Fix Git Codex Review invocation

Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/14726

Add SstFileReader::ParseTableIteratorKey() so callers of NewTableIterator() have a public way to decode raw table keys without duplicating RocksDB internal-key layout. The implementation delegates to the existing public ParsedEntryInfo parser using the reader comparator.

Reviewed By: pdillinger

Differential Revision: D104584393

fbshipit-source-id: 98e21c4d6676fbba69e533376b3da67539dd8fad

.github/scripts/parse-claude-review.js

@@ -18,7 +18,7 @@ module.exports = function parseClaude({executionFile, conclusion, meta}) {
     if (meta.trigger !== 'auto') {
       return `*Requested by @${meta.reviewer}*`;
     }
-    const shortSha = meta.headSha.substring(0, 7);
+    const shortSha = meta.headSha ? meta.headSha.substring(0, 7) : 'unknown';
     if (meta.autoMode === 'early') {
       return `*Auto-triggered after CI reached the early-review threshold — reviewing commit ${
           shortSha}*`;

.github/scripts/parse-codex-review.js

@@ -22,7 +22,7 @@ module.exports = function parseCodex(
     if (meta.trigger !== 'auto') {
       return `*Requested by @${meta.reviewer}*`;
     }
-    const shortSha = meta.headSha.substring(0, 7);
+    const shortSha = meta.headSha ? meta.headSha.substring(0, 7) : 'unknown';
     if (meta.autoMode === 'early') {
       return `*Auto-triggered after CI reached the early-review threshold — reviewing commit ${
           shortSha}*`;

.github/workflows/ai-review-analysis.yml

@@ -531,7 +531,9 @@ jobs:
               core.warning(
                 `Could not parse classifier output: ${error.message}`);
             }
-            const budget = label === 'simple' ? '16000' : '32000';
+            // Keep Claude's thinking budget below the action's effective
+            // max_tokens so the API has room for the final review text.
+            const budget = label === 'simple' ? '16000' : '24000';
             core.info(
               `Complexity classified as: ${label} ` +
               `-> MAX_THINKING_TOKENS=${budget}`
@@ -549,13 +551,30 @@ jobs:
           CODEX_HOME: /tmp/codex-home
           CODEX_MODEL: ${{ inputs.default_model }}
         run: |
+          if [ -z "${OPENAI_API_KEY}" ]; then
+            echo "OPENAI_API_KEY not configured, skipping Codex review"
+            printf 'complex\n' > codex-complexity-output.txt
+            exit 0
+          fi
           mkdir -p "${CODEX_HOME}"
+          set +e
           codex exec \
             -m "${CODEX_MODEL}" \
             -c model_reasoning_effort="medium" \
             --dangerously-bypass-approvals-and-sandbox \
             --output-last-message codex-complexity-output.txt \
             - < /tmp/complexity_prompt.txt > codex-complexity.log 2>&1
+          exit_code=$?
+          set -e
+          if [ "${exit_code}" -ne 0 ]; then
+            echo "Codex complexity classification failed with exit code ${exit_code}"
+            echo "Defaulting to complex review budget"
+            if [ -s codex-complexity.log ]; then
+              echo "Last 100 lines of codex-complexity.log:"
+              tail -100 codex-complexity.log
+            fi
+            printf 'complex\n' > codex-complexity-output.txt
+          fi
 
       - name: Pick thinking budget (Codex)
         if: >-
@@ -695,14 +714,14 @@ jobs:
           CODEX_HOME: /tmp/codex-home
           CODEX_MODEL: ${{ inputs.default_model }}
           CODEX_REASONING_EFFORT: ${{ steps.budget_codex.outputs.effort }}
-          BASE_SHA: ${{ steps.pr_info.outputs.base_sha }}
-          PR_TITLE: ${{ steps.pr_info.outputs.pr_title }}
         run: |
+          if [ -z "${OPENAI_API_KEY}" ]; then
+            echo "OPENAI_API_KEY not configured, skipping Codex review"
+            exit 0
+          fi
           mkdir -p "${CODEX_HOME}"
           set +e
-          codex exec review \
-            --base "${BASE_SHA}" \
-            --title "${PR_TITLE}" \
+          codex exec \
             -m "${CODEX_MODEL}" \
             -c model_reasoning_effort="${CODEX_REASONING_EFFORT}" \
             --dangerously-bypass-approvals-and-sandbox \
@@ -853,6 +872,7 @@ jobs:
     runs-on: ubuntu-latest
     permissions:
       contents: read
+      pull-requests: read
 
     if: >-
       github.event_name == 'workflow_dispatch' ||
@@ -1093,9 +1113,19 @@ jobs:
             }
             const override = (process.env.OVERRIDE || '').trim();
             if (override && /^\d+$/.test(override)) {
-              core.info(`Using thinking_budget override: ${override}`);
+              const maxSafeBudget = 24000;
+              const requested = parseInt(override, 10);
+              const budget = Math.min(requested, maxSafeBudget).toString();
+              if (budget !== override) {
+                core.info(
+                  `Clamping thinking_budget override from ${override} ` +
+                  `to ${budget} so max_tokens can exceed thinking tokens.`
+                );
+              } else {
+                core.info(`Using thinking_budget override: ${budget}`);
+              }
               core.setOutput('label', 'override');
-              core.setOutput('tokens', override);
+              core.setOutput('tokens', budget);
               return;
             }
             let label = 'complex';
@@ -1113,7 +1143,9 @@ jobs:
               core.warning(
                 `Could not parse classifier output: ${error.message}`);
             }
-            const budget = label === 'simple' ? '16000' : '32000';
+            // Keep Claude's thinking budget below the action's effective
+            // max_tokens so the API has room for the final review text.
+            const budget = label === 'simple' ? '16000' : '24000';
             core.info(
               `Complexity classified as: ${label} ` +
               `-> MAX_THINKING_TOKENS=${budget}`
@@ -1131,13 +1163,30 @@ jobs:
           CODEX_HOME: /tmp/codex-home
           CODEX_MODEL: ${{ inputs.selected_model != '' && inputs.selected_model || inputs.default_model }}
         run: |
+          if [ -z "${OPENAI_API_KEY}" ]; then
+            echo "OPENAI_API_KEY not configured, skipping Codex review"
+            printf 'complex\n' > codex-complexity-output.txt
+            exit 0
+          fi
           mkdir -p "${CODEX_HOME}"
+          set +e
           codex exec \
             -m "${CODEX_MODEL}" \
             -c model_reasoning_effort="medium" \
             --dangerously-bypass-approvals-and-sandbox \
             --output-last-message codex-complexity-output.txt \
             - < /tmp/complexity_prompt.txt > codex-complexity.log 2>&1
+          exit_code=$?
+          set -e
+          if [ "${exit_code}" -ne 0 ]; then
+            echo "Codex complexity classification failed with exit code ${exit_code}"
+            echo "Defaulting to complex review budget"
+            if [ -s codex-complexity.log ]; then
+              echo "Last 100 lines of codex-complexity.log:"
+              tail -100 codex-complexity.log
+            fi
+            printf 'complex\n' > codex-complexity-output.txt
+          fi
 
       - name: Pick thinking budget (Codex)
         if: steps.request.outputs.skip != 'true' && inputs.provider == 'codex'
@@ -1287,8 +1336,6 @@ jobs:
           CODEX_HOME: /tmp/codex-home
           CODEX_MODEL: ${{ inputs.selected_model != '' && inputs.selected_model || inputs.default_model }}
           CODEX_REASONING_EFFORT: ${{ steps.manual_budget_codex.outputs.effort }}
-          BASE_SHA: ${{ steps.pr_info.outputs.base_sha }}
-          PR_TITLE: ${{ steps.pr_info.outputs.pr_title }}
           COMMAND_TYPE: ${{ steps.request.outputs.type }}
         run: |
           mkdir -p "${CODEX_HOME}"
@@ -1301,9 +1348,7 @@ jobs:
               --output-last-message codex-review-output.md \
               - < /tmp/prompt.txt > codex-review-execution.log 2>&1
           else
-            codex exec review \
-              --base "${BASE_SHA}" \
-              --title "${PR_TITLE}" \
+            codex exec \
               -m "${CODEX_MODEL}" \
               -c model_reasoning_effort="${CODEX_REASONING_EFFORT}" \
               --dangerously-bypass-approvals-and-sandbox \

.github/workflows/claude-review.yml

@@ -30,12 +30,11 @@ on:
         required: false
         type: choice
         options:
-          - claude-opus-4-7
           - claude-opus-4-6
           - claude-sonnet-4-6
-        default: claude-opus-4-7
+        default: claude-opus-4-6
       thinking_budget:
-        description: Override MAX_THINKING_TOKENS (blank = auto-classify)
+        description: Override MAX_THINKING_TOKENS (blank = auto-classify, capped at 24000)
         required: false
         type: string
         default: ""
@@ -61,7 +60,7 @@ jobs:
       query_command: /claude-query
       result_artifact_name: claude-review-result
       comment_file: claude-review-comment.md
-      default_model: claude-opus-4-7
+      default_model: claude-opus-4-6
       selected_model: ${{ github.event_name == 'workflow_dispatch' && inputs.model || '' }}
       classifier_model: claude-sonnet-4-6
       recovery_model: claude-sonnet-4-6

Makefile

@@ -384,6 +384,10 @@ endif
 # TSAN doesn't work well with jemalloc. If we're compiling with TSAN, we should use regular malloc.
 ifdef COMPILE_WITH_TSAN
 	DISABLE_JEMALLOC=1
+	# Use a suppressions file instead of the process-wide TSAN default
+	# suppressions hook, which belongs to the final application.
+	TSAN_OPTIONS?=suppressions=$(CURDIR)/tools/tsan_suppressions.txt
+	export TSAN_OPTIONS
 	EXEC_LDFLAGS += -fsanitize=thread
 	PLATFORM_CCFLAGS += -fsanitize=thread -fPIC -DFOLLY_SANITIZE_THREAD
 	PLATFORM_CXXFLAGS += -fsanitize=thread -fPIC -DFOLLY_SANITIZE_THREAD

build_tools/check-sources.sh

@@ -37,7 +37,7 @@ if [ "$?" != "1" ]; then
   BAD=1
 fi
 
-git grep -n -P "[\x80-\xFF]" -- ':!docs' ':!*.md'
+LC_ALL=C git grep -n $'[\x80-\xff]' -- ':!docs' ':!*.md' ':!.github'
 if [ "$?" != "1" ]; then
   echo '^^^^ Use only ASCII characters in source files'
   BAD=1

crash_test.mk

@@ -7,6 +7,13 @@ DB_STRESS_CMD?=./db_stress
 
 include common.mk
 
+ifdef COMPILE_WITH_TSAN
+	# Keep direct `make -f crash_test.mk COMPILE_WITH_TSAN=1 ...` runs
+	# aligned with the main Makefile's TSAN runtime options.
+	TSAN_OPTIONS?=suppressions=$(CURDIR)/tools/tsan_suppressions.txt
+	export TSAN_OPTIONS
+endif
+
 CRASHTEST_MAKE=$(MAKE) -f crash_test.mk
 CRASHTEST_PY=$(PYTHON) -u tools/db_crashtest.py --stress_cmd=$(DB_STRESS_CMD) --cleanup_cmd='$(DB_CLEANUP_CMD)' --destroy_db_initially=1
 

db/blob/blob_file_partition_manager.cc

@@ -786,13 +786,19 @@ Status BlobFilePartitionManager::ResolveBlobDirectWriteIndex(
     }
   }
 
-  Status s = version != nullptr ? Status::Corruption("Invalid blob file number")
-                                : Status::NotFound();
-
   if (blob_file_cache == nullptr) {
-    return s;
+    return version != nullptr ? Status::Corruption("Invalid blob file number")
+                              : Status::NotFound();
   }
 
+  if (read_options.read_tier == kBlockCacheTier) {
+    // The direct-write fallback below may need to open the blob file reader,
+    // which `kBlockCacheTier` forbids. Keep the normal Version-backed path
+    // above eligible for cache-only hits.
+    return Status::Incomplete("Cannot read blob(s): no disk I/O allowed");
+  }
+
+  Status s;
   CacheHandleGuard<BlobFileReader> reader;
   s = blob_file_cache->GetBlobFileReader(read_options, blob_idx.file_number(),
                                          &reader,

db/blob/blob_write_batch_transformer.cc

@@ -278,7 +278,7 @@ Status BlobWriteBatchTransformer::MergeCF(uint32_t column_family_id,
 Status BlobWriteBatchTransformer::PutBlobIndexCF(uint32_t column_family_id,
                                                  const Slice& key,
                                                  const Slice& value) {
-  // Already a blob index — pass through unchanged.
+  // Already a blob index -- pass through unchanged.
   return WriteBatchInternal::PutBlobIndex(output_batch_, column_family_id, key,
                                           value);
 }

db/blob/blob_write_batch_transformer.h

@@ -35,7 +35,7 @@ struct BlobDirectWriteSettings {
   CompressionType compression_type = kNoCompression;
   // Compression options for newly written blob records.
   CompressionOptions compression_opts;
-  // Raw pointer — the Cache is owned by ColumnFamilyOptions and outlives all
+  // Raw pointer -- the Cache is owned by ColumnFamilyOptions and outlives all
   // settings snapshots. Using raw avoids 2 atomic ref-count ops per Put().
   Cache* blob_cache = nullptr;
   // Blob-cache prepopulation policy for direct-write records.

db/blob/db_blob_basic_test.cc

@@ -1620,6 +1620,47 @@ TEST_P(DBBlobBasicIOErrorTest, GetBlob_IOError) {
   SyncPoint::GetInstance()->ClearAllCallBacks();
 }
 
+TEST_P(DBBlobBasicIOErrorTest, GetEntityMergeWithBlobBaseIOError) {
+  // Goal: verify GetEntity preserves injected blob-read IOErrors when merge
+  // reads a blob-backed base value, instead of laundering them into Corruption.
+  // The test writes a blob-backed base value plus a merge operand, then injects
+  // an IOError at blob read time and checks both GetEntity and Get see it.
+  Options options;
+  options.env = fault_injection_env_.get();
+  options.enable_blob_files = true;
+  options.min_blob_size = 0;
+  options.merge_operator = MergeOperators::CreateStringAppendOperator();
+
+  Reopen(options);
+
+  constexpr char key[] = "key";
+  constexpr char base_value[] = "base_value";
+
+  ASSERT_OK(Put(key, base_value));
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(Merge(key, "merge_operand"));
+  ASSERT_OK(Flush());
+
+  SyncPoint::GetInstance()->SetCallBack(sync_point_, [this](void* /* arg */) {
+    fault_injection_env_->SetFilesystemActive(false,
+                                              Status::IOError(sync_point_));
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  PinnableWideColumns entity_result;
+  Status s = db_->GetEntity(ReadOptions(), db_->DefaultColumnFamily(), key,
+                            &entity_result);
+  ASSERT_TRUE(s.IsIOError()) << "Expected IOError but got: " << s.ToString();
+
+  PinnableSlice get_result;
+  s = db_->Get(ReadOptions(), db_->DefaultColumnFamily(), key, &get_result);
+  ASSERT_TRUE(s.IsIOError()) << "Expected IOError but got: " << s.ToString();
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
 TEST_P(DBBlobBasicIOErrorMultiGetTest, MultiGetBlobs_IOError) {
   Options options = GetDefaultOptions();
   options.env = fault_injection_env_.get();

db/blob/db_blob_index_test.cc

@@ -128,6 +128,16 @@ class DBBlobIndexTest : public DBTestBase {
         columns, s, is_blob_index, value_found);
   }
 
+  bool MaybeResolveMemtableBlobValueForTest(const Slice& key,
+                                            const BlobFetcher* blob_fetcher,
+                                            PinnableSlice* value,
+                                            PinnableWideColumns* columns,
+                                            Status* s, bool* is_blob_index,
+                                            bool* value_found = nullptr) {
+    return DBImpl::MaybeResolveMemtableBlobValue(
+        key, blob_fetcher, value, columns, s, is_blob_index, value_found);
+  }
+
   Options GetTestOptions() {
     Options options;
     options.env = CurrentOptions().env;
@@ -271,6 +281,55 @@ TEST_F(DBBlobIndexTest,
   ASSERT_EQ(static_cast<char>(BlobIndex::Type::kUnknown), blob_index.front());
 }
 
+TEST_F(DBBlobIndexTest,
+       MaybeResolveMemtableBlobValueWithoutFetcherFailsClosed) {
+  // Goal: if a readonly/secondary memtable hit produces a blob-backed payload
+  // but no BlobFetcher is available, the helper must fail closed instead of
+  // handing raw blob-index bytes back to the caller as if they were the value.
+  std::string blob_index;
+  BlobIndex::EncodeBlob(&blob_index, /*file_number=*/123, /*offset=*/456,
+                        /*size=*/789, kNoCompression);
+
+  PinnableSlice value;
+  value.GetSelf()->assign(blob_index.data(), blob_index.size());
+  value.PinSelf();
+
+  Status s = Status::OK();
+  bool is_blob_index = true;
+  ASSERT_TRUE(MaybeResolveMemtableBlobValueForTest(
+      Slice("key"), /*blob_fetcher=*/nullptr, &value, /*columns=*/nullptr, &s,
+      &is_blob_index));
+
+  ASSERT_TRUE(s.IsNotSupported()) << s.ToString();
+  ASSERT_TRUE(value.empty());
+  ASSERT_FALSE(is_blob_index);
+}
+
+TEST_F(DBBlobIndexTest, ReadOnlyGetImplReturnsBlobIndexWhenRequested) {
+  // Goal: cover the internal read-only GetImpl contract when the caller
+  // explicitly asks for raw blob-index bytes via `is_blob_index`. Recovery
+  // keeps the blob index in the memtable, and the read-only path must preserve
+  // the encoded index instead of eagerly resolving or rejecting it.
+  Options options = GetTestOptions();
+
+  DestroyAndReopen(options);
+
+  std::string blob_index;
+  BlobIndex::EncodeInlinedTTL(&blob_index, /*expiration=*/9876543210, "blob");
+
+  WriteBatch batch;
+  ASSERT_OK(PutBlobIndex(&batch, "blob_key", blob_index));
+  ASSERT_OK(Write(&batch));
+
+  Close();
+  options.avoid_flush_during_recovery = true;
+  ASSERT_OK(ReadOnlyReopen(options));
+
+  bool is_blob_index = false;
+  ASSERT_EQ(blob_index, GetImpl("blob_key", &is_blob_index));
+  ASSERT_TRUE(is_blob_index);
+}
+
 class PlainBlobValueFilterV3 : public CompactionFilter {
  public:
   PlainBlobValueFilterV3(std::atomic<int>* filter_call_count,
@@ -2138,7 +2197,7 @@ TEST_F(DBBlobIndexTest, EntityBlobFilterV3Remove) {
       db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(), key, columns));
   ASSERT_OK(Flush());
 
-  // Compact to trigger the filter — entity should be removed
+  // Compact to trigger the filter -- entity should be removed
   ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
 
   // Key should be gone

db/column_family_test.cc

@@ -734,6 +734,42 @@ TEST_P(ColumnFamilyTest, AddDrop) {
               std::vector<std::string>({"default", "four", "three"}));
 }
 
+TEST_P(ColumnFamilyTest, EmptyNameRejected) {
+  // Creating a column family with an empty-string name should be rejected with
+  // InvalidArgument. Empty strings are reserved by various RocksDB APIs and
+  // serialization formats to mean "unknown/unspecified column family" (e.g.
+  // TablePropertiesCollectorFactory::Context::kUnknownColumnFamily, table
+  // properties).
+  //
+  // Prior behavior was silently broken: CreateColumnFamily("") would return
+  // OK and a usable handle, but the CF was not persisted in the manifest --
+  // data written to it was lost on DB reopen.
+  Open();
+  ColumnFamilyHandle* handle = nullptr;
+  Status s =
+      db_->CreateColumnFamily(column_family_options_, std::string(), &handle);
+  ASSERT_TRUE(s.IsInvalidArgument()) << s.ToString();
+  ASSERT_EQ(handle, nullptr);
+
+  // Same via the bulk-by-names API. The first valid name should still get
+  // created; the failure should come on the empty one.
+  std::vector<ColumnFamilyHandle*> handles;
+  s = db_->CreateColumnFamilies(column_family_options_,
+                                {"ok_name", "", "another"}, &handles);
+  ASSERT_TRUE(s.IsInvalidArgument()) << s.ToString();
+  // Clean up the partially-created handle ("ok_name") before the failure.
+  ASSERT_EQ(handles.size(), 1U);
+  ASSERT_OK(db_->DropColumnFamily(handles[0]));
+  ASSERT_OK(db_->DestroyColumnFamilyHandle(handles[0]));
+  handles.clear();
+
+  // Same via the bulk-by-descriptor API.
+  s = db_->CreateColumnFamilies(
+      {ColumnFamilyDescriptor("", column_family_options_)}, &handles);
+  ASSERT_TRUE(s.IsInvalidArgument()) << s.ToString();
+  ASSERT_TRUE(handles.empty());
+}
+
 TEST_P(ColumnFamilyTest, BulkAddDrop) {
   constexpr int kNumCF = 1000;
   ColumnFamilyOptions cf_options;

db/compaction/compaction_iterator.h

@@ -75,7 +75,7 @@ class CompactionBlobResolver : public WideColumnBlobResolver {
   bool track_resolve_error_ = false;
 
   // Cache for resolved blob values to avoid re-fetching. Uses a vector of
-  // (column_index, PinnableSlice) pairs — typical entities have few blob
+  // (column_index, PinnableSlice) pairs -- typical entities have few blob
   // columns (<5), making linear scan cheaper than hash map overhead.
   std::vector<std::pair<size_t, std::unique_ptr<PinnableSlice>>>
       resolved_cache_;

db/compaction/compaction_picker_fifo.cc

@@ -251,7 +251,7 @@ Compaction* FIFOCompactionPicker::PickSizeCompaction(
     // When using blob-aware sizing, use proportional estimation (same
     // principle as EstimateTotalDataForSST): each SST "owns"
     // effective_size / num_files of total data. This is an approximation
-    // — individual SSTs may reference different amounts of blob data,
+    // -- individual SSTs may reference different amounts of blob data,
     // but uniform distribution is a reasonable estimate for FIFO dropping.
     uint64_t remaining_size = effective_size;
     const uint64_t num_files = last_level_files.size();
@@ -477,12 +477,12 @@ Compaction* FIFOCompactionPicker::PickIntraL0Compaction(
     if (fifo_opts.max_data_files_size == 0) {
       ROCKS_LOG_BUFFER(
           log_buffer,
-          "[%s] FIFO kv-ratio compaction: skipping — "
+          "[%s] FIFO kv-ratio compaction: skipping -- "
           "max_data_files_size is 0, cannot compute target file size. ",
           cf_name.c_str());
     } else if (fifo_opts.max_data_files_size < fifo_opts.max_table_files_size) {
       ROCKS_LOG_BUFFER(log_buffer,
-                       "[%s] FIFO kv-ratio compaction: skipping — "
+                       "[%s] FIFO kv-ratio compaction: skipping -- "
                        "max_data_files_size (%" PRIu64
                        ") < max_table_files_size "
                        "(%" PRIu64 ").",
@@ -547,7 +547,7 @@ Compaction* FIFOCompactionPicker::PickRatioBasedIntraL0Compaction(
   for (int level = 1; level < vstorage->num_levels(); ++level) {
     if (!vstorage->LevelFiles(level).empty()) {
       ROCKS_LOG_BUFFER(log_buffer,
-                       "[%s] FIFO kv-ratio compaction: skipping — non-L0 "
+                       "[%s] FIFO kv-ratio compaction: skipping -- non-L0 "
                        "level %d still has %" ROCKSDB_PRIszt
                        " files (migration in progress)",
                        cf_name.c_str(), level,
@@ -587,7 +587,7 @@ Compaction* FIFOCompactionPicker::PickRatioBasedIntraL0Compaction(
   // once per flush or compaction completion, so no caching is needed.
   uint64_t target = 0;
   if (mutable_cf_options.max_compaction_bytes > 0) {
-    // User explicitly set max_compaction_bytes — use it as target
+    // User explicitly set max_compaction_bytes -- use it as target
     target = mutable_cf_options.max_compaction_bytes;
   } else {
     // Auto-calculate from capacity and observed SST/blob ratio
@@ -642,7 +642,7 @@ Compaction* FIFOCompactionPicker::PickRatioBasedIntraL0Compaction(
   //   of linear) write amplification.
 
   // Build tier boundaries from smallest to largest.
-  // Stop at 10KB minimum — SST files of most workloads are larger than
+  // Stop at 10KB minimum -- SST files of most workloads are larger than
   // this, so lower boundaries would only waste CPU scanning L0 files.
   // Files smaller than the lowest boundary simply merge at that boundary.
   static constexpr uint64_t kMinTierBoundary = 10 * 1024;  // 10KB
@@ -651,7 +651,7 @@ Compaction* FIFOCompactionPicker::PickRatioBasedIntraL0Compaction(
     boundaries.push_back(b);
   }
   if (boundaries.empty()) {
-    // target itself is below kMinTierBoundary — use target as the
+    // target itself is below kMinTierBoundary -- use target as the
     // sole boundary so we can still compact at the target size.
     boundaries.push_back(target);
   }
@@ -668,7 +668,7 @@ Compaction* FIFOCompactionPicker::PickRatioBasedIntraL0Compaction(
         continue;
       }
 
-      // Found a file < boundary — collect contiguous batch
+      // Found a file < boundary -- collect contiguous batch
       std::vector<FileMetaData*> batch;
       uint64_t accumulated = 0;
       size_t pos = scan;
@@ -713,7 +713,7 @@ Compaction* FIFOCompactionPicker::PickRatioBasedIntraL0Compaction(
         return c;
       }
 
-      // This batch wasn't enough — advance past it
+      // This batch wasn't enough -- advance past it
       scan = pos;
     }
   }

db/compaction/compaction_picker_test.cc

@@ -1101,7 +1101,7 @@ TEST_F(CompactionPickerTest, ReadTriggeredSkipsLastLevel) {
   Add(0, 1U, "150", "200", kFileSize, 0, 500, 550);
   Add(4, 3U, "301", "350", kFileSize, 0, 101, 150);
 
-  // File 3 is at the last non-empty level — should NOT be marked for
+  // File 3 is at the last non-empty level -- should NOT be marked for
   // read-triggered compaction. Bottommost file cleanup is handled
   // separately by ComputeBottommostFilesMarkedForCompaction().
   file_map_[3U].first->stats.num_collapsible_entry_reads_sampled.store(

db/compaction/compaction_service_job.cc

@@ -846,11 +846,16 @@ static std::unordered_map<std::string, OptionTypeInfo>
          {offsetof(struct InternalStats::CompactionStats, count),
           OptionType::kUInt64T, OptionVerificationType::kNormal,
           OptionTypeFlags::kNone}},
-        {"counts", OptionTypeInfo::Array<
-                       int, static_cast<int>(CompactionReason::kNumOfReasons)>(
-                       offsetof(struct InternalStats::CompactionStats, counts),
-                       OptionVerificationType::kNormal, OptionTypeFlags::kNone,
-                       {0, OptionType::kInt})},
+        {"counts",
+         OptionTypeInfo::Array<
+             /* In release 11.2, a new compaction reason was added. This broken
+              * the reader and writer. To unblock release 11.1, we temporarily
+              * reduce the count array size to the old one. TODO add a proper
+              * serialization and deserialization method. */
+             int, static_cast<int>(CompactionReason::kNumOfReasons) - 1>(
+             offsetof(struct InternalStats::CompactionStats, counts),
+             OptionVerificationType::kNormal, OptionTypeFlags::kNone,
+             {0, OptionType::kInt})},
 };
 
 static std::unordered_map<std::string, OptionTypeInfo>

db/db_basic_test.cc

@@ -102,6 +102,757 @@ TEST_F(DBBasicTest, OpenWhenOpen) {
   ASSERT_TRUE(strstr(s.getState(), "lock ") != nullptr);
 }
 
