@nxtedition/rocksdb 10.0.13 → 10.0.15

package/deps/rocksdb/rocksdb/db/compact_files_test.cc

@@ -34,8 +34,8 @@ class CompactFilesTest : public testing::Test {
 // A class which remembers the name of each flushed file.
 class FlushedFileCollector : public EventListener {
  public:
-  FlushedFileCollector() {}
-  ~FlushedFileCollector() override {}
+  FlushedFileCollector() = default;
+  ~FlushedFileCollector() override = default;
 
   void OnFlushCompleted(DB* /*db*/, const FlushJobInfo& info) override {
     std::lock_guard<std::mutex> lock(mutex_);
@@ -45,7 +45,7 @@ class FlushedFileCollector : public EventListener {
   std::vector<std::string> GetFlushedFiles() {
     std::lock_guard<std::mutex> lock(mutex_);
     std::vector<std::string> result;
-    for (auto fname : flushed_files_) {
+    for (const auto& fname : flushed_files_) {
       result.push_back(fname);
     }
     return result;
@@ -159,7 +159,9 @@ TEST_F(CompactFilesTest, MultipleLevel) {
   // Compact files except the file in L3
   std::vector<std::string> files;
   for (int i = 0; i < 6; ++i) {
-    if (i == 3) continue;
+    if (i == 3) {
+      continue;
+    }
     for (auto& file : meta.levels[i].files) {
       files.push_back(file.db_path + "/" + file.name);
     }
@@ -228,7 +230,7 @@ TEST_F(CompactFilesTest, ObsoleteFiles) {
   ASSERT_OK(static_cast_with_check<DBImpl>(db)->TEST_WaitForCompact());
 
   // verify all compaction input files are deleted
-  for (auto fname : l0_files) {
+  for (const auto& fname : l0_files) {
     ASSERT_EQ(Status::NotFound(), env_->FileExists(fname));
   }
   delete db;
@@ -439,6 +441,50 @@ TEST_F(CompactFilesTest, SentinelCompressionType) {
   }
 }
 
+TEST_F(CompactFilesTest, CompressionWithBlockAlign) {
+  Options options;
+  options.compression = CompressionType::kNoCompression;
+  options.create_if_missing = true;
+  options.disable_auto_compactions = true;
+
+  std::shared_ptr<FlushedFileCollector> collector =
+      std::make_shared<FlushedFileCollector>();
+  options.listeners.push_back(collector);
+
+  {
+    BlockBasedTableOptions bbto;
+    bbto.block_align = true;
+    options.table_factory.reset(NewBlockBasedTableFactory(bbto));
+  }
+
+  std::unique_ptr<DB> db;
+  {
+    DB* _db = nullptr;
+    ASSERT_OK(DB::Open(options, db_name_, &_db));
+    db.reset(_db);
+  }
+
+  ASSERT_OK(db->Put(WriteOptions(), "key", "val"));
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  // Ensure background work is fully finished including listener callbacks
+  // before accessing listener state.
+  ASSERT_OK(
+      static_cast_with_check<DBImpl>(db.get())->TEST_WaitForBackgroundWork());
+  auto l0_files = collector->GetFlushedFiles();
+  ASSERT_EQ(1, l0_files.size());
+
+  // We can run this test even without Snappy support because we expect the
+  // `CompactFiles()` to fail before actually invoking Snappy compression.
+  CompactionOptions compaction_opts;
+  compaction_opts.compression = CompressionType::kSnappyCompression;
+  ASSERT_TRUE(db->CompactFiles(compaction_opts, l0_files, 1 /* output_level */)
+                  .IsInvalidArgument());
+
+  compaction_opts.compression = CompressionType::kDisableCompressionOption;
+  ASSERT_OK(db->CompactFiles(compaction_opts, l0_files, 1 /* output_level */));
+}
+
 TEST_F(CompactFilesTest, GetCompactionJobInfo) {
   Options options;
   options.create_if_missing = true;
@@ -492,4 +538,3 @@ int main(int argc, char** argv) {
   ::testing::InitGoogleTest(&argc, argv);
   return RUN_ALL_TESTS();
 }
-

package/deps/rocksdb/rocksdb/db/compaction/compaction.cc

@@ -13,6 +13,7 @@
 #include <vector>
 
 #include "db/column_family.h"
+#include "db/dbformat.h"
 #include "logging/logging.h"
 #include "rocksdb/compaction_filter.h"
 #include "rocksdb/sst_partitioner.h"
@@ -21,9 +22,6 @@
 
 namespace ROCKSDB_NAMESPACE {
 
-const uint64_t kRangeTombstoneSentinel =
-    PackSequenceAndType(kMaxSequenceNumber, kTypeRangeDeletion);
-
 int sstableKeyCompare(const Comparator* uc, const Slice& a, const Slice& b) {
   auto c = uc->CompareWithoutTimestamp(ExtractUserKey(a), ExtractUserKey(b));
   if (c != 0) {
@@ -115,6 +113,42 @@ void Compaction::GetBoundaryKeys(
   }
 }
 
