMaxTimestamp(std::string* result, const Slice& key,
                                       size_t ts_sz);
 
+// `key` is a user key with timestamp. Append the user key without timestamp
+// and the maximal timestamp to *result.
+extern void AppendUserKeyWithMaxTimestamp(std::string* result, const Slice& key,
+                                          size_t ts_sz);
+
 // Attempt to parse an internal key from "internal_key".  On success,
 // stores the parsed data in "*result", and returns true.
 //
@@ -290,6 +295,10 @@ class InternalKey {
   InternalKey(const Slice& _user_key, SequenceNumber s, ValueType t) {
     AppendInternalKey(&rep_, ParsedInternalKey(_user_key, s, t));
   }
+  InternalKey(const Slice& _user_key, SequenceNumber s, ValueType t, Slice ts) {
+    AppendInternalKeyWithDifferentTimestamp(
+        &rep_, ParsedInternalKey(_user_key, s, t), ts);
+  }
 
   // sets the internal key to be bigger or equal to all internal keys with this
   // user key
@@ -324,11 +333,24 @@ class InternalKey {
     SetFrom(ParsedInternalKey(_user_key, s, t));
   }
 
+  void Set(const Slice& _user_key_with_ts, SequenceNumber s, ValueType t,
+           const Slice& ts) {
+    ParsedInternalKey pik = ParsedInternalKey(_user_key_with_ts, s, t);
+    // Should not call pik.SetTimestamp() directly as it overwrites the buffer
+    // containing _user_key.
+    SetFrom(pik, ts);
+  }
+
   void SetFrom(const ParsedInternalKey& p) {
     rep_.clear();
     AppendInternalKey(&rep_, p);
   }
 
+  void SetFrom(const ParsedInternalKey& p, const Slice& ts) {
+    rep_.clear();
+    AppendInternalKeyWithDifferentTimestamp(&rep_, p, ts);
+  }
+
   void Clear() { rep_.clear(); }
 
   // The underlying representation.
@@ -518,7 +540,9 @@ class IterKey {
 
   bool IsKeyPinned() const { return (key_ != buf_); }
 
-  // user_key does not have timestamp.
+  // If `ts` is provided, user_key should not contain timestamp,
+  // and `ts` is appended after user_key.
+  // TODO: more efficient storage for timestamp.
   void SetInternalKey(const Slice& key_prefix, const Slice& user_key,
                       SequenceNumber s,
                       ValueType value_type = kValueTypeForSeek,
@@ -671,16 +695,38 @@ extern Status ReadRecordFromWriteBatch(Slice* input, char* tag,
 
 // When user call DeleteRange() to delete a range of keys,
 // we will store a serialized RangeTombstone in MemTable and SST.
-// the struct here is a easy-understood form
+// the struct here is an easy-understood form
 // start/end_key_ is the start/end user key of the range to be deleted
 struct RangeTombstone {
   Slice start_key_;
   Slice end_key_;
   SequenceNumber seq_;
+  // TODO: we should optimize the storage here when user-defined timestamp
+  //  is NOT enabled: they currently take up (16 + 32 + 32) bytes per tombstone.
+  Slice ts_;
+  std::string pinned_start_key_;
+  std::string pinned_end_key_;
+
   RangeTombstone() = default;
   RangeTombstone(Slice sk, Slice ek, SequenceNumber sn)
       : start_key_(sk), end_key_(ek), seq_(sn) {}
 
+  // User-defined timestamp is enabled, `sk` and `ek` should be user key
+  // with timestamp, `ts` will replace the timestamps in `sk` and
+  // `ek`.
+  RangeTombstone(Slice sk, Slice ek, SequenceNumber sn, Slice ts)
+      : seq_(sn), ts_(ts) {
+    assert(!ts.empty());
+    pinned_start_key_.reserve(sk.size());
+    pinned_start_key_.append(sk.data(), sk.size() - ts.size());
+    pinned_start_key_.append(ts.data(), ts.size());
+    pinned_end_key_.reserve(ek.size());
+    pinned_end_key_.append(ek.data(), ek.size() - ts.size());
+    pinned_end_key_.append(ts.data(), ts.size());
+    start_key_ = pinned_start_key_;
+    end_key_ = pinned_end_key_;
+  }
+
   RangeTombstone(ParsedInternalKey parsed_key, Slice value) {
     start_key_ = parsed_key.user_key;
     seq_ = parsed_key.sequence;
@@ -690,8 +736,7 @@ struct RangeTombstone {
   // be careful to use Serialize(), allocates new memory
   std::pair<InternalKey, Slice> Serialize() const {
     auto key = InternalKey(start_key_, seq_, kTypeRangeDeletion);
-    Slice value = end_key_;
-    return std::make_pair(std::move(key), std::move(value));
+    return std::make_pair(std::move(key), end_key_);
   }
 
   // be careful to use SerializeKey(), allocates new memory
@@ -707,6 +752,16 @@ struct RangeTombstone {
   //
   // be careful to use SerializeEndKey(), allocates new memory
   InternalKey SerializeEndKey() const {
+    if (!ts_.empty()) {
+      static constexpr char kTsMax[] = "\xff\xff\xff\xff\xff\xff\xff\xff\xff";
+      if (ts_.size() <= strlen(kTsMax)) {
+        return InternalKey(end_key_, kMaxSequenceNumber, kTypeRangeDeletion,
+                           Slice(kTsMax, ts_.size()));
+      } else {
+        return InternalKey(end_key_, kMaxSequenceNumber, kTypeRangeDeletion,
+                           std::string(ts_.size(), '\xff'));
+      }
+    }
     return InternalKey(end_key_, kMaxSequenceNumber, kTypeRangeDeletion);
   }
 };

db/dbformat_test.cc

@@ -206,6 +206,7 @@ TEST_F(FormatTest, RangeTombstoneSerializeEndKey) {
 }  // namespace ROCKSDB_NAMESPACE
 