+namespace {
+// Helper that captures per-branch SkippedNoopEdit counts for tests.
+struct RecoveryOptimizationCounters {
+  std::atomic<int> setup_dbid{0};
+  std::atomic<int> per_cf{0};
+  std::atomic<int> wal_deletion{0};
+  std::atomic<int> next_file_number{0};
+
+  void Install() {
+    auto* sp = ROCKSDB_NAMESPACE::SyncPoint::GetInstance();
+    sp->SetCallBack("DBImpl::Recovery:SkippedNoopEdit:SetupDBId",
+                    [this](void*) { setup_dbid.fetch_add(1); });
+    sp->SetCallBack("DBImpl::Recovery:SkippedNoopEdit:PerCF",
+                    [this](void*) { per_cf.fetch_add(1); });
+    sp->SetCallBack("DBImpl::Recovery:SkippedNoopEdit:WalDeletion",
+                    [this](void*) { wal_deletion.fetch_add(1); });
+    sp->SetCallBack("DBImpl::Recovery:SkippedNoopEdit:NextFileNumber",
+                    [this](void*) { next_file_number.fetch_add(1); });
+    sp->EnableProcessing();
+  }
+  void Uninstall() {
+    auto* sp = ROCKSDB_NAMESPACE::SyncPoint::GetInstance();
+    sp->DisableProcessing();
+    sp->ClearAllCallBacks();
+  }
+};
+}  // namespace
+
+// optimize_manifest_for_recovery=true: a clean reopen of a flushed DB must
+// append fewer individual records to the MANIFEST than the default-off path.
+// Verified by counting AddRecord calls (one per VersionEdit written to the
+// MANIFEST log).
+TEST_F(DBBasicTest,
+       OptimizeManifestForRecoveryReducesManifestWritesOnCleanReopen) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  auto* sp = ROCKSDB_NAMESPACE::SyncPoint::GetInstance();
+
+  // Measure MANIFEST records with optimize_manifest_for_recovery OFF (default).
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k1", "v1"));
+  ASSERT_OK(Flush());
+  Close();
+
+  std::atomic<int> records_off{0};
+  std::atomic<int> next_file_skips_off{0};
+  sp->SetCallBack("VersionSet::ProcessManifestWrites:AddRecord",
+                  [&](void*) { records_off.fetch_add(1); });
+  sp->SetCallBack("DBImpl::Recovery:SkippedNoopEdit:NextFileNumber",
+                  [&](void*) { next_file_skips_off.fetch_add(1); });
+  sp->EnableProcessing();
+  Reopen(options);
+  sp->DisableProcessing();
+  sp->ClearAllCallBacks();
+  ASSERT_EQ("v1", Get("k1"));
+  int off_count = records_off.load();
+  ASSERT_GT(off_count, 0);
+  ASSERT_EQ(0, next_file_skips_off.load());
+  Close();
+
+  // Measure MANIFEST records with optimize_manifest_for_recovery ON.
+  options.optimize_manifest_for_recovery = true;
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k2", "v2"));
+  ASSERT_OK(Flush());
+  Close();
+
+  std::atomic<int> records_on{0};
+  std::atomic<int> next_file_skips_on{0};
+  sp->SetCallBack("VersionSet::ProcessManifestWrites:AddRecord",
+                  [&](void*) { records_on.fetch_add(1); });
+  sp->SetCallBack("DBImpl::Recovery:SkippedNoopEdit:NextFileNumber",
+                  [&](void*) { next_file_skips_on.fetch_add(1); });
+  sp->EnableProcessing();
+  Reopen(options);
+  sp->DisableProcessing();
+  sp->ClearAllCallBacks();
+  ASSERT_EQ("v2", Get("k2"));
+  int on_count = records_on.load();
+  ASSERT_GT(next_file_skips_on.load(), 0);
+  ASSERT_LT(on_count, off_count);
+}
+
+// When the per-CF log_number actually advances, the per-CF skip must NOT
+// fire -- the edit carries real information and must be emitted.
+TEST_F(DBBasicTest, OptimizeManifestForRecoveryEmitsPerCFWhenLogAdvances) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.optimize_manifest_for_recovery = true;
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  // Write data and close WITHOUT flushing -- the WAL has un-replayed
+  // records, so on reopen recovery flushes the memtable and the per-CF
+  // edit's log_number must advance past the prior WAL.
+  Close();
+
+  RecoveryOptimizationCounters counters;
+  counters.Install();
+  Reopen(options);
+  counters.Uninstall();
+
+  ASSERT_EQ(0, counters.per_cf.load());
+  ASSERT_EQ("v", Get("k"));
+}
+
+// With track_and_verify_wals_in_manifest=true, the wal_deletion edit
+// must STILL be emitted (the DeleteWalsBefore record is required).
+TEST_F(DBBasicTest, OptimizeManifestForRecoveryPreservesWalTracking) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.optimize_manifest_for_recovery = true;
+  options.track_and_verify_wals_in_manifest = true;
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  ASSERT_OK(Flush());
+  Close();
+
+  RecoveryOptimizationCounters counters;
+  counters.Install();
+  Reopen(options);
+  counters.Uninstall();
+
+  ASSERT_EQ(0, counters.wal_deletion.load());
+}
+
+// best_efforts_recovery requires a fresh MANIFEST + CURRENT to be
+// produced on every open (the salvage contract). Even with the option
+// on, none of the SkippedNoopEdit branches must fire under
+// best_efforts_recovery -- otherwise CURRENT can be left missing or
+// stale (regression caught by DBBasicTest.RecoverWithNoCurrentFile and
+// DBTest2.BestEffortsRecoveryWithSstUniqueIdVerification under an
+// option-on default).
+TEST_F(DBBasicTest, OptimizeManifestForRecoveryDisabledByBestEffortsRecovery) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.optimize_manifest_for_recovery = true;
+  options.best_efforts_recovery = true;
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  ASSERT_OK(Flush());
+  Close();
+
+  RecoveryOptimizationCounters counters;
+  counters.Install();
+  Reopen(options);
+  counters.Uninstall();
+
+  ASSERT_EQ(0, counters.setup_dbid.load());
+  ASSERT_EQ(0, counters.per_cf.load());
+  ASSERT_EQ(0, counters.wal_deletion.load());
+  ASSERT_EQ(0, counters.next_file_number.load());
+  ASSERT_EQ("v", Get("k"));
+}
+
+// Multi-column-family coverage: with one CF flushed and another not,
+// the un-flushed CF's edit must still be emitted.
+TEST_F(DBBasicTest, OptimizeManifestForRecoveryMultiCF) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.optimize_manifest_for_recovery = true;
+  CreateAndReopenWithCF({"pikachu", "raichu"}, options);
+  ASSERT_OK(Put(0, "k", "v0"));
+  ASSERT_OK(Put(1, "k", "v1"));
+  ASSERT_OK(Put(2, "k", "v2"));
+  ASSERT_OK(Flush(0));
+  // CF 1 and 2 keep dirty memtables -- recovery must emit their per-CF
+  // edits (log_number advance from WAL replay).
+  Close();
+
+  RecoveryOptimizationCounters counters;
+  counters.Install();
+  ReopenWithColumnFamilies({"default", "pikachu", "raichu"}, options);
+  counters.Uninstall();
+
+  ASSERT_EQ("v0", Get(0, "k"));
+  ASSERT_EQ("v1", Get(1, "k"));
+  ASSERT_EQ("v2", Get(2, "k"));
+}
+
+// MaybeUpdateNextFileNumber's seed value (next_file_number_) makes the
+// post-loop comparison trivially true on every clean recovery, causing a
+// no-op SetNextFile edit to be appended to the MANIFEST. With
+// optimize_manifest_for_recovery=true, that emission is gated and
+// the SkippedNoopEdit:NextFileNumber sync point fires in its place.
+TEST_F(DBBasicTest, OptimizeManifestForRecoverySkipsNextFileNumber) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.optimize_manifest_for_recovery = true;
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  ASSERT_OK(Flush());
+  Close();
+
+  RecoveryOptimizationCounters counters;
+  counters.Install();
+  Reopen(options);
+  counters.Uninstall();
+
+  ASSERT_EQ(1, counters.next_file_number.load());
+}
+
+// With option=true and a synthetic on-disk file whose number is at or
+// above next_file_number_, MaybeUpdateNextFileNumber MUST emit the
+// SetNextFile edit and advance the counter past the synthetic number.
+TEST_F(DBBasicTest,
+       OptimizeManifestForRecoveryEmitsNextFileNumberWhenJustified) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.optimize_manifest_for_recovery = true;
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  ASSERT_OK(Flush());
+
+  const uint64_t next_before_close =
+      dbfull()->GetVersionSet()->current_next_file_number();
+  Close();
+
+  // Drop an empty .sst with a number well above next_file_number into
+  // the DB dir. MaybeUpdateNextFileNumber must observe it and advance.
+  const uint64_t synthetic_number = next_before_close + 100;
+  const std::string synthetic_sst =
+      dbname_ + "/" + MakeTableFileName("", synthetic_number);
+  ASSERT_OK(WriteStringToFile(env_, "" /*data*/, synthetic_sst,
+                              /*should_sync=*/true));
+
+  RecoveryOptimizationCounters counters;
+  counters.Install();
+  Reopen(options);
+  counters.Uninstall();
+
+  ASSERT_EQ(0, counters.next_file_number.load());
+  ASSERT_GT(dbfull()->GetVersionSet()->current_next_file_number(),
+            synthetic_number);
+}
+
+// optimize_manifest_for_recovery=true: after a clean Put + Flush + Close, the
+// next Open's min_log_number_to_keep equals (max_wal_before_close + 1),
+// proving the close-time MANIFEST write took effect.
+TEST_F(DBBasicTest, OptimizeManifestForRecoveryAdvancesWalMarkersOnClose) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.optimize_manifest_for_recovery = true;
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  ASSERT_OK(Flush());
+
+  uint64_t max_wal_before_close = 0;
+  auto* sp = ROCKSDB_NAMESPACE::SyncPoint::GetInstance();
+  sp->SetCallBack("DBImpl::CloseHelper:CapturedMaxWal", [&](void* arg) {
+    max_wal_before_close = *static_cast<uint64_t*>(arg);
+  });
+  sp->EnableProcessing();
+  Close();
+  sp->DisableProcessing();
+  sp->ClearAllCallBacks();
+  ASSERT_GT(max_wal_before_close, 0u);
+
+  Reopen(options);
+  ASSERT_EQ(max_wal_before_close + 1,
+            dbfull()->GetVersionSet()->min_log_number_to_keep());
+  ASSERT_EQ("v", Get("k"));
+}
+
+// Shared-option matrix: default-off and disable-before-close should both fall
+// back to recovery-time MANIFEST work, while enable-through-close should avoid
+// MANIFEST appends on a clean reopen.
+TEST_F(DBBasicTest, OptimizeManifestForRecoveryCleanReopenMatrix) {
+  struct TestCase {
+    const char* name;
+    bool enable_on_open;
+    bool disable_before_close;
+    bool expect_close_write;
+  };
+  const TestCase test_cases[] = {
+      {"default_off", false, false, false},
+      {"enabled", true, false, true},
+      {"disabled_before_close", true, true, false},
+  };
+
+  auto* sp = ROCKSDB_NAMESPACE::SyncPoint::GetInstance();
+  for (const auto& test_case : test_cases) {
+    SCOPED_TRACE(test_case.name);
+    Options options = CurrentOptions();
+    options.create_if_missing = true;
+    options.optimize_manifest_for_recovery = test_case.enable_on_open;
+    DestroyAndReopen(options);
+    ASSERT_OK(Put("k", test_case.name));
+    ASSERT_OK(Flush());
+
+    if (test_case.disable_before_close) {
+      ASSERT_OK(dbfull()->SetDBOptions(
+          {{"optimize_manifest_for_recovery", "false"}}));
+    }
+
+    std::atomic<int> entered{0};
+    sp->SetCallBack("DBImpl::CloseHelper:WriteWalMarkersOnCloseEntered",
+                    [&](void*) { entered.fetch_add(1); });
+    sp->EnableProcessing();
+    Close();
+    sp->DisableProcessing();
+    sp->ClearAllCallBacks();
+
+    std::atomic<int> records{0};
+    RecoveryOptimizationCounters counters;
+    counters.Install();
+    sp->SetCallBack("VersionSet::ProcessManifestWrites:AddRecord",
+                    [&](void*) { records.fetch_add(1); });
+    Reopen(options);
+    counters.Uninstall();
+
+    if (test_case.expect_close_write) {
+      ASSERT_EQ(1, entered.load());
+      ASSERT_EQ(0, records.load());
+      ASSERT_GT(counters.per_cf.load(), 0);
+    } else {
+      ASSERT_EQ(0, entered.load());
+      ASSERT_GT(records.load(), 0);
+    }
+    ASSERT_EQ(test_case.name, Get("k"));
+    Close();
+  }
+}
+
+// allow_2pc=true: even with the option on, the close-time write must NOT
+// advance MinLogNumberToKeep / DeleteWalsBefore (2pc requires the WAL to
+// remain replayable for uncommitted prepared transactions).
+TEST_F(DBBasicTest, OptimizeManifestForRecoveryRespectsTwoPC) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.optimize_manifest_for_recovery = true;
+  options.allow_2pc = true;
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  ASSERT_OK(Flush());
+
+  const uint64_t min_log_before_close =
+      dbfull()->GetVersionSet()->min_log_number_to_keep();
+  Close();
+  Reopen(options);
+  ASSERT_EQ(min_log_before_close,
+            dbfull()->GetVersionSet()->min_log_number_to_keep());
+  ASSERT_EQ("v", Get("k"));
+}
+
+// Mixed-CF emptiness: on reopen, recovery should still have MANIFEST work to
+// do for the dirty CF while skipping per-CF recovery edits for the empty CFs
+// whose markers were persisted at close.
+TEST_F(DBBasicTest, OptimizeManifestForRecoverySkipsNonEmptyCF) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.optimize_manifest_for_recovery = true;
+  CreateAndReopenWithCF({"empty_cf", "dirty_cf"}, options);
+  ASSERT_OK(Put(1, "k", "v1"));
+  ASSERT_OK(Flush(1));
+  ASSERT_OK(Put(2, "k", "v2"));
+  Close();
+
+  RecoveryOptimizationCounters counters;
+  std::atomic<int> records{0};
+  auto* sp = ROCKSDB_NAMESPACE::SyncPoint::GetInstance();
+  counters.Install();
+  sp->SetCallBack("VersionSet::ProcessManifestWrites:AddRecord",
+                  [&](void*) { records.fetch_add(1); });
+  ReopenWithColumnFamilies({"default", "empty_cf", "dirty_cf"}, options);
+  counters.Uninstall();
+
+  ASSERT_GT(counters.per_cf.load(), 0);
+  ASSERT_GT(records.load(), 0);
+  ASSERT_EQ("v1", Get(1, "k"));
+  ASSERT_EQ("v2", Get(2, "k"));
+}
+
+// track_and_verify_wals_in_manifest=true: the close-time write must leave
+// recovery with only the mandatory WAL-tracking MANIFEST work. The per-CF
+// recovery edits should still skip.
+TEST_F(DBBasicTest, OptimizeManifestForRecoveryEmitsWalDeletionWhenTracking) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.optimize_manifest_for_recovery = true;
+  options.track_and_verify_wals_in_manifest = true;
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  ASSERT_OK(Flush());
+  Close();
+
+  RecoveryOptimizationCounters counters;
+  std::atomic<int> records{0};
+  auto* sp = ROCKSDB_NAMESPACE::SyncPoint::GetInstance();
+  counters.Install();
+  sp->SetCallBack("VersionSet::ProcessManifestWrites:AddRecord",
+                  [&](void*) { records.fetch_add(1); });
+  Reopen(options);
+  counters.Uninstall();
+
+  ASSERT_GT(counters.per_cf.load(), 0);
+  ASSERT_EQ(1, records.load());
+  ASSERT_GT(dbfull()->GetVersionSet()->GetWalSet().GetMinWalNumberToKeep(), 0u);
+  ASSERT_EQ("v", Get("k"));
+}
+
+// Regression for the close-time marker path: if a new WAL is created while an
+// otherwise-empty user CF stays empty, Close() must reserve the next file
+// number before persisting SetLogNumber(cur_wal + 1) for that CF.
+TEST_F(DBBasicTest, OptimizeManifestForRecoveryImmediateCloseAfterWarmReopen) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.optimize_manifest_for_recovery = true;
+  CreateAndReopenWithCF({"empty_cf"}, options);
+
+  ASSERT_OK(Put(0, "k", "v"));
+  const uint64_t wal_before_switch = dbfull()->TEST_LogfileNumber();
+  ASSERT_OK(dbfull()->TEST_SwitchMemtable());
+  ASSERT_GT(dbfull()->TEST_LogfileNumber(), wal_before_switch);
+
+  const uint64_t expected_new_log_num = dbfull()->TEST_LogfileNumber() + 1;
+  ASSERT_EQ(expected_new_log_num,
+            dbfull()->GetVersionSet()->current_next_file_number());
+
+  Close();
+  ReopenWithColumnFamilies({"default", "empty_cf"}, options);
+
+  ASSERT_GT(dbfull()->GetVersionSet()->current_next_file_number(),
+            expected_new_log_num);
+  ASSERT_EQ("v", Get(0, "k"));
+}
+
+// Dropped-CF safety: dropped column families must NOT have markers written for
+// them at close time (the !IsDropped() guard protects against attaching the
+// global edit to a dropped CF, which would later trip MANIFEST replay
+// assertions).
+TEST_F(DBBasicTest, OptimizeManifestForRecoverySkipsDroppedCF) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.optimize_manifest_for_recovery = true;
+  CreateAndReopenWithCF({"to_drop", "keeper"}, options);
+  ASSERT_OK(Put(1, "k", "v1"));
+  ASSERT_OK(Put(2, "k", "v2"));
+  ASSERT_OK(Flush(1));
+  ASSERT_OK(Flush(2));
+  ASSERT_OK(db_->DropColumnFamily(handles_[1]));
+  Close();
+
+  ReopenWithColumnFamilies({"default", "keeper"}, options);
+  ASSERT_EQ("v2", Get(1, "k"));
+}
+
+// reuse_manifest_on_open=true: the next LogAndApply after Recover must
+// append to the existing MANIFEST file instead of allocating a fresh
+// one. Force a write after reopen and verify the MANIFEST file number
+// stays unchanged.
+TEST_F(DBBasicTest, ReuseManifestOnOpenAppendsToExistingFile) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.reuse_manifest_on_open = true;
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  ASSERT_OK(Flush());
+  Close();
+
+  Reopen(options);
+  const uint64_t manifest_after_reopen =
+      dbfull()->TEST_Current_Manifest_FileNo();
+  ASSERT_OK(Put("k2", "v2"));
+  ASSERT_OK(Flush());
+  const uint64_t manifest_after_flush =
+      dbfull()->TEST_Current_Manifest_FileNo();
+
+  EXPECT_EQ(manifest_after_reopen, manifest_after_flush);
+  EXPECT_EQ("v", Get("k"));
+  EXPECT_EQ("v2", Get("k2"));
+}
+
+// Default off: ReopenManifestForAppend must NOT be invoked.
+TEST_F(DBBasicTest, ReuseManifestOnOpenDefaultOffSkipsReopen) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  // reuse_manifest_on_open defaults to false
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  ASSERT_OK(Flush());
+  Close();
+
+  std::atomic<int> reopened{0};
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "VersionSet::ReopenManifestForAppend:Reopened",
+      [&](void* /*arg*/) { reopened.fetch_add(1); });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  Reopen(options);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  ASSERT_EQ(0, reopened.load());
+}
+
+// Regression for the WritableFileWriter::filesize_=0 bug fixed via
+// constructor-time initial_file_size: after ReopenManifestForAppend binds a
+// writer to the existing MANIFEST, GetFileSize() must return the on-disk
+// size, not 0.
+// (If it returned 0, the size-limit check in ProcessManifestWrites would
+// compare against 0 and Close-time Truncate could shrink the file.)
+TEST_F(DBBasicTest, ReuseManifestOnOpenAdoptsOnDiskSize) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.reuse_manifest_on_open = true;
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  ASSERT_OK(Flush());
+  Close();
+
+  // Capture the on-disk MANIFEST size before reopen.
+  Reopen(options);
+  const uint64_t manifest_no = dbfull()->TEST_Current_Manifest_FileNo();
+  const std::string manifest_path = DescriptorFileName(dbname_, manifest_no);
+  uint64_t on_disk_size = 0;
+  ASSERT_OK(env_->GetFileSize(manifest_path, &on_disk_size));
+  ASSERT_GT(on_disk_size, 0u);
+
+  // Force a write to flush the writer's buffer; verify Close doesn't
+  // shrink the file (which would happen if Truncate(filesize_=0) ran).
+  ASSERT_OK(Put("k2", "v2"));
+  ASSERT_OK(Flush());
+  Close();
+
+  uint64_t after_close_size = 0;
+  ASSERT_OK(env_->GetFileSize(manifest_path, &after_close_size));
+  EXPECT_GE(after_close_size, on_disk_size);
+}
+
+// MANIFEST rotation continues to work under reuse: when the file grows
+// past tuned_max_manifest_file_size_, ProcessManifestWrites must rotate
+// to a fresh MANIFEST (same as legacy). Use the rotation SyncPoint to
+// observe rotation directly, since tuned_max_manifest_file_size_ is
+// auto-derived and not directly = max_manifest_file_size.
+TEST_F(DBBasicTest, ReuseManifestOnOpenStillRotatesOnSizeCap) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.reuse_manifest_on_open = true;
+  options.max_manifest_file_size = 1;
+  options.max_manifest_space_amp_pct = 0;  // disable amp-based tuning
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  ASSERT_OK(Flush());
+  Close();
+
+  std::atomic<int> rotations{0};
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "VersionSet::ProcessManifestWrites:BeforeNewManifest",
+      [&](void* /*arg*/) { rotations.fetch_add(1); });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  Reopen(options);
+  ASSERT_OK(Put("k2", "v2"));
+  ASSERT_OK(Flush());
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  // The Flush after Reopen must have triggered a rotation given the
+  // tiny size cap -- proves the size-driven rotation path still runs
+  // when descriptor_log_ is bound by reuse.
+  EXPECT_GT(rotations.load(), 0);
+  EXPECT_EQ("v", Get("k"));
+  EXPECT_EQ("v2", Get("k2"));
+}
+
+// Multi-CF reuse: append-mode MANIFEST must correctly handle edits
+// from multiple CFs after reopen.
+TEST_F(DBBasicTest, ReuseManifestOnOpenMultiCF) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.reuse_manifest_on_open = true;
+  CreateAndReopenWithCF({"alpha", "beta"}, options);
+  ASSERT_OK(Put(0, "k", "v0"));
+  ASSERT_OK(Put(1, "k", "v1"));
+  ASSERT_OK(Put(2, "k", "v2"));
+  ASSERT_OK(Flush(0));
+  ASSERT_OK(Flush(1));
+  ASSERT_OK(Flush(2));
+  Close();
+
+  ReopenWithColumnFamilies({"default", "alpha", "beta"}, options);
+  // Trigger more writes post-reopen on each CF -- these get appended to
+  // the reused MANIFEST.
+  ASSERT_OK(Put(0, "k2", "v0b"));
+  ASSERT_OK(Put(1, "k2", "v1b"));
+  ASSERT_OK(Put(2, "k2", "v2b"));
+  ASSERT_OK(Flush(0));
+  ASSERT_OK(Flush(1));
+  ASSERT_OK(Flush(2));
+
+  EXPECT_EQ("v0", Get(0, "k"));
+  EXPECT_EQ("v1", Get(1, "k"));
+  EXPECT_EQ("v2", Get(2, "k"));
+  EXPECT_EQ("v0b", Get(0, "k2"));
+  EXPECT_EQ("v1b", Get(1, "k2"));
+  EXPECT_EQ("v2b", Get(2, "k2"));
+}
+
+// Disabled under best_efforts_recovery: that mode rebuilds CURRENT and
+// MANIFEST as the side-effect of LogAndApplyForRecovery emitting an
+// edit; reusing the prior MANIFEST contradicts the salvage contract.
+TEST_F(DBBasicTest, ReuseManifestOnOpenDisabledByBestEffortsRecovery) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.reuse_manifest_on_open = true;
+  options.best_efforts_recovery = true;
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  ASSERT_OK(Flush());
+  Close();
+
+  std::atomic<int> reopened{0};
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "VersionSet::ReopenManifestForAppend:Reopened",
+      [&](void* /*arg*/) { reopened.fetch_add(1); });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  Reopen(options);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  ASSERT_EQ(0, reopened.load());
+  ASSERT_EQ("v", Get("k"));
+}
+
+// Full-stack composition: optimize_manifest_for_recovery plus
+// reuse_manifest_on_open. Verifies that Close writes recovery markers,
+// Reopen skips clean-recovery MANIFEST edits, and the MANIFEST is reused
+// instead of recreated for the next metadata update.
+TEST_F(DBBasicTest, ReuseManifestOnOpenFullStackComposition) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.optimize_manifest_for_recovery = true;
+  options.reuse_manifest_on_open = true;
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  ASSERT_OK(Flush());
+  Close();
+
+  // Capture the MANIFEST file number before reopen.
+  std::atomic<int> reopened{0};
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "VersionSet::ReopenManifestForAppend:Reopened",
+      [&](void* /*arg*/) { reopened.fetch_add(1); });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  Reopen(options);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  // All three composing: reuse fired, data is intact, no fresh MANIFEST.
+  EXPECT_EQ(1, reopened.load());
+  EXPECT_EQ("v", Get("k"));
+}
+
+// Direct unit test for WritableFileWriter's initial_file_size parameter:
+// verifies the visible size accessors report the existing on-disk size
+// immediately, rather than the constructor's zero default.
+TEST_F(DBBasicTest, WritableFileWriterInitialFileSizeAdoptsExistingSize) {
+  Env* env = Env::Default();
+  std::string fname = test::PerThreadDBPath("set_file_size_test");
+  ASSERT_OK(env->CreateDirIfMissing(test::TmpDir(env)));
+
+  // Create a file with some bytes, close it.
+  {
+    std::unique_ptr<WritableFile> raw;
+    ASSERT_OK(env->NewWritableFile(fname, &raw, EnvOptions()));
+    ASSERT_OK(raw->Append("hello world"));
+    ASSERT_OK(raw->Close());
+  }
+
+  // Reopen and wrap in a WritableFileWriter without initial_file_size --
+  // GetFileSize() should be 0 (the constructor's default).
+  std::unique_ptr<FSWritableFile> fs_file;
+  ASSERT_OK(env->GetFileSystem()->ReopenWritableFile(
+      fname, FileOptions(), &fs_file, /*dbg=*/nullptr));
+  std::unique_ptr<WritableFileWriter> writer(
+      new WritableFileWriter(std::move(fs_file), fname, FileOptions()));
+  EXPECT_EQ(0u, writer->GetFileSize());
+
+  // Reopen again and seed the writer's size accounting from the existing
+  // bytes. GetFileSize and GetFlushedSize must reflect it immediately.
+  ASSERT_OK(env->GetFileSystem()->ReopenWritableFile(
+      fname, FileOptions(), &fs_file, /*dbg=*/nullptr));
+  writer.reset(new WritableFileWriter(
+      std::move(fs_file), fname, FileOptions(),
+      /*clock=*/nullptr, /*io_tracer=*/nullptr, /*stats=*/nullptr,
+      Histograms::HISTOGRAM_ENUM_MAX, /*listeners=*/{},
+      /*file_checksum_gen_factory=*/nullptr,
+      /*perform_data_verification=*/false,
+      /*buffered_data_with_checksum=*/false,
+      /*initial_file_size=*/11));
+  EXPECT_EQ(11u, writer->GetFileSize());
+  EXPECT_EQ(11u, writer->GetFlushedSize());
+
+  ASSERT_OK(env->DeleteFile(fname));
+}
+
+// Tail corruption: appending garbage bytes to the MANIFEST after a
+// clean close must prevent reuse -- the physical size exceeds the
+// last valid record end.
+TEST_F(DBBasicTest, ReuseManifestOnOpenSkipsOnTailCorruption) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.reuse_manifest_on_open = true;
+  DestroyAndReopen(options);
+  ASSERT_OK(Put("k", "v"));
+  ASSERT_OK(Flush());
+  Close();
+
+  // Find the MANIFEST file and append garbage to it.
+  std::string manifest_path;
+  {
+    std::vector<std::string> files;
+    ASSERT_OK(env_->GetChildren(dbname_, &files));
+    for (const auto& f : files) {
+      uint64_t number;
+      FileType type;
+      if (ParseFileName(f, &number, &type) && type == kDescriptorFile) {
+        manifest_path = dbname_ + "/" + f;
+        break;
+      }
+    }
+  }
+  ASSERT_FALSE(manifest_path.empty());
+  {
+    std::string contents;
+    ASSERT_OK(ReadFileToString(env_, manifest_path, &contents));
+    contents.append("garbage!");
+    ASSERT_OK(WriteStringToFile(env_, contents, manifest_path));
+  }
+
+  std::atomic<int> reopened{0};
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "VersionSet::ReopenManifestForAppend:Reopened",
+      [&](void* /*arg*/) { reopened.fetch_add(1); });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  Reopen(options);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  ASSERT_EQ(0, reopened.load());
+  ASSERT_EQ("v", Get("k"));
+}
+
 TEST_F(DBBasicTest, EnableDirectIOWithZeroBuf) {
   if (!IsDirectIOSupported()) {
     ROCKSDB_GTEST_BYPASS("Direct IO not supported");

db/db_block_cache_test.cc

@@ -542,7 +542,7 @@ TEST_F(DBBlockCacheTest, WarmCacheWithDataBlocksDuringCompaction) {
   EXPECT_TRUE(tracking_cache->HasPriority(Cache::Priority::BOTTOM));
   EXPECT_FALSE(tracking_cache->HasPriority(Cache::Priority::LOW));
 
-  // Compaction output is in cache — reads should have zero misses.
+  // Compaction output is in cache -- reads should have zero misses.
   auto data_miss_before =
       options.statistics->getTickerCount(BLOCK_CACHE_DATA_MISS);
   ASSERT_EQ(value + "2", Get("key"));

db/db_compaction_test.cc

@@ -11833,7 +11833,7 @@ TEST_F(DBCompactionTest, RoundRobinCleanCutWithSharedBoundary) {
 // 2. A post-verification step fails (injected here via sync point), setting
 //    compact_->status to error while each subcompaction's status stays OK.
 // 3. SubcompactionState::Cleanup checks individual status (OK) and skips
-//    ReleaseObsolete — the cache entries leak.
+//    ReleaseObsolete -- the cache entries leak.
 // 4. FaultInjectionTestFS injects metadata read errors, causing GetChildren
 //    to fail in FindObsoleteFiles.
 // 5. Close()'s FindObsoleteFiles also fails to find the orphan for the same
@@ -11864,7 +11864,7 @@ TEST_F(DBCompactionTest, LeakedTableCacheEntryOnCompactionFailure) {
   // fail while individual subcompaction statuses stay OK (so Cleanup skips
   // ReleaseObsolete). The filesystem deactivation makes GetChildren fail
   // in FindObsoleteFiles, preventing the backstop from evicting the leaked
-  // cache entries — matching the crash test's metadata read fault injection.
+  // cache entries -- matching the crash test's metadata read fault injection.
   std::atomic<bool> inject_error{true};
   ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
       "CompactionJob::Run():AfterVerifyOutputFiles", [&](void* arg) {
@@ -11876,7 +11876,7 @@ TEST_F(DBCompactionTest, LeakedTableCacheEntryOnCompactionFailure) {
   // Enable metadata read fault injection on the bg compaction thread after
   // the compaction job finishes but before FindObsoleteFiles runs. This
   // makes GetChildren fail (metadata read), matching crash test's
-  // --open_metadata_read_fault_one_in=8. Only metadata reads fail —
+  // --open_metadata_read_fault_one_in=8. Only metadata reads fail --
   // logging and other IO operations continue normally.
   ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
       "BackgroundCallCompaction:1", [&](void*) {
@@ -11889,7 +11889,7 @@ TEST_F(DBCompactionTest, LeakedTableCacheEntryOnCompactionFailure) {
 
   ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
 
-  // Trigger compaction — fails after VerifyOutputFiles.
+  // Trigger compaction -- fails after VerifyOutputFiles.
   Status s = dbfull()->TEST_CompactRange(0, nullptr, nullptr);
   ASSERT_NOK(s);
 
@@ -11971,7 +11971,7 @@ TEST_F(DBCompactionTest, LeakedTableCacheEntryOnInstallFailure) {
 
   ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
 
-  // Trigger compaction — Run() succeeds but Install() fails.
+  // Trigger compaction -- Run() succeeds but Install() fails.
   Status s = dbfull()->TEST_CompactRange(0, nullptr, nullptr);
   ASSERT_NOK(s);
 
@@ -12032,7 +12032,7 @@ TEST_F(DBCompactionTest, ObsoleteFileTableCacheEntryWithConcurrentRef) {
       TableCache::Lookup(table_cache, target_file_number);
   ASSERT_NE(concurrent_handle, nullptr);
 
-  // Call ReleaseObsolete directly — this is what PurgeObsoleteFiles calls
+  // Call ReleaseObsolete directly -- this is what PurgeObsoleteFiles calls
   // when a file becomes obsolete. Pass nullptr for the handle to make
   // ReleaseObsolete do its own Lookup internally.
   TableCache::ReleaseObsolete(table_cache, target_file_number,
@@ -12049,6 +12049,88 @@ TEST_F(DBCompactionTest, ObsoleteFileTableCacheEntryWithConcurrentRef) {
   ASSERT_EQ(leaked, nullptr);
 }
 
+// Regression test for the atomic flush path leaking table cache entries when
+// InstallMemtableAtomicFlushResults (MANIFEST write) fails.
+// 1. atomic_flush is enabled with two column families.
+// 2. Both CFs have data, a flush is triggered.
+// 3. FlushJob::Run() succeeds for both (files are added to table cache by
+//    BuildTable), but write_manifest=false so no install happens in Run().
+// 4. The combined MANIFEST write (InstallMemtableAtomicFlushResults) fails
+//    due to injected I/O error.
+// 5. Without the fix, the files remain in the table cache but are not in
+//    any Version, causing TEST_VerifyNoObsoleteFilesCached to fire during
+//    Close() (especially when metadata_read_fault_one_in disables the
+//    FindObsoleteFiles backstop).
+TEST_F(DBCompactionTest, LeakedTableCacheEntryOnAtomicFlushInstallFailure) {
+  auto fault_fs = std::make_shared<FaultInjectionTestFS>(env_->GetFileSystem());
+  std::unique_ptr<Env> fault_env(NewCompositeEnv(fault_fs));
+
+  Options options = CurrentOptions();
+  options.env = fault_env.get();
+  options.atomic_flush = true;
+  options.create_if_missing = true;
+  options.disable_auto_compactions = true;
+  // Enough buffer so auto-flush doesn't trigger prematurely.
+  options.write_buffer_size = 64 << 20;
+  DestroyAndReopen(options);
+
+  // Create a second column family.
+  CreateAndReopenWithCF({"pikachu"}, options);
+  ASSERT_EQ(2, handles_.size());
+
+  // Write data to both CFs.
+  ASSERT_OK(Put(0, "key0", std::string(1024, 'a')));
+  ASSERT_OK(Put(1, "key1", std::string(1024, 'b')));
+
+  // Inject a write error during the combined MANIFEST write that happens
+  // inside InstallMemtableAtomicFlushResults (via LogAndApply).
+  std::atomic<bool> inject_error{true};
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "VersionSet::LogAndApply:WriteManifest", [&](void*) {
+        if (inject_error.exchange(false)) {
+          fault_fs->SetThreadLocalErrorContext(
+              FaultInjectionIOType::kWrite, /*seed=*/0, /*one_in=*/1,
+              /*retryable=*/false, /*has_data_loss=*/false);
+          fault_fs->EnableThreadLocalErrorInjection(
+              FaultInjectionIOType::kWrite);
+        }
+      });
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  // Trigger atomic flush - Run() succeeds but MANIFEST write fails.
+  FlushOptions flush_opts;
+  flush_opts.wait = true;
+  Status s = db_->Flush(flush_opts, handles_);
+  ASSERT_NOK(s);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  // Disable the write fault so Close() can proceed.
+  fault_fs->DisableThreadLocalErrorInjection(FaultInjectionIOType::kWrite);
+
+  // Enable metadata read fault injection on the main thread so that
+  // Close()'s FindObsoleteFiles full-scan backstop cannot find the orphan
+  // files (simulating metadata_read_fault_one_in from the stress test).
+  fault_fs->SetThreadLocalErrorContext(
+      FaultInjectionIOType::kMetadataRead, /*seed=*/0, /*one_in=*/1,
+      /*retryable=*/false, /*has_data_loss=*/false);
+  fault_fs->EnableThreadLocalErrorInjection(
+      FaultInjectionIOType::kMetadataRead);
+
+  // Destroy column family handles before closing.
+  for (auto h : handles_) {
+    ASSERT_OK(db_->DestroyColumnFamilyHandle(h));
+  }
+  handles_.clear();
+
+  // Close the DB. Without the fix, TEST_VerifyNoObsoleteFilesCached fires.
+  s = db_->Close();
+  ASSERT_OK(s);
+  db_ = nullptr;
+}
+
 TEST_F(DBCompactionTest, VerifyFileChecksumOnCompactionOutput) {
   Options options = CurrentOptions();
   options.disable_auto_compactions = true;

db/db_flush_test.cc

@@ -1852,6 +1852,8 @@ TEST_F(DBFlushTest, MemPurgeCorrectLogNumberAndSSTFileCreation) {
     }
   }
 
+  ASSERT_OK(WaitForFlushCallbacks());
+
   // Check that there was at least one mempurge
   uint32_t expected_min_mempurge_count = 1;
   // Check that there was no SST files created during flush.
@@ -1872,6 +1874,8 @@ TEST_F(DBFlushTest, MemPurgeCorrectLogNumberAndSSTFileCreation) {
     ASSERT_EQ(Get(key), value);
   }
 
+  ASSERT_OK(WaitForFlushCallbacks());
+
   // Check that there was at least one SST files created during flush.
   expected_sst_count = 1;
   EXPECT_GE(sst_count.load(), expected_sst_count);
@@ -1891,6 +1895,9 @@ TEST_F(DBFlushTest, MemPurgeCorrectLogNumberAndSSTFileCreation) {
   // Extra check of database consistency.
   ASSERT_EQ(Get(key), value);
 
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
   Close();
 }
 
@@ -3883,7 +3890,7 @@ TEST_F(DBFlushTest, LeakedTableCacheEntryOnFlushInstallFailure) {
 
   ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
 
-  // Trigger flush — BuildTable succeeds but LogAndApply fails.
+  // Trigger flush -- BuildTable succeeds but LogAndApply fails.
   Status s = Flush();
   ASSERT_NOK(s);
 

db/db_impl/db_impl.cc

@@ -15,6 +15,7 @@
 
 #include <cinttypes>
 #include <cstdio>
+#include <deque>
 #include <map>
 #include <memory>
 #include <optional>
@@ -29,6 +30,7 @@
 
 #include "db/arena_wrapped_db_iter.h"
 #include "db/attribute_group_iterator_impl.h"
+#include "db/blob/blob_fetcher.h"
 #include "db/blob/blob_file_partition_manager.h"
 #include "db/blob/blob_index.h"
 #include "db/builder.h"
@@ -481,6 +483,18 @@ void DBImpl::WaitForBackgroundWork() {
   }
 }
 
+void DBImpl::WaitForAsyncFileOpen() {
+  if (!immutable_db_options_.open_files_async) {
+    return;
+  }
+  InstrumentedMutexLock l(&mutex_);
+  TEST_SYNC_POINT("DBImpl::WaitForAsyncFileOpen::BeforeWait");
+  while (bg_async_file_open_state_ == AsyncFileOpenState::kScheduled &&
+         !shutting_down_.load(std::memory_order_acquire)) {
+    bg_cv_.Wait();
+  }
+}
+
 // Will lock the mutex_,  will wait for completion if wait is true
 void DBImpl::CancelAllBackgroundWork(bool wait) {
   ROCKS_LOG_INFO(immutable_db_options_.info_log,
@@ -636,6 +650,105 @@ void DBImpl::UnregisterBlobDirectWriteColumnFamily() {
   assert(false);
 }
 
+Status DBImpl::MaybeWriteWalMarkersToManifestOnClose() {
+  mutex_.AssertHeld();
+  if (!mutable_db_options_.optimize_manifest_for_recovery ||
+      !opened_successfully_ || versions_ == nullptr || logs_.empty()) {
+    return Status::OK();
+  }
+
+  TEST_SYNC_POINT("DBImpl::CloseHelper:WriteWalMarkersOnCloseEntered");
+  const ReadOptions read_options(Env::IOActivity::kUnknown);
+  const WriteOptions write_options(Env::IOActivity::kUnknown);
+  uint64_t max_wal_number = logs_.back().number;
+  TEST_SYNC_POINT_CALLBACK("DBImpl::CloseHelper:CapturedMaxWal",
+                           &max_wal_number);
+  const uint64_t new_log_num = max_wal_number + 1;
+
+  // std::deque keeps VersionEdit pointers stable across emplace_back.
+  std::deque<VersionEdit> edits;
+  autovector<ColumnFamilyData*> cfds;
+  autovector<autovector<VersionEdit*>> edit_lists;
+  VersionEdit global_edit;
+  bool persist_global_edit = false;
+  bool reserved_log_number = false;
+
+  bool all_cfs_empty = true;
+  for (auto* cfd : *versions_->GetColumnFamilySet()) {
+    if (cfd->IsDropped()) {
+      continue;
+    }
+    // cfd->IsEmpty() checks both mem and imm. An immutable memtable queued for
+    // flush still pins an older WAL; advancing past it would drop unflushed
+    // records on the next open.
+    if (!cfd->IsEmpty()) {
+      all_cfs_empty = false;
+      continue;
+    }
+    if (new_log_num > cfd->GetLogNumber()) {
+      if (!reserved_log_number) {
+        const uint64_t next_file_number = versions_->current_next_file_number();
+        if (next_file_number <= new_log_num) {
+          versions_->FetchAddFileNumber(new_log_num + 1 - next_file_number);
+          global_edit.SetNextFile(versions_->current_next_file_number());
+          persist_global_edit = true;
+        }
+        reserved_log_number = true;
+      }
+      edits.emplace_back();
+      VersionEdit& edit = edits.back();
+      edit.SetLogNumber(new_log_num);
+      edit.SetColumnFamily(cfd->GetID());
+      cfds.push_back(cfd);
+      edit_lists.emplace_back();
+      edit_lists.back().push_back(&edit);
+    }
+  }
+
+  if (all_cfs_empty && !allow_2pc()) {
+    if (new_log_num > versions_->min_log_number_to_keep()) {
+      global_edit.SetMinLogNumberToKeep(new_log_num);
+      persist_global_edit = true;
+    }
+    // Mirror the monotonicity guard in db_impl_files.cc:
+    // PrecomputeWalDeletionEdit emits DeleteWalsBefore only when the
+    // marker actually advances the WalSet's existing min. Skipping
+    // this guard would let a stale-replay edit regress WalSet state.
+    if (immutable_db_options_.track_and_verify_wals_in_manifest &&
+        new_log_num > versions_->GetWalSet().GetMinWalNumberToKeep()) {
+      global_edit.DeleteWalsBefore(new_log_num);
+      persist_global_edit = true;
+    }
+  }
+
+  if (persist_global_edit) {
+    ColumnFamilyData* default_cfd =
+        versions_->GetColumnFamilySet()->GetDefault();
+    if (default_cfd && !default_cfd->IsDropped()) {
+      bool attached_to_existing = false;
+      for (size_t i = 0; i < cfds.size(); ++i) {
+        if (cfds[i] == default_cfd) {
+          edit_lists[i].push_back(&global_edit);
+          attached_to_existing = true;
+          break;
+        }
+      }
+      if (!attached_to_existing) {
+        cfds.push_back(default_cfd);
+        edit_lists.emplace_back();
+        edit_lists.back().push_back(&global_edit);
+      }
+    }
+  }
+
+  if (cfds.empty()) {
+    return Status::OK();
+  }
+  return versions_->LogAndApply(cfds, read_options, write_options, edit_lists,
+                                &mutex_, directories_.GetDbDir(),
+                                /*new_descriptor_log=*/false);
+}
+
 Status DBImpl::CloseHelper() {
   // Guarantee that there is no background error recovery in progress before
   // continuing with the shutdown
@@ -740,6 +853,17 @@ Status DBImpl::CloseHelper() {
     job_context.Clean();
     mutex_.Lock();
   }
+  // Best-effort warm-reopen optimization: after obsolete-file cleanup and
+  // before logs_.clear(), persist close-time WAL markers that can reduce
+  // recovery work on the next open. This matters most when some CFs have no
+  // unflushed writes and/or the newest WAL is empty, since recovery would
+  // otherwise append marker-only MANIFEST edits on reopen.
+  Status manifest_s = MaybeWriteWalMarkersToManifestOnClose();
+  if (!manifest_s.ok()) {
+    ROCKS_LOG_WARN(immutable_db_options_.info_log,
+                   "optimize_manifest_for_recovery close-time write failed: %s",
+                   manifest_s.ToString().c_str());
+  }
   {
     InstrumentedMutexLock lock(&wal_write_mutex_);
     for (auto l : wals_to_free_) {
@@ -2748,6 +2872,148 @@ Status DBImpl::ResolveDirectWriteWideColumns(const ReadOptions& read_options,
   return status;
 }
 