+void Compaction::GetBoundaryInternalKeys(
+    VersionStorageInfo* vstorage,
+    const std::vector<CompactionInputFiles>& inputs, InternalKey* smallest_key,
+    InternalKey* largest_key, int exclude_level) {
+  bool initialized = false;
+  const InternalKeyComparator* icmp = vstorage->InternalComparator();
+  for (size_t i = 0; i < inputs.size(); ++i) {
+    if (inputs[i].files.empty() || inputs[i].level == exclude_level) {
+      continue;
+    }
+    if (inputs[i].level == 0) {
+      // we need to consider all files on level 0
+      for (const auto* f : inputs[i].files) {
+        if (!initialized || icmp->Compare(f->smallest, *smallest_key) < 0) {
+          *smallest_key = f->smallest;
+        }
+        if (!initialized || icmp->Compare(f->largest, *largest_key) > 0) {
+          *largest_key = f->largest;
+        }
+        initialized = true;
+      }
+    } else {
+      // we only need to consider the first and last file
+      if (!initialized ||
+          icmp->Compare(inputs[i].files[0]->smallest, *smallest_key) < 0) {
+        *smallest_key = inputs[i].files[0]->smallest;
+      }
+      if (!initialized ||
+          icmp->Compare(inputs[i].files.back()->largest, *largest_key) > 0) {
+        *largest_key = inputs[i].files.back()->largest;
+      }
+      initialized = true;
+    }
+  }
+}
+
 std::vector<CompactionInputFiles> Compaction::PopulateWithAtomicBoundaries(
     VersionStorageInfo* vstorage, std::vector<CompactionInputFiles> inputs) {
   const Comparator* ucmp = vstorage->InternalComparator()->user_comparator();
@@ -126,7 +160,9 @@ std::vector<CompactionInputFiles> Compaction::PopulateWithAtomicBoundaries(
     AtomicCompactionUnitBoundary cur_boundary;
     size_t first_atomic_idx = 0;
     auto add_unit_boundary = [&](size_t to) {
-      if (first_atomic_idx == to) return;
+      if (first_atomic_idx == to) {
+        return;
+      }
       for (size_t k = first_atomic_idx; k < to; k++) {
         inputs[i].atomic_compaction_unit_boundaries.push_back(cur_boundary);
       }
@@ -316,11 +352,9 @@ Compaction::Compaction(
 
   // for the non-bottommost levels, it tries to build files match the target
   // file size, but not guaranteed. It could be 2x the size of the target size.
-  max_output_file_size_ =
-      bottommost_level_ || grandparents_.empty() ||
-              !_immutable_options.level_compaction_dynamic_file_size
-          ? target_output_file_size_
-          : 2 * target_output_file_size_;
+  max_output_file_size_ = bottommost_level_ || grandparents_.empty()
+                              ? target_output_file_size_
+                              : 2 * target_output_file_size_;
 
 #ifndef NDEBUG
   for (size_t i = 1; i < inputs_.size(); ++i) {
@@ -399,9 +433,14 @@ void Compaction::PopulatePenultimateLevelOutputRange() {
     }
   }
 
-  GetBoundaryKeys(input_vstorage_, inputs_,
-                  &penultimate_level_smallest_user_key_,
-                  &penultimate_level_largest_user_key_, exclude_level);
+  // FIXME: should make use of `penultimate_output_range_type_`.
+  // FIXME: when last level's input range does not overlap with
+  //  penultimate level, and penultimate level input is empty,
+  //  this call will not set penultimate_level_smallest_ or
+  //  penultimate_level_largest_. No keys will be compacted up.
+  GetBoundaryInternalKeys(input_vstorage_, inputs_,
+                          &penultimate_level_smallest_,
+                          &penultimate_level_largest_, exclude_level);
 }
 