 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
   ::testing::InitGoogleTest(&argc, argv);
   RegisterCustomObjects(argc, argv);
   return RUN_ALL_TESTS();

db/experimental.cc

@@ -94,7 +94,9 @@ Status UpdateManifestForFilesState(
 
         std::unique_ptr<FSSequentialFile> f;
         FileOptions fopts;
-        fopts.temperature = lf->temperature;
+        // Use kUnknown to signal the FileSystem to search all tiers for the
+        // file.
+        fopts.temperature = Temperature::kUnknown;
 
         IOStatus file_ios =
             fs->NewSequentialFile(fname, fopts, &f, /*dbg*/ nullptr);

db/external_sst_file_basic_test.cc

@@ -1472,7 +1472,7 @@ TEST_P(ExternalSSTFileBasicTest, IngestFileWithBadBlockChecksum) {
   SyncPoint::GetInstance()->DisableProcessing();
   SyncPoint::GetInstance()->ClearAllCallBacks();
   SyncPoint::GetInstance()->SetCallBack(
-      "BlockBasedTableBuilder::WriteRawBlock:TamperWithChecksum",
+      "BlockBasedTableBuilder::WriteMaybeCompressedBlock:TamperWithChecksum",
       change_checksum);
   SyncPoint::GetInstance()->EnableProcessing();
   int file_id = 0;

db/external_sst_file_ingestion_job.cc

@@ -339,9 +339,30 @@ Status ExternalSstFileIngestionJob::Prepare(
 Status ExternalSstFileIngestionJob::NeedsFlush(bool* flush_needed,
                                                SuperVersion* super_version) {
   autovector<Range> ranges;
-  for (const IngestedFileInfo& file_to_ingest : files_to_ingest_) {
-    ranges.emplace_back(file_to_ingest.smallest_internal_key.user_key(),
-                        file_to_ingest.largest_internal_key.user_key());
+  autovector<std::string> keys;
+  size_t ts_sz = cfd_->user_comparator()->timestamp_size();
+  if (ts_sz) {
+    // Check all ranges [begin, end] inclusively. Add maximum
+    // timestamp to include all `begin` keys, and add minimal timestamp to
+    // include all `end` keys.
+    for (const IngestedFileInfo& file_to_ingest : files_to_ingest_) {
+      std::string begin_str;
+      std::string end_str;
+      AppendUserKeyWithMaxTimestamp(
+          &begin_str, file_to_ingest.smallest_internal_key.user_key(), ts_sz);
+      AppendKeyWithMinTimestamp(
+          &end_str, file_to_ingest.largest_internal_key.user_key(), ts_sz);
+      keys.emplace_back(std::move(begin_str));
+      keys.emplace_back(std::move(end_str));
+    }
+    for (size_t i = 0; i < files_to_ingest_.size(); ++i) {
+      ranges.emplace_back(keys[2 * i], keys[2 * i + 1]);
+    }
+  } else {
+    for (const IngestedFileInfo& file_to_ingest : files_to_ingest_) {
+      ranges.emplace_back(file_to_ingest.smallest_internal_key.user_key(),
+                          file_to_ingest.largest_internal_key.user_key());
+    }
   }
   Status status = cfd_->RangesOverlapWithMemtables(
       ranges, super_version, db_options_.allow_data_in_errors, flush_needed);

db/fault_injection_test.cc

@@ -631,6 +631,7 @@ INSTANTIATE_TEST_CASE_P(
 }  // namespace ROCKSDB_NAMESPACE
 
 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
   ::testing::InitGoogleTest(&argc, argv);
   RegisterCustomObjects(argc, argv);
   return RUN_ALL_TESTS();

db/filename_test.cc

@@ -237,6 +237,7 @@ TEST_F(FileNameTest, NormalizePath) {
 }  // namespace ROCKSDB_NAMESPACE
 
 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
   ::testing::InitGoogleTest(&argc, argv);
   return RUN_ALL_TESTS();
 }

db/flush_job.cc

@@ -420,9 +420,11 @@ Status FlushJob::MemPurge() {
   // Place iterator at the First (meaning most recent) key node.
   iter->SeekToFirst();
 
+  const std::string* const full_history_ts_low = &(cfd_->GetFullHistoryTsLow());
   std::unique_ptr<CompactionRangeDelAggregator> range_del_agg(
       new CompactionRangeDelAggregator(&(cfd_->internal_comparator()),
-                                       existing_snapshots_));
+                                       existing_snapshots_,
+                                       full_history_ts_low));
   for (auto& rd_iter : range_del_iters) {
     range_del_agg->AddTombstones(std::move(rd_iter));
   }
@@ -479,8 +481,7 @@ Status FlushJob::MemPurge() {
         ioptions->enforce_single_del_contracts,
         /*manual_compaction_canceled=*/kManualCompactionCanceledFalse,
         /*compaction=*/nullptr, compaction_filter.get(),
-        /*shutting_down=*/nullptr, ioptions->info_log,
-        &(cfd_->GetFullHistoryTsLow()));
+        /*shutting_down=*/nullptr, ioptions->info_log, full_history_ts_low);
 
     // Set earliest sequence number in the new memtable
     // to be equal to the earliest sequence number of the

db/flush_job_test.cc

@@ -739,6 +739,7 @@ TEST_F(FlushJobTimestampTest, NoKeyExpired) {
 }  // namespace ROCKSDB_NAMESPACE
 
 int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
   ::testing::InitGoogleTest(&argc, argv);
   return RUN_ALL_TESTS();
 }