+bool DBImpl::MaybeResolveMemtableBlobValue(const Slice& key,
+                                           const BlobFetcher* blob_fetcher,
+                                           PinnableSlice* value,
+                                           PinnableWideColumns* columns,
+                                           Status* s, bool* is_blob_index,
+                                           bool* value_found) {
+  if (!s->ok() || (!value && !columns)) {
+    return false;
+  }
+
+  auto reset_outputs = [&]() {
+    if (value != nullptr) {
+      value->Reset();
+    }
+    if (columns != nullptr) {
+      columns->Reset();
+    }
+  };
+  auto clear_blob_state = [&]() {
+    if (is_blob_index != nullptr) {
+      *is_blob_index = false;
+    }
+  };
+
+  const bool needs_plain_value_resolution =
+      is_blob_index != nullptr && *is_blob_index;
+  const bool needs_wide_column_resolution =
+      columns != nullptr && !columns->unresolved_blob_column_indices_.empty();
+  if (!needs_plain_value_resolution && !needs_wide_column_resolution) {
+    return false;
+  }
+
+  if (blob_fetcher == nullptr) {
+    reset_outputs();
+    *s = Status::NotSupported(
+        "Encountered blob-backed memtable value without blob fetcher.");
+    clear_blob_state();
+    return true;
+  }
+
+  if (needs_plain_value_resolution) {
+    Slice blob_index_slice;
+    std::string blob_index_storage;
+    if (value != nullptr) {
+      if (value->size() > 0) {
+        blob_index_slice = Slice(value->data(), value->size());
+      } else {
+        blob_index_slice = Slice(*(value->GetSelf()));
+      }
+    } else {
+      assert(columns != nullptr);
+
+      const WideColumns& plain_value_columns = columns->columns();
+      assert(plain_value_columns.size() == 1);
+      assert(plain_value_columns.front().name() == kDefaultWideColumnName);
+      blob_index_slice = plain_value_columns.front().value();
+    }
+
+    PinnableSlice resolved_value;
+    PinnableSlice* target = value != nullptr ? value : &resolved_value;
+    if (value != nullptr) {
+      // BlobIndex::DecodeFrom can retain Slices into the encoded bytes for
+      // inlined blob indices, so take an owned copy before resetting `value`
+      // and reusing the same PinnableSlice as the output target.
+      blob_index_storage.assign(blob_index_slice.data(),
+                                blob_index_slice.size());
+      blob_index_slice = Slice(blob_index_storage);
+      value->Reset();
+    }
+
+    *s = blob_fetcher->FetchBlob(key, blob_index_slice,
+                                 nullptr /* prefetch_buffer */, target,
+                                 nullptr /* bytes_read */);
+    if (s->ok() && columns != nullptr) {
+      columns->SetPlainValue(std::move(*target));
+    } else if (!s->ok()) {
+      reset_outputs();
+      if (s->IsIncomplete() && value_found != nullptr) {
+        *value_found = false;
+      }
+    }
+
+    clear_blob_state();
+    return true;
+  }
+
+  assert(columns != nullptr);
+
+  std::string resolved_entity;
+  bool resolved = false;
+  *s = WideColumnSerialization::ResolveEntityBlobColumns(
+      columns->value_, key, blob_fetcher, nullptr /* prefetch_buffers */,
+      resolved_entity, resolved, nullptr /* total_bytes_read */,
+      nullptr /* num_blobs_resolved */);
+  if (s->ok()) {
+    assert(resolved);
+    if (resolved) {
+      *s = columns->SetWideColumnValue(std::move(resolved_entity));
+    }
+  }
+  if (!s->ok()) {
+    reset_outputs();
+    if (s->IsIncomplete() && value_found != nullptr) {
+      *value_found = false;
+    }
+  }
+
+  clear_blob_state();
+  return true;
+}
+
+void DBImpl::PostprocessMemtableValueRead(
+    const Slice& key, const std::string* timestamp,
+    bool resolve_blob_backed_memtable_value,
+    const BlobFetcher* memtable_blob_fetcher, PinnableSlice* value,
+    PinnableWideColumns* columns, Status* s, bool* is_blob_index,
+    bool* value_found) {
+  if (resolve_blob_backed_memtable_value) {
+    std::string blob_lookup_key_storage;
+    const bool value_resolved = MaybeResolveMemtableBlobValue(
+        GetBlobLookupUserKey(key, timestamp, &blob_lookup_key_storage),
+        memtable_blob_fetcher, value, columns, s, is_blob_index, value_found);
+    if (!value_resolved && value != nullptr && s->ok()) {
+      value->PinSelf();
+    }
+    return;
+  }
+
+  if (s->ok()) {
+    if (value != nullptr) {
+      value->PinSelf();
+    }
+  } else {
+    if (value != nullptr) {
+      value->Reset();
+    }
+    if (columns != nullptr) {
+      columns->Reset();
+    }
+  }
+}
+
 bool DBImpl::MaybeResolveDirectWriteValue(
     const ReadOptions& read_options, const Slice& key,
     bool resolve_direct_write_value, const Version* current,
@@ -2963,6 +3229,14 @@ Status DBImpl::GetImpl(const ReadOptions& read_options, const Slice& key,
   }
   const bool resolve_direct_write_value =
       partition_mgr != nullptr && (is_blob_ptr == &is_blob_index);
+  std::optional<BlobFetcher> memtable_blob_fetcher;
+  if (partition_mgr != nullptr) {
+    memtable_blob_fetcher.emplace(sv->current, read_options,
+                                  cfd->blob_file_cache(),
+                                  /*allow_write_path_fallback=*/true);
+  }
+  const BlobFetcher* memtable_blob_fetcher_ptr =
+      memtable_blob_fetcher ? &*memtable_blob_fetcher : nullptr;
   std::string blob_lookup_key_storage;
   auto get_blob_lookup_key = [&]() -> Slice {
     return GetBlobLookupUserKey(key, timestamp, &blob_lookup_key_storage);
@@ -2990,19 +3264,21 @@ Status DBImpl::GetImpl(const ReadOptions& read_options, const Slice& key,
                        get_impl_options.columns, timestamp, &s, &merge_context,
                        &max_covering_tombstone_seq, read_options,
                        false /* immutable_memtable */,
-                       get_impl_options.callback, is_blob_ptr)) {
+                       get_impl_options.callback, is_blob_ptr,
+                       /*do_merge=*/true, memtable_blob_fetcher_ptr)) {
         done = true;
         maybe_resolve_memtable_value();
 
         RecordTick(stats_, MEMTABLE_HIT);
       } else if ((s.ok() || s.IsMergeInProgress()) &&
-                 sv->imm->Get(
-                     lkey,
-                     get_impl_options.value ? get_impl_options.value->GetSelf()
-                                            : nullptr,
-                     get_impl_options.columns, timestamp, &s, &merge_context,
-                     &max_covering_tombstone_seq, read_options,
-                     get_impl_options.callback, is_blob_ptr)) {
+                 sv->imm->Get(lkey,
+                              get_impl_options.value
+                                  ? get_impl_options.value->GetSelf()
+                                  : nullptr,
+                              get_impl_options.columns, timestamp, &s,
+                              &merge_context, &max_covering_tombstone_seq,
+                              read_options, get_impl_options.callback,
+                              is_blob_ptr, memtable_blob_fetcher_ptr)) {
         done = true;
         maybe_resolve_memtable_value();
 
@@ -3014,14 +3290,14 @@ Status DBImpl::GetImpl(const ReadOptions& read_options, const Slice& key,
       if (sv->mem->Get(lkey, /*value=*/nullptr, /*columns=*/nullptr,
                        /*timestamp=*/nullptr, &s, &merge_context,
                        &max_covering_tombstone_seq, read_options,
-                       false /* immutable_memtable */, nullptr, nullptr,
-                       false)) {
+                       false /* immutable_memtable */, nullptr, nullptr, false,
+                       memtable_blob_fetcher_ptr)) {
         done = true;
         RecordTick(stats_, MEMTABLE_HIT);
       } else if ((s.ok() || s.IsMergeInProgress()) &&
-                 sv->imm->GetMergeOperands(lkey, &s, &merge_context,
-                                           &max_covering_tombstone_seq,
-                                           read_options)) {
+                 sv->imm->GetMergeOperands(
+                     lkey, &s, &merge_context, &max_covering_tombstone_seq,
+                     read_options, memtable_blob_fetcher_ptr)) {
         done = true;
         RecordTick(stats_, MEMTABLE_HIT);
       }
@@ -3717,6 +3993,14 @@ Status DBImpl::MultiGetImpl(
       (*sorted_keys)[start_key]->column_family);
   ColumnFamilyData* cfd = cfh->cfd();
   auto* partition_mgr = cfd->blob_partition_manager();
+  std::optional<BlobFetcher> memtable_blob_fetcher;
+  if (partition_mgr != nullptr) {
+    memtable_blob_fetcher.emplace(super_version->current, read_options,
+                                  cfd->blob_file_cache(),
+                                  /*allow_write_path_fallback=*/true);
+  }
+  const BlobFetcher* memtable_blob_fetcher_ptr =
+      memtable_blob_fetcher ? &*memtable_blob_fetcher : nullptr;
   // Clear the timestamps for returning results so that we can distinguish
   // between tombstone or key that has never been written
   for (size_t i = start_key; i < start_key + num_keys; ++i) {
@@ -3763,9 +4047,11 @@ Status DBImpl::MultiGetImpl(
          has_unpersisted_data_.load(std::memory_order_relaxed));
     if (!skip_memtable) {
       super_version->mem->MultiGet(read_options, &range, callback,
-                                   false /* immutable_memtable */);
+                                   false /* immutable_memtable */,
+                                   memtable_blob_fetcher_ptr);
       if (!range.empty()) {
-        super_version->imm->MultiGet(read_options, &range, callback);
+        super_version->imm->MultiGet(read_options, &range, callback,
+                                     memtable_blob_fetcher_ptr);
       }
       if (!range.empty()) {
         uint64_t left = range.KeysLeft();
@@ -4142,9 +4428,21 @@ Status DBImpl::CreateColumnFamilyImpl(const ReadOptions& read_options,
                                       const std::string& column_family_name,
                                       ColumnFamilyHandle** handle) {
   options_mutex_.AssertHeld();
-  Status s;
   *handle = nullptr;
 
+  // The empty string is reserved as a "no/unknown column family name" sentinel
+  // in various APIs and serialization formats (e.g.
+  // TablePropertiesCollectorFactory::Context::kUnknownColumnFamily, table
+  // properties), so reject it to avoid ambiguity. Note: prior to this check,
+  // CreateColumnFamily("") would silently "succeed" but the CF was not
+  // persisted in the manifest, causing data loss across DB reopens.
+  if (column_family_name.empty()) {
+    return Status::InvalidArgument(
+        "Column family name cannot be the empty string");
+  }
+
+  Status s;
+
   DBOptions db_options =
       BuildDBOptions(immutable_db_options_, mutable_db_options_);
   s = ColumnFamilyData::ValidateOptions(db_options, cf_options);
@@ -6200,12 +6498,20 @@ Status DBImpl::GetLatestSequenceForKey(
 
   *seq = kMaxSequenceNumber;
   *found_record_for_key = false;
+  std::optional<BlobFetcher> memtable_blob_fetcher;
+  if (cfd->blob_partition_manager() != nullptr) {
+    memtable_blob_fetcher.emplace(sv->current, read_options,
+                                  cfd->blob_file_cache(),
+                                  /*allow_write_path_fallback=*/true);
+  }
+  const BlobFetcher* memtable_blob_fetcher_ptr =
+      memtable_blob_fetcher ? &*memtable_blob_fetcher : nullptr;
 
   // Check if there is a record for this key in the latest memtable
   sv->mem->Get(lkey, /*value=*/nullptr, /*columns=*/nullptr, timestamp, &s,
                &merge_context, &max_covering_tombstone_seq, seq, read_options,
                false /* immutable_memtable */, nullptr /*read_callback*/,
-               is_blob_index);
+               is_blob_index, /*do_merge=*/true, memtable_blob_fetcher_ptr);
 
   if (!(s.ok() || s.IsNotFound() || s.IsMergeInProgress())) {
     // unexpected error reading memtable.
@@ -6238,7 +6544,8 @@ Status DBImpl::GetLatestSequenceForKey(
   // Check if there is a record for this key in the immutable memtables
   sv->imm->Get(lkey, /*value=*/nullptr, /*columns=*/nullptr, timestamp, &s,
                &merge_context, &max_covering_tombstone_seq, seq, read_options,
-               nullptr /*read_callback*/, is_blob_index);
+               nullptr /*read_callback*/, is_blob_index,
+               memtable_blob_fetcher_ptr);
 
   if (!(s.ok() || s.IsNotFound() || s.IsMergeInProgress())) {
     // unexpected error reading memtable.
@@ -6271,7 +6578,7 @@ Status DBImpl::GetLatestSequenceForKey(
   sv->imm->GetFromHistory(lkey, /*value=*/nullptr, /*columns=*/nullptr,
                           timestamp, &s, &merge_context,
                           &max_covering_tombstone_seq, seq, read_options,
-                          is_blob_index);
+                          is_blob_index, memtable_blob_fetcher_ptr);
 
   if (!(s.ok() || s.IsNotFound() || s.IsMergeInProgress())) {
     // unexpected error reading memtable.
@@ -7273,6 +7580,7 @@ Status DBImpl::ReserveFileNumbersBeforeIngestion(
 Status DBImpl::GetCreationTimeOfOldestFile(uint64_t* creation_time) {
   if (mutable_db_options_.max_open_files == -1) {
     uint64_t oldest_time = std::numeric_limits<uint64_t>::max();
+    std::once_flag waited_for_async_open;
     for (auto cfd : *versions_->GetColumnFamilySet()) {
       if (!cfd->IsDropped()) {
         uint64_t ctime;
@@ -7280,6 +7588,17 @@ Status DBImpl::GetCreationTimeOfOldestFile(uint64_t* creation_time) {
           SuperVersion* sv = GetAndRefSuperVersion(cfd);
           Version* version = sv->current;
           version->GetCreationTimeOfOldestFile(&ctime);
+          // For modern DBs, manifest carries file_creation_time and the
+          // first call returns the real value. We only need to wait for
+          // BGWorkAsyncFileOpen on legacy DBs whose manifest lacks
+          // file_creation_time -- those rely on the pinned reader, which is
+          // null until async file open completes.
+          if (ctime == 0 && immutable_db_options_.open_files_async) {
+            std::call_once(waited_for_async_open, [&]() {
+              WaitForAsyncFileOpen();
+              version->GetCreationTimeOfOldestFile(&ctime);
+            });
+          }
           ReturnAndCleanupSuperVersion(cfd, sv);
         }
 

db/db_impl/db_impl.h

@@ -1096,6 +1096,7 @@ class DBImpl : public DB {
     superversions_to_free_queue_.push_back(sv);
   }
 
+  void EnableTrackPublishedSeqInSnapshotContext();
   void SetSnapshotChecker(SnapshotChecker* snapshot_checker);
 
   // Fill JobContext with snapshot information needed by flush and compaction.
@@ -1854,6 +1855,11 @@ class DBImpl : public DB {
   // Background work function for async file opening.
   static void BGWorkAsyncFileOpen(void* arg);
 
+  // Block the until any in-flight async file open work has
+  // completed. No-op when open_files_async is false. Returns early if
+  // shutdown begins.
+  void WaitForAsyncFileOpen();
+
   void InvokeWalFilterIfNeededOnColumnFamilyToWalNumberMap();
 
   // Return true to proceed with current WAL record whose content is stored in
@@ -1869,6 +1875,7 @@ class DBImpl : public DB {
 
  private:
   friend class DB;
+  friend class DBImplReadOnly;
   friend class DBImplSecondary;
   friend class ErrorHandler;
   friend class InternalStats;
@@ -2743,6 +2750,8 @@ class DBImpl : public DB {
 
   Status MaybeReleaseTimestampedSnapshotsAndCheck();
 
+  Status MaybeWriteWalMarkersToManifestOnClose();
+
   Status CloseHelper();
 
   void WaitForBackgroundWork();
@@ -2944,6 +2953,29 @@ class DBImpl : public DB {
       bool resolve_direct_write_value, const Version* current,
       ColumnFamilyData* cfd, PinnableSlice* value, PinnableWideColumns* columns,
       Status* s, bool* is_blob_index, bool* value_found = nullptr);
+  // Resolves memtable read results that still carry blob references through
+  // either a raw blob-index payload in `value` or unresolved blob columns in
+  // `columns`. Unlike the direct-write helper above, this path only depends on
+  // a BlobFetcher and therefore works for read-only/secondary DBs.
+  static bool MaybeResolveMemtableBlobValue(const Slice& key,
+                                            const BlobFetcher* blob_fetcher,
+                                            PinnableSlice* value,
+                                            PinnableWideColumns* columns,
+                                            Status* s, bool* is_blob_index,
+                                            bool* value_found = nullptr);
+  // Completes read-only/secondary memtable Get()/GetEntity() hits by resolving
+  // blob-backed payloads when `resolve_blob_backed_memtable_value` is true,
+  // pinning plain values on success, and clearing outputs on error. When the
+  // caller explicitly requested raw blob indices via
+  // `GetImplOptions::is_blob_index`, this helper leaves that payload
+  // untouched. `memtable_blob_fetcher` may be null when blob support is
+  // disabled for the column family.
+  static void PostprocessMemtableValueRead(
+      const Slice& key, const std::string* timestamp,
+      bool resolve_blob_backed_memtable_value,
+      const BlobFetcher* memtable_blob_fetcher, PinnableSlice* value,
+      PinnableWideColumns* columns, Status* s, bool* is_blob_index,
+      bool* value_found = nullptr);
 
   template <typename IterType, typename ImplType,
             typename ErrorIteratorFuncType>
@@ -3348,6 +3380,13 @@ class DBImpl : public DB {
   // Callback for compaction to check if a key is visible to a snapshot.
   // REQUIRES: mutex held
   std::unique_ptr<SnapshotChecker> snapshot_checker_;
+  // When set, InitSnapshotContext() appends GetLastPublishedSequence() to the
+  // job's snapshot_seqs (if not already present) so that flush/compaction
+  // preserves the published-sequence boundary even when no explicit user
+  // snapshot exists there. Unlike taking a real ManagedSnapshot, this just
+  // pins the seqno into the snapshot list: no allocation, no snapshot list
+  // mutation, and no SnapshotChecker.
+  bool track_published_seq_in_snapshot_context_ = false;
 
   // Callback for when the cached_recoverable_state_ is written to memtable
   // Only to be set during initialization

db/db_impl/db_impl_compaction_flush.cc

@@ -874,6 +874,24 @@ Status DBImpl::AtomicFlushMemTablesToOutputFiles(
         directories_.GetDbDir(), log_buffer);
   }
 
+  if (!s.ok()) {
+    // If the atomic flush's combined MANIFEST write failed, the output files
+    // are cached in the table cache (added by BuildTable during FlushJob::Run)
+    // but never installed into any Version. Evict them now so that
+    // TEST_VerifyNoObsoleteFilesCached does not fire during Close() when the
+    // FindObsoleteFiles full-scan backstop is disabled by
+    // metadata_read_fault_one_in.
+    for (int i = 0; i != num_cfs; ++i) {
+      if (exec_status[i].first && exec_status[i].second.ok() &&
+          !switched_to_mempurge[i] && file_meta[i].fd.GetFileSize() > 0) {
+        TableCache::ReleaseObsolete(
+            cfds[i]->table_cache()->get_cache().get(),
+            file_meta[i].fd.GetNumber(), nullptr /*handle*/,
+            all_mutable_cf_options[i].uncache_aggressiveness);
+      }
+    }
+  }
+
   for (int i = 0; i != num_cfs; ++i) {
     auto* mgr = cfds[i]->blob_partition_manager();
     if (mgr == nullptr || prepared_blob_generations[i] == 0) {
@@ -5114,6 +5132,11 @@ void DBImpl::MarkAsGrabbedForPurge(uint64_t file_number) {
   files_grabbed_for_purge_.insert(file_number);
 }
 
+void DBImpl::EnableTrackPublishedSeqInSnapshotContext() {
+  InstrumentedMutexLock l(&mutex_);
+  track_published_seq_in_snapshot_context_ = true;
+}
+
 void DBImpl::SetSnapshotChecker(SnapshotChecker* snapshot_checker) {
   InstrumentedMutexLock l(&mutex_);
   // snapshot_checker_ should only set once. If we need to set it multiple
@@ -5147,6 +5170,19 @@ void DBImpl::InitSnapshotContext(JobContext* job_context) {
   SequenceNumber earliest_write_conflict_snapshot = kMaxSequenceNumber;
   std::vector<SequenceNumber> snapshot_seqs =
       snapshots_.GetAll(&earliest_write_conflict_snapshot);
+
+  // The SnapshotChecker path above already adds a managed snapshot to
+  // snapshot_seqs.
+  if (track_published_seq_in_snapshot_context_ && !snapshot_checker) {
+    // Pin the published-sequence boundary into snapshot_seqs without taking
+    // a real snapshot. This prevents flush/compaction from collapsing a version
+    // visible at the published boundary into a newer unpublished seqno.
+    const SequenceNumber published_seq = GetLastPublishedSequence();
+    if (snapshot_seqs.empty() || snapshot_seqs.back() < published_seq) {
+      snapshot_seqs.push_back(published_seq);
+    }
+  }
+  TEST_SYNC_POINT("DBImpl::InitSnapshotContext:BeforeInit");
   job_context->InitSnapshotContext(
       snapshot_checker, std::move(managed_snapshot),
       earliest_write_conflict_snapshot, std::move(snapshot_seqs));

db/db_impl/db_impl_files.cc

@@ -1186,8 +1186,9 @@ std::set<std::string> DBImpl::CollectAllDBPaths() {
 
 Status DBImpl::MaybeUpdateNextFileNumber(RecoveryContext* recovery_ctx) {
   mutex_.AssertHeld();
-  uint64_t next_file_number = versions_->current_next_file_number();
-  uint64_t largest_file_number = next_file_number;
+  const uint64_t prev_next_file_number = versions_->current_next_file_number();
+  uint64_t largest_file_number = prev_next_file_number;
+  bool on_disk_file_advanced_counter = false;
   Status s;
   for (const auto& path : CollectAllDBPaths()) {
     std::vector<std::string> files;
@@ -1201,7 +1202,15 @@ Status DBImpl::MaybeUpdateNextFileNumber(RecoveryContext* recovery_ctx) {
       if (!ParseFileName(fname, &number, &type)) {
         continue;
       }
-      const std::string normalized_fpath = path + kFilePathSeparator + fname;
+      std::string normalized_fpath = path;
+      normalized_fpath += kFilePathSeparator;
+      normalized_fpath.append(fname);
+      // Use >= so a crashed-mid-allocation file at exactly
+      // prev_next_file_number triggers the advance -- otherwise the next
+      // NewFileNumber() would collide with it.
+      if (number >= prev_next_file_number) {
+        on_disk_file_advanced_counter = true;
+      }
       largest_file_number = std::max(largest_file_number, number);
       if ((type == kTableFile || type == kBlobFile)) {
         recovery_ctx->existing_data_files_.push_back(normalized_fpath);
@@ -1212,16 +1221,31 @@ Status DBImpl::MaybeUpdateNextFileNumber(RecoveryContext* recovery_ctx) {
     return s;
   }
 
-  if (largest_file_number >= next_file_number) {
-    versions_->next_file_number_.store(largest_file_number + 1);
-  }
+  // Legacy bug: initializing largest_file_number from prev_next_file_number
+  // made the post-loop check trivially true and emitted a noop SetNextFile
+  // on every recovery.
+  // Preserved verbatim by default (some tests pin file numbers); skipped
+  // only when the option is on AND we're not in salvage mode (which
+  // requires the MANIFEST rewrite this emit triggers).
+  const bool optimize_manifest_for_recovery =
+      mutable_db_options_.optimize_manifest_for_recovery &&
+      !immutable_db_options_.best_efforts_recovery;
 
-  VersionEdit edit;
-  edit.SetNextFile(versions_->next_file_number_.load());
-  assert(versions_->GetColumnFamilySet());
-  ColumnFamilyData* default_cfd = versions_->GetColumnFamilySet()->GetDefault();
-  assert(default_cfd);
-  recovery_ctx->UpdateVersionEdits(default_cfd, edit);
+  if (!optimize_manifest_for_recovery || on_disk_file_advanced_counter) {
+    if (largest_file_number >= prev_next_file_number) {
+      versions_->next_file_number_.store(largest_file_number + 1);
+    }
+    TEST_SYNC_POINT("DBImpl::MaybeUpdateNextFileNumber:EmitEdit");
+    VersionEdit edit;
+    edit.SetNextFile(versions_->next_file_number_.load());
+    assert(versions_->GetColumnFamilySet());
+    ColumnFamilyData* default_cfd =
+        versions_->GetColumnFamilySet()->GetDefault();
+    assert(default_cfd);
+    recovery_ctx->UpdateVersionEdits(default_cfd, edit);
+  } else {
+    TEST_SYNC_POINT("DBImpl::Recovery:SkippedNoopEdit:NextFileNumber");
+  }
   return s;
 }
 }  // namespace ROCKSDB_NAMESPACE

db/db_impl/db_impl_follower.cc

@@ -295,7 +295,8 @@ Status DB::OpenAsFollower(
   impl->versions_.reset(new ReactiveVersionSet(
       dbname, &impl->immutable_db_options_, impl->mutable_db_options_,
       impl->file_options_, impl->table_cache_.get(),
-      impl->write_buffer_manager_, &impl->write_controller_, impl->io_tracer_));
+      impl->write_buffer_manager_, &impl->write_controller_, impl->io_tracer_,
+      impl->db_id_, impl->db_session_id_));
   impl->column_family_memtables_.reset(
       new ColumnFamilyMemTablesImpl(impl->versions_->GetColumnFamilySet()));
   impl->wal_in_db_path_ = impl->immutable_db_options_.IsWalDirSameAsDBPath();

db/db_impl/db_impl_open.cc

@@ -680,8 +680,18 @@ Status DBImpl::Recover(
   } else if (immutable_db_options_.write_dbid_to_manifest && recovery_ctx) {
     VersionEdit edit;
     s = SetupDBId(write_options, read_only, is_new_db, is_retry, &edit);
-    recovery_ctx->UpdateVersionEdits(
-        versions_->GetColumnFamilySet()->GetDefault(), edit);
+    // best_efforts_recovery rebuilds CURRENT/MANIFEST as the side-effect
+    // of LogAndApplyForRecovery emitting an edit, so the optimization is
+    // disabled there (see optimize_manifest_for_recovery doc).
+    const bool optimize_manifest_for_recovery =
+        mutable_db_options_.optimize_manifest_for_recovery &&
+        !immutable_db_options_.best_efforts_recovery;
+    if (!optimize_manifest_for_recovery || edit.HasDbId()) {
+      recovery_ctx->UpdateVersionEdits(
+          versions_->GetColumnFamilySet()->GetDefault(), edit);
+    } else {
+      TEST_SYNC_POINT("DBImpl::Recovery:SkippedNoopEdit:SetupDBId");
+    }
   } else {
     s = SetupDBId(write_options, read_only, is_new_db, is_retry, nullptr);
   }
@@ -770,6 +780,10 @@ Status DBImpl::Recover(
       // otherwise, in the future, if WAL tracking is enabled again,
       // since the WALs deleted when WAL tracking is disabled are not persisted
       // into MANIFEST, WAL check may fail.
+      //
+      // Intentionally NOT gated by optimize_manifest_for_recovery:
+      // this edit is safety-critical for a future re-enable of WAL tracking,
+      // even though it can look noop-ish under the current session.
       VersionEdit edit;
       WalNumber max_wal_number =
           versions_->GetWalSet().GetWals().rbegin()->first;
@@ -983,7 +997,11 @@ Status DBImpl::InitPersistStatsColumnFamily() {
 
 Status DBImpl::LogAndApplyForRecovery(const RecoveryContext& recovery_ctx) {
   mutex_.AssertHeld();
-  assert(versions_->descriptor_log_ == nullptr);
+  // descriptor_log_ is normally null after Recover, but when
+  // reuse_manifest_on_open is set VersionSet::Recover may have already
+  // bound a log::Writer to the existing MANIFEST for append.
+  assert(versions_->descriptor_log_ == nullptr ||
+         immutable_db_options_.reuse_manifest_on_open);
   const ReadOptions read_options(Env::IOActivity::kDBOpen);
   const WriteOptions write_options(Env::IOActivity::kDBOpen);
 
@@ -1870,25 +1888,47 @@ Status DBImpl::MaybeFlushFinalMemtableOrRestoreActiveLogFiles(
       versions_->MarkFileNumberUsed(max_wal_number + 1);
       assert(recovery_ctx != nullptr);
 
+      const bool optimize_manifest_for_recovery =
+          mutable_db_options_.optimize_manifest_for_recovery &&
+          !immutable_db_options_.best_efforts_recovery;
       for (auto* cfd : *versions_->GetColumnFamilySet()) {
         auto iter = version_edits->find(cfd->GetID());
         assert(iter != version_edits->end());
-        recovery_ctx->UpdateVersionEdits(cfd, iter->second);
+        const VersionEdit& cf_edit = iter->second;
+        if (!optimize_manifest_for_recovery ||
+            cf_edit.ShouldEmitPerColumnFamilyRecoveryEdit(
+                cfd->GetLogNumber())) {
+          recovery_ctx->UpdateVersionEdits(cfd, cf_edit);
+        } else {
+          TEST_SYNC_POINT("DBImpl::Recovery:SkippedNoopEdit:PerCF");
+        }
       }
 
       if (flushed || !data_seen) {
+        const uint64_t new_min_log = max_wal_number + 1;
         VersionEdit wal_deletion;
+        bool emit_wal_deletion = false;
         if (immutable_db_options_.track_and_verify_wals_in_manifest) {
-          wal_deletion.DeleteWalsBefore(max_wal_number + 1);
+          // Determining whether DeleteWalsBefore actually shrinks WalSet
+          // membership requires WalSet state outside this site, so emit
+          // unconditionally (pre-existing behavior).
+          wal_deletion.DeleteWalsBefore(new_min_log);
+          emit_wal_deletion = true;
         }
-        if (!allow_2pc()) {
-          // In non-2pc mode, flushing the memtables of the column families
-          // means we can advance min_log_number_to_keep.
-          wal_deletion.SetMinLogNumberToKeep(max_wal_number + 1);
+        const bool min_log_advances =
+            new_min_log > versions_->min_log_number_to_keep();
+        if (!allow_2pc() &&
+            (!optimize_manifest_for_recovery || min_log_advances)) {
+          wal_deletion.SetMinLogNumberToKeep(new_min_log);
+          emit_wal_deletion = true;
         }
         assert(versions_->GetColumnFamilySet() != nullptr);
-        recovery_ctx->UpdateVersionEdits(
-            versions_->GetColumnFamilySet()->GetDefault(), wal_deletion);
+        if (!optimize_manifest_for_recovery || emit_wal_deletion) {
+          recovery_ctx->UpdateVersionEdits(
+              versions_->GetColumnFamilySet()->GetDefault(), wal_deletion);
+        } else {
+          TEST_SYNC_POINT("DBImpl::Recovery:SkippedNoopEdit:WalDeletion");
+        }
       }
     }
   }

db/db_impl/db_impl_readonly.cc

@@ -5,7 +5,10 @@
 
 #include "db/db_impl/db_impl_readonly.h"
 
+#include <optional>
+
 #include "db/arena_wrapped_db_iter.h"
+#include "db/blob/blob_fetcher.h"
 #include "db/db_impl/compacted_db_impl.h"
 #include "db/db_impl/db_impl.h"
 #include "db/manifest_ops.h"
@@ -63,13 +66,25 @@ Status DBImplReadOnly::GetImpl(const ReadOptions& read_options,
 
   const Comparator* ucmp = get_impl_options.column_family->GetComparator();
   assert(ucmp);
-  std::string* ts =
-      ucmp->timestamp_size() > 0 ? get_impl_options.timestamp : nullptr;
   SequenceNumber snapshot = versions_->LastSequence();
   GetWithTimestampReadCallback read_cb(snapshot);
   auto cfh = static_cast_with_check<ColumnFamilyHandleImpl>(
       get_impl_options.column_family);
   auto cfd = cfh->cfd();
+  bool is_blob_index = false;
+  bool* is_blob_ptr = get_impl_options.is_blob_index;
+  std::string timestamp_storage;
+  std::string* ts = nullptr;
+  if (ucmp->timestamp_size() > 0) {
+    ts = get_impl_options.timestamp != nullptr
+             ? get_impl_options.timestamp
+             : (get_impl_options.get_value ? &timestamp_storage : nullptr);
+  }
+  if (!is_blob_ptr && get_impl_options.get_value) {
+    is_blob_ptr = &is_blob_index;
+  }
+  const bool resolve_blob_backed_memtable_value =
+      get_impl_options.get_value && (is_blob_ptr == &is_blob_index);
   if (tracer_) {
     InstrumentedMutexLock lock(&trace_mutex_);
     if (tracer_) {
@@ -95,6 +110,18 @@ Status DBImplReadOnly::GetImpl(const ReadOptions& read_options,
   SequenceNumber max_covering_tombstone_seq = 0;
   LookupKey lkey(key, snapshot, read_options.timestamp);
   PERF_TIMER_STOP(get_snapshot_time);
+  std::optional<BlobFetcher> memtable_blob_fetcher;
+  if (cfd->ioptions().enable_blob_direct_write ||
+      cfd->GetLatestMutableCFOptions().enable_blob_files) {
+    // Recovered memtables can still contain older blob references after
+    // mutable blob-file settings change, so keep blob resolution available
+    // whenever either blob knob indicates it may be needed.
+    memtable_blob_fetcher.emplace(super_version->current, read_options,
+                                  cfd->blob_file_cache(),
+                                  /*allow_write_path_fallback=*/true);
+  }
+  const BlobFetcher* memtable_blob_fetcher_ptr =
+      memtable_blob_fetcher ? &*memtable_blob_fetcher : nullptr;
 
   // Look up starts here
   if (super_version->mem->Get(
@@ -102,11 +129,12 @@ Status DBImplReadOnly::GetImpl(const ReadOptions& read_options,
           get_impl_options.value ? get_impl_options.value->GetSelf() : nullptr,
           get_impl_options.columns, ts, &s, &merge_context,
           &max_covering_tombstone_seq, read_options,
-          false /* immutable_memtable */, &read_cb,
-          /*is_blob_index=*/nullptr, /*do_merge=*/get_impl_options.get_value)) {
-    if (get_impl_options.value) {
-      get_impl_options.value->PinSelf();
-    }
+          false /* immutable_memtable */, &read_cb, is_blob_ptr,
+          /*do_merge=*/get_impl_options.get_value, memtable_blob_fetcher_ptr)) {
+    DBImpl::PostprocessMemtableValueRead(
+        key, ts, resolve_blob_backed_memtable_value, memtable_blob_fetcher_ptr,
+        get_impl_options.value, get_impl_options.columns, &s, &is_blob_index,
+        get_impl_options.value_found);
     RecordTick(stats_, MEMTABLE_HIT);
   } else {
     PERF_TIMER_GUARD(get_from_output_files_time);

db/db_impl/db_impl_secondary.cc

@@ -6,8 +6,10 @@
 #include "db/db_impl/db_impl_secondary.h"
 
 #include <cinttypes>
+#include <optional>
 
 #include "db/arena_wrapped_db_iter.h"
+#include "db/blob/blob_fetcher.h"
 #include "db/log_reader.h"
 #include "db/log_writer.h"
 #include "db/merge_context.h"
@@ -363,13 +365,25 @@ Status DBImplSecondary::GetImpl(const ReadOptions& read_options,
 
   const Comparator* ucmp = get_impl_options.column_family->GetComparator();
   assert(ucmp);
-  std::string* ts =
-      ucmp->timestamp_size() > 0 ? get_impl_options.timestamp : nullptr;
   SequenceNumber snapshot = versions_->LastSequence();
   GetWithTimestampReadCallback read_cb(snapshot);
   auto cfh = static_cast_with_check<ColumnFamilyHandleImpl>(
       get_impl_options.column_family);
   auto cfd = cfh->cfd();
+  bool is_blob_index = false;
+  bool* is_blob_ptr = get_impl_options.is_blob_index;
+  std::string timestamp_storage;
+  std::string* ts = nullptr;
+  if (ucmp->timestamp_size() > 0) {
+    ts = get_impl_options.timestamp != nullptr
+             ? get_impl_options.timestamp
+             : (get_impl_options.get_value ? &timestamp_storage : nullptr);
+  }
+  if (!is_blob_ptr && get_impl_options.get_value) {
+    is_blob_ptr = &is_blob_index;
+  }
+  const bool resolve_blob_backed_memtable_value =
+      get_impl_options.get_value && (is_blob_ptr == &is_blob_index);
   if (tracer_) {
     InstrumentedMutexLock lock(&trace_mutex_);
     if (tracer_) {
@@ -395,6 +409,18 @@ Status DBImplSecondary::GetImpl(const ReadOptions& read_options,
   LookupKey lkey(key, snapshot, read_options.timestamp);
   PERF_TIMER_STOP(get_snapshot_time);
   bool done = false;
+  std::optional<BlobFetcher> memtable_blob_fetcher;
+  if (cfd->ioptions().enable_blob_direct_write ||
+      cfd->GetLatestMutableCFOptions().enable_blob_files) {
+    // Catch-up can rebuild older blob references into memtables after mutable
+    // blob-file settings change, so keep blob resolution available whenever
+    // either blob knob indicates it may be needed.
+    memtable_blob_fetcher.emplace(super_version->current, read_options,
+                                  cfd->blob_file_cache(),
+                                  /*allow_write_path_fallback=*/true);
+  }
+  const BlobFetcher* memtable_blob_fetcher_ptr =
+      memtable_blob_fetcher ? &*memtable_blob_fetcher : nullptr;
 
   // Look up starts here
   if (get_impl_options.get_value) {
@@ -404,12 +430,14 @@ Status DBImplSecondary::GetImpl(const ReadOptions& read_options,
                                    : nullptr,
             get_impl_options.columns, ts, &s, &merge_context,
             &max_covering_tombstone_seq, read_options,
-            false /* immutable_memtable */, &read_cb,
-            /*is_blob_index=*/nullptr, /*do_merge=*/true)) {
+            false /* immutable_memtable */, &read_cb, is_blob_ptr,
+            /*do_merge=*/true, memtable_blob_fetcher_ptr)) {
       done = true;
-      if (get_impl_options.value) {
-        get_impl_options.value->PinSelf();
-      }
+      DBImpl::PostprocessMemtableValueRead(
+          key, ts, resolve_blob_backed_memtable_value,
+          memtable_blob_fetcher_ptr, get_impl_options.value,
+          get_impl_options.columns, &s, &is_blob_index,
+          get_impl_options.value_found);
       RecordTick(stats_, MEMTABLE_HIT);
     } else if ((s.ok() || s.IsMergeInProgress()) &&
                super_version->imm->Get(
@@ -417,11 +445,14 @@ Status DBImplSecondary::GetImpl(const ReadOptions& read_options,
                    get_impl_options.value ? get_impl_options.value->GetSelf()
                                           : nullptr,
                    get_impl_options.columns, ts, &s, &merge_context,
-                   &max_covering_tombstone_seq, read_options, &read_cb)) {
+                   &max_covering_tombstone_seq, read_options, &read_cb,
+                   is_blob_ptr, memtable_blob_fetcher_ptr)) {
       done = true;
-      if (get_impl_options.value) {
-        get_impl_options.value->PinSelf();
-      }
+      DBImpl::PostprocessMemtableValueRead(
+          key, ts, resolve_blob_backed_memtable_value,
+          memtable_blob_fetcher_ptr, get_impl_options.value,
+          get_impl_options.columns, &s, &is_blob_index,
+          get_impl_options.value_found);
       RecordTick(stats_, MEMTABLE_HIT);
     }
   } else {
@@ -433,13 +464,14 @@ Status DBImplSecondary::GetImpl(const ReadOptions& read_options,
             get_impl_options.columns, ts, &s, &merge_context,
             &max_covering_tombstone_seq, read_options,
             false /* immutable_memtable */, &read_cb,
-            /*is_blob_index=*/nullptr, /*do_merge=*/false)) {
+            /*is_blob_index=*/nullptr, /*do_merge=*/false,
+            memtable_blob_fetcher_ptr)) {
       done = true;
       RecordTick(stats_, MEMTABLE_HIT);
     } else if ((s.ok() || s.IsMergeInProgress()) &&
-               super_version->imm->GetMergeOperands(lkey, &s, &merge_context,
-                                                    &max_covering_tombstone_seq,
-                                                    read_options)) {
+               super_version->imm->GetMergeOperands(
+                   lkey, &s, &merge_context, &max_covering_tombstone_seq,
+                   read_options, memtable_blob_fetcher_ptr)) {
       done = true;
       RecordTick(stats_, MEMTABLE_HIT);
     }
@@ -457,7 +489,7 @@ Status DBImplSecondary::GetImpl(const ReadOptions& read_options,
         ts, &s, &merge_context, &max_covering_tombstone_seq, &pinned_iters_mgr,
         /*value_found*/ nullptr,
         /*key_exists*/ nullptr, /*seq*/ nullptr, &read_cb, /*is_blob*/ nullptr,
-        /*do_merge*/ true);
+        /*do_merge=*/get_impl_options.get_value);
     RecordTick(stats_, MEMTABLE_MISS);
   }
   {
@@ -777,7 +809,8 @@ Status DBImplSecondary::OpenAsSecondaryImpl(
   impl->versions_.reset(new ReactiveVersionSet(
       dbname, &impl->immutable_db_options_, impl->mutable_db_options_,
       impl->file_options_, impl->table_cache_.get(),
-      impl->write_buffer_manager_, &impl->write_controller_, impl->io_tracer_));
+      impl->write_buffer_manager_, &impl->write_controller_, impl->io_tracer_,
+      impl->db_id_, impl->db_session_id_));
   impl->column_family_memtables_.reset(
       new ColumnFamilyMemTablesImpl(impl->versions_->GetColumnFamilySet()));
   impl->wal_in_db_path_ = impl->immutable_db_options_.IsWalDirSameAsDBPath();
@@ -1553,7 +1586,7 @@ Status DB::OpenAndCompact(
     }
   }
 
-  // 5. Open db As Secondary (skip WAL recovery — remote compaction only
+  // 5. Open db As Secondary (skip WAL recovery -- remote compaction only
   //    needs LSM state from MANIFEST, not memtable data from WAL replay)
   std::unique_ptr<DB> db;
   std::vector<ColumnFamilyHandle*> handles;

db/db_impl/db_impl_write.cc

@@ -1546,6 +1546,10 @@ Status DBImpl::WriteImpl(
       // Note: if we are to resume after non-OK statuses we need to revisit how
       // we react to non-OK statuses here.
       if (w.status.ok()) {  // Don't publish a partial batch write
+        TEST_SYNC_POINT(
+            "DBImpl::WriteImpl:AfterWBWIIngestBeforeSetLastSequence:pause");
+        TEST_SYNC_POINT(
+            "DBImpl::WriteImpl:AfterWBWIIngestBeforeSetLastSequence:resume");
         versions_->SetLastSequence(last_sequence);
       }
     }

db/db_iter.cc

@@ -623,7 +623,7 @@ bool DBIter::FindNextUserEntryInternal(bool skipping_saved_key) {
               PERF_COUNTER_ADD(internal_delete_skipped_count, 1);
               MarkMemtableForFlushForPerOpTrigger(mem_hidden_op_scanned);
               // Track contiguous tombstones for range conversion.
-              // Skip if outside seek prefix — the top-of-loop check
+              // Skip if outside seek prefix -- the top-of-loop check
               // flushed any pending run, but we must also avoid starting
               // a new run outside the prefix.
               if (min_tombstones_for_range_conversion_ > 0 &&
@@ -1097,7 +1097,7 @@ void DBIter::PrevInternal() {
     if (min_tombstones_for_range_conversion_ > 0 &&
         range_tomb_end_key_.Size() > 0 && timestamp_lb_ == nullptr) {
       if (!valid_ && found_visible && PrefixCheck(saved_key_without_ts)) {
-        // Key was deleted and is within the seek prefix — track it.
+        // Key was deleted and is within the seek prefix -- track it.
         TrackContiguousTombstone(saved_key_.GetUserKey(),
                                  /*always_update_first_key=*/true);
       } else if (valid_) {
@@ -1141,7 +1141,7 @@ void DBIter::PrevInternal() {
 // Sets ikey_ to the last visible entry's internal key.  When found_visible
 // is false, ikey_ is not updated and may contain stale data.
 // Sets found_visible to true if at least one entry passed the IsVisible()
-// check (seqno <= snapshot).  When false, no entry was visible — the key
+// check (seqno <= snapshot).  When false, no entry was visible -- the key
 // does not exist at this snapshot and should not be treated as a tombstone.
 // Returns false if an error occurred, and !status().ok() and !valid_.
 //
@@ -1196,7 +1196,7 @@ bool DBIter::FindValueForCurrentKey(bool& found_visible) {
       break;
     }
 
-    // Entry survived the visibility check — at least one visible version
+    // Entry survived the visibility check -- at least one visible version
     // exists for this user key.
     found_visible = true;
 

db/db_iter.h

@@ -589,7 +589,7 @@ class DBIter final : public Iterator {
 
   // If enough contiguous tombstones have been tracked, insert a range
   // tombstone [first_key, end_key) into the mutable memtable.
-  // end_key is the exclusive upper bound — typically the next live key.
+  // end_key is the exclusive upper bound -- typically the next live key.
   void MaybeInsertRangeTombstone(const Slice& end_key);
   void ResetContiguousTombstoneTracking() {
     contiguous_tombstone_count_ = 0;

db/db_iterator_test.cc

@@ -4151,7 +4151,7 @@ TEST_P(DBIteratorTest, PrefixSameAsStartSeekToNonInDomainKey) {
   ASSERT_OK(Put("abc2", "v2"));
   ASSERT_OK(Put("abc3", "v3"));
 
-  // Seek to "ab" (2 bytes) — out-of-domain for FixedPrefixTransform(3).
+  // Seek to "ab" (2 bytes) -- out-of-domain for FixedPrefixTransform(3).
   // ShouldSetPrefix returns false for out-of-domain targets, so no prefix
   // constraint is set. The seek should find "abc1" and iteration should
   // proceed without prefix_same_as_start enforcement.
@@ -5820,13 +5820,13 @@ TEST_P(ReadPathRangeTombstoneTest, ExhaustedIteratorWithBounds) {
   // Both directions encounter two tombstone runs (a-d and g-j).
   ASSERT_EQ(attempted_insert_ranges_.size(), 2);
   if (Forward()) {
-    // Forward: sees a-d tombstones first → live e terminates → [a, e).
-    // Then g-j → exhaustion past j, saved_key_="j" fallback → [g, j),
+    // Forward: sees a-d tombstones first -> live e terminates -> [a, e).
+    // Then g-j -> exhaustion past j, saved_key_="j" fallback -> [g, j),
     // covering g,h,i. j remains a point tombstone.
     AssertRange(0, "a", "e");
     AssertRange(1, "g", "j");
   } else {
-    // Reverse: sees j-g tombstones first, then f,e live, then d-a → exhausts.
+    // Reverse: sees j-g tombstones first, then f,e live, then d-a -> exhausts.
     AssertRange(0, "g", "z");
     AssertRange(1, "a", "e");
   }
@@ -5863,8 +5863,8 @@ TEST_P(ReadPathRangeTombstoneTest, ExhaustedIteratorNoBounds) {
   VerifyIteration({"e", "f"});
 
   if (Forward()) {
-    // Forward: a-d run terminates at live e → [a, e). g-j run exhausts
-    // past j → saved_key_ fallback → [g, j), covering g,h,i with j as a
+    // Forward: a-d run terminates at live e -> [a, e). g-j run exhausts
+    // past j -> saved_key_ fallback -> [g, j), covering g,h,i with j as a
     // point tombstone.
     ASSERT_EQ(attempted_insert_ranges_.size(), 2);
     AssertRange(0, "a", "e");
@@ -6061,7 +6061,7 @@ TEST_P(ReadPathRangeTombstoneTest, PrefixFilterDefaultReadOptions) {
 }
 
 TEST_P(ReadPathRangeTombstoneTest, PrefixFilterTotalOrderSeek) {
-  // total_order_seek=true disables prefix filtering — all files visible.
+  // total_order_seek=true disables prefix filtering -- all files visible.
   // The delete run spans from prefix 'b' into prefix 'c', terminated by
   // live key "cb" from L0. The tombstone [ba, cb) crosses prefix boundaries,
   // which is safe because total_order_seek makes all files visible.
@@ -6103,7 +6103,7 @@ TEST_P(ReadPathRangeTombstoneTest, PrefixFilterTotalOrderSeek) {
     ASSERT_OK(Flush());
 
     // Memtable: contiguous deletes spanning prefixes 'b' and 'c'.
-    // No live key between "bb" and "cb" — the run crosses prefix boundaries.
+    // No live key between "bb" and "cb" -- the run crosses prefix boundaries.
     ASSERT_OK(put("aa", "below"));
     ASSERT_OK(del("ba"));
     ASSERT_OK(del("bb"));
@@ -6311,7 +6311,7 @@ TEST_P(ReadPathRangeTombstoneTest, PrefixFilterOutOfDomainSeek) {
   ro.prefix_same_as_start = true;
   auto it = std::unique_ptr<Iterator>(db_->NewIterator(ro));
   if (Forward()) {
-    // Seek with a 1-byte key — out-of-domain for FixedPrefixTransform(4).
+    // Seek with a 1-byte key -- out-of-domain for FixedPrefixTransform(4).
     // prefix_same_as_start cannot set a prefix bound for this target.
     it->Seek("b");
     while (it->Valid()) {
@@ -6674,7 +6674,7 @@ TEST_P(ReadPathRangeTombstoneTest, ExhaustedWithUDT) {
 
   ASSERT_EQ(attempted_insert_ranges_.size(), 1);
   if (Forward()) {
-    // Forward exhaustion past h: saved_key_="h"+ts fallback → [e+ts, h+ts).
+    // Forward exhaustion past h: saved_key_="h"+ts fallback -> [e+ts, h+ts).
     AssertRange(0, std::string("e") + ts, std::string("h") + ts);
   } else {
     // Reverse exhaustion: range is [a+ts, e+ts).
@@ -6687,9 +6687,9 @@ TEST_P(ReadPathRangeTombstoneTest, ExhaustedWithUDT) {
 
 TEST_P(ReadPathRangeTombstoneTest, SeekForPrevTombstone) {
   // SeekForPrev lands directly on tombstones. No upper bound. Forward
-  // exhaustion past h falls back to saved_key_="h" → [e, h), covering
+  // exhaustion past h falls back to saved_key_="h" -> [e, h), covering
   // e,f,g with h as a point tombstone.
-  // Reverse: SeekForPrev("h") → Delete(h,g,f,e) → live "d" → range [e, h).
+  // Reverse: SeekForPrev("h") -> Delete(h,g,f,e) -> live "d" -> range [e, h).
   for (bool use_udt : {false, true}) {
     SCOPED_TRACE(use_udt ? "with UDT" : "without UDT");
     Options options = CurrentOptions();
@@ -6727,7 +6727,7 @@ TEST_P(ReadPathRangeTombstoneTest, SeekForPrevTombstone) {
         keys.push_back(iter->key().ToString());
       }
     } else {
-      // SeekForPrev("h") lands on Delete(h), traverses g,f,e → finds "d".
+      // SeekForPrev("h") lands on Delete(h), traverses g,f,e -> finds "d".
       for (iter->SeekForPrev("h"); iter->Valid(); iter->Prev()) {
         keys.push_back(iter->key().ToString());
       }
@@ -6738,7 +6738,7 @@ TEST_P(ReadPathRangeTombstoneTest, SeekForPrevTombstone) {
 
     ASSERT_EQ(attempted_insert_ranges_.size(), 1);
     if (Forward()) {
-      // Forward exhausts past h → saved_key_="h" fallback → [e, h),
+      // Forward exhausts past h -> saved_key_="h" fallback -> [e, h),
       // covering e,f,g with h as a point tombstone.
       if (use_udt) {
         AssertRange(0, std::string("e") + ts, std::string("h") + ts);
@@ -6761,11 +6761,11 @@ TEST_P(ReadPathRangeTombstoneTest, SeekForPrevTombstone) {
 
 TEST_P(ReadPathRangeTombstoneTest, UpperBoundTombstone) {
   // iterate_upper_bound lands past the data. Both directions see tombstones
-  // e-h. Forward exhausts naturally past h (no key in DB ≥ upper) so the
+  // e-h. Forward exhausts naturally past h (no key in DB >= upper) so the
   // forward path falls back to saved_key_ ("h") as the exclusive end_key,
-  // covering n-1 of n deletes — [e, h). The remaining tombstone for h
+  // covering n-1 of n deletes -- [e, h). The remaining tombstone for h
   // stays as a point delete. Reverse captures the live key "d" before the
-  // run and uses it via range_tomb_end_key_ — [e, i) when SeekToLast
+  // run and uses it via range_tomb_end_key_ -- [e, i) when SeekToLast
   // delegates to SeekForPrev("i").
   for (bool use_udt : {false, true}) {
     SCOPED_TRACE(use_udt ? "with UDT" : "without UDT");
@@ -6837,8 +6837,8 @@ TEST_P(ReadPathRangeTombstoneTest, UpperBoundTombstone) {
 
 TEST_P(ReadPathRangeTombstoneTest, LowerBoundTruncatesReverse) {
   // Keys a-j, delete a-h. Lower bound "e" truncates reverse iteration
-  // mid-tombstone-run. Forward: tombstones e-h ended by live key i → [e, i).
-  // Reverse: tombstones h,g,f,e then lower_bound hit → flush [e, i).
+  // mid-tombstone-run. Forward: tombstones e-h ended by live key i -> [e, i).
+  // Reverse: tombstones h,g,f,e then lower_bound hit -> flush [e, i).
   for (bool use_udt : {false, true}) {
     SCOPED_TRACE(use_udt ? "with UDT" : "without UDT");
     Options options = CurrentOptions();
@@ -6975,14 +6975,14 @@ TEST_P(ReadPathRangeTombstoneTest, InvisibleKeysDontBreakTombstoneRun) {
   ASSERT_OK(Put("f", "vf"));
   ASSERT_OK(Flush());
 
-  // Delete b, d — visible tombstones.
+  // Delete b, d -- visible tombstones.
   ASSERT_OK(Delete("b"));
   ASSERT_OK(Delete("d"));
 
   // Snapshot S.  At S: a(live), b(del), d(del), f(live).
   const Snapshot* snap = db_->GetSnapshot();
 
-  // Write c, e AFTER the snapshot — invisible at S.
+  // Write c, e AFTER the snapshot -- invisible at S.
   // At S the iterator sees: a(live), b(del), c(invis), d(del),
   //   e(invis), f(live).
   ASSERT_OK(Put("c", "vc"));
@@ -6994,7 +6994,7 @@ TEST_P(ReadPathRangeTombstoneTest, InvisibleKeysDontBreakTombstoneRun) {
   VerifyIteration({"a", "f"}, ro);
 
   // Invisible keys c and e should not break the tombstone run.
-  // 2 tombstones (b, d) ≥ threshold → range [b, f).
+  // 2 tombstones (b, d) >= threshold -> range [b, f).
   ASSERT_EQ(attempted_insert_ranges_.size(), 1);
   AssertRange(0, "b", "f");
 

db/db_secondary_test.cc

@@ -7,9 +7,12 @@
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file. See the AUTHORS file for names of contributors.
 
+#include "db/blob/blob_index.h"
 #include "db/db_impl/db_impl_secondary.h"
 #include "db/db_test_util.h"
 #include "db/db_with_timestamp_test_util.h"
+#include "db/wide/wide_column_test_util.h"
+#include "db/write_batch_internal.h"
 #include "port/stack_trace.h"
 #include "rocksdb/utilities/transaction_db.h"
 #include "test_util/sync_point.h"
@@ -357,8 +360,7 @@ TEST_F(DBSecondaryTest, GetMergeOperands) {
   const Status s = db_secondary_->GetMergeOperands(
       ReadOptions(), cfh, "k1", values.data(), &merge_operands_info,
       &number_of_operands);
-  ASSERT_NOK(s);
-  ASSERT_TRUE(s.IsMergeInProgress());
+  ASSERT_OK(s);
 
   ASSERT_EQ(number_of_operands, 4);
   ASSERT_EQ(values[0].ToString(), "v1");
@@ -367,6 +369,197 @@ TEST_F(DBSecondaryTest, GetMergeOperands) {
   ASSERT_EQ(values[3].ToString(), "v4");
 }
 
+TEST_F(DBSecondaryTest, GetMergeOperandsWithBlobBackedEntityDefaultColumn) {
+  // Goal: exercise the secondary read path with a blob-backed V2 entity base
+  // in SST and a newer merge operand applied through catch-up. The secondary
+  // DB must resolve the blob-backed default column for both Get() and
+  // GetMergeOperands() while combining the newer memtable merge with the older
+  // SST-backed base entity.
+  Options options = GetDefaultOptions();
+  options.create_if_missing = true;
+  options.enable_blob_files = true;
+  options.min_blob_size = 50;
+  options.disable_auto_compactions = true;
+  options.merge_operator = MergeOperators::CreateStringAppendOperator("|");
+  options.env = env_;
+  Reopen(options);
+
+  const std::string key = "secondary_blob_entity";
+  const std::string default_value(100, 'd');
+  const std::string large_value(120, 'l');
+  const std::string merge_operand = "suffix";
+  const std::string expected_merged = default_value + "|" + merge_operand;
+  WideColumns columns{{kDefaultWideColumnName, default_value},
+                      {"col_large", large_value},
+                      {"meta", "inline"}};
+
+  ASSERT_OK(
+      db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(), key, columns));
+  ASSERT_OK(Flush());
+  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+
+  options.max_open_files = -1;
+  OpenSecondary(options);
+  ASSERT_OK(db_->Merge(WriteOptions(), db_->DefaultColumnFamily(), key,
+                       merge_operand));
+  ASSERT_OK(db_secondary_->TryCatchUpWithPrimary());
+
+  auto* cfh = db_secondary_->DefaultColumnFamily();
+
+  {
+    PinnableSlice result;
+    ASSERT_OK(db_secondary_->Get(ReadOptions(), cfh, key, &result));
+    ASSERT_EQ(result, expected_merged);
+  }
+
+  {
+    GetMergeOperandsOptions get_merge_opts;
+    get_merge_opts.expected_max_number_of_operands = 2;
+
+    std::array<PinnableSlice, 2> merge_operands;
+    int number_of_operands = 0;
+
+    ASSERT_OK(db_secondary_->GetMergeOperands(
+        ReadOptions(), cfh, key, merge_operands.data(), &get_merge_opts,
+        &number_of_operands));
+    ASSERT_EQ(number_of_operands, 2);
+    ASSERT_EQ(merge_operands[0], default_value);
+    ASSERT_EQ(merge_operands[1], merge_operand);
+  }
+}
+
+TEST_F(DBSecondaryTest, GetImplReturnsBlobIndexWhenRequested) {
+  // Goal: cover the internal secondary GetImpl contract when the caller
+  // explicitly asks for raw blob-index bytes via `is_blob_index`. Catch-up
+  // rebuilds the blob index into the secondary memtable, and this path must
+  // preserve the encoded index instead of eagerly resolving or rejecting it.
+  Options options = GetDefaultOptions();
+  options.create_if_missing = true;
+  options.env = env_;
+
+  Reopen(options);
+
+  options.max_open_files = -1;
+  OpenSecondary(options);
+
+  std::string blob_index;
+  BlobIndex::EncodeInlinedTTL(&blob_index, /*expiration=*/9876543210, "blob");
+
+  WriteBatch batch;
+  ASSERT_OK(WriteBatchInternal::PutBlobIndex(
+      &batch, db_->DefaultColumnFamily()->GetID(), "blob_key", blob_index));
+  ASSERT_OK(db_->Write(WriteOptions(), &batch));
+  ASSERT_OK(db_secondary_->TryCatchUpWithPrimary());
+
+  PinnableSlice value;
+  bool is_blob_index = false;
+  DBImpl::GetImplOptions get_impl_options;
+  get_impl_options.column_family = db_secondary_->DefaultColumnFamily();
+  get_impl_options.value = &value;
+  get_impl_options.is_blob_index = &is_blob_index;
+
+  ASSERT_OK(db_secondary_full()->GetImpl(ReadOptions(), Slice("blob_key"),
+                                         get_impl_options));
+  ASSERT_TRUE(is_blob_index);
+  ASSERT_EQ(value, blob_index);
+}
+
+TEST_F(DBSecondaryTest,
+       GetAndGetEntityWithBlobBackedDefaultColumnDirectWriteMemtable) {
+  // Goal: cover the secondary memtable path after catch-up replays a blob
+  // direct-write entity from the primary WAL. The test checks both `Get()`,
+  // which must resolve the blob-backed default column, and `GetEntity()`,
+  // which must eagerly resolve all unresolved blob columns instead of exposing
+  // encoded blob indices to the caller.
+  Options options =
+      wide_column_test_util::GetDirectWriteOptions(GetDefaultOptions());
+  options.create_if_missing = true;
+  options.min_blob_size = 50;
+  options.env = env_;
+
+  Reopen(options);
+
+  options.max_open_files = -1;
+  OpenSecondary(options);
+
+  const std::string key = "secondary_direct_write_memtable_entity";
+  const std::string default_value =
+      wide_column_test_util::GenerateLargeValue(100, 'd');
+  const std::string large_value =
+      wide_column_test_util::GenerateLargeValue(120, 'l');
+  const std::string small_value = wide_column_test_util::GenerateSmallValue();
+  WideColumns columns{{kDefaultWideColumnName, default_value},
+                      {"col_large", large_value},
+                      {"col_small", small_value}};
+
+  ASSERT_OK(
+      db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(), key, columns));
+  ASSERT_OK(db_secondary_->TryCatchUpWithPrimary());
+
+  auto* cfh = db_secondary_->DefaultColumnFamily();
+
+  {
+    PinnableSlice result;
+    ASSERT_OK(db_secondary_->Get(ReadOptions(), cfh, key, &result));
+    ASSERT_EQ(result, default_value);
+  }
+
+  {
+    PinnableWideColumns result;
+    ASSERT_OK(db_secondary_->GetEntity(ReadOptions(), cfh, key, &result));
+    ASSERT_EQ(result.columns(), columns);
+  }
+}
+
+TEST_F(DBSecondaryTest, SecondaryDirectWriteMemtableBlobBlockCacheTier) {
+  // Goal: cover the secondary memtable path under kBlockCacheTier. Catch-up
+  // rebuilds the blob-backed direct-write entity from WAL, so serving Get() or
+  // GetEntity() would require blob I/O through the memtable resolution path,
+  // which must return Incomplete instead of issuing that I/O.
+  Options options =
+      wide_column_test_util::GetDirectWriteOptions(GetDefaultOptions());
+  options.create_if_missing = true;
+  options.min_blob_size = 50;
+  options.env = env_;
+
+  Reopen(options);
+
+  options.max_open_files = -1;
+  OpenSecondary(options);
+
+  const std::string key = "secondary_direct_write_memtable_block_cache_tier";
+  const std::string default_value =
+      wide_column_test_util::GenerateLargeValue(100, 'd');
+  const std::string large_value =
+      wide_column_test_util::GenerateLargeValue(120, 'l');
+  const std::string small_value = wide_column_test_util::GenerateSmallValue();
+  WideColumns columns{{kDefaultWideColumnName, default_value},
+                      {"col_large", large_value},
+                      {"col_small", small_value}};
+
+  ASSERT_OK(
+      db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(), key, columns));
+  ASSERT_OK(db_secondary_->TryCatchUpWithPrimary());
+
+  auto* cfh = db_secondary_->DefaultColumnFamily();
+  ReadOptions read_opts;
+  read_opts.read_tier = kBlockCacheTier;
+
+  {
+    PinnableSlice result;
+    const Status s = db_secondary_->Get(read_opts, cfh, key, &result);
+    ASSERT_TRUE(s.IsIncomplete()) << s.ToString();
+    ASSERT_TRUE(result.empty());
+  }
+
+  {
+    PinnableWideColumns result;
+    const Status s = db_secondary_->GetEntity(read_opts, cfh, key, &result);
+    ASSERT_TRUE(s.IsIncomplete()) << s.ToString();
+    ASSERT_TRUE(result.columns().empty());
+  }
+}
+
 TEST_F(DBSecondaryTest, InternalCompactionCompactedFiles) {
   Options options;
   options.env = env_;

db/db_test.cc

@@ -8203,6 +8203,98 @@ TEST_P(OpenFilesAsyncTest, BeforeRead) {
   }
 }
 
+// For modern DBs (manifest carries file_creation_time),
+// GetCreationTimeOfOldestFile returns the real value without waiting for async
+// file open.
+TEST_P(OpenFilesAsyncTest, GetCreationTimeOfOldestFileSkipsWaitForModernDB) {
+  if (max_open_files_ != -1) {
+    return;
+  }
+
+  Options options = CurrentOptions();
+  ASSERT_NO_FATAL_FAILURE(SetupData(options));
+
+  // If WaitForAsyncFileOpen is entered, the test fails -- modern DBs must
+  // never need the wait.
+  std::atomic<bool> waited{false};
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::WaitForAsyncFileOpen::BeforeWait",
+      [&](void* /*arg*/) { waited.store(true); });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  OpenTestDB(options);
+
+  uint64_t creation_time = std::numeric_limits<uint64_t>::max();
+  ASSERT_OK(dbfull()->GetCreationTimeOfOldestFile(&creation_time));
+  EXPECT_GT(creation_time, 0);
+  EXPECT_LT(creation_time, std::numeric_limits<uint64_t>::max());
+  EXPECT_FALSE(waited.load());
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+  Close();
+}
+
+// For legacy DBs (manifest lacks file_creation_time, value comes from the
+// pinned reader), GetCreationTimeOfOldestFile must block until
+// BGWorkAsyncFileOpen pins the readers; otherwise it returns 0 (the "info
+// unavailable" sentinel) instead of the real value.
+TEST_P(OpenFilesAsyncTest,
+       GetCreationTimeOfOldestFileBlocksOnAsyncOpenForLegacyDB) {
+  if (max_open_files_ != -1) {
+    return;
+  }
+
+  Options options = CurrentOptions();
+  ASSERT_NO_FATAL_FAILURE(SetupData(options));
+
+  // Simulate a legacy DB: force TryGetFileCreationTime to return
+  // kUnknownFileCreationTime until BGWorkAsyncFileOpen completes (i.e., what
+  // would happen if the manifest had no file_creation_time and the pinned
+  // reader was the only source).
+  std::atomic<bool> async_open_done{false};
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::BGWorkAsyncFileOpen:Done",
+      [&](void* /*arg*/) { async_open_done.store(true); });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "Version::GetCreationTimeOfOldestFile::FileCreationTime", [&](void* arg) {
+        if (!async_open_done.load()) {
+          *static_cast<uint64_t*>(arg) = kUnknownFileCreationTime;
+        }
+      });
+
+  // Dependency chain: caller's first pass returns 0, enters
+  // WaitForAsyncFileOpen -> main thread confirms wait -> main thread releases
+  // async open -> background worker completes -> caller wakes -> second pass
+  // sees the real value.
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+      {{"DBImpl::WaitForAsyncFileOpen::BeforeWait",
+        "OpenFilesAsyncTest::CreationTime::CallerBlocked"},
+       {"OpenFilesAsyncTest::CreationTime::CallerBlocked",
+        "OpenFilesAsyncTest::CreationTime::ReleaseAsyncOpen"},
+       {"OpenFilesAsyncTest::CreationTime::ReleaseAsyncOpen",
+        "DBImpl::BGWorkAsyncFileOpen::Start"}});
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  OpenTestDB(options);
+
+  uint64_t creation_time = std::numeric_limits<uint64_t>::max();
+  port::Thread caller([&]() {
+    ASSERT_OK(dbfull()->GetCreationTimeOfOldestFile(&creation_time));
+  });
+
+  TEST_SYNC_POINT("OpenFilesAsyncTest::CreationTime::CallerBlocked");
+  TEST_SYNC_POINT("OpenFilesAsyncTest::CreationTime::ReleaseAsyncOpen");
+  caller.join();
+
+  EXPECT_GT(creation_time, 0);
+  EXPECT_LT(creation_time, std::numeric_limits<uint64_t>::max());
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+  Close();
+}
+
 TEST_P(OpenFilesAsyncTest, Shutdown) {
   Options options = CurrentOptions();
   ASSERT_NO_FATAL_FAILURE(SetupData(options));

db/db_test2.cc

@@ -107,7 +107,7 @@ TEST_F(DBTest2, ReadOnlyDBWalInDbPathInitialized) {
   ASSERT_OK(Put("key2", "value2"));
   Close();
 
-  // Reopen as read-only — wal_in_db_path_ must be properly initialized.
+  // Reopen as read-only -- wal_in_db_path_ must be properly initialized.
   // Before the fix, closing this DB would read an uninitialized bool in
   // DeleteObsoleteFileImpl, which UBSan catches as undefined behavior.
   std::unique_ptr<DB> db_ptr;
@@ -115,7 +115,7 @@ TEST_F(DBTest2, ReadOnlyDBWalInDbPathInitialized) {
   std::string value;
   ASSERT_OK(db_ptr->Get(ReadOptions(), "key1", &value));
   ASSERT_EQ("value1", value);
-  // Close the read-only DB — this triggers PurgeObsoleteFiles which reads
+  // Close the read-only DB -- this triggers PurgeObsoleteFiles which reads
   // wal_in_db_path_. Under UBSan, an uninitialized bool here would fail.
   db_ptr.reset();
 
@@ -7510,6 +7510,65 @@ TEST_F(DBTest2, GetLatestSeqAndTsForKey) {
   ASSERT_EQ(0, options.statistics->getTickerCount(GET_HIT_L0));
 }
 
+TEST_F(DBTest2,
+       GetLatestSequenceForKeyFromHistoryWithBlobBackedWideColumnEntity) {
+  // Goal: exercise the memtable history lookup in GetLatestSequenceForKey()
+  // after blob direct write stores a blob-backed V2 entity in a flushed
+  // memtable. Using cache_only=true ensures the lookup succeeds only if the
+  // history path can resolve the wide-column base value under the newer merge.
+  Destroy(last_options_);
+
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.enable_blob_files = true;
+  options.enable_blob_direct_write = true;
+  options.min_blob_size = 50;
+  options.allow_concurrent_memtable_write = false;
+  options.blob_direct_write_partitions = 1;
+  options.max_write_buffer_size_to_maintain = 64 << 10;
+  options.merge_operator = MergeOperators::CreateStringAppendOperator("|");
+
+  Reopen(options);
+
+  const std::string key = "history_blob_entity";
+  const std::string default_value(100, 'd');
+  const std::string merge_operand = "suffix";
+  WideColumns columns{{kDefaultWideColumnName, default_value},
+                      {"meta", "inline"}};
+
+  ASSERT_OK(
+      db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(), key, columns));
+  ASSERT_OK(db_->Merge(WriteOptions(), db_->DefaultColumnFamily(), key,
+                       merge_operand));
+  const SequenceNumber expected_seq = dbfull()->GetLatestSequenceNumber();
+
+  ASSERT_OK(Flush());
+  ASSERT_FALSE(GetBlobFileNumbers().empty());
+
+  uint64_t num_immutable_memtables = 0;
+  ASSERT_TRUE(db_->GetIntProperty(DB::Properties::kNumImmutableMemTable,
+                                  &num_immutable_memtables));
+  ASSERT_EQ(num_immutable_memtables, 0);
+
+  auto* cfhi = static_cast_with_check<ColumnFamilyHandleImpl>(
+      dbfull()->DefaultColumnFamily());
+  assert(cfhi);
+  assert(cfhi->cfd());
+  SuperVersion* sv = cfhi->cfd()->GetSuperVersion();
+
+  SequenceNumber seq = kMaxSequenceNumber;
+  bool found_record_for_key = false;
+  bool is_blob_index = false;
+  const Status s = dbfull()->GetLatestSequenceForKey(
+      sv, key, /*cache_only=*/true, /*lower_bound_seq=*/0, &seq,
+      /*timestamp=*/nullptr, &found_record_for_key, &is_blob_index);
+
+  ASSERT_OK(s);
+  ASSERT_TRUE(found_record_for_key);
+  ASSERT_EQ(expected_seq, seq);
+  ASSERT_FALSE(is_blob_index);
+}
+
 #if defined(ZSTD)
 TEST_F(DBTest2, ZSTDChecksum) {
   // Verify that corruption during decompression is caught.
@@ -8173,7 +8232,7 @@ TEST_F(DBTest2, FastSstOpenDisableAfterMetadataPersisted) {
   ASSERT_EQ(1, test_fs->GetMetadataRetrievedCount());
   db.reset();
 
-  // Reopen with fast_sst_open DISABLED — metadata is in the MANIFEST
+  // Reopen with fast_sst_open DISABLED -- metadata is in the MANIFEST
   // but should NOT be passed to the filesystem
   options.fast_sst_open = false;
   test_fs->ResetCounters();

db/dbformat_test.cc

@@ -218,7 +218,7 @@ TEST_P(IterKeySwapTest, SwapAndDestroy) {
         b_use_secondary] = GetParam();
 
   std::string expected_a, expected_b;
-  // Storage for pinned keys — must outlive the IterKeys.
+  // Storage for pinned keys -- must outlive the IterKeys.
   std::string a_storage, b_storage;
   IterKey a;
   expected_a =
@@ -236,7 +236,7 @@ TEST_P(IterKeySwapTest, SwapAndDestroy) {
     // After swap: a has b's old data, b has a's old data.
     ASSERT_EQ(a.GetUserKey().ToString(), expected_b);
     ASSERT_EQ(b.GetUserKey().ToString(), expected_a);
-  }  // b destroyed here — must not corrupt a's data
+  }  // b destroyed here -- must not corrupt a's data
 
   // a must still hold valid data after b's destruction.
   ASSERT_EQ(a.GetUserKey().ToString(), expected_b);

db/listener_test.cc

@@ -1696,7 +1696,7 @@ TEST_F(EventListenerTest, BackgroundJobPressure) {
                  Env::Priority::LOW);
   sleeping_task.WaitUntilSleeping();
 
-  // Phase 1: No pressure — 3 SST files (below slowdown trigger=4).
+  // Phase 1: No pressure -- 3 SST files (below slowdown trigger=4).
   for (int i = 0; i < 3; i++) {
     ASSERT_OK(Put("k" + std::to_string(i), std::string(100, 'x')));
     ASSERT_OK(Flush());
@@ -1714,7 +1714,7 @@ TEST_F(EventListenerTest, BackgroundJobPressure) {
   // MaybeScheduleFlushOrCompaction() runs before the pressure callback and
   // may schedule new flush work, making flush counts non-deterministic.
 
-  // Phase 2: Build pressure — flush past slowdown trigger (4 L0 SST files).
+  // Phase 2: Build pressure -- flush past slowdown trigger (4 L0 SST files).
   // Compaction is blocked, so L0 SST files pile up.
   listener->Reset();
   {
@@ -1750,7 +1750,7 @@ TEST_F(EventListenerTest, BackgroundJobPressure) {
   ASSERT_TRUE(found_compaction_scheduled);
   ASSERT_TRUE(found_high_proximity);
 
-  // Phase 3: Relieve pressure — unblock compaction, wait for completion.
+  // Phase 3: Relieve pressure -- unblock compaction, wait for completion.
   listener->Reset();
   sleeping_task.WakeUp();
   sleeping_task.WaitUntilDone();

db/log_writer.cc

@@ -22,9 +22,10 @@ namespace ROCKSDB_NAMESPACE::log {
 
 Writer::Writer(std::unique_ptr<WritableFileWriter>&& dest, uint64_t log_number,
                bool recycle_log_files, bool manual_flush,
-               CompressionType compression_type, bool track_and_verify_wals)
+               CompressionType compression_type, bool track_and_verify_wals,
+               size_t initial_block_offset)
     : dest_(std::move(dest)),
-      block_offset_(0),
+      block_offset_(initial_block_offset),
       log_number_(log_number),
       recycle_log_files_(recycle_log_files),
       // Header size varies depending on whether we are recycling or not.

db/log_writer.h

@@ -75,15 +75,20 @@ namespace log {
 class Writer {
  public:
   // Create a writer that will append data to "*dest".
-  // "*dest" must be initially empty.
-  // "*dest" must remain live while this Writer is in use.
+  // "*dest" must remain live while this Writer is in use. By default
+  // "*dest" is expected to be empty (initial_block_offset = 0). When
+  // resuming append into an existing log file (e.g.
+  // VersionSet::ReopenManifestForAppend), pass the offset within the
+  // current 32 KiB block at which writes should resume so that record
+  // framing aligns to the existing block layout.
   // TODO(hx235): separate WAL related parameters from general `Reader`
   // parameters
   explicit Writer(std::unique_ptr<WritableFileWriter>&& dest,
                   uint64_t log_number, bool recycle_log_files,
                   bool manual_flush = false,
                   CompressionType compressionType = kNoCompression,
-                  bool track_and_verify_wals = false);
+                  bool track_and_verify_wals = false,
+                  size_t initial_block_offset = 0);
   // No copying allowed
   Writer(const Writer&) = delete;
   void operator=(const Writer&) = delete;

db/memtable.cc

@@ -15,6 +15,7 @@
 #include <memory>
 #include <optional>
 
+#include "db/blob/blob_fetcher.h"
 #include "db/blob/blob_file_partition_manager.h"
 #include "db/blob/blob_index.h"
 #include "db/dbformat.h"
@@ -70,6 +71,67 @@ Status GetDefaultColumnBlobIndexSlice(Slice entity, Slice* blob_index_slice) {
   return status;
 }
 
+Status PushWideColumnEntityDefaultOperand(const Slice& user_key,
+                                          const Slice& entity,
+                                          MergeContext* merge_context,
+                                          bool operand_pinned,
+                                          const BlobFetcher* blob_fetcher) {
+  assert(merge_context != nullptr);
+
+  Slice entity_ref = entity;
+  Slice value_of_default;
+  Status status = WideColumnSerialization::GetValueOfDefaultColumn(
+      entity_ref, value_of_default);
+  if (status.ok()) {
+    merge_context->PushOperand(value_of_default, operand_pinned);
+    return status;
+  }
+  if (!status.IsNotSupported()) {
+    return status;
+  }
+  if (blob_fetcher == nullptr) {
+    return Status::Corruption(
+        "Cannot resolve blob-backed default column without a blob fetcher");
+  }
+
+  PinnableSlice resolved_default;
+  bool resolved = false;
+  status = WideColumnSerialization::GetValueOfDefaultColumnResolvingBlobs(
+      entity, user_key, blob_fetcher, resolved_default, resolved);
+  if (status.ok()) {
+    // Resolved blob values are backed by this stack-local PinnableSlice, so
+    // copy them into MergeContext instead of pinning their storage.
+    merge_context->PushOperand(Slice(resolved_default), false);
+  }
+  return status;
+}
+
+Status MergeWithWideColumnEntityBaseValue(
+    const Slice& user_key, const Slice& entity,
+    const MergeOperator* merge_operator, MergeContext* merge_context,
+    Logger* logger, Statistics* statistics, SystemClock* clock,
+    std::string* value, PinnableWideColumns* columns,
+    const BlobFetcher* blob_fetcher) {
+  assert(merge_context != nullptr);
+
+  std::string resolved_entity;
+  Slice effective_entity;
+  Status status = WideColumnSerialization::ResolveEntityForMerge(
+      entity, user_key, blob_fetcher, nullptr /* prefetch_buffers */,
+      resolved_entity, effective_entity);
+  if (!status.ok()) {
+    return status;
+  }
+
+  // `op_failure_scope` (an output parameter) is not provided (set to nullptr)
+  // since a failure must be propagated regardless of its value.
+  return MergeHelper::TimedFullMerge(
+      merge_operator, user_key, MergeHelper::kWideBaseValue, effective_entity,
+      merge_context->GetOperands(), logger, statistics, clock,
+      /* update_num_ops_stats */ true, /* op_failure_scope */ nullptr, value,
+      columns);
+}
+
 }  // namespace
 
 ImmutableMemTableOptions::ImmutableMemTableOptions(
@@ -1254,6 +1316,7 @@ struct Saver {
   bool inplace_update_support;
   bool do_merge;
   SystemClock* clock;
+  const BlobFetcher* blob_fetcher;
 
   ReadCallback* callback_;
   bool* is_blob_index;
@@ -1424,27 +1487,17 @@ static bool SaveValue(void* arg, const char* entry) {
           // Preserve the value with the goal of returning it as part of
           // raw merge operands to the user
 
-          Slice value_of_default;
-          *(s->status) = WideColumnSerialization::GetValueOfDefaultColumn(
-              v, value_of_default);
-
-          if (s->status->ok()) {
-            merge_context->PushOperand(
-                value_of_default,
-                s->inplace_update_support == false /* operand_pinned */);
-          }
+          *(s->status) = PushWideColumnEntityDefaultOperand(
+              s->key->user_key(), v, merge_context,
+              s->inplace_update_support == false /* operand_pinned */,
+              s->blob_fetcher);
         } else if (*(s->merge_in_progress)) {
           assert(s->do_merge);
 
           if (s->value || s->columns) {
-            // `op_failure_scope` (an output parameter) is not provided (set
-            // to nullptr) since a failure must be propagated regardless of
-            // its value.
-            *(s->status) = MergeHelper::TimedFullMerge(
-                merge_operator, s->key->user_key(), MergeHelper::kWideBaseValue,
-                v, merge_context->GetOperands(), s->logger, s->statistics,
-                s->clock, /* update_num_ops_stats */ true,
-                /* op_failure_scope */ nullptr, s->value, s->columns);
+            *(s->status) = MergeWithWideColumnEntityBaseValue(
+                s->key->user_key(), v, merge_operator, merge_context, s->logger,
+                s->statistics, s->clock, s->value, s->columns, s->blob_fetcher);
           }
         } else if (s->value) {
           Slice value_of_default;
@@ -1520,7 +1573,8 @@ bool MemTable::Get(const LookupKey& key, std::string* value,
                    SequenceNumber* max_covering_tombstone_seq,
                    SequenceNumber* seq, const ReadOptions& read_opts,
                    bool immutable_memtable, ReadCallback* callback,
-                   bool* is_blob_index, bool do_merge) {
+                   bool* is_blob_index, bool do_merge,
+                   const BlobFetcher* blob_fetcher) {
   // The sequence number is updated synchronously in version_set.h
   if (IsEmpty()) {
     // Avoiding recording stats for speed.
@@ -1578,7 +1632,7 @@ bool MemTable::Get(const LookupKey& key, std::string* value,
     }
     GetFromTable(key, *max_covering_tombstone_seq, do_merge, callback,
                  is_blob_index, value, columns, timestamp, s, merge_context,
-                 seq, &found_final_value, &merge_in_progress);
+                 seq, &found_final_value, &merge_in_progress, blob_fetcher);
   }
 
   // No change to value, since we have not yet found a Put/Delete
@@ -1601,7 +1655,8 @@ void MemTable::GetFromTable(const LookupKey& key,
                             PinnableWideColumns* columns,
                             std::string* timestamp, Status* s,
                             MergeContext* merge_context, SequenceNumber* seq,
-                            bool* found_final_value, bool* merge_in_progress) {
+                            bool* found_final_value, bool* merge_in_progress,
+                            const BlobFetcher* blob_fetcher) {
   Saver saver;
   saver.status = s;
   saver.found_final_value = found_final_value;
@@ -1619,6 +1674,7 @@ void MemTable::GetFromTable(const LookupKey& key,
   saver.inplace_update_support = moptions_.inplace_update_support;
   saver.statistics = moptions_.statistics;
   saver.clock = clock_;
+  saver.blob_fetcher = blob_fetcher;
   saver.callback_ = callback;
   saver.is_blob_index = is_blob_index;
   saver.do_merge = do_merge;
@@ -1648,7 +1704,8 @@ Status MemTable::ValidateKey(const char* key, bool allow_data_in_errors) {
 }
 
 void MemTable::MultiGet(const ReadOptions& read_options, MultiGetRange* range,
-                        ReadCallback* callback, bool immutable_memtable) {
+                        ReadCallback* callback, bool immutable_memtable,
+                        const BlobFetcher* blob_fetcher) {
   // The sequence number is updated synchronously in version_set.h
   if (IsEmpty()) {
     // Avoiding recording stats for speed.
@@ -1740,6 +1797,7 @@ void MemTable::MultiGet(const ReadOptions& read_options, MultiGetRange* range,
       saver.inplace_update_support = moptions_.inplace_update_support;
       saver.statistics = moptions_.statistics;
       saver.clock = clock_;
+      saver.blob_fetcher = blob_fetcher;
       saver.callback_ = callback;
       saver.is_blob_index = &iter->is_blob_index;
       saver.do_merge = true;
@@ -1831,7 +1889,7 @@ void MemTable::MultiGet(const ReadOptions& read_options, MultiGetRange* range,
           *(iter->lkey), iter->max_covering_tombstone_seq, true, callback,
           &iter->is_blob_index, iter->value ? iter->value->GetSelf() : nullptr,
           iter->columns, iter->timestamp, iter->s, &(iter->merge_context),
-          &dummy_seq, &found_final_value, &merge_in_progress);
+          &dummy_seq, &found_final_value, &merge_in_progress, blob_fetcher);
 
       if (!found_final_value && merge_in_progress) {
         if (iter->s->ok()) {

db/memtable.h

@@ -39,6 +39,7 @@
 
 namespace ROCKSDB_NAMESPACE {
 
+class BlobFetcher;
 class BlobFilePartitionManager;
 struct FlushJobInfo;
 class Mutex;
@@ -236,25 +237,27 @@ class ReadOnlyMemTable {
                    SequenceNumber* max_covering_tombstone_seq,
                    SequenceNumber* seq, const ReadOptions& read_opts,
                    bool immutable_memtable, ReadCallback* callback = nullptr,
-                   bool* is_blob_index = nullptr, bool do_merge = true) = 0;
+                   bool* is_blob_index = nullptr, bool do_merge = true,
+                   const BlobFetcher* blob_fetcher = nullptr) = 0;
   bool Get(const LookupKey& key, std::string* value,
            PinnableWideColumns* columns, std::string* timestamp, Status* s,
            MergeContext* merge_context,
            SequenceNumber* max_covering_tombstone_seq,
            const ReadOptions& read_opts, bool immutable_memtable,
            ReadCallback* callback = nullptr, bool* is_blob_index = nullptr,
-           bool do_merge = true) {
+           bool do_merge = true, const BlobFetcher* blob_fetcher = nullptr) {
     SequenceNumber seq;
     return Get(key, value, columns, timestamp, s, merge_context,
                max_covering_tombstone_seq, &seq, read_opts, immutable_memtable,
-               callback, is_blob_index, do_merge);
+               callback, is_blob_index, do_merge, blob_fetcher);
   }
 
   // @param immutable_memtable Whether this memtable is immutable. Used
   // internally by NewRangeTombstoneIterator(). See comment above
   // NewRangeTombstoneIterator() for more detail.
   virtual void MultiGet(const ReadOptions& read_options, MultiGetRange* range,
-                        ReadCallback* callback, bool immutable_memtable) = 0;
+                        ReadCallback* callback, bool immutable_memtable,
+                        const BlobFetcher* blob_fetcher = nullptr) = 0;
 
   // Get total number of entries in the mem table.
   // REQUIRES: external synchronization to prevent simultaneous
@@ -320,7 +323,7 @@ class ReadOnlyMemTable {
   virtual uint64_t ApproximateOldestKeyTime() const = 0;
 
   // Inserts a range tombstone [start_key, end_key) that is logically redundant
-  // — it is derived from existing point tombstones observed during iteration
+  // -- it is derived from existing point tombstones observed during iteration
   // and does not delete any data that isn't already deleted. This is a
   // best-effort optimization. It allows future reads to skip iterating over
   // continuous single deletion tombstones.
@@ -688,10 +691,12 @@ class MemTable final : public ReadOnlyMemTable {
            SequenceNumber* max_covering_tombstone_seq, SequenceNumber* seq,
            const ReadOptions& read_opts, bool immutable_memtable,
            ReadCallback* callback = nullptr, bool* is_blob_index = nullptr,
-           bool do_merge = true) override;
+           bool do_merge = true,
+           const BlobFetcher* blob_fetcher = nullptr) override;
 
   void MultiGet(const ReadOptions& read_options, MultiGetRange* range,
-                ReadCallback* callback, bool immutable_memtable) override;
+                ReadCallback* callback, bool immutable_memtable,
+                const BlobFetcher* blob_fetcher = nullptr) override;
 
   // If `key` exists in current memtable with type value_type and the existing
   // value is at least as large as the new value, updates it in-place. Otherwise
@@ -991,7 +996,8 @@ class MemTable final : public ReadOnlyMemTable {
                     std::string* value, PinnableWideColumns* columns,
                     std::string* timestamp, Status* s,
                     MergeContext* merge_context, SequenceNumber* seq,
-                    bool* found_final_value, bool* merge_in_progress);
+                    bool* found_final_value, bool* merge_in_progress,
+                    const BlobFetcher* blob_fetcher);
 
   // Always returns non-null and assumes certain pre-checks (e.g.,
   // is_range_del_table_empty_) are done. This is only valid during the lifetime

db/memtable_list.cc

@@ -108,18 +108,19 @@ bool MemTableListVersion::Get(const LookupKey& key, std::string* value,
                               MergeContext* merge_context,
                               SequenceNumber* max_covering_tombstone_seq,
                               SequenceNumber* seq, const ReadOptions& read_opts,
-                              ReadCallback* callback, bool* is_blob_index) {
+                              ReadCallback* callback, bool* is_blob_index,
+                              const BlobFetcher* blob_fetcher) {
   return GetFromList(&memlist_, key, value, columns, timestamp, s,
                      merge_context, max_covering_tombstone_seq, seq, read_opts,
-                     callback, is_blob_index);
+                     callback, is_blob_index, blob_fetcher);
 }
 
 void MemTableListVersion::MultiGet(const ReadOptions& read_options,
-                                   MultiGetRange* range,
-                                   ReadCallback* callback) {
+                                   MultiGetRange* range, ReadCallback* callback,
+                                   const BlobFetcher* blob_fetcher) {
   for (auto memtable : memlist_) {
     memtable->MultiGet(read_options, range, callback,
-                       true /* immutable_memtable */);
+                       true /* immutable_memtable */, blob_fetcher);
     if (range->empty()) {
       return;
     }
@@ -128,12 +129,13 @@ void MemTableListVersion::MultiGet(const ReadOptions& read_options,
 
 bool MemTableListVersion::GetMergeOperands(
     const LookupKey& key, Status* s, MergeContext* merge_context,
-    SequenceNumber* max_covering_tombstone_seq, const ReadOptions& read_opts) {
+    SequenceNumber* max_covering_tombstone_seq, const ReadOptions& read_opts,
+    const BlobFetcher* blob_fetcher) {
   for (ReadOnlyMemTable* memtable : memlist_) {
     bool done = memtable->Get(
         key, /*value=*/nullptr, /*columns=*/nullptr, /*timestamp=*/nullptr, s,
         merge_context, max_covering_tombstone_seq, read_opts,
-        true /* immutable_memtable */, nullptr, nullptr, false);
+        true /* immutable_memtable */, nullptr, nullptr, false, blob_fetcher);
     if (done) {
       return true;
     }
@@ -145,10 +147,11 @@ bool MemTableListVersion::GetFromHistory(
     const LookupKey& key, std::string* value, PinnableWideColumns* columns,
     std::string* timestamp, Status* s, MergeContext* merge_context,
     SequenceNumber* max_covering_tombstone_seq, SequenceNumber* seq,
-    const ReadOptions& read_opts, bool* is_blob_index) {
+    const ReadOptions& read_opts, bool* is_blob_index,
+    const BlobFetcher* blob_fetcher) {
   return GetFromList(&memlist_history_, key, value, columns, timestamp, s,
                      merge_context, max_covering_tombstone_seq, seq, read_opts,
-                     nullptr /*read_callback*/, is_blob_index);
+                     nullptr /*read_callback*/, is_blob_index, blob_fetcher);
 }
 
 bool MemTableListVersion::GetFromList(
@@ -156,17 +159,18 @@ bool MemTableListVersion::GetFromList(
     std::string* value, PinnableWideColumns* columns, std::string* timestamp,
     Status* s, MergeContext* merge_context,
     SequenceNumber* max_covering_tombstone_seq, SequenceNumber* seq,
-    const ReadOptions& read_opts, ReadCallback* callback, bool* is_blob_index) {
+    const ReadOptions& read_opts, ReadCallback* callback, bool* is_blob_index,
+    const BlobFetcher* blob_fetcher) {
   *seq = kMaxSequenceNumber;
 
   for (auto& memtable : *list) {
     assert(memtable->IsFragmentedRangeTombstonesConstructed());
     SequenceNumber current_seq = kMaxSequenceNumber;
 
-    bool done =
-        memtable->Get(key, value, columns, timestamp, s, merge_context,
-                      max_covering_tombstone_seq, &current_seq, read_opts,
-                      true /* immutable_memtable */, callback, is_blob_index);
+    bool done = memtable->Get(key, value, columns, timestamp, s, merge_context,
+                              max_covering_tombstone_seq, &current_seq,
+                              read_opts, true /* immutable_memtable */,
+                              callback, is_blob_index, true, blob_fetcher);
     if (*seq == kMaxSequenceNumber) {
       // Store the most recent sequence number of any operation on this key.
       // Since we only care about the most recent change, we only need to

db/memtable_list.h

@@ -61,29 +61,33 @@ class MemTableListVersion {
            MergeContext* merge_context,
            SequenceNumber* max_covering_tombstone_seq, SequenceNumber* seq,
            const ReadOptions& read_opts, ReadCallback* callback = nullptr,
-           bool* is_blob_index = nullptr);
+           bool* is_blob_index = nullptr,
+           const BlobFetcher* blob_fetcher = nullptr);
 
   bool Get(const LookupKey& key, std::string* value,
            PinnableWideColumns* columns, std::string* timestamp, Status* s,
            MergeContext* merge_context,
            SequenceNumber* max_covering_tombstone_seq,
            const ReadOptions& read_opts, ReadCallback* callback = nullptr,
-           bool* is_blob_index = nullptr) {
+           bool* is_blob_index = nullptr,
+           const BlobFetcher* blob_fetcher = nullptr) {
     SequenceNumber seq;
     return Get(key, value, columns, timestamp, s, merge_context,
                max_covering_tombstone_seq, &seq, read_opts, callback,
-               is_blob_index);
+               is_blob_index, blob_fetcher);
   }
 
   void MultiGet(const ReadOptions& read_options, MultiGetRange* range,
-                ReadCallback* callback);
+                ReadCallback* callback,
+                const BlobFetcher* blob_fetcher = nullptr);
 
   // Returns all the merge operands corresponding to the key by searching all
   // memtables starting from the most recent one.
   bool GetMergeOperands(const LookupKey& key, Status* s,
                         MergeContext* merge_context,
                         SequenceNumber* max_covering_tombstone_seq,
-                        const ReadOptions& read_opts);
+                        const ReadOptions& read_opts,
+                        const BlobFetcher* blob_fetcher = nullptr);
 
   // Similar to Get(), but searches the Memtable history of memtables that
   // have already been flushed.  Should only be used from in-memory only
@@ -94,17 +98,19 @@ class MemTableListVersion {
                       Status* s, MergeContext* merge_context,
                       SequenceNumber* max_covering_tombstone_seq,
                       SequenceNumber* seq, const ReadOptions& read_opts,
-                      bool* is_blob_index = nullptr);
+                      bool* is_blob_index = nullptr,
+                      const BlobFetcher* blob_fetcher = nullptr);
   bool GetFromHistory(const LookupKey& key, std::string* value,
                       PinnableWideColumns* columns, std::string* timestamp,
                       Status* s, MergeContext* merge_context,
                       SequenceNumber* max_covering_tombstone_seq,
                       const ReadOptions& read_opts,
-                      bool* is_blob_index = nullptr) {
+                      bool* is_blob_index = nullptr,
+                      const BlobFetcher* blob_fetcher = nullptr) {
     SequenceNumber seq;
     return GetFromHistory(key, value, columns, timestamp, s, merge_context,
                           max_covering_tombstone_seq, &seq, read_opts,
-                          is_blob_index);
+                          is_blob_index, blob_fetcher);
   }
 
   Status AddRangeTombstoneIterators(const ReadOptions& read_opts, Arena* arena,
@@ -188,7 +194,8 @@ class MemTableListVersion {
                    SequenceNumber* max_covering_tombstone_seq,
                    SequenceNumber* seq, const ReadOptions& read_opts,
                    ReadCallback* callback = nullptr,
-                   bool* is_blob_index = nullptr);
+                   bool* is_blob_index = nullptr,
+                   const BlobFetcher* blob_fetcher = nullptr);
 
   void AddMemTable(ReadOnlyMemTable* m);
 

db/version_edit.cc

@@ -30,7 +30,7 @@ PinnedTableReader& PinnedTableReader::operator=(
   TableReader* r = other.reader_.load(std::memory_order_acquire);
   // Only read handle_ when reader_ is non-null. Pin() writes handle_ before
   // reader_ (with release), so a non-null reader_ guarantees handle_ is stable.
-  // If reader_ is null, Pin() may be in progress — avoid reading handle_.
+  // If reader_ is null, Pin() may be in progress -- avoid reading handle_.
   handle_ = (r != nullptr) ? other.handle_ : nullptr;
   reader_.store(r, std::memory_order_release);
   return *this;
@@ -109,6 +109,17 @@ Status FileMetaData::UpdateBoundaries(const Slice& key, const Slice& value,
 
 void VersionEdit::Clear() { *this = VersionEdit(); }
 
+bool VersionEdit::ShouldEmitPerColumnFamilyRecoveryEdit(
+    uint64_t current_log_number) const {
+  return (HasLogNumber() && GetLogNumber() > current_log_number) ||
+         NumEntries() > 0 || HasComparatorName() || HasPrevLogNumber() ||
+         HasNextFile() || HasMaxColumnFamily() || HasMinLogNumberToKeep() ||
+         HasLastSequence() || !GetCompactCursors().empty() || HasDbId() ||
+         IsColumnFamilyManipulation() || IsInAtomicGroup() ||
+         HasFullHistoryTsLow() || HasPersistUserDefinedTimestamps() ||
+         HasSubcompactionProgress();
+}
+
 bool VersionEdit::EncodeTo(std::string* dst,
                            std::optional<size_t> ts_sz) const {
   assert(!IsNoManifestWriteDummy());

db/version_edit.h

@@ -1016,6 +1016,10 @@ class VersionEdit {
     subcompaction_progress_.Clear();
   }
 
+  // Recovery-time per-column-family edits only need to be written when they
+  // advance the CF's log number or carry some other manifest state.
+  bool ShouldEmitPerColumnFamilyRecoveryEdit(uint64_t current_log_number) const;
+
   // return true on success.
   // `ts_sz` is the size in bytes for the user-defined timestamp contained in
   // a user key. This argument is optional because it's only required for

db/version_edit_handler.cc

@@ -35,6 +35,7 @@ void VersionEditHandlerBase::Iterate(log::Reader& reader,
     VersionEdit edit;
     s = edit.DecodeFrom(record);
     if (s.ok()) {
+      last_valid_record_end_ = reader.LastRecordEnd();
       s = read_buffer_.AddEdit(&edit);
     }
     if (s.ok()) {
@@ -460,6 +461,7 @@ void VersionEditHandler::CheckIterationResult(const log::Reader& reader,
   if (s->ok()) {
     version_set_->manifest_file_size_ = reader.GetReadOffset();
     assert(version_set_->manifest_file_size_ > 0);
+    version_set_->manifest_last_valid_record_end_ = last_valid_record_end_;
     version_set_->next_file_number_.store(version_edit_params_.GetNextFile() +
                                           1);
     SequenceNumber last_seq = version_edit_params_.GetLastSequence();

db/version_edit_handler.h

@@ -31,6 +31,8 @@ class VersionEditHandlerBase {
 
   AtomicGroupReadBuffer& GetReadBuffer() { return read_buffer_; }
 
+  uint64_t GetLastValidRecordEnd() const { return last_valid_record_end_; }
+
  protected:
   explicit VersionEditHandlerBase(const ReadOptions& read_options,
                                   uint64_t max_read_size)
@@ -52,6 +54,10 @@ class VersionEditHandlerBase {
 
   const ReadOptions& read_options_;
 
+  // File offset at the end of the last successfully completed and
+  // decoded logical record.
+  uint64_t last_valid_record_end_ = 0;
+
  private:
   AtomicGroupReadBuffer read_buffer_;
   const uint64_t max_manifest_read_size_;

db/version_set.cc

@@ -2243,8 +2243,10 @@ void Version::GetCreationTimeOfOldestFile(uint64_t* creation_time) {
   uint64_t oldest_time = std::numeric_limits<uint64_t>::max();
   for (int level = 0; level < storage_info_.num_non_empty_levels_; level++) {
     for (FileMetaData* meta : storage_info_.LevelFiles(level)) {
-      assert(meta->fd.pinned_reader.Get() != nullptr);
       uint64_t file_creation_time = meta->TryGetFileCreationTime();
+      TEST_SYNC_POINT_CALLBACK(
+          "Version::GetCreationTimeOfOldestFile::FileCreationTime",
+          &file_creation_time);
       if (file_creation_time == kUnknownFileCreationTime) {
         *creation_time = 0;
         return;
@@ -4217,10 +4219,10 @@ void VersionStorageInfo::ComputeFilesMarkedForReadTriggeredCompaction(
     return;
   }
 
-  // Skip files at the last non-empty level — there is no lower level to
+  // Skip files at the last non-empty level -- there is no lower level to
   // compact into. Exception: L0 files are allowed even when L0 is the last
   // non-empty level, because in single-level universal or FIFO compaction, L0
-  // files can be compacted together (L0 → L0).
+  // files can be compacted together (L0 -> L0).
   int last_non_empty = num_non_empty_levels_ - 1;
 
   for (int level = 0; level < num_levels(); level++) {
@@ -5618,6 +5620,7 @@ VersionSet::VersionSet(
       prev_log_number_(0),
       current_version_number_(0),
       manifest_file_size_(0),
+      manifest_last_valid_record_end_(0),
       last_compacted_manifest_file_size_(0),
       file_options_(storage_options),
       block_cache_tracer_(block_cache_tracer),
@@ -5689,7 +5692,7 @@ Status VersionSet::Close(FSDirectory* db_dir, InstrumentedMutex* mu) {
           content_manifest_name, fs_->OptimizeForManifestRead(file_options_),
           &manifest_file, nullptr);
       if (!content_io_s.ok()) {
-        // Surface I/O errors to the caller — users who call DB::Close() and
+        // Surface I/O errors to the caller -- users who call DB::Close() and
         // check the status should know about filesystem problems.
         s = content_io_s;
         ROCKS_LOG_ERROR(db_options_->info_log,
@@ -5741,7 +5744,7 @@ Status VersionSet::Close(FSDirectory* db_dir, InstrumentedMutex* mu) {
       corrupt_io_s.PermitUncheckedError();
       io_error_info.io_status.PermitUncheckedError();
       if (content_check == 0) {
-        // First check failed — rewrite and verify again
+        // First check failed -- rewrite and verify again
         ROCKS_LOG_ERROR(db_options_->info_log,
                         "MANIFEST content verification on Close failed, "
                         "filename %s, rewriting manifest\n",
@@ -5752,7 +5755,7 @@ Status VersionSet::Close(FSDirectory* db_dir, InstrumentedMutex* mu) {
         s = LogAndApply(cfd, ReadOptions(), WriteOptions(), &recovery_edit, mu,
                         db_dir);
       } else {
-        // Rewritten manifest is also corrupt — likely a recurring filesystem
+        // Rewritten manifest is also corrupt -- likely a recurring filesystem
         // issue. Surface it so DB::Close() callers can detect the problem.
         ROCKS_LOG_ERROR(db_options_->info_log,
                         "MANIFEST content verification on Close failed again "
@@ -5827,6 +5830,7 @@ void VersionSet::Reset() {
   current_version_number_ = 0;
   manifest_writers_.clear();
   manifest_file_size_ = 0;
+  manifest_last_valid_record_end_ = 0;
   last_compacted_manifest_file_size_ = 0;
   TuneMaxManifestFileSize();
   obsolete_files_.clear();
@@ -6161,11 +6165,7 @@ Status VersionSet::ProcessManifestWrites(
       }
     }
   } else {
-    FileOptions opt_file_opts = fs_->OptimizeForManifestWrite(file_options_);
-    // DB option (in file_options_) takes precedence when not kUnknown
-    if (file_options_.temperature != Temperature::kUnknown) {
-      opt_file_opts.temperature = file_options_.temperature;
-    }
+    FileOptions opt_file_opts = GetFileOptionsForManifestWrite();
     mu->Unlock();
     TEST_SYNC_POINT("VersionSet::LogAndApply:WriteManifestStart");
     TEST_SYNC_POINT_CALLBACK("VersionSet::LogAndApply:WriteManifest", nullptr);
@@ -6200,16 +6200,9 @@ Status VersionSet::ProcessManifestWrites(
       io_s = NewWritableFile(fs_.get(), descriptor_fname, &descriptor_file,
                              opt_file_opts);
       if (io_s.ok()) {
-        descriptor_file->SetPreallocationBlockSize(manifest_preallocation_size);
-        FileTypeSet tmp_set = db_options_->checksum_handoff_file_types;
-        std::unique_ptr<WritableFileWriter> file_writer(new WritableFileWriter(
-            std::move(descriptor_file), descriptor_fname, opt_file_opts, clock_,
-            io_tracer_, nullptr, Histograms::HISTOGRAM_ENUM_MAX /* hist_type */,
-            db_options_->listeners, nullptr,
-            tmp_set.Contains(FileType::kDescriptorFile),
-            tmp_set.Contains(FileType::kDescriptorFile)));
-        new_desc_log_ptr.reset(
-            new log::Writer(std::move(file_writer), 0, false));
+        new_desc_log_ptr =
+            CreateManifestWriter(std::move(descriptor_file), descriptor_fname,
+                                 opt_file_opts, manifest_preallocation_size);
         raw_desc_log_ptr = new_desc_log_ptr.get();
         s = WriteCurrentStateToManifest(write_options, curr_state,
                                         wal_additions, raw_desc_log_ptr, io_s);
@@ -6259,6 +6252,7 @@ Status VersionSet::ProcessManifestWrites(
         }
         ++idx;
 #endif /* !NDEBUG */
+        TEST_SYNC_POINT("VersionSet::ProcessManifestWrites:AddRecord");
         io_s = raw_desc_log_ptr->AddRecord(write_options, record);
         if (!io_s.ok()) {
           s = io_s;
@@ -6680,6 +6674,106 @@ Status VersionSet::LogAndApplyHelper(ColumnFamilyData* cfd,
   return builder ? builder->Apply(edit) : Status::OK();
 }
 
+FileOptions VersionSet::GetFileOptionsForManifestWrite() const {
+  FileOptions opt_file_opts = fs_->OptimizeForManifestWrite(file_options_);
+  // DB-configured temperature wins over the FS's default (when not kUnknown).
+  if (file_options_.temperature != Temperature::kUnknown) {
+    opt_file_opts.temperature = file_options_.temperature;
+  }
+  return opt_file_opts;
+}
+
+std::unique_ptr<log::Writer> VersionSet::CreateManifestWriter(
+    std::unique_ptr<FSWritableFile> file, const std::string& fname,
+    const FileOptions& opts, uint64_t preallocation_size,
+    uint64_t initial_file_size) const {
+  file->SetPreallocationBlockSize(preallocation_size);
+  FileTypeSet tmp_set = db_options_->checksum_handoff_file_types;
+  auto file_writer = std::make_unique<WritableFileWriter>(
+      std::move(file), fname, opts, clock_, io_tracer_,
+      /*stats=*/nullptr, Histograms::HISTOGRAM_ENUM_MAX, db_options_->listeners,
+      /*file_checksum_gen_factory=*/nullptr,
+      tmp_set.Contains(FileType::kDescriptorFile),
+      tmp_set.Contains(FileType::kDescriptorFile), initial_file_size);
+  const size_t block_offset = initial_file_size % log::kBlockSize;
+  return std::make_unique<log::Writer>(
+      std::move(file_writer), /*log_number=*/0, /*recycle_log_files=*/false,
+      /*manual_flush=*/false, kNoCompression,
+      /*track_and_verify_wals=*/false, block_offset);
+}
+
+Status VersionSet::ReopenManifestForAppend(const std::string& manifest_path) {
+  assert(db_options_->reuse_manifest_on_open);
+  assert(manifest_last_valid_record_end_ > 0);
+
+  // Disabled under best_efforts_recovery: that mode rebuilds the
+  // MANIFEST + CURRENT from scratch via the side-effect of
+  // LogAndApplyForRecovery emitting an edit. Reusing the (possibly
+  // stale) prior MANIFEST contradicts the salvage contract.
+  if (db_options_->best_efforts_recovery) {
+    return Status::OK();
+  }
+
+  FileOptions opt_file_opts = GetFileOptionsForManifestWrite();
+
+  // Bail if the on-disk size diverges from what Recover's Reader
+  // consumed -- likely a torn tail from a prior crash; mid-block append
+  // would mis-frame the record stream.
+  uint64_t physical_size = 0;
+  IOStatus stat_s = fs_->GetFileSize(manifest_path, IOOptions(), &physical_size,
+                                     /*dbg=*/nullptr);
+  if (!stat_s.ok() || physical_size != manifest_last_valid_record_end_) {
+    ROCKS_LOG_WARN(db_options_->info_log,
+                   "reuse_manifest_on_open: physical size %" PRIu64
+                   " != last valid record end %" PRIu64
+                   " (tail corruption?); falling back to fresh MANIFEST",
+                   physical_size, manifest_last_valid_record_end_);
+    return Status::OK();
+  }
+
+  // Direct-writes path is hard to reason about for an already-populated
+  // file (alignment + tail-pad). POSIX's OptimizeForManifestWrite forces
+  // it off, but custom FileSystems may not.
+  if (opt_file_opts.use_direct_writes) {
+    ROCKS_LOG_WARN(db_options_->info_log,
+                   "reuse_manifest_on_open: direct writes enabled; "
+                   "falling back to fresh MANIFEST");
+    return Status::OK();
+  }
+
+  std::unique_ptr<FSWritableFile> manifest_file;
+  IOStatus io_s = fs_->ReopenWritableFile(manifest_path, opt_file_opts,
+                                          &manifest_file, /*dbg=*/nullptr);
+  if (!io_s.ok()) {
+    ROCKS_LOG_WARN(db_options_->info_log,
+                   "Failed to reopen MANIFEST for append: %s",
+                   io_s.ToString().c_str());
+    return Status::OK();
+  }
+
+  const uint64_t reopened_size =
+      manifest_file->GetFileSize(opt_file_opts.io_options, /*dbg=*/nullptr);
+  if (reopened_size != manifest_last_valid_record_end_) {
+    ROCKS_LOG_WARN(db_options_->info_log,
+                   "reuse_manifest_on_open: reopened handle size %" PRIu64
+                   " != last valid record end %" PRIu64
+                   "; falling back to fresh MANIFEST",
+                   reopened_size, manifest_last_valid_record_end_);
+    return Status::OK();
+  }
+
+  descriptor_log_ = CreateManifestWriter(
+      std::move(manifest_file), manifest_path, opt_file_opts,
+      manifest_preallocation_size_, reopened_size);
+
+  ROCKS_LOG_INFO(db_options_->info_log,
+                 "Reusing existing MANIFEST file: %s (valid data size: %" PRIu64
+                 ")",
+                 manifest_path.c_str(), manifest_last_valid_record_end_);
+  TEST_SYNC_POINT("VersionSet::ReopenManifestForAppend:Reopened");
+  return Status::OK();
+}
+
 Status VersionSet::Recover(
     const std::vector<ColumnFamilyDescriptor>& column_families, bool read_only,
     std::string* db_id, bool no_error_if_files_missing, bool is_retry,
@@ -6767,6 +6861,11 @@ Status VersionSet::Recover(
     }
   }
 
+  if (s.ok() && !read_only && db_options_->reuse_manifest_on_open &&
+      manifest_last_valid_record_end_ > 0) {
+    s = ReopenManifestForAppend(manifest_path);
+  }
+
   return s;
 }
 
@@ -8064,11 +8163,12 @@ ReactiveVersionSet::ReactiveVersionSet(
     const MutableDBOptions& mutable_db_options,
     const FileOptions& _file_options, Cache* table_cache,
     WriteBufferManager* write_buffer_manager, WriteController* write_controller,
-    const std::shared_ptr<IOTracer>& io_tracer)
+    const std::shared_ptr<IOTracer>& io_tracer, const std::string& db_id,
+    const std::string& db_session_id)
     : VersionSet(dbname, imm_db_options, mutable_db_options, _file_options,
                  table_cache, write_buffer_manager, write_controller,
-                 /*block_cache_tracer=*/nullptr, io_tracer, /*db_id*/ "",
-                 /*db_session_id*/ "", /*daily_offpeak_time_utc*/ "",
+                 /*block_cache_tracer=*/nullptr, io_tracer, db_id,
+                 db_session_id, /*daily_offpeak_time_utc*/ "",
                  /*error_handler=*/nullptr, /*unchanging=*/false) {}
 
 ReactiveVersionSet::~ReactiveVersionSet() = default;

db/version_set.h

@@ -1719,6 +1719,33 @@ class VersionSet {
       const std::unordered_map<uint32_t, MutableCFState>& curr_state,
       const VersionEdit& wal_additions, log::Writer* log, IOStatus& io_s);
 
+  // Reopen the existing MANIFEST file for append at the end of Recover()
+  // when reuse_manifest_on_open is set, so the next LogAndApply appends
+  // to it instead of creating a fresh MANIFEST. Falls back to OK
+  // (descriptor_log_ stays null, next LogAndApply creates a fresh
+  // MANIFEST as before) if ReopenWritableFile fails.
+  Status ReopenManifestForAppend(const std::string& manifest_path);
+
+  // FileOptions for MANIFEST writes -- applies the FS's
+  // OptimizeForManifestWrite tuning, then re-applies the user-configured
+  // temperature so a custom FS can't override it.
+  FileOptions GetFileOptionsForManifestWrite() const;
+
+  // Create a log::Writer over the given FSWritableFile with all standard
+  // MANIFEST setup applied (preallocation block size, checksum-handoff
+  // classification, listeners, etc.). preallocation_size should be a
+  // snapshot of manifest_preallocation_size_ taken under the DB mutex
+  // (the fresh-path call site reads the field before releasing mu).
+  // When initial_file_size > 0 the writer is treated as resuming over
+  // an already-populated file: WritableFileWriter is constructed with the
+  // existing size so size accounting is correct, and the log writer's
+  // initial_block_offset aligns to the existing file's tail within the
+  // current 32 KiB block.
+  std::unique_ptr<log::Writer> CreateManifestWriter(
+      std::unique_ptr<FSWritableFile> file, const std::string& fname,
+      const FileOptions& opts, uint64_t preallocation_size,
+      uint64_t initial_file_size = 0) const;
+
   void AppendVersion(ColumnFamilyData* column_family_data, Version* v);
 
   ColumnFamilyData* CreateColumnFamily(const ColumnFamilyOptions& cf_options,
@@ -1788,6 +1815,17 @@ class VersionSet {
   // Current size of manifest file
   uint64_t manifest_file_size_;
 
+  // File offset at the end of the last successfully completed logical
+  // record during MANIFEST recovery. Unlike manifest_file_size_ (the
+  // reader's I/O high-water mark, which includes any tolerated tail
+  // garbage), this value points to the byte after the last valid record.
+  // Used by ReopenManifestForAppend to detect intra-block tail
+  // corruption that doesn't extend the physical file size.
+  // manifest_file_size_ is kept separate because it is used for
+  // rotation decisions (ProcessManifestWrites), close-time verification
+  // (Close), and backup metadata.
+  uint64_t manifest_last_valid_record_end_;
+
   // Size of the populated manifest file last time it was re-written from
   // scratch.
   uint64_t last_compacted_manifest_file_size_;
@@ -1860,7 +1898,9 @@ class ReactiveVersionSet : public VersionSet {
                      const FileOptions& _file_options, Cache* table_cache,
                      WriteBufferManager* write_buffer_manager,
                      WriteController* write_controller,
-                     const std::shared_ptr<IOTracer>& io_tracer);
+                     const std::shared_ptr<IOTracer>& io_tracer,
+                     const std::string& db_id,
+                     const std::string& db_session_id);
 
   ~ReactiveVersionSet() override;
 

db/version_set_test.cc

@@ -1203,7 +1203,7 @@ class VersionSetTestBase {
     reactive_versions_ = std::make_shared<ReactiveVersionSet>(
         dbname_, &imm_db_options_, mutable_db_options_, env_options_,
         table_cache_.get(), &write_buffer_manager_, &write_controller_,
-        nullptr);
+        /*io_tracer=*/nullptr, /*db_id=*/"", /*db_session_id=*/"");
     imm_db_options_.db_paths.emplace_back(dbname_,
                                           std::numeric_limits<uint64_t>::max());
   }
@@ -2440,7 +2440,7 @@ TEST_F(VersionSetTest, ManifestContentValidationOnCloseClean) {
   // Verify content validation actually ran
   ASSERT_TRUE(content_validation_ran);
 
-  // No corruption — counter should be zero
+  // No corruption -- counter should be zero
   ASSERT_EQ(0, stats->getTickerCount(MANIFEST_VALIDATION_FAILURE_COUNT));
 
   // Manifest path should be unchanged (no rotation)
@@ -2513,7 +2513,7 @@ TEST_F(VersionSetTest, ManifestContentValidationOnCloseCorruptRecord) {
 
 TEST_F(VersionSetTest, ManifestContentValidationOnCloseDisabled) {
   // Default (option disabled), corrupt manifest after writer close,
-  // verify no rotation occurred — corrupt manifest persists.
+  // verify no rotation occurred -- corrupt manifest persists.
   NewDB();
   auto stats = CreateDBStatistics();
   imm_db_options_.statistics = stats;
@@ -2538,7 +2538,7 @@ TEST_F(VersionSetTest, ManifestContentValidationOnCloseDisabled) {
 
   ASSERT_FALSE(content_validation_ran);
 
-  // Validation disabled — counter should be zero
+  // Validation disabled -- counter should be zero
   ASSERT_EQ(0, stats->getTickerCount(MANIFEST_VALIDATION_FAILURE_COUNT));
 
   // Manifest path should be unchanged (no rotation since validation is off)
@@ -2585,7 +2585,7 @@ TEST_F(VersionSetTest, ManifestContentValidationOnCloseSizeCheckFails) {
   SyncPoint::GetInstance()->DisableProcessing();
 
   // Size check caught the issue; content validation on rewritten manifest
-  // should pass — no content validation failure recorded
+  // should pass -- no content validation failure recorded
   ASSERT_EQ(0, stats->getTickerCount(MANIFEST_VALIDATION_FAILURE_COUNT));
 
   // Size check should have triggered rotation
@@ -2697,7 +2697,7 @@ TEST_F(VersionSetTest, ManifestContentValidationOnCloseOpenFails) {
 
   ASSERT_TRUE(close_s.IsIOError()) << close_s.ToString();
 
-  // File couldn't be opened — no content validation ran
+  // File couldn't be opened -- no content validation ran
   ASSERT_EQ(0, stats->getTickerCount(MANIFEST_VALIDATION_FAILURE_COUNT));
 }
 

db/wide/db_wide_basic_test.cc

@@ -30,6 +30,10 @@ class DBWideBasicTest : public DBTestBase {
     return wide_column_test_util::GetOptionsForBlobTest(GetDefaultOptions());
   }
 
+  Options GetDirectWriteOptions() {
+    return wide_column_test_util::GetDirectWriteOptions(GetDefaultOptions());
+  }
+
   // Helper: runs the EntityBlobAfterFlush test logic with the given options.
   void RunEntityBlobAfterFlush(const Options& options);
 
@@ -2159,6 +2163,57 @@ TEST_F(DBWideBasicTest, MultiGetEntityWithBlobResolution) {
   }
 }
 
+TEST_F(DBWideBasicTest,
+       MultiGetBlobBackedEntityDirectWriteMemtableBatchLookup) {
+  // Goal: force the memtable batch MultiGet optimization to read a direct-write
+  // blob-backed entity from the active memtable. One key exercises merge
+  // resolution against a blob-backed entity base, which requires Saver to carry
+  // the BlobFetcher through the batched callback path. The second key proves
+  // the non-merge default-column path still resolves correctly in the same
+  // batched request.
+  Options options = GetDirectWriteOptions();
+  options.min_blob_size = 50;
+  options.memtable_batch_lookup_optimization = true;
+  options.merge_operator = MergeOperators::CreateStringAppendOperator("|");
+
+  DestroyAndReopen(options);
+
+  const std::string merge_key = "multiget_direct_write_merge_key";
+  const std::string plain_key = "multiget_direct_write_plain_key";
+  const std::string default_value = GenerateLargeValue(100, 'd');
+  const std::string other_default_value = GenerateLargeValue(110, 'p');
+  const std::string large_value = GenerateLargeValue(120, 'l');
+  const std::string small_value = GenerateSmallValue();
+  const std::string merge_operand = "tail";
+
+  WideColumns merge_columns{{kDefaultWideColumnName, default_value},
+                            {"col_large", large_value},
+                            {"col_small", small_value}};
+  WideColumns plain_columns{{kDefaultWideColumnName, other_default_value},
+                            {"col_large", large_value},
+                            {"col_small", small_value}};
+
+  ASSERT_OK(db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(),
+                           merge_key, merge_columns));
+  ASSERT_OK(db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(),
+                           plain_key, plain_columns));
+  ASSERT_OK(db_->Merge(WriteOptions(), db_->DefaultColumnFamily(), merge_key,
+                       merge_operand));
+
+  std::array<Slice, 2> keys{{merge_key, plain_key}};
+  std::array<PinnableSlice, 2> values;
+  std::array<Status, 2> statuses;
+
+  db_->MultiGet(ReadOptions(), db_->DefaultColumnFamily(), keys.size(),
+                keys.data(), values.data(), statuses.data());
+
+  ASSERT_OK(statuses[0]);
+  ASSERT_EQ(values[0], default_value + "|" + merge_operand);
+
+  ASSERT_OK(statuses[1]);
+  ASSERT_EQ(values[1], other_default_value);
+}
+
 void DBWideBasicTest::RunEntityBlobAfterFlush(const Options& options) {
   Reopen(options);
 
@@ -2776,7 +2831,7 @@ TEST_F(DBWideBasicTest, EntityBlobBlockCacheTierGet) {
   // Reopen to clear any caches
   Reopen(options);
 
-  // Read with kBlockCacheTier — blob is not in cache
+  // Read with kBlockCacheTier -- blob is not in cache
   ReadOptions read_opts;
   read_opts.read_tier = kBlockCacheTier;
   PinnableSlice result;
@@ -2909,6 +2964,347 @@ TEST_F(DBWideBasicTest, MergeEntityWithBlobColumns) {
   }
 }
 
+TEST_F(DBWideBasicTest, GetMergeOperandsWithBlobBackedEntityDefaultColumn) {
+  // Goal: cover both GetMergeOperands code paths that read a compacted V2
+  // wide-column entity whose default column was moved to a blob file. The
+  // first read exercises the base-value path directly, then a merge operand is
+  // added so the second read exercises the merge-plus-base path.
+  Options options = GetBlobTestOptions();
+  options.min_blob_size = 50;
+  options.merge_operator = MergeOperators::CreateStringAppendOperator("|");
+
+  DestroyAndReopen(options);
+
+  const std::string key = "merge_operands_blob_entity";
+  const std::string default_value = GenerateLargeValue(100, 'd');
+  const std::string large_value = GenerateLargeValue(120, 'l');
+  const std::string small_value = GenerateSmallValue();
+  const std::string merge_operand = "suffix";
+
+  WideColumns columns{{kDefaultWideColumnName, default_value},
+                      {"col_large", large_value},
+                      {"col_small", small_value}};
+  ASSERT_OK(
+      db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(), key, columns));
+  ASSERT_OK(Flush());
+  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  ASSERT_FALSE(GetBlobFileNumbers().empty());
+
+  {
+    // The compacted V2 entity becomes the first operand when there are no
+    // newer merge operands above it.
+    GetMergeOperandsOptions get_merge_opts;
+    get_merge_opts.expected_max_number_of_operands = 1;
+
+    std::array<PinnableSlice, 2> merge_operands;
+    int number_of_operands = 0;
+
+    ASSERT_OK(db_->GetMergeOperands(ReadOptions(), db_->DefaultColumnFamily(),
+                                    key, merge_operands.data(), &get_merge_opts,
+                                    &number_of_operands));
+    ASSERT_EQ(number_of_operands, 1);
+    ASSERT_EQ(merge_operands[0], default_value);
+  }
+
+  ASSERT_OK(db_->Merge(WriteOptions(), db_->DefaultColumnFamily(), key,
+                       merge_operand));
+
+  {
+    // After a merge is added, GetMergeOperands must still resolve the blob
+    // backed base default column while traversing the older base entry below
+    // the newer merge operand.
+    GetMergeOperandsOptions get_merge_opts;
+    get_merge_opts.expected_max_number_of_operands = 2;
+
+    std::array<PinnableSlice, 2> merge_operands;
+    int number_of_operands = 0;
+
+    ASSERT_OK(db_->GetMergeOperands(ReadOptions(), db_->DefaultColumnFamily(),
+                                    key, merge_operands.data(), &get_merge_opts,
+                                    &number_of_operands));
+    ASSERT_EQ(number_of_operands, 2);
+    ASSERT_EQ(merge_operands[0], default_value);
+    ASSERT_EQ(merge_operands[1], merge_operand);
+  }
+}
+
+TEST_F(DBWideBasicTest,
+       GetMergeOperandsWithBlobBackedEntityDefaultColumnDirectWriteMemtable) {
+  // Goal: cover the pre-flush memtable path when blob direct write stores a
+  // blob-backed V2 entity in memory. The first GetMergeOperands call exercises
+  // the no-merge base-operand path. After adding a merge, DB::Get exercises
+  // merge resolution against the same memtable entity, and a second
+  // GetMergeOperands call exercises the merge-plus-base path.
+  Options options = GetDirectWriteOptions();
+  options.min_blob_size = 50;
+  options.merge_operator = MergeOperators::CreateStringAppendOperator("|");
+
+  DestroyAndReopen(options);
+
+  const std::string key = "direct_write_merge_operands_blob_entity";
+  const std::string default_value = GenerateLargeValue(100, 'd');
+  const std::string large_value = GenerateLargeValue(120, 'l');
+  const std::string small_value = GenerateSmallValue();
+  const std::string merge_operand = "suffix";
+  const std::string expected_merged = default_value + "|" + merge_operand;
+
+  WideColumns columns{{kDefaultWideColumnName, default_value},
+                      {"col_large", large_value},
+                      {"col_small", small_value}};
+  ASSERT_OK(
+      db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(), key, columns));
+
+  {
+    GetMergeOperandsOptions get_merge_opts;
+    get_merge_opts.expected_max_number_of_operands = 1;
+
+    std::array<PinnableSlice, 2> merge_operands;
+    int number_of_operands = 0;
+
+    ASSERT_OK(db_->GetMergeOperands(ReadOptions(), db_->DefaultColumnFamily(),
+                                    key, merge_operands.data(), &get_merge_opts,
+                                    &number_of_operands));
+    ASSERT_EQ(number_of_operands, 1);
+    ASSERT_EQ(merge_operands[0], default_value);
+  }
+
+  ASSERT_OK(db_->Merge(WriteOptions(), db_->DefaultColumnFamily(), key,
+                       merge_operand));
+
+  {
+    PinnableSlice result;
+    ASSERT_OK(
+        db_->Get(ReadOptions(), db_->DefaultColumnFamily(), key, &result));
+    ASSERT_EQ(result, expected_merged);
+  }
+
+  {
+    GetMergeOperandsOptions get_merge_opts;
+    get_merge_opts.expected_max_number_of_operands = 2;
+
+    std::array<PinnableSlice, 2> merge_operands;
+    int number_of_operands = 0;
+
+    ASSERT_OK(db_->GetMergeOperands(ReadOptions(), db_->DefaultColumnFamily(),
+                                    key, merge_operands.data(), &get_merge_opts,
+                                    &number_of_operands));
+    ASSERT_EQ(number_of_operands, 2);
+    ASSERT_EQ(merge_operands[0], default_value);
+    ASSERT_EQ(merge_operands[1], merge_operand);
+  }
+}
+
+TEST_F(DBWideBasicTest,
+       GetAndGetEntityWithBlobBackedDefaultColumnDirectWriteMemtableReadOnly) {
+  // Goal: cover the read-only reopen path after writing a blob-backed
+  // direct-write entity without manually flushing it first. The test checks
+  // both `Get()`, which must resolve the default-column blob reference, and
+  // `GetEntity()`, which must eagerly resolve the entity's unresolved blob
+  // columns before returning them to the caller.
+  Options options = GetDirectWriteOptions();
+  options.min_blob_size = 50;
+
+  DestroyAndReopen(options);
+
+  const std::string key = "readonly_direct_write_memtable_entity";
+  const std::string default_value = GenerateLargeValue(100, 'd');
+  const std::string large_value = GenerateLargeValue(120, 'l');
+  const std::string small_value = GenerateSmallValue();
+  WideColumns columns{{kDefaultWideColumnName, default_value},
+                      {"col_large", large_value},
+                      {"col_small", small_value}};
+
+  ASSERT_OK(
+      db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(), key, columns));
+
+  Close();
+  options.avoid_flush_during_recovery = true;
+  ASSERT_OK(ReadOnlyReopen(options));
+
+  {
+    PinnableSlice result;
+    ASSERT_OK(
+        db_->Get(ReadOptions(), db_->DefaultColumnFamily(), key, &result));
+    ASSERT_EQ(result, default_value);
+  }
+
+  {
+    PinnableWideColumns result;
+    ASSERT_OK(db_->GetEntity(ReadOptions(), db_->DefaultColumnFamily(), key,
+                             &result));
+    ASSERT_EQ(result.columns(), columns);
+  }
+}
+
+TEST_F(DBWideBasicTest, ReadOnlyDirectWriteMemtableBlobBlockCacheTier) {
+  // Goal: cover the read-only memtable path under kBlockCacheTier. The test
+  // keeps a blob-backed direct-write entity in WAL-backed recovery state so
+  // resolving either Get() or GetEntity() would require blob I/O, which must
+  // surface as Incomplete rather than silently reading the blob file.
+  Options options = GetDirectWriteOptions();
+  options.min_blob_size = 50;
+
+  DestroyAndReopen(options);
+
+  const std::string key = "readonly_direct_write_memtable_block_cache_tier";
+  const std::string default_value = GenerateLargeValue(100, 'd');
+  const std::string large_value = GenerateLargeValue(120, 'l');
+  const std::string small_value = GenerateSmallValue();
+  WideColumns columns{{kDefaultWideColumnName, default_value},
+                      {"col_large", large_value},
+                      {"col_small", small_value}};
+
+  ASSERT_OK(
+      db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(), key, columns));
+
+  Close();
+  options.avoid_flush_during_recovery = true;
+  ASSERT_OK(ReadOnlyReopen(options));
+
+  ReadOptions read_opts;
+  read_opts.read_tier = kBlockCacheTier;
+
+  {
+    PinnableSlice result;
+    const Status s =
+        db_->Get(read_opts, db_->DefaultColumnFamily(), key, &result);
+    ASSERT_TRUE(s.IsIncomplete()) << s.ToString();
+    ASSERT_TRUE(result.empty());
+  }
+
+  {
+    PinnableWideColumns result;
+    const Status s =
+        db_->GetEntity(read_opts, db_->DefaultColumnFamily(), key, &result);
+    ASSERT_TRUE(s.IsIncomplete()) << s.ToString();
+    ASSERT_TRUE(result.columns().empty());
+  }
+}
+
+TEST_F(DBWideBasicTest,
+       GetMergeOperandsWithBlobBackedEntityDefaultColumnReadOnly) {
+  // Goal: exercise OpenForReadOnly on a blob-backed V2 entity base in SST with
+  // a newer merge operand in a separate SST. The read-only DB must resolve the
+  // blob-backed default column for both Get() and GetMergeOperands() without
+  // relying on mutable memtable state.
+  Options options = GetBlobTestOptions();
+  options.min_blob_size = 50;
+  options.merge_operator = MergeOperators::CreateStringAppendOperator("|");
+
+  DestroyAndReopen(options);
+
+  const std::string key = "readonly_merge_operands_blob_entity";
+  const std::string default_value = GenerateLargeValue(100, 'd');
+  const std::string large_value = GenerateLargeValue(120, 'l');
+  const std::string small_value = GenerateSmallValue();
+  const std::string merge_operand = "suffix";
+  const std::string expected_merged = default_value + "|" + merge_operand;
+
+  WideColumns columns{{kDefaultWideColumnName, default_value},
+                      {"col_large", large_value},
+                      {"col_small", small_value}};
+  ASSERT_OK(
+      db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(), key, columns));
+  ASSERT_OK(Flush());
+  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  ASSERT_FALSE(GetBlobFileNumbers().empty());
+
+  ASSERT_OK(db_->Merge(WriteOptions(), db_->DefaultColumnFamily(), key,
+                       merge_operand));
+  ASSERT_OK(Flush());
+
+  Close();
+  ASSERT_OK(ReadOnlyReopen(options));
+
+  {
+    PinnableSlice result;
+    ASSERT_OK(
+        db_->Get(ReadOptions(), db_->DefaultColumnFamily(), key, &result));
+    ASSERT_EQ(result, expected_merged);
+  }
+
+  {
+    GetMergeOperandsOptions get_merge_opts;
+    get_merge_opts.expected_max_number_of_operands = 2;
+
+    std::array<PinnableSlice, 2> merge_operands;
+    int number_of_operands = 0;
+
+    ASSERT_OK(db_->GetMergeOperands(ReadOptions(), db_->DefaultColumnFamily(),
+                                    key, merge_operands.data(), &get_merge_opts,
+                                    &number_of_operands));
+    ASSERT_EQ(number_of_operands, 2);
+    ASSERT_EQ(merge_operands[0], default_value);
+    ASSERT_EQ(merge_operands[1], merge_operand);
+  }
+}
+
+TEST_F(DBWideBasicTest,
+       GetMergeOperandsWithBlobBackedEntityDefaultColumnBlockCacheTier) {
+  // Goal: prove the SST-backed GetMergeOperands path propagates Incomplete
+  // when resolving a blob-backed default column would require I/O. One key
+  // covers the direct base-entity path, and the other covers the merge-plus-
+  // base path with the merge operand flushed to a newer SST.
+  Options options = GetBlobTestOptions();
+  options.min_blob_size = 50;
+  options.merge_operator = MergeOperators::CreateStringAppendOperator("|");
+
+  DestroyAndReopen(options);
+
+  const std::string base_key = "merge_operands_blob_entity_cache_base";
+  const std::string merge_key = "merge_operands_blob_entity_cache_merge";
+  const std::string default_value = GenerateLargeValue(100, 'd');
+  const std::string large_value = GenerateLargeValue(120, 'l');
+  const std::string small_value = GenerateSmallValue();
+
+  WideColumns columns{{kDefaultWideColumnName, default_value},
+                      {"col_large", large_value},
+                      {"col_small", small_value}};
+  ASSERT_OK(db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(), base_key,
+                           columns));
+  ASSERT_OK(db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(),
+                           merge_key, columns));
+  ASSERT_OK(Flush());
+  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+
+  ASSERT_OK(db_->Merge(WriteOptions(), db_->DefaultColumnFamily(), merge_key,
+                       "tail"));
+  ASSERT_OK(Flush());
+
+  // Reopen to clear any cached blob readers/values so kBlockCacheTier must
+  // report Incomplete instead of succeeding through a warm cache.
+  Reopen(options);
+
+  ReadOptions read_opts;
+  read_opts.read_tier = kBlockCacheTier;
+
+  {
+    GetMergeOperandsOptions get_merge_opts;
+    get_merge_opts.expected_max_number_of_operands = 1;
+
+    std::array<PinnableSlice, 2> merge_operands;
+    int number_of_operands = 0;
+
+    const Status s = db_->GetMergeOperands(
+        read_opts, db_->DefaultColumnFamily(), base_key, merge_operands.data(),
+        &get_merge_opts, &number_of_operands);
+    ASSERT_TRUE(s.IsIncomplete()) << s.ToString();
+  }
+
+  {
+    GetMergeOperandsOptions get_merge_opts;
+    get_merge_opts.expected_max_number_of_operands = 2;
+
+    std::array<PinnableSlice, 2> merge_operands;
+    int number_of_operands = 0;
+
+    const Status s = db_->GetMergeOperands(
+        read_opts, db_->DefaultColumnFamily(), merge_key, merge_operands.data(),
+        &get_merge_opts, &number_of_operands);
+    ASSERT_TRUE(s.IsIncomplete()) << s.ToString();
+  }
+}
+
 TEST_F(DBWideBasicTest, MergeEntityWithBlobColumnsNoDefault) {
   // Test: Merge on a V2 entity without a default column. The merge result
   // should produce a valid entity with all original columns plus a default
@@ -2998,7 +3394,7 @@ TEST_F(DBWideBasicTest, MergeEntityWithBlobColumnsBlockCacheTier) {
   // Reopen to clear caches
   Reopen(options);
 
-  // Read with kBlockCacheTier — blob not in cache, merge needs blob resolution
+  // Read with kBlockCacheTier -- blob not in cache, merge needs blob resolution
   ReadOptions read_opts;
   read_opts.read_tier = kBlockCacheTier;
   PinnableSlice result;

db/wide/read_path_blob_resolver.cc

@@ -59,7 +59,7 @@ Status ReadPathBlobResolver::ResolveColumn(size_t column_index,
         blob_resolver_util::FindBlobColumn(blob_columns_, column_index);
 
     if (blob_index_ptr == nullptr) {
-      // Inline column — return the value directly
+      // Inline column -- return the value directly
       *resolved_value = (*columns_)[column_index].value();
     } else {
       // Check if already resolved

db/wide/wide_column_serialization.h

@@ -153,6 +153,8 @@ class WideColumnSerialization {
   // value() is initially the serialized BlobIndex bytes from the entity.
   // blob_columns: receives (column_index, blob_index) pairs identifying which
   // entries in `columns` are blob references and their decoded BlobIndex data.
+  // For valid input, this must agree with HasBlobColumns() about whether blob
+  // column references are present.
   static Status DeserializeV2(
       Slice& input, std::vector<WideColumn>& columns,
       std::vector<std::pair<size_t, BlobIndex>>& blob_columns);
@@ -160,6 +162,8 @@ class WideColumnSerialization {
   // Check if the serialized entity has any blob column references.
   // Sets *has_blob_columns to true if version >= 2 and at least one column
   // has blob type; false otherwise.
+  // For valid input, this must agree with DeserializeV2() about whether blob
+  // column references are present.
   // Returns Status::Corruption on decode errors.
   static Status HasBlobColumns(const Slice& input, bool& has_blob_columns);
 

db/wide/wide_column_serialization_test.cc

@@ -391,7 +391,7 @@ static BlobIndex MakeBlobIndex(uint64_t file_number, uint64_t offset,
   return bi;
 }
 
-// Helper: V2 serialize → DeserializeV2 round-trip, returning
+// Helper: V2 serialize -> DeserializeV2 round-trip, returning
 // deserialized columns and blob column info.
 static void V2SerializeAndDeserialize(
     const std::vector<std::pair<std::string, std::string>>& columns,
@@ -798,7 +798,7 @@ TEST_F(WideColumnSerializationTest, RandomizedSerializeDeserializeRoundTrip) {
       }
     }
 
-    // V2 serialize → DeserializeV2 round-trip
+    // V2 serialize -> DeserializeV2 round-trip
     std::string serialized;
     std::vector<WideColumn> deserialized;
     std::vector<std::pair<size_t, BlobIndex>> blob_out;

db_stress_tool/db_stress_common.h

@@ -279,7 +279,9 @@ DECLARE_bool(use_full_merge_v1);
 DECLARE_int32(sync_wal_one_in);
 DECLARE_bool(avoid_unnecessary_blocking_io);
 DECLARE_bool(write_dbid_to_manifest);
+DECLARE_bool(optimize_manifest_for_recovery);
 DECLARE_bool(write_identity_file);
+DECLARE_bool(reuse_manifest_on_open);
 DECLARE_bool(avoid_flush_during_recovery);
 DECLARE_bool(enforce_write_buffer_manager_during_recovery);
 DECLARE_uint64(max_write_batch_group_size_bytes);

db_stress_tool/db_stress_env_wrapper.h

@@ -197,10 +197,10 @@ class DbStressFSWrapper : public FileSystemWrapper {
       assert(!file_opts.file_checksum_func_name.empty());
       if (file_opts.file_checksum_func_name == kUnknownFileChecksumFuncName ||
           file_opts.file_checksum_func_name == kNoFileChecksumFuncName) {
-        // No checksum available — checksum value must be empty
+        // No checksum available -- checksum value must be empty
         assert(file_opts.file_checksum.empty());
       } else {
-        // A real checksum function — checksum value must be present
+        // A real checksum function -- checksum value must be present
         assert(!file_opts.file_checksum.empty());
       }
     }

db_stress_tool/db_stress_gflags.cc

@@ -1133,10 +1133,19 @@ DEFINE_bool(write_dbid_to_manifest,
             ROCKSDB_NAMESPACE::Options().write_dbid_to_manifest,
             "Write DB_ID to manifest");
 
+DEFINE_bool(optimize_manifest_for_recovery,
+            ROCKSDB_NAMESPACE::Options().optimize_manifest_for_recovery,
+            "Reduce recovery work after a clean shutdown");
+
 DEFINE_bool(write_identity_file,
             ROCKSDB_NAMESPACE::Options().write_identity_file,
             "Write DB_ID to IDENTITY file");
 
+DEFINE_bool(reuse_manifest_on_open,
+            ROCKSDB_NAMESPACE::Options().reuse_manifest_on_open,
+            "Reopen existing MANIFEST for append after recovery instead of "
+            "creating a fresh one");
+
 DEFINE_bool(avoid_flush_during_recovery,
             ROCKSDB_NAMESPACE::Options().avoid_flush_during_recovery,
             "Avoid flush during recovery");

db_stress_tool/db_stress_test_base.cc

@@ -3747,6 +3747,10 @@ void StressTest::PrintEnv() const {
           static_cast<int>(FLAGS_avoid_unnecessary_blocking_io));
   fprintf(stdout, "Write DB ID to manifest   : %d\n",
           static_cast<int>(FLAGS_write_dbid_to_manifest));
+  fprintf(stdout, "Optimize manifest recovery: %d\n",
+          static_cast<int>(FLAGS_optimize_manifest_for_recovery));
+  fprintf(stdout, "Reuse manifest on open    : %d\n",
+          static_cast<int>(FLAGS_reuse_manifest_on_open));
   fprintf(stdout, "Max Write Batch Group Size: %" PRIu64 "\n",
           FLAGS_max_write_batch_group_size_bytes);
   fprintf(stdout, "Use dynamic level         : %d\n",
@@ -4666,7 +4670,9 @@ void InitializeOptionsFromFlags(
   options.manual_wal_flush = FLAGS_manual_wal_flush_one_in > 0 ? true : false;
   options.avoid_unnecessary_blocking_io = FLAGS_avoid_unnecessary_blocking_io;
   options.write_dbid_to_manifest = FLAGS_write_dbid_to_manifest;
+  options.optimize_manifest_for_recovery = FLAGS_optimize_manifest_for_recovery;
   options.write_identity_file = FLAGS_write_identity_file;
+  options.reuse_manifest_on_open = FLAGS_reuse_manifest_on_open;
   options.avoid_flush_during_recovery = FLAGS_avoid_flush_during_recovery;
   options.enforce_write_buffer_manager_during_recovery =
       FLAGS_enforce_write_buffer_manager_during_recovery;

db_stress_tool/no_batched_ops_stress.cc

@@ -2900,7 +2900,7 @@ class NonBatchedOpsStressTest : public StressTest {
       op_logs += "N";
     }
 
-    // backward scan — skip when backward iteration is not supported
+    // backward scan -- skip when backward iteration is not supported
     if (FLAGS_test_backward_scan) {
       key_str = Key(ub - 1);
       iter->SeekForPrev(key_str);

docs/_data/authors.yml

@@ -88,7 +88,7 @@ hx235:
 
 pdillinger:
   full_name: Peter Dillinger
-  fbid: 513108
+  fbid: 593944283
 
 alanpaxton:
   full_name: Alan Paxton
@@ -101,3 +101,7 @@ anand1976:
 
 poojam23:
   full_name: Pooja Malik
+
+joshkang97:
+  full_name: Josh Kang
+  fbid: 180489168460795

docs/_posts/2026-05-04-interpolation-search.markdown

@@ -0,0 +1,118 @@
+---
+title: Interpolation search for SST index blocks
+layout: post
+author: joshkang97
+category: blog
+---
+
+For workloads with uniformly distributed keys, RocksDB now supports **interpolation search** for SST index blocks as an alternative to the default binary search.
+
+## The idea
+
+Binary search always splits the remaining range in half:
+
+```
+mid = low + (high - low) / 2
+```
+
+That's Θ(log n) probes regardless of the data. Interpolation search instead estimates where the target should land based on its value relative to the current boundaries:
+
+```
+probe = low + (target - key[low]) * (high - low) / (key[high] - key[low])
+```
+
+On uniformly distributed keys, that's expected O(log log n) probes. The canonical example: for an index block with restart keys `0, 1, 2, ..., 1023` and a seek for `900`, binary search needs about 10 hops; interpolation search lands on it in 1.
+
+The catch is that pure interpolation search degrades to O(n) on badly skewed data.
+
+## Turning a key into a number
+
+The interpolation formula needs numeric values, but index keys are variable-length byte slices. RocksDB extracts a `uint64_t` per key by reading the first 8 bytes after the common prefix shared by the block's boundary keys, in big-endian, and zero-pads to the right if the remaining bytes are too short.
+
+```cpp
+inline uint64_t ReadBe64FromKey(Slice s, bool is_user_key, size_t offset) {
+  // ... strip internal seq/type bytes if needed ...
+  if (s.size() - offset >= 8) {
+    uint64_t val;
+    memcpy(&val, s.data() + offset, sizeof(val));
+    return port::kLittleEndian ? EndianSwapValue(val) : val;
+  }
+  // pad short tails with zeros on the right (preserves bytewise order)
+}
+```
+
+Big-endian + zero-pad preserves bytewise ordering, so the linear interpolation formula stays consistent with the comparator. This is also why the feature requires `BytewiseComparator`.
+
+Two distinct keys can still collapse to the same `uint64_t` once you go past the first 8 non-shared bytes. To avoid a divide-by-zero, we simply fall back to binary search in that case.
+
+## How to enable it
+
+To force interpolation search on every index block:
+
+```cpp
+rocksdb::BlockBasedTableOptions table_options;
+table_options.index_block_search_type =
+    rocksdb::BlockBasedTableOptions::kInterpolation;
+```
+
+## kAuto: per-block selection
+
+`kAuto` is the recommended way to use the feature. It chooses the search algorithm for each index block automatically, based on a uniformity hint written into the block footer at SST construction time:
+
+```cpp
+table_options.index_block_search_type =
+    rocksdb::BlockBasedTableOptions::kAuto;
+table_options.uniform_cv_threshold = 0.2;
+```
+
+When `uniform_cv_threshold >= 0`, the SST writer scans each index block's restart keys and computes the **coefficient of variation (CV)** of the gaps between consecutive numeric key values:
+
+```
+gap[i] = key_value[i + 1] - key_value[i]
+CV     = stddev(gap) / mean(gap)
+```
+
+Lower CV means the gaps are more uniform — and the more likely interpolation search will outperform binary search. The CV is computed incrementally with Welford's online algorithm, so the scan is one pass over the restart points.
+
+If `CV < uniform_cv_threshold`, RocksDB sets an `is_uniform` bit in the block footer. At read time, `kAuto` resolves to `kInterpolation` only when that bit is set *and* the comparator is bytewise; otherwise it uses `kBinary`.
+
+### Write overhead
+
+Computing the `is_uniform` bit is a cheap operation as it is only computed for the index blocks in a SST file. CPU profiling of `db_bench -benchmarks=fillseq,compact -compression_type=none -disable_wal=1` attributes only ~0.08% of write-path CPU to `ScanForUniformity`.
+
+After a few more releases, we plan to enable kAuto and uniform_cv_threshold by default.
+
+## Benchmarks
+
+Setup — populate a DB and force a single-level shape so all reads hit the same index structure, then measure point-read throughput:
+
+```
+# Build a release binary
+make clean && DEBUG_LEVEL=0 make db_bench
+
+# Load + compact, varying the index_shortening_mode
+./db_bench -benchmarks=fillrandom,compact \
+           -index_shortening_mode=1
+
+# Then point-read against the populated DB
+./db_bench -use_existing_db=true -benchmarks=readrandom \
+           -index_block_search_type=binary_search   # or interpolation_search / auto_search
+```
+
+`index_shortening_mode=1` (`kShortenSeparators`) keeps the file's last index key intact, which preserves a roughly uniform numeric distribution for the benchmark.
+
+Results, averaged over multiple runs:
+
+| Mode | ops/s | vs `binary_search` |
+|---|---|---|
+| `binary_search` | 335,749 | baseline |
+| `interpolation_search` | 366,598 | **+9.2%** |
+| `auto_search` | 366,832 | **+9.2%** |
+
+## Compatibility
+
+The `is_uniform` bit reuses a previously-reserved bit in the data block footer. SSTs written by older RocksDB never set it and decode as `is_uniform = false`, so they read with binary search under `kAuto`. However, after the bit is set, if read by an older RocksDB version < 11.0.0, it will read it as a corruption error.
+
+## Future work
+
+Some future opportunities can involve extending interpolation search to data blocks, as well as supporting other comparators such as `ReverseBytewiseComparator`.

env/env_test.cc

@@ -1892,7 +1892,7 @@ TEST_F(EnvPosixTest, SupportedOpsNoAsyncIOOnIOUringInitFailure) {
   int64_t thread_supported_ops2 = 0;
   std::thread t([&]() {
     fs->SupportedOps(thread_supported_ops);
-    // Second call on the same thread — cached failure, no retry.
+    // Second call on the same thread -- cached failure, no retry.
     fs->SupportedOps(thread_supported_ops2);
   });
   t.join();
@@ -1903,7 +1903,7 @@ TEST_F(EnvPosixTest, SupportedOpsNoAsyncIOOnIOUringInitFailure) {
   // Second call should also lack kAsyncIO.
   ASSERT_EQ(thread_supported_ops2 & (1 << FSSupportedOps::kAsyncIO), 0);
   ASSERT_NE(thread_supported_ops2 & (1 << FSSupportedOps::kFSPrefetch), 0);
-  // CreateIOUring must have been called exactly once — not retried.
+  // CreateIOUring must have been called exactly once -- not retried.
   ASSERT_EQ(create_count, 1);
 
   SyncPoint::GetInstance()->DisableProcessing();

file/file_util.cc

@@ -221,7 +221,7 @@ IOStatus GenerateOneFileChecksum(
     std::unique_ptr<FSRandomAccessFile> r_file;
     FileOptions fopts = file_options;
     if (fopts.file_checksum.empty()) {
-      // No expected checksum is known — this is a from-scratch computation.
+      // No expected checksum is known -- this is a from-scratch computation.
       fopts.file_checksum_func_name = kNoFileChecksumFuncName;
     }
     io_s = fs->NewRandomAccessFile(file_path, fopts, &r_file, nullptr);

file/prefetch_test.cc

@@ -3899,7 +3899,7 @@ TEST_P(FSBufferPrefetchTest, FSBufferPrefetchStatsInternals) {
                            &result, &s, for_compaction);
   // Platforms that don't have IO uring may not support async IO.
   // With the ReadAsync sync fallback, s will be OK even when async IO is
-  // unavailable — detect by checking if the second buffer has an async read
+  // unavailable -- detect by checking if the second buffer has an async read
   // in progress.
   if (use_async_prefetch && s.IsNotSupported()) {
     return;

file/writable_file_writer.h

@@ -169,6 +169,9 @@ class WritableFileWriter {
   Temperature temperature_;
 
  public:
+  // `initial_file_size` seeds size accounting when wrapping a reopened file
+  // that will continue appending from its current end. Leave at 0 for newly
+  // created files and reuse-from-offset-0 paths.
   WritableFileWriter(
       std::unique_ptr<FSWritableFile>&& file, const std::string& _file_name,
       const FileOptions& options, SystemClock* clock = nullptr,
@@ -178,21 +181,21 @@ class WritableFileWriter {
       const std::vector<std::shared_ptr<EventListener>>& listeners = {},
       FileChecksumGenFactory* file_checksum_gen_factory = nullptr,
       bool perform_data_verification = false,
-      bool buffered_data_with_checksum = false)
+      bool buffered_data_with_checksum = false, uint64_t initial_file_size = 0)
       : file_name_(_file_name),
         writable_file_(std::move(file), io_tracer, _file_name),
         clock_(clock),
         buf_(),
         max_buffer_size_(options.writable_file_max_buffer_size),
-        filesize_(0),
-        flushed_size_(0),
-        next_write_offset_(0),
+        filesize_(initial_file_size),
+        flushed_size_(initial_file_size),
+        next_write_offset_(initial_file_size),
         pending_sync_(false),
         seen_error_(false),
 #ifndef NDEBUG
         seen_injected_error_(false),
 #endif  // NDEBUG
-        last_sync_size_(0),
+        last_sync_size_(initial_file_size),
         bytes_per_sync_(options.bytes_per_sync),
         rate_limiter_(options.rate_limiter),
         stats_(stats),

include/rocksdb/advanced_cache.h

@@ -152,7 +152,12 @@ class Cache : public Customizable {
 
     // For helpers without SecondaryCache support
     explicit CacheItemHelper(CacheEntryRole _role, DeleterFn _del_cb = nullptr)
-        : CacheItemHelper(_role, _del_cb, nullptr, nullptr, nullptr, this) {}
+        : del_cb(_del_cb),
+          size_cb(nullptr),
+          saveto_cb(nullptr),
+          create_cb(nullptr),
+          role(_role),
+          without_secondary_compat(this) {}
 
     // For helpers with SecondaryCache support
     explicit CacheItemHelper(CacheEntryRole _role, DeleterFn _del_cb,

include/rocksdb/advanced_options.h

@@ -946,7 +946,7 @@ struct AdvancedColumnFamilyOptions {
   // (estimated_reads / file_size) exceeds this threshold. This helps reduce
   // read amplification for hot keys by compacting frequently-read files.
   //
-  // Only "collapsible" reads are counted — lookups that return NotFound
+  // Only "collapsible" reads are counted -- lookups that return NotFound
   // (bloom filter false positive), Delete/SingleDeletion (tombstone), or
   // Merge (partial result). These are reads where the file contributed no
   // final value and compaction would eliminate the wasted work.
@@ -972,15 +972,15 @@ struct AdvancedColumnFamilyOptions {
   //     r * S * B = 2 * (1 + F) * S
   //     r = 2 * (1 + F) / B
   //
-  //   With F = 10, B = 4096:  r = 22 / 4096 ≈ 0.005.
+  //   With F = 10, B = 4096:  r = 22 / 4096 ~= 0.005.
   //
-  // With a block-cache hit rate h (0 ≤ h < 1), each collapsible read
+  // With a block-cache hit rate h (0 <= h < 1), each collapsible read
   // only costs (1 - h) * B bytes of actual disk IO, so:
   //     r = 2 * (1 + F) / ((1 - h) * B)
   //
-  //   h = 0   → r ≈ 0.005
-  //   h = 0.5 → r ≈ 0.01
-  //   h = 0.9 → r ≈ 0.05
+  //   h = 0   -> r ~= 0.005
+  //   h = 0.5 -> r ~= 0.01
+  //   h = 0.9 -> r ~= 0.05
   //
   // A recommended starting point is 0.01, which avoids triggering
   // compactions that cost more IO than they save for most cache-friendly

include/rocksdb/compaction_filter.h

@@ -287,6 +287,10 @@ class CompactionFilter : public Customizable {
   // may not realize there is a write conflict and may allow a Transaction to
   // Commit that should have failed.  Instead, it is better to implement any
   // Merge filtering inside the MergeOperator.
+  //
+  // Note: Under WritePrepared, WriteUnprepared, and WriteCommitted
+  // TransactionDB, this method is skipped for operands at or below the latest
+  // published sequence, typically every operand in the input.
   virtual bool FilterMergeOperand(int /*level*/, const Slice& /*key*/,
                                   const Slice& /*operand*/) const {
     return false;
@@ -416,7 +420,7 @@ class CompactionFilter : public Customizable {
 
   // Returns true if the filter overrides FilterV4 and can handle
   // WideColumnBlobResolver (lazy blob loading). When false (default),
-  // FilterV4 delegates to FilterV3 with fully resolved column values —
+  // FilterV4 delegates to FilterV3 with fully resolved column values --
   // blob columns are eagerly fetched before the filter is called, ensuring
   // backward compatibility with FilterV3-only filters.
   //

include/rocksdb/db.h

@@ -1868,6 +1868,13 @@ class DB {
   // and may not have file_creation_time property. In both the cases
   // file_creation_time is considered 0 which means this API will return
   // creation_time = 0 as there wouldn't be a timestamp lower than 0.
+  //
+  // NOTE: When the DB was opened with open_files_async = true, this API may
+  // block to wait for background SST file loading to complete, but only when
+  // necessary (i.e., when a file's manifest does not carry file_creation_time
+  // and the value must come from the SST reader). Modern DBs return the real
+  // value immediately. It is possible to have an incomplete result if the DB
+  // is closed while files are still being opened in the background.
   virtual Status GetCreationTimeOfOldestFile(uint64_t* creation_time) = 0;
 
   // Note: this API is not yet consistent with WritePrepared transactions.

include/rocksdb/listener.h

@@ -556,7 +556,7 @@ struct IOErrorInfo {
   uint64_t offset;
 };
 
-// EXPERIMENTAL — under active development, fields may change.
+// EXPERIMENTAL -- under active development, fields may change.
 // Point-in-time snapshot of background job pressure for one DB: how busy
 // compaction and flush are, and how close the DB is to write-stalling.
 struct BackgroundJobPressure {

include/rocksdb/options.h

@@ -1001,6 +1001,36 @@ struct DBOptions {
   // This option is mutable with SetDBOptions().
   bool verify_manifest_content_on_close = false;
 
+  // EXPERIMENTAL: If true, RocksDB can reduce recovery work after a clean
+  // shutdown, which may reduce DB::Open latency on warm reopens, especially on
+  // storage where metadata appends are expensive.
+  //
+  // Best-effort optimization: if it is disabled or unavailable, RocksDB falls
+  // back to the standard recovery path.
+  //
+  // Temporary rollout / kill switch for an optimization that is intended to be
+  // correct and eventually always enabled. Mutable via SetDBOptions().
+  bool optimize_manifest_for_recovery = false;
+
+  // EXPERIMENTAL: If true, DB::Open can try to reuse the existing MANIFEST
+  // for the first post-open metadata update instead of creating a fresh one.
+  // This can reduce warm-open latency for DBs whose MANIFEST is expensive to
+  // rebuild.
+  //
+  // Best-effort optimization: even when enabled, RocksDB may still create a
+  // fresh MANIFEST if the FileSystem does not support reopening the existing
+  // MANIFEST for append, or if RocksDB decides reuse is unsafe. That fallback
+  // is normal behavior.
+  //
+  // With very small `max_manifest_file_size` settings, the reused MANIFEST can
+  // still rotate earlier than expected after open, because RocksDB may keep a
+  // conservative auto-tuned rotation threshold until it later refreshes its
+  // compacted-size estimate.
+  //
+  // Temporary rollout / kill switch while this optimization is being
+  // validated.
+  bool reuse_manifest_on_open = false;
+
   // This option mostly replaces max_manifest_file_size to control an auto-tuned
   // balance of manifest write amplification and space amplification. A new
   // manifest file is created with the "compacted" contents of the old one when
@@ -2922,7 +2952,7 @@ struct SizeApproximationOptions {
   // blob file data in the key range. When enabled, the total blob file size
   // is prorated by the ratio of SST data in the range to the total SST data:
   //
-  //   blob_size_in_range ≈ total_blob_size * (sst_in_range / total_sst)
+  //   blob_size_in_range ~= total_blob_size * (sst_in_range / total_sst)
   //
   // Limitations of this approximation:
   // - Assumes blob data is distributed proportionally to SST data, which

include/rocksdb/sst_file_reader.h

@@ -9,6 +9,7 @@
 #include "rocksdb/options.h"
 #include "rocksdb/slice.h"
 #include "rocksdb/table_properties.h"
+#include "rocksdb/types.h"
 
 namespace ROCKSDB_NAMESPACE {
 
@@ -56,6 +57,12 @@ class SstFileReader {
   // capabilities like filling cache, read ahead are disabled.
   std::unique_ptr<Iterator> NewTableIterator();
 
+  // Parses a raw key returned by NewTableIterator().
+  // The Slice fields in parsed_key point into raw_table_key and are valid only
+  // while raw_table_key remains valid.
+  Status ParseTableIteratorKey(const Slice& raw_table_key,
+                               ParsedEntryInfo* parsed_key) const;
+
   std::shared_ptr<const TableProperties> GetTableProperties() const;
 
   // Verifies whether there is corruption in this table.

include/rocksdb/user_defined_index.h

@@ -43,7 +43,7 @@ class UserDefinedIndexBuilder {
     kOther = 3,   // Other types (e.g., blob reference, wide-column entity).
                   // The value format is type-specific and may not be the
                   // actual user data.
-    kTypeMax,     // Sentinel — must be last. Value may change across releases.
+    kTypeMax,     // Sentinel -- must be last. Value may change across releases.
   };
 
   // File offset and size of the data block
@@ -106,7 +106,7 @@ class UserDefinedIndexBuilder {
   // (e.g., trie-based indexes) can leave this as a no-op.
   //
   // @key:   The user key (without sequence number or type suffix).
-  // @type:  The entry type — kValue (Put), kDelete, kMerge, or kOther.
+  // @type:  The entry type -- kValue (Put), kDelete, kMerge, or kOther.
   //         For kDelete entries, the value may be empty. For kOther, the
   //         value format is type-specific and may not be actual user data.
   // @value: The associated value (may be empty for deletions).

memtable/inlineskiplist_test.cc

@@ -559,7 +559,7 @@ TEST_F(InlineSkipTest, MultiGetDuplicateKeysWithCallbackWalk) {
   }
 
   // Callback that walks forward through multiple entries before stopping.
-  // This simulates the Merge operand accumulation in SaveValue — the callback
+  // This simulates the Merge operand accumulation in SaveValue -- the callback
   // returns true for entries until it reaches one >= stop_at, simulating
   // walking through merge chain entries.
   struct WalkingCallbackArg {
@@ -1068,11 +1068,11 @@ static void ConcurrentMultiGetWorker(void* arg) {
     // Validate all results: each found result must equal its query key
     // (since all queried keys were inserted, an exact match is expected).
     // However, due to concurrent inserts, a key sampled from the ring
-    // might not yet be visible — so we only check that if a result is
+    // might not yet be visible -- so we only check that if a result is
     // found, it equals the query key (i.e., no wrong key returned).
     for (size_t j = 0; j < batch_size; j++) {
       if (results[j] != 0 && results[j] != query_keys[j]) {
-        // Got a result that doesn't match the query — either a bug or
+        // Got a result that doesn't match the query -- either a bug or
         // the exact key wasn't inserted and we got the next one. Since
         // all our query keys are inserted, this means the result should
         // be >= query. For the just-inserted key we already checked

memtable/wbwi_memtable.cc

@@ -48,13 +48,18 @@ bool WBWIMemTable::Get(const LookupKey& key, std::string* value,
                        SequenceNumber* max_covering_tombstone_seq,
                        SequenceNumber* out_seq, const ReadOptions&,
                        bool immutable_memtable, ReadCallback* callback,
-                       bool* is_blob_index, bool do_merge) {
+                       bool* is_blob_index, bool do_merge,
+                       const BlobFetcher* blob_fetcher) {
   assert(s->ok() || s->IsMergeInProgress());
   (void)immutable_memtable;
   (void)timestamp;
   (void)columns;
+  (void)blob_fetcher;
   assert(immutable_memtable);
   assert(!timestamp);  // TODO: support UDT
+  // IngestWriteBatchWithIndex() is rejected while any live column family has
+  // blob direct write enabled, so WBWI should never need blob resolution.
+  assert(blob_fetcher == nullptr);
   assert(assigned_seqno_.upper_bound != kMaxSequenceNumber);
   assert(assigned_seqno_.lower_bound != kMaxSequenceNumber);
   // WBWI does not support DeleteRange yet.
@@ -154,10 +159,15 @@ bool WBWIMemTable::Get(const LookupKey& key, std::string* value,
 
 void WBWIMemTable::MultiGet(const ReadOptions& read_options,
                             MultiGetRange* range, ReadCallback* callback,
-                            bool immutable_memtable) {
+                            bool immutable_memtable,
+                            const BlobFetcher* blob_fetcher) {
   (void)immutable_memtable;
+  (void)blob_fetcher;
   // Should only be used as immutable memtable.
   assert(immutable_memtable);
+  // IngestWriteBatchWithIndex() is rejected while any live column family has
+  // blob direct write enabled, so WBWI should never need blob resolution.
+  assert(blob_fetcher == nullptr);
   // TODO: reuse the InternalIterator created in Get().
   for (auto iter = range->begin(); iter != range->end(); ++iter) {
     SequenceNumber dummy_seq = 0;

memtable/wbwi_memtable.h

@@ -134,10 +134,12 @@ class WBWIMemTable final : public ReadOnlyMemTable {
            SequenceNumber* max_covering_tombstone_seq, SequenceNumber* seq,
            const ReadOptions& read_opts, bool immutable_memtable,
            ReadCallback* callback = nullptr, bool* is_blob_index = nullptr,
-           bool do_merge = true) override;
+           bool do_merge = true,
+           const BlobFetcher* blob_fetcher = nullptr) override;
 
   void MultiGet(const ReadOptions& read_options, MultiGetRange* range,
-                ReadCallback* callback, bool immutable_memtable) override;
+                ReadCallback* callback, bool immutable_memtable,
+                const BlobFetcher* blob_fetcher = nullptr) override;
 
   uint64_t NumEntries() const override { return num_entries_; }
 

options/db_options.cc

@@ -140,6 +140,10 @@ static std::unordered_map<std::string, OptionTypeInfo>
          {offsetof(struct MutableDBOptions, verify_manifest_content_on_close),
           OptionType::kBoolean, OptionVerificationType::kNormal,
           OptionTypeFlags::kMutable}},
+        {"optimize_manifest_for_recovery",
+         {offsetof(struct MutableDBOptions, optimize_manifest_for_recovery),
+          OptionType::kBoolean, OptionVerificationType::kNormal,
+          OptionTypeFlags::kMutable}},
         {"daily_offpeak_time_utc",
          {offsetof(struct MutableDBOptions, daily_offpeak_time_utc),
           OptionType::kString, OptionVerificationType::kNormal,
@@ -438,6 +442,10 @@ static std::unordered_map<std::string, OptionTypeInfo>
          {offsetof(struct ImmutableDBOptions, write_dbid_to_manifest),
           OptionType::kBoolean, OptionVerificationType::kNormal,
           OptionTypeFlags::kNone}},
+        {"reuse_manifest_on_open",
+         {offsetof(struct ImmutableDBOptions, reuse_manifest_on_open),
+          OptionType::kBoolean, OptionVerificationType::kNormal,
+          OptionTypeFlags::kNone}},
         {"write_identity_file",
          {offsetof(struct ImmutableDBOptions, write_identity_file),
           OptionType::kBoolean, OptionVerificationType::kNormal,
@@ -806,6 +814,7 @@ ImmutableDBOptions::ImmutableDBOptions(const DBOptions& options)
       prefix_seek_opt_in_only(options.prefix_seek_opt_in_only),
       persist_stats_to_disk(options.persist_stats_to_disk),
       write_dbid_to_manifest(options.write_dbid_to_manifest),
+      reuse_manifest_on_open(options.reuse_manifest_on_open),
       write_identity_file(options.write_identity_file),
       log_readahead_size(options.log_readahead_size),
       file_checksum_gen_factory(options.file_checksum_gen_factory),
@@ -985,6 +994,8 @@ void ImmutableDBOptions::Dump(Logger* log) const {
                    persist_stats_to_disk);
   ROCKS_LOG_HEADER(log, "                Options.write_dbid_to_manifest: %d",
                    write_dbid_to_manifest);
+  ROCKS_LOG_HEADER(log, "                Options.reuse_manifest_on_open: %d",
+                   reuse_manifest_on_open);
   ROCKS_LOG_HEADER(log, "                Options.write_identity_file: %d",
                    write_identity_file);
   ROCKS_LOG_HEADER(
@@ -1073,6 +1084,7 @@ MutableDBOptions::MutableDBOptions(const DBOptions& options)
       manifest_preallocation_size(options.manifest_preallocation_size),
       verify_manifest_content_on_close(
           options.verify_manifest_content_on_close),
+      optimize_manifest_for_recovery(options.optimize_manifest_for_recovery),
       fast_sst_open(options.fast_sst_open),
       daily_offpeak_time_utc(options.daily_offpeak_time_utc),
       max_compaction_trigger_wakeup_seconds(
@@ -1131,6 +1143,8 @@ void MutableDBOptions::Dump(Logger* log) const {
       manifest_preallocation_size);
   ROCKS_LOG_HEADER(log, "      Options.verify_manifest_content_on_close: %d",
                    verify_manifest_content_on_close);
+  ROCKS_LOG_HEADER(log, "         Options.optimize_manifest_for_recovery: %d",
+                   optimize_manifest_for_recovery);
   ROCKS_LOG_HEADER(log, "Options.daily_offpeak_time_utc: %s",
                    daily_offpeak_time_utc.c_str());
   ROCKS_LOG_HEADER(log,

options/db_options.h

@@ -88,6 +88,7 @@ struct ImmutableDBOptions {
   bool prefix_seek_opt_in_only;
   bool persist_stats_to_disk;
   bool write_dbid_to_manifest;
+  bool reuse_manifest_on_open;
   bool write_identity_file;
   size_t log_readahead_size;
   std::shared_ptr<FileChecksumGenFactory> file_checksum_gen_factory;
@@ -149,6 +150,7 @@ struct MutableDBOptions {
   int max_manifest_space_amp_pct;
   size_t manifest_preallocation_size;
   bool verify_manifest_content_on_close;
+  bool optimize_manifest_for_recovery;
   bool fast_sst_open;
   std::string daily_offpeak_time_utc;
   uint64_t max_compaction_trigger_wakeup_seconds;

options/options_helper.cc

@@ -173,6 +173,7 @@ void BuildDBOptions(const ImmutableDBOptions& immutable_db_options,
       immutable_db_options.avoid_unnecessary_blocking_io;
   options.write_dbid_to_manifest = immutable_db_options.write_dbid_to_manifest;
   options.write_identity_file = immutable_db_options.write_identity_file;
+  options.reuse_manifest_on_open = immutable_db_options.reuse_manifest_on_open;
   options.prefix_seek_opt_in_only =
       immutable_db_options.prefix_seek_opt_in_only;
   options.log_readahead_size = immutable_db_options.log_readahead_size;
@@ -192,6 +193,8 @@ void BuildDBOptions(const ImmutableDBOptions& immutable_db_options,
       immutable_db_options.enforce_single_del_contracts;
   options.verify_manifest_content_on_close =
       mutable_db_options.verify_manifest_content_on_close;
+  options.optimize_manifest_for_recovery =
+      mutable_db_options.optimize_manifest_for_recovery;
   options.daily_offpeak_time_utc = mutable_db_options.daily_offpeak_time_utc;
   options.max_compaction_trigger_wakeup_seconds =
       mutable_db_options.max_compaction_trigger_wakeup_seconds;

options/options_settable_test.cc

@@ -473,6 +473,7 @@ TEST_F(OptionsSettableTest, DBOptionsAllFieldsSettable) {
       "avoid_unnecessary_blocking_io=false;"
       "log_readahead_size=0;"
       "write_dbid_to_manifest=false;"
+      "optimize_manifest_for_recovery=false;"
       "best_efforts_recovery=false;"
       "max_bgerror_resume_count=2;"
       "bgerror_resume_retry_interval=1000000;"
@@ -489,10 +490,12 @@ TEST_F(OptionsSettableTest, DBOptionsAllFieldsSettable) {
       "wal_write_temperature=kHot;"
       "background_close_inactive_wals=true;"
       "write_dbid_to_manifest=true;"
+      "optimize_manifest_for_recovery=true;"
       "write_identity_file=true;"
       "verify_manifest_content_on_close=false;"
       "prefix_seek_opt_in_only=true;"
-      "fast_sst_open=true;",
+      "fast_sst_open=true;"
+      "reuse_manifest_on_open=true;",
       new_options));
 
   ASSERT_EQ(unset_bytes_base, NumUnsetBytes(new_options_ptr, sizeof(DBOptions),

port/stack_trace.cc

@@ -434,14 +434,14 @@ void InstallStackTraceHandler() {
   // Ignore SIGPIPE so that broken-pipe writes (e.g. to a closed stdout)
   // return EPIPE instead of killing the process.
   signal(SIGPIPE, SIG_IGN);
-  // Crash signals — invoke full stack trace + ring buffer
+  // Crash signals -- invoke full stack trace + ring buffer
   signal(SIGILL, StackTraceHandler);
   signal(SIGSEGV, StackTraceHandler);
   signal(SIGBUS, StackTraceHandler);
   signal(SIGABRT, StackTraceHandler);
   signal(SIGFPE, StackTraceHandler);
   signal(SIGQUIT, StackTraceHandler);
-  // Termination signals — print ring buffer only, no stack trace
+  // Termination signals -- print ring buffer only, no stack trace
   signal(SIGTERM, TerminationHandler);
   signal(SIGINT, TerminationHandler);
   atexit(AtExit);

table/block_based/block_based_table_iterator.cc

@@ -41,7 +41,7 @@ void BlockBasedTableIterator::SeekImpl(const Slice* target,
     return;
   }
 
-  // MultiScan requires an explicit seek key — SeekToFirst() is not supported
+  // MultiScan requires an explicit seek key -- SeekToFirst() is not supported
   if (multi_scan_read_set_ && !target) {
     multi_scan_status_ = Status::InvalidArgument("No seek key for MultiScan");
     RecordTick(table_->GetStatistics(), MULTISCAN_SEEK_ERRORS);
@@ -683,7 +683,7 @@ void BlockBasedTableIterator::FindBlockForward() {
         if (multi_scan_index_iter_ &&
             multi_scan_index_iter_->IsScanRangeExhausted()) {
           if (multi_scan_index_iter_->HasMoreScanRanges()) {
-            // More ranges remain — signal out-of-bound so DBIter/LevelIter
+            // More ranges remain -- signal out-of-bound so DBIter/LevelIter
             // will trigger the next Seek for the next scan range.
             is_out_of_bound_ = true;
           }

table/block_based/block_based_table_reader_test.cc

@@ -1087,7 +1087,7 @@ class BlockBasedTableReaderMultiScanAsyncIOTest
     : public BlockBasedTableReaderMultiScanTest {};
 
 // TODO: test no block cache case
-TEST_P(BlockBasedTableReaderMultiScanAsyncIOTest, MultiScanPrepare) {
+TEST_P(BlockBasedTableReaderMultiScanAsyncIOTest, DISABLED_MultiScanPrepare) {
   auto param = GetParam();
   auto fill_cache = param.fill_cache;
   auto use_async_io = param.use_async_io;

table/block_based/data_block_footer.h

@@ -52,7 +52,7 @@ struct DataBlockFooter {
   // limit is adequate because a 4GiB block (maximum due to 32-bit block size)
   // with restart_interval=1 and minimum entries (12 bytes: 3 varint bytes +
   // 9-byte internal key + empty value) plus 4-byte restart offsets = 16 bytes
-  // per restart, fits at most (2^32 - 4) / 16 ≈ 268 million restarts.
+  // per restart, fits at most (2^32 - 4) / 16 ~= 268 million restarts.
   static constexpr uint32_t kMaxNumRestarts = (1u << 28) - 1;
 
   // Maximum encoded length of a DataBlockFooter (for buffer sizing).

table/block_based/multi_scan_index_iterator.cc

@@ -65,7 +65,7 @@ void MultiScanIndexIterator::Seek(const Slice& target) {
 
   // Enforce forward-only seek
   if (!prev_seek_key_.empty() && icomp_.Compare(target, prev_seek_key_) <= 0) {
-    // Seek key is not moving forward — keep current position
+    // Seek key is not moving forward -- keep current position
     return;
   }
   prev_seek_key_ = target.ToString();
@@ -87,11 +87,12 @@ void MultiScanIndexIterator::Seek(const Slice& target) {
   // The following diagram explains different seek targets seeking at various
   // positions on the table, while the next_scan_idx_ points to PreparedRange 2.
   //
-  // next_scan_idx_: ------------------┐
-  //                                   ▼
+  // next_scan_idx_: ------------------+
+  //                                   |
+  //                                   v
   // table:     : __[PreparedRange 1]__[PreparedRange 2]__[PreparedRange 3]__
-  // Seek target: <----- Case 1 ------>▲<------------- Case 2 -------------->
-  //                                   │
+  // Seek target: <----- Case 1 ------>^<------------- Case 2 -------------->
+  //                                   |
   //                                 Case 3
   //
   // Case 1: seek before the start of next prepared range. This could happen
@@ -103,7 +104,7 @@ void MultiScanIndexIterator::Seek(const Slice& target) {
   if (compare_next_scan_start_result < 0) {
     // Case 1: Seek before the start of the next prepared range
     if (next_scan_idx_ == 0) {
-      // Should not happen — seek before first prepared range
+      // Should not happen -- seek before first prepared range
       assert(false && "Seek target before the first prepared range");
       valid_ = false;
       return;
@@ -122,7 +123,7 @@ void MultiScanIndexIterator::Seek(const Slice& target) {
       valid_ = false;
       return;
     }
-    // Seek within a gap — advance to the right scan range and find block
+    // Seek within a gap -- advance to the right scan range and find block
     SeekToBlock(&user_seek_target);
   } else if (compare_next_scan_start_result > 0) {
     // Case 2: Seek after the start of the next prepared range
@@ -220,7 +221,7 @@ void MultiScanIndexIterator::SeekToBlockIdx(size_t block_idx) {
 void MultiScanIndexIterator::SetExhausted() {
   scan_range_exhausted_ = true;
   if (next_scan_idx_ < block_index_ranges_per_scan_.size()) {
-    // More ranges remain — signal out-of-bound for current range.
+    // More ranges remain -- signal out-of-bound for current range.
     valid_ = true;
     // Position at the start of the next range so that the next Seek()
     // can find it. We need to be "valid" so that FindBlockForward sets
@@ -232,7 +233,7 @@ void MultiScanIndexIterator::SetExhausted() {
     }
     valid_ = false;
   } else {
-    // Last range — natural EOF. Don't set out-of-bound so LevelIterator
+    // Last range -- natural EOF. Don't set out-of-bound so LevelIterator
     // advances to the next file.
     valid_ = false;
   }

table/block_based/multi_scan_index_iterator.h

@@ -58,7 +58,7 @@ class MultiScanIndexIterator : public InternalIteratorBase<IndexValue> {
   // Move to the first block of the first scan range.
   void SeekToFirst() override;
 
-  // Not supported — sets valid_ = false.
+  // Not supported -- sets valid_ = false.
   void SeekForPrev(const Slice& target) override;
   void SeekToLast() override;
   void Prev() override;

table/external_table.cc

@@ -5,6 +5,7 @@
 
 #include "rocksdb/external_table.h"
 
+#include "db/dbformat.h"
 #include "logging/logging.h"
 #include "rocksdb/table.h"
 #include "table/block_based/block.h"
@@ -118,7 +119,7 @@ class ExternalTableIteratorAdapter : public InternalIterator {
 
   Slice key() const override {
     if (iterator_) {
-      return Slice(*key_.const_rep());
+      return key_.GetInternalKey();
     }
     return Slice();
   }
@@ -142,7 +143,7 @@ class ExternalTableIteratorAdapter : public InternalIterator {
 
  private:
   std::unique_ptr<ExternalTableIterator> iterator_;
-  InternalKey key_;
+  IterKey key_;
   bool valid_;
   Status status_;
   IterateResult result_;
@@ -152,8 +153,8 @@ class ExternalTableIteratorAdapter : public InternalIterator {
       valid_ = iterator_->Valid();
       status_ = iterator_->status();
       if (valid_ && status_.ok()) {
-        key_.Set(res.has_value() ? res.value() : iterator_->key(), 0,
-                 ValueType::kTypeValue);
+        key_.SetInternalKey(res.has_value() ? res.value() : iterator_->key(),
+                            /*s=*/0, ValueType::kTypeValue);
       }
     }
   }

table/get_context.cc

@@ -127,6 +127,10 @@ Status GetContext::SaveWideColumnEntityToPinnable(const Slice& user_key,
       pinnable_val_->PinSelf(value_of_default);
     }
   } else if (status.IsNotSupported()) {
+    if (blob_fetcher_ == nullptr) {
+      return Status::Corruption(
+          "Cannot resolve blob-backed default column without a blob fetcher");
+    }
     // Default column is a blob reference, so resolve it into the output value.
     bool resolved = false;
     status = WideColumnSerialization::GetValueOfDefaultColumnResolvingBlobs(
@@ -140,14 +144,28 @@ Status GetContext::SaveWideColumnEntityToColumns(const Slice& user_key,
                                                  Cleanable* value_pinner) {
   assert(columns_ != nullptr);
 
+  bool has_blob_columns = false;
+  Status status =
+      WideColumnSerialization::HasBlobColumns(entity, has_blob_columns);
+  if (!status.ok()) {
+    return status;
+  }
+  if (!has_blob_columns) {
+    return columns_->SetWideColumnValue(entity, value_pinner);
+  }
   std::vector<WideColumn> entity_columns;
   std::vector<std::pair<size_t, BlobIndex>> blob_cols;
   Slice entity_ref = entity;
-  Status status = WideColumnSerialization::DeserializeV2(
-      entity_ref, entity_columns, blob_cols);
+  status = WideColumnSerialization::DeserializeV2(entity_ref, entity_columns,
+                                                  blob_cols);
   if (status.ok()) {
-    if (LIKELY(blob_cols.empty())) {
-      return columns_->SetWideColumnValue(entity, value_pinner);
+    // HasBlobColumns() and DeserializeV2() must agree on whether this entity
+    // contains blob-valued columns.
+    assert(!blob_cols.empty());
+    if (blob_fetcher_ == nullptr) {
+      return Status::Corruption(
+          "Cannot resolve blob-backed wide-column entity without a blob "
+          "fetcher");
     }
 
     // TODO: Add lazy resolution support for GetEntity point lookups. This
@@ -201,6 +219,37 @@ Status GetContext::SaveWideColumnEntityToColumns(const Slice& user_key,
   return status;
 }
 
+Status GetContext::PushWideColumnEntityDefaultOperand(const Slice& user_key,
+                                                      const Slice& entity,
+                                                      Cleanable* value_pinner) {
+  Slice value_of_default;
+  Slice entity_ref = entity;
+  Status status = WideColumnSerialization::GetValueOfDefaultColumn(
+      entity_ref, value_of_default);
+  if (status.ok()) {
+    push_operand(value_of_default, value_pinner);
+    return status;
+  }
+  if (!status.IsNotSupported()) {
+    return status;
+  }
+  if (blob_fetcher_ == nullptr) {
+    return Status::Corruption(
+        "Cannot resolve blob-backed default column without a blob fetcher");
+  }
+
+  PinnableSlice resolved_default;
+  bool resolved = false;
+  status = WideColumnSerialization::GetValueOfDefaultColumnResolvingBlobs(
+      entity, user_key, blob_fetcher_, resolved_default, resolved);
+  if (status.ok()) {
+    // Resolved blob values are backed by this stack-local PinnableSlice, so
+    // copy them into MergeContext instead of pinning their storage.
+    push_operand(Slice(resolved_default), nullptr);
+  }
+  return status;
+}
+
 void GetContext::SaveValue(const Slice& value, SequenceNumber /*seq*/,
                            Cleanable* value_pinner) {
   assert(state_ == kNotFound);
@@ -329,7 +378,7 @@ bool GetContext::SaveValue(const ParsedInternalKey& parsed_key,
                            const Slice& value, bool* matched,
                            Status* read_status, Cleanable* value_pinner) {
   assert(matched);
-  assert((state_ != kMerge && parsed_key.type != kTypeMerge) ||
+  assert((State() != kMerge && parsed_key.type != kTypeMerge) ||
          merge_context_ != nullptr);
   if (ucmp_->EqualWithoutTimestamp(parsed_key.user_key, user_key_)) {
     *matched = true;
@@ -400,7 +449,7 @@ bool GetContext::SaveValue(const ParsedInternalKey& parsed_key,
       case kTypeValuePreferredSeqno:
       case kTypeBlobIndex:
       case kTypeWideColumnEntity:
-        assert(state_ == kNotFound || state_ == kMerge);
+        assert(State() == kNotFound || State() == kMerge);
         if (type == kTypeValuePreferredSeqno) {
           unpacked_value = ParsePackedValueForValue(value);
         }
@@ -416,7 +465,7 @@ bool GetContext::SaveValue(const ParsedInternalKey& parsed_key,
           *is_blob_index_ = (type == kTypeBlobIndex);
         }
 
-        if (kNotFound == state_) {
+        if (State() == kNotFound) {
           state_ = kFound;
           if (do_merge_) {
             if (type == kTypeBlobIndex && ucmp_->timestamp_size() != 0) {
@@ -431,8 +480,10 @@ bool GetContext::SaveValue(const ParsedInternalKey& parsed_key,
                 if (!s.ok()) {
                   if (s.IsIncomplete()) {
                     MarkKeyMayExist();
+                    *read_status = s;
                   } else {
                     state_ = kCorrupt;
+                    *read_status = s;
                   }
                   return false;
                 }
@@ -456,8 +507,10 @@ bool GetContext::SaveValue(const ParsedInternalKey& parsed_key,
                 if (!s.ok()) {
                   if (s.IsIncomplete()) {
                     MarkKeyMayExist();
+                    *read_status = s;
                   } else {
                     state_ = kCorrupt;
+                    *read_status = s;
                   }
                   return false;
                 }
@@ -478,23 +531,24 @@ bool GetContext::SaveValue(const ParsedInternalKey& parsed_key,
               Slice blob_value(pin_val);
               push_operand(blob_value, nullptr);
             } else if (type == kTypeWideColumnEntity) {
-              Slice value_copy = unpacked_value;
-              Slice value_of_default;
-
-              if (!WideColumnSerialization::GetValueOfDefaultColumn(
-                       value_copy, value_of_default)
-                       .ok()) {
-                state_ = kCorrupt;
+              const Status s = PushWideColumnEntityDefaultOperand(
+                  parsed_key.user_key, unpacked_value, value_pinner);
+              if (!s.ok()) {
+                if (s.IsIncomplete()) {
+                  MarkKeyMayExist();
+                  *read_status = s;
+                } else {
+                  state_ = kCorrupt;
+                  *read_status = s;
+                }
                 return false;
               }
-
-              push_operand(value_of_default, value_pinner);
             } else {
               assert(type == kTypeValue || type == kTypeValuePreferredSeqno);
               push_operand(unpacked_value, value_pinner);
             }
           }
-        } else if (kMerge == state_) {
+        } else if (State() == kMerge) {
           assert(merge_operator_ != nullptr);
           if (type == kTypeBlobIndex) {
             PinnableSlice pin_val;
@@ -505,7 +559,11 @@ bool GetContext::SaveValue(const ParsedInternalKey& parsed_key,
             Slice blob_value(pin_val);
             state_ = kFound;
             if (do_merge_) {
-              MergeWithPlainBaseValue(blob_value);
+              const Status s = MergeWithPlainBaseValue(blob_value);
+              if (!s.ok()) {
+                *read_status = s;
+                return false;
+              }
             } else {
               // It means this function is called as part of DB GetMergeOperands
               // API and the current value should be part of
@@ -516,29 +574,38 @@ bool GetContext::SaveValue(const ParsedInternalKey& parsed_key,
             state_ = kFound;
 
             if (do_merge_) {
-              MergeWithWideColumnBaseValue(unpacked_value);
+              const Status s = MergeWithWideColumnBaseValue(unpacked_value);
+              if (!s.ok()) {
+                *read_status = s;
+                return false;
+              }
             } else {
               // It means this function is called as part of DB GetMergeOperands
               // API and the current value should be part of
               // merge_context_->operand_list
-              Slice value_copy = unpacked_value;
-              Slice value_of_default;
-
-              if (!WideColumnSerialization::GetValueOfDefaultColumn(
-                       value_copy, value_of_default)
-                       .ok()) {
-                state_ = kCorrupt;
+              const Status s = PushWideColumnEntityDefaultOperand(
+                  parsed_key.user_key, unpacked_value, value_pinner);
+              if (!s.ok()) {
+                if (s.IsIncomplete()) {
+                  MarkKeyMayExist();
+                  *read_status = s;
+                } else {
+                  state_ = kCorrupt;
+                  *read_status = s;
+                }
                 return false;
               }
-
-              push_operand(value_of_default, value_pinner);
             }
           } else {
             assert(type == kTypeValue || type == kTypeValuePreferredSeqno);
 
             state_ = kFound;
             if (do_merge_) {
-              MergeWithPlainBaseValue(unpacked_value);
+              const Status s = MergeWithPlainBaseValue(unpacked_value);
+              if (!s.ok()) {
+                *read_status = s;
+                return false;
+              }
             } else {
               // It means this function is called as part of DB GetMergeOperands
               // API and the current value should be part of
@@ -555,13 +622,17 @@ bool GetContext::SaveValue(const ParsedInternalKey& parsed_key,
       case kTypeRangeDeletion:
         // TODO(noetzli): Verify correctness once merge of single-deletes
         // is supported
-        assert(state_ == kNotFound || state_ == kMerge);
-        if (kNotFound == state_) {
+        assert(State() == kNotFound || State() == kMerge);
+        if (State() == kNotFound) {
           state_ = kDeleted;
-        } else if (kMerge == state_) {
+        } else if (State() == kMerge) {
           state_ = kFound;
           if (do_merge_) {
-            MergeWithNoBaseValue();
+            const Status s = MergeWithNoBaseValue();
+            if (!s.ok()) {
+              *read_status = s;
+              return false;
+            }
           }
           // If do_merge_ = false then the current value shouldn't be part of
           // merge_context_->operand_list
@@ -569,7 +640,7 @@ bool GetContext::SaveValue(const ParsedInternalKey& parsed_key,
         return false;
 
       case kTypeMerge:
-        assert(state_ == kNotFound || state_ == kMerge);
+        assert(State() == kNotFound || State() == kMerge);
         state_ = kMerge;
         // value_pinner is not set from plain_table_reader.cc for example.
         push_operand(value, value_pinner);
@@ -579,7 +650,11 @@ bool GetContext::SaveValue(const ParsedInternalKey& parsed_key,
             merge_operator_->ShouldMerge(
                 merge_context_->GetOperandsDirectionBackward())) {
           state_ = kFound;
-          MergeWithNoBaseValue();
+          const Status s = MergeWithNoBaseValue();
+          if (!s.ok()) {
+            *read_status = s;
+            return false;
+          }
           return false;
         }
         if (merge_context_->get_merge_operands_options != nullptr &&
@@ -597,26 +672,27 @@ bool GetContext::SaveValue(const ParsedInternalKey& parsed_key,
     }
   }
 
-  // state_ could be Corrupt, merge or notfound
+  // State() could be Corrupt, merge or notfound
   return false;
 }
 
-void GetContext::PostprocessMerge(const Status& merge_status) {
+Status GetContext::PostprocessMerge(const Status& merge_status) {
   if (!merge_status.ok()) {
     if (merge_status.subcode() == Status::SubCode::kMergeOperatorFailed) {
       state_ = kMergeOperatorFailed;
     } else {
       state_ = kCorrupt;
     }
-    return;
+    return merge_status;
   }
 
   if (LIKELY(pinnable_val_ != nullptr)) {
     pinnable_val_->PinSelf();
   }
+  return Status::OK();
 }
 
-void GetContext::MergeWithNoBaseValue() {
+Status GetContext::MergeWithNoBaseValue() {
   assert(do_merge_);
   assert(pinnable_val_ || columns_);
   assert(!pinnable_val_ || !columns_);
@@ -628,10 +704,10 @@ void GetContext::MergeWithNoBaseValue() {
       merge_context_->GetOperands(), logger_, statistics_, clock_,
       /* update_num_ops_stats */ true, /* op_failure_scope */ nullptr,
       pinnable_val_ ? pinnable_val_->GetSelf() : nullptr, columns_);
-  PostprocessMerge(s);
+  return PostprocessMerge(s);
 }
 
-void GetContext::MergeWithPlainBaseValue(const Slice& value) {
+Status GetContext::MergeWithPlainBaseValue(const Slice& value) {
   assert(do_merge_);
   assert(pinnable_val_ || columns_);
   assert(!pinnable_val_ || !columns_);
@@ -643,10 +719,10 @@ void GetContext::MergeWithPlainBaseValue(const Slice& value) {
       merge_context_->GetOperands(), logger_, statistics_, clock_,
       /* update_num_ops_stats */ true, /* op_failure_scope */ nullptr,
       pinnable_val_ ? pinnable_val_->GetSelf() : nullptr, columns_);
-  PostprocessMerge(s);
+  return PostprocessMerge(s);
 }
 
-void GetContext::MergeWithWideColumnBaseValue(const Slice& entity) {
+Status GetContext::MergeWithWideColumnBaseValue(const Slice& entity) {
   assert(do_merge_);
   assert(pinnable_val_ || columns_);
   assert(!pinnable_val_ || !columns_);
@@ -655,16 +731,16 @@ void GetContext::MergeWithWideColumnBaseValue(const Slice& entity) {
   // supports V1 format.
   std::string resolved_entity;
   Slice effective_entity;
-  const Status s_resolve = WideColumnSerialization::ResolveEntityForMerge(
+  Status s_resolve = WideColumnSerialization::ResolveEntityForMerge(
       entity, user_key_, blob_fetcher_, nullptr /* prefetch_buffers */,
       resolved_entity, effective_entity);
   if (!s_resolve.ok()) {
     if (s_resolve.IsIncomplete()) {
       MarkKeyMayExist();
-      return;
+      return s_resolve;
     }
     state_ = kCorrupt;
-    return;
+    return s_resolve;
   }
 
   // `op_failure_scope` (an output parameter) is not provided (set to nullptr)
@@ -674,7 +750,7 @@ void GetContext::MergeWithWideColumnBaseValue(const Slice& entity) {
       merge_context_->GetOperands(), logger_, statistics_, clock_,
       /* update_num_ops_stats */ true, /* op_failure_scope */ nullptr,
       pinnable_val_ ? pinnable_val_->GetSelf() : nullptr, columns_);
-  PostprocessMerge(s);
+  return PostprocessMerge(s);
 }
 
 bool GetContext::GetBlobValue(const Slice& user_key, const Slice& blob_index,
@@ -741,11 +817,16 @@ Status replayGetContextLog(const Slice& replay_log, const Slice& user_key,
 
     (void)ret;
 
-    Status read_status;
-    get_context->SaveValue(ikey, value, &dont_care, &read_status, value_pinner);
+    Status read_status = Status::OK();
+    const bool keep_replaying = get_context->SaveValue(
+        ikey, value, &dont_care, &read_status, value_pinner);
     if (!read_status.ok()) {
       return read_status;
     }
+    if (!keep_replaying) {
+      // SaveValue() reached a terminal state for this row-cache replay.
+      break;
+    }
   }
   return Status::OK();
 }

table/get_context.h

@@ -7,6 +7,7 @@
 #include <string>
 
 #include "db/read_callback.h"
+#include "rocksdb/status.h"
 #include "rocksdb/types.h"
 
 namespace ROCKSDB_NAMESPACE {
@@ -19,7 +20,6 @@ class MergeOperator;
 class PinnableWideColumns;
 class PinnedIteratorsManager;
 class Statistics;
-class Status;
 class SystemClock;
 struct ParsedInternalKey;
 
@@ -121,6 +121,10 @@ class GetContext {
              PinnedIteratorsManager* _pinned_iters_mgr = nullptr,
              ReadCallback* callback = nullptr, bool* is_blob_index = nullptr,
              uint64_t tracing_get_id = 0, BlobFetcher* blob_fetcher = nullptr);
+  // emplace-only; default construction and move assignment are intentionally
+  // disabled.
+  GetContext(GetContext&&) noexcept = default;
+  GetContext& operator=(GetContext&&) noexcept = delete;
 
   GetContext() = delete;
 
@@ -207,16 +211,19 @@ class GetContext {
   Status SaveWideColumnEntityToColumns(const Slice& user_key,
                                        const Slice& entity,
                                        Cleanable* value_pinner);
+  Status PushWideColumnEntityDefaultOperand(const Slice& user_key,
+                                            const Slice& entity,
+                                            Cleanable* value_pinner);
 
   // Helper method that postprocesses the results of merge operations, e.g. it
   // sets the state correctly upon merge errors.
-  void PostprocessMerge(const Status& merge_status);
+  Status PostprocessMerge(const Status& merge_status);
 
   // The following methods perform the actual merge operation for the
   // no base value/plain base value/wide-column base value cases.
-  void MergeWithNoBaseValue();
-  void MergeWithPlainBaseValue(const Slice& value);
-  void MergeWithWideColumnBaseValue(const Slice& entity);
+  Status MergeWithNoBaseValue();
+  Status MergeWithPlainBaseValue(const Slice& value);
+  Status MergeWithWideColumnBaseValue(const Slice& entity);
 
   bool GetBlobValue(const Slice& user_key, const Slice& blob_index,
                     PinnableSlice* blob_value, Status* read_status);

table/sst_file_reader.cc

@@ -14,6 +14,7 @@
 #include "rocksdb/env.h"
 #include "rocksdb/file_checksum.h"
 #include "rocksdb/file_system.h"
+#include "rocksdb/utilities/types_util.h"
 #include "table/get_context.h"
 #include "table/table_builder.h"
 #include "table/table_iterator.h"
@@ -157,6 +158,8 @@ std::vector<Status> SstFileReader::MultiGet(
           statuses[i] = Status::MergeInProgress();
           break;
         case GetContext::kCorrupt:
+          statuses[i] = Status::Corruption();
+          break;
         case GetContext::kUnexpectedBlobIndex:
         case GetContext::kMergeOperatorFailed:
           statuses[i] = Status::Corruption();
@@ -218,6 +221,8 @@ Status SstFileReader::Get(const ReadOptions& roptions, const Slice& key,
         status = Status::MergeInProgress();
         break;
       case GetContext::kCorrupt:
+        status = Status::Corruption();
+        break;
       case GetContext::kUnexpectedBlobIndex:
       case GetContext::kMergeOperatorFailed:
         status = Status::Corruption();
@@ -272,6 +277,11 @@ std::unique_ptr<Iterator> SstFileReader::NewTableIterator() {
   return std::make_unique<TableIterator>(internal_iter);
 }
 
+Status SstFileReader::ParseTableIteratorKey(const Slice& raw_table_key,
+                                            ParsedEntryInfo* parsed_key) const {
+  return ParseEntry(raw_table_key, rep_->options.comparator, parsed_key);
+}
+
 std::shared_ptr<const TableProperties> SstFileReader::GetTableProperties()
     const {
   return rep_->table_reader->GetTableProperties();

table/sst_file_reader_test.cc

@@ -47,8 +47,9 @@ class SstFileReaderTest : public testing::Test {
   }
 
   ~SstFileReaderTest() {
-    Status s = env_->DeleteFile(sst_name_);
-    EXPECT_OK(s);
+    if (env_->FileExists(sst_name_).ok()) {
+      EXPECT_OK(env_->DeleteFile(sst_name_));
+    }
   }
 
   void CreateFile(const std::string& file_name,
@@ -118,6 +119,105 @@ TEST_F(SstFileReaderTest, Basic) {
   CreateFileAndCheck(keys);
 }
 
+TEST_F(SstFileReaderTest, ParseTableIteratorKey) {
+  // Verify that callers of the raw table iterator can decode key metadata
+  // through the public SstFileReader API instead of duplicating dbformat.h.
+  std::vector<std::string> keys = {EncodeAsString(0), EncodeAsString(1),
+                                   EncodeAsString(2)};
+  CreateFile(sst_name_, keys);
+
+  SstFileReader reader(options_);
+  ASSERT_OK(reader.Open(sst_name_));
+  std::unique_ptr<Iterator> iter = reader.NewTableIterator();
+  ASSERT_NE(iter, nullptr);
+  iter->SeekToFirst();
+
+  ASSERT_TRUE(iter->Valid());
+  ParsedEntryInfo parsed_key;
+  ASSERT_OK(reader.ParseTableIteratorKey(iter->key(), &parsed_key));
+  EXPECT_EQ(parsed_key.user_key, keys[0]);
+  EXPECT_EQ(parsed_key.sequence, 0);
+  EXPECT_EQ(parsed_key.type, kEntryPut);
+  iter->Next();
+
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_OK(reader.ParseTableIteratorKey(iter->key(), &parsed_key));
+  EXPECT_EQ(parsed_key.user_key, keys[1]);
+  EXPECT_EQ(parsed_key.sequence, 0);
+  EXPECT_EQ(parsed_key.type, kEntryMerge);
+  iter->Next();
+
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_OK(reader.ParseTableIteratorKey(iter->key(), &parsed_key));
+  EXPECT_EQ(parsed_key.user_key, keys[2]);
+  EXPECT_EQ(parsed_key.sequence, 0);
+  EXPECT_EQ(parsed_key.type, kEntryDelete);
+  iter->Next();
+
+  EXPECT_FALSE(iter->Valid());
+}
+
+TEST_F(SstFileReaderTest, ParseTableIteratorKeyFromDbGeneratedSst) {
+  // Verify that the parser decodes real DB sequence numbers, not only the
+  // zero sequence numbers generated by SstFileWriter-created external files.
+  Options options = options_;
+  options.create_if_missing = true;
+  options.disable_auto_compactions = true;
+  std::string db_name = test::PerThreadDBPath("parse_table_iterator_key_db");
+  std::unique_ptr<DB> db;
+  ASSERT_OK(DB::Open(options, db_name, &db));
+
+  std::vector<std::string> keys = {EncodeAsString(0), EncodeAsString(1)};
+  ASSERT_OK(db->Put(WriteOptions(), keys[0], "value0"));
+  ASSERT_OK(db->Put(WriteOptions(), keys[1], "value1"));
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  std::vector<std::string> live_files;
+  uint64_t manifest_file_size = 0;
+  ASSERT_OK(db->GetLiveFiles(live_files, &manifest_file_size));
+  std::string flushed_file;
+  for (const auto& live_file : live_files) {
+    if (live_file.substr(live_file.size() - 4) == ".sst") {
+      ASSERT_TRUE(flushed_file.empty()) << flushed_file << " " << live_file;
+      flushed_file = db_name + live_file;
+    }
+  }
+  ASSERT_FALSE(flushed_file.empty());
+  db.reset();
+
+  SstFileReader reader(options_);
+  ASSERT_OK(reader.Open(flushed_file));
+  std::unique_ptr<Iterator> iter = reader.NewTableIterator();
+  ASSERT_NE(iter, nullptr);
+  iter->SeekToFirst();
+
+  for (size_t i = 0; i < keys.size(); ++i) {
+    ASSERT_TRUE(iter->Valid());
+    ParsedEntryInfo parsed_key;
+    ASSERT_OK(reader.ParseTableIteratorKey(iter->key(), &parsed_key));
+    EXPECT_EQ(parsed_key.user_key, keys[i]);
+    EXPECT_EQ(parsed_key.sequence, i + 1);
+    EXPECT_EQ(parsed_key.type, kEntryPut);
+    iter->Next();
+  }
+
+  EXPECT_FALSE(iter->Valid());
+
+  ASSERT_OK(DestroyDB(db_name, options));
+}
+
+TEST_F(SstFileReaderTest, ParseTableIteratorKeyRejectsShortKey) {
+  // Verify that invalid raw table keys fail through the public parser before
+  // callers can accidentally interpret a missing internal-key trailer.
+  CreateFile(sst_name_,
+             {EncodeAsString(0), EncodeAsString(1), EncodeAsString(2)});
+  SstFileReader reader(options_);
+  ParsedEntryInfo parsed_key;
+
+  const Status status = reader.ParseTableIteratorKey("short", &parsed_key);
+  EXPECT_TRUE(status.IsInvalidArgument()) << status.ToString();
+}
+
 TEST_F(SstFileReaderTest, Uint64Comparator) {
   options_.comparator = test::Uint64Comparator();
   std::vector<std::string> keys;
@@ -864,6 +964,47 @@ TEST_P(SstFileReaderTableGetTest, Basic) {
   Close();
 }
 
+TEST_P(SstFileReaderTableGetTest, BlobBackedWideColumnDefaultWithoutFetcher) {
+  // Goal: cover the SstFileReader path where GetContext has no BlobFetcher.
+  // The test writes a blob-backed wide-column entity, compacts it so the SST
+  // stores blob references, and then verifies SstFileReader reports a clean
+  // Corruption status instead of dereferencing a null BlobFetcher.
+  Options options = CurrentOptions();
+  options.disable_auto_compactions = true;
+  options.enable_blob_files = true;
+  options.min_blob_size = 50;
+
+  DestroyAndReopen(options);
+
+  const std::string key = "blob_backed_entity";
+  const std::string default_value(100, 'd');
+  WideColumns columns{{kDefaultWideColumnName, default_value},
+                      {"meta", "inline"}};
+
+  ASSERT_OK(
+      db_->PutEntity(WriteOptions(), db_->DefaultColumnFamily(), key, columns));
+  ASSERT_OK(Flush());
+  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+
+  std::vector<LiveFileMetaData> files;
+  dbfull()->GetLiveFilesMetaData(&files);
+  ASSERT_EQ(files.size(), 1);
+  std::string file_name = files[0].directory + "/" + files[0].relative_filename;
+
+  SstFileReader reader(options);
+  ASSERT_OK(reader.Open(file_name));
+
+  std::vector<Slice> keys = {key};
+  std::vector<std::string> values;
+  auto statuses = DoGet(reader, keys, &values);
+
+  ASSERT_TRUE(statuses[0].IsCorruption()) << statuses[0].ToString();
+  ASSERT_NE(statuses[0].ToString().find("blob fetcher"), std::string::npos)
+      << statuses[0].ToString();
+
+  Close();
+}
+
 INSTANTIATE_TEST_CASE_P(SingleAndMulti, SstFileReaderTableGetTest,
                         testing::Bool());
 

table/table_test.cc

@@ -8633,7 +8633,7 @@ TEST_P(UserDefinedIndexTest, ValueTypeMappingViaDBFlush) {
   // ValueTypes by writing various operation types via the DB API, flushing,
   // and inspecting what the TestUserDefinedIndexBuilder received.
   if (is_reverse_comparator_) {
-    // Skip for reverse comparator — the key ordering makes this test
+    // Skip for reverse comparator -- the key ordering makes this test
     // unnecessarily complex and the mapping logic is comparator-independent.
     ROCKSDB_GTEST_BYPASS("Skipped for reverse comparator");
     return;
@@ -8735,7 +8735,7 @@ TEST_P(UserDefinedIndexTest, CompactionWithSnapshotsAndUDI) {
   ASSERT_OK(db->Delete(WriteOptions(), "key_bb"));
   ASSERT_OK(db->Flush(FlushOptions()));
 
-  // Compact L0 → L1. With the snapshot held, both versions of key_aa
+  // Compact L0 -> L1. With the snapshot held, both versions of key_aa
   // and the delete tombstone for key_bb must be preserved in the compaction
   // output. The UDI builder receives multiple entries for key_aa.
   ASSERT_OK(db->CompactRange(CompactRangeOptions(), nullptr, nullptr));
@@ -8747,7 +8747,7 @@ TEST_P(UserDefinedIndexTest, CompactionWithSnapshotsAndUDI) {
   const auto& log = user_defined_index_factory->key_type_log_;
   ASSERT_FALSE(log.empty());
 
-  // Count total occurrences of key_aa across all builders — at least 4:
+  // Count total occurrences of key_aa across all builders -- at least 4:
   // flush1 (v1) + flush2 (v2) + compaction (v2, v1).
   int key_aa_count = 0;
   int key_bb_count = 0;

tools/db_crashtest.py

@@ -30,6 +30,10 @@ _NO_SPACE_SUBSTRINGS = (
     "enospc",
 )
 _OUTPUT_PATH_RE = re.compile(r"(/[^\s]+)")
+_TSAN_OPTIONS_ENV_VAR = "TSAN_OPTIONS"
+_TSAN_SUPPRESSIONS_FILE = os.path.abspath(
+    os.path.join(os.path.dirname(__file__), "tsan_suppressions.txt")
+)
 
 
 def get_random_seed(override):
@@ -51,6 +55,16 @@ def quote_arg_for_display(arg):
     return f"{flag}={shlex.quote(value)}"
 
 
+def stress_cmd_env():
+    env = os.environ.copy()
+    if (
+        _TSAN_OPTIONS_ENV_VAR not in env
+        and os.path.exists(_TSAN_SUPPRESSIONS_FILE)
+    ):
+        env[_TSAN_OPTIONS_ENV_VAR] = "suppressions=" + _TSAN_SUPPRESSIONS_FILE
+    return env
+
+
 def early_argument_parsing_before_main():
     parser = argparse.ArgumentParser()
     parser.add_argument(
@@ -1905,7 +1919,9 @@ def diagnostic_paths(finalized_params):
 
 
 def execute_cmd(cmd, timeout=None, timeout_pstack=False):
-    child = subprocess.Popen(cmd, stderr=subprocess.PIPE, stdout=subprocess.PIPE)
+    child = subprocess.Popen(
+        cmd, stderr=subprocess.PIPE, stdout=subprocess.PIPE, env=stress_cmd_env()
+    )
     print(
         "Running db_stress with pid=%d: %s\n\n"
         % (child.pid, " ".join(quote_arg_for_display(arg) for arg in cmd))
@@ -1920,7 +1936,7 @@ def execute_cmd(cmd, timeout=None, timeout_pstack=False):
         hit_timeout = True
         if timeout_pstack:
             os.system("pstack %d" % pid)
-        child.terminate()  # SIGTERM — triggers TerminationHandler
+        child.terminate()  # SIGTERM -- triggers TerminationHandler
         try:
             outs, errs = child.communicate(timeout=3)
             print("TERMINATED %d\n" % child.pid)
@@ -2006,7 +2022,7 @@ def cleanup_after_success(dbname):
         if parts[0] in ["--env_uri", "--fs_uri"]:
             cleanup_cmd_parts.append(arg)
     print("Running DB cleanup command - %s\n" % " ".join(cleanup_cmd_parts))
-    ret = subprocess.call(cleanup_cmd_parts)
+    ret = subprocess.call(cleanup_cmd_parts, env=stress_cmd_env())
     if ret != 0:
         print("ERROR: DB cleanup returned error %d\n" % ret)
         sys.exit(2)

tools/db_crashtest_test.py

@@ -16,6 +16,7 @@ import unittest
 _DB_CRASHTEST_PATH = os.path.join(os.path.dirname(__file__), "db_crashtest.py")
 _TEST_DIR_ENV_VAR = "TEST_TMPDIR"
 _TEST_EXPECTED_DIR_ENV_VAR = "TEST_TMPDIR_EXPECTED"
+_TSAN_OPTIONS_ENV_VAR = "TSAN_OPTIONS"
 
 
 def load_db_crashtest_module():
@@ -40,6 +41,7 @@ class DBCrashTestTest(unittest.TestCase):
         )
         self.old_test_tmpdir = os.environ.get(_TEST_DIR_ENV_VAR)
         self.old_test_expected_tmpdir = os.environ.get(_TEST_EXPECTED_DIR_ENV_VAR)
+        self.old_tsan_options = os.environ.get(_TSAN_OPTIONS_ENV_VAR)
         os.environ[_TEST_DIR_ENV_VAR] = self.test_tmpdir
         os.environ.pop(_TEST_EXPECTED_DIR_ENV_VAR, None)
 
@@ -54,6 +56,11 @@ class DBCrashTestTest(unittest.TestCase):
         else:
             os.environ[_TEST_EXPECTED_DIR_ENV_VAR] = self.old_test_expected_tmpdir
 
+        if self.old_tsan_options is None:
+            os.environ.pop(_TSAN_OPTIONS_ENV_VAR, None)
+        else:
+            os.environ[_TSAN_OPTIONS_ENV_VAR] = self.old_tsan_options
+
         shutil.rmtree(self.test_tmpdir)
 
     def load_db_crashtest(self):
@@ -66,6 +73,28 @@ class DBCrashTestTest(unittest.TestCase):
             params.update(overrides)
         return params
 
+    def test_stress_cmd_env_defaults_tsan_suppressions(self):
+        os.environ.pop(_TSAN_OPTIONS_ENV_VAR, None)
+        db_crashtest = self.load_db_crashtest()
+
+        env = db_crashtest.stress_cmd_env()
+
+        self.assertEqual(
+            "suppressions="
+            + os.path.abspath(
+                os.path.join(os.path.dirname(__file__), "tsan_suppressions.txt")
+            ),
+            env[_TSAN_OPTIONS_ENV_VAR],
+        )
+
+    def test_stress_cmd_env_preserves_tsan_options(self):
+        os.environ[_TSAN_OPTIONS_ENV_VAR] = "halt_on_error=1"
+        db_crashtest = self.load_db_crashtest()
+
+        env = db_crashtest.stress_cmd_env()
+
+        self.assertEqual("halt_on_error=1", env[_TSAN_OPTIONS_ENV_VAR])
+
     def test_setup_expected_values_dir_preserves_existing_contents(self):
         os.makedirs(self.expected_dir)
         marker = os.path.join(self.expected_dir, "marker")

tools/run_clang_tidy.py

@@ -205,7 +205,7 @@ def collect_changed_lines(repo_root, diff_base_arg=None):
         )
         merge_into(all_changed, parse_diff_for_changed_lines(diff_staged))
 
-        # Untracked files — treat every line as changed
+        # Untracked files -- treat every line as changed
         untracked_out, _ = run_cmd(
             ["git", "ls-files", "--others", "--exclude-standard"], cwd=repo_root
         )
@@ -510,7 +510,7 @@ def main():
     jobs = args.jobs or os.cpu_count() or 4
 
     # ------------------------------------------------------------------
-    # Step 1 — detect changes
+    # Step 1 -- detect changes
     # ------------------------------------------------------------------
     log("=" * 70)
     log("Step 1: Detecting changes")
@@ -529,7 +529,7 @@ def main():
         log(f"    {f}  ({len(changed_lines[f])} lines)")
 
     # ------------------------------------------------------------------
-    # Step 2 — select compilable files present in compile_commands.json
+    # Step 2 -- select compilable files present in compile_commands.json
     # ------------------------------------------------------------------
     db_files = load_compile_db(compile_db_path, repo_root)
     cc_changed = sorted(
@@ -549,7 +549,7 @@ def main():
     log("=" * 70)
 
     # ------------------------------------------------------------------
-    # Step 3 — run clang-tidy in parallel via ThreadPoolExecutor
+    # Step 3 -- run clang-tidy in parallel via ThreadPoolExecutor
     # ------------------------------------------------------------------
     all_raw_output = []
     all_filtered = []
@@ -603,7 +603,7 @@ def main():
         log(f"\nFull clang-tidy output saved to {args.output}")
 
     # ------------------------------------------------------------------
-    # Step 4 — report filtered results
+    # Step 4 -- report filtered results
     # ------------------------------------------------------------------
     log(f"\n{'=' * 70}")
     log(f"Step 3: Results  (wall time {wall_time:.1f}s)")

tools/tsan_suppressions.txt

@@ -0,0 +1,20 @@
+# ThreadSanitizer suppressions for known third-party false positives.
+#
+# Use with:
+#   TSAN_OPTIONS="suppressions=/path/to/rocksdb/tools/tsan_suppressions.txt"
+#
+# If the final application already has a TSAN suppressions file, merge these
+# entries into that application-owned file.
+#
+# liburing's io_uring_mmap() uses __sys_mmap, which bypasses TSAN's mmap
+# interceptor, then immediately reads from the mapped memory in
+# io_uring_setup_ring_pointers(). When the kernel reuses a virtual address that
+# was previously mmap'd through libc and tracked by TSAN, TSAN can report a race
+# between the old mapping's write and the new io_uring read. RocksDB annotates
+# the io_uring mappings after io_uring_queue_init() returns, but that is too
+# late for reads that happen inside liburing during initialization.
+#
+# Revisit these suppressions when RocksDB upgrades to a liburing version that no
+# longer uses the raw mmap syscall for io_uring_mmap().
+race:io_uring_mmap
+race:io_uring_setup_ring_pointers

unreleased_history/behavior_changes/write_committed_merge_operand_filtering

@@ -0,0 +1 @@
+WriteCommitted TransactionDB now matches WritePrepared and WriteUnprepared compaction filtering in both single- and two-write-queue modes: a compaction filter's FilterMergeOperand will not be invoked on merge operands at or below the latest published sequence number.

unreleased_history/bug_fixes/blob_backed_wide_column_merge_reads.md

@@ -0,0 +1,3 @@
+Fixed blob-backed wide-column merge reads to preserve correct status
+propagation and resolution across memtable, read-only, and secondary DB
+paths.

unreleased_history/bug_fixes/get_creation_time_of_oldest_file_async_open

@@ -0,0 +1 @@
+Fixed a bug where `DB::GetCreationTimeOfOldestFile()` could return inaccurate results instead of the real creation time when called shortly after opening a legacy DB (one whose manifest lacks `file_creation_time`) with `open_files_async = true`. The API now waits for background SST file loading to complete only when needed; modern DBs are unaffected.