 Compaction::~Compaction() {
@@ -426,33 +465,40 @@ bool Compaction::OverlapPenultimateLevelOutputRange(
   if (!SupportsPerKeyPlacement()) {
     return false;
   }
+
+  // See FIXME in Compaction::PopulatePenultimateLevelOutputRange().
+  // We do not compact any key up in this case.
+  if (penultimate_level_smallest_.size() == 0 ||
+      penultimate_level_largest_.size() == 0) {
+    return false;
+  }
+
   const Comparator* ucmp =
       input_vstorage_->InternalComparator()->user_comparator();
 
   return ucmp->CompareWithoutTimestamp(
-             smallest_key, penultimate_level_largest_user_key_) <= 0 &&
+             smallest_key, penultimate_level_largest_.user_key()) <= 0 &&
          ucmp->CompareWithoutTimestamp(
-             largest_key, penultimate_level_smallest_user_key_) >= 0;
+             largest_key, penultimate_level_smallest_.user_key()) >= 0;
 }
 
 // key includes timestamp if user-defined timestamp is enabled.
-bool Compaction::WithinPenultimateLevelOutputRange(const Slice& key) const {
+bool Compaction::WithinPenultimateLevelOutputRange(
+    const ParsedInternalKey& ikey) const {
   if (!SupportsPerKeyPlacement()) {
     return false;
   }
 
-  if (penultimate_level_smallest_user_key_.empty() ||
-      penultimate_level_largest_user_key_.empty()) {
+  if (penultimate_level_smallest_.size() == 0 ||
+      penultimate_level_largest_.size() == 0) {
     return false;
   }
 
-  const Comparator* ucmp =
-      input_vstorage_->InternalComparator()->user_comparator();
+  const InternalKeyComparator* icmp = input_vstorage_->InternalComparator();
 
-  return ucmp->CompareWithoutTimestamp(
-             key, penultimate_level_smallest_user_key_) >= 0 &&
-         ucmp->CompareWithoutTimestamp(
-             key, penultimate_level_largest_user_key_) <= 0;
+  // op_type of a key can change during compaction, e.g. Merge -> Put.
+  return icmp->CompareKeySeq(ikey, penultimate_level_smallest_.Encode()) >= 0 &&
+         icmp->CompareKeySeq(ikey, penultimate_level_largest_.Encode()) <= 0;
 }
 
 bool Compaction::InputCompressionMatchesOutput() const {
@@ -709,7 +755,9 @@ int InputSummary(const std::vector<FileMetaData*>& files, char* output,
     AppendHumanBytes(files.at(i)->fd.GetFileSize(), sztxt, 16);
     ret = snprintf(output + write, sz, "%" PRIu64 "(%s) ",
                    files.at(i)->fd.GetNumber(), sztxt);
-    if (ret < 0 || ret >= sz) break;
+    if (ret < 0 || ret >= sz) {
+      break;
+    }
     write += ret;
   }
   // if files.size() is non-zero, overwrite the last space

package/deps/rocksdb/rocksdb/db/compaction/compaction.h

@@ -353,14 +353,6 @@ class Compaction {
 
   Slice GetLargestUserKey() const { return largest_user_key_; }
 
-  Slice GetPenultimateLevelSmallestUserKey() const {
-    return penultimate_level_smallest_user_key_;
-  }
-
-  Slice GetPenultimateLevelLargestUserKey() const {
-    return penultimate_level_largest_user_key_;
-  }
-
   PenultimateOutputRangeType GetPenultimateOutputRangeType() const {
     return penultimate_output_range_type_;
   }
@@ -383,10 +375,8 @@ class Compaction {
   // per_key_placement feature, which is safe to place the key to the
   // penultimate level. different compaction strategy has different rules.
   // If per_key_placement is not supported, always return false.
-  // TODO: currently it doesn't support moving data from the last level to the
-  //  penultimate level
   //  key includes timestamp if user-defined timestamp is enabled.
-  bool WithinPenultimateLevelOutputRange(const Slice& key) const;
+  bool WithinPenultimateLevelOutputRange(const ParsedInternalKey& ikey) const;
 
   CompactionReason compaction_reason() const { return compaction_reason_; }
 
@@ -456,6 +446,13 @@ class Compaction {
                               Slice* smallest_key, Slice* largest_key,
                               int exclude_level = -1);
 
+  // get the smallest and largest internal key present in files to be compacted
+  static void GetBoundaryInternalKeys(
+      VersionStorageInfo* vstorage,
+      const std::vector<CompactionInputFiles>& inputs,
+      InternalKey* smallest_key, InternalKey* largest_key,
+      int exclude_level = -1);
+
   // populate penultimate level output range, which will be used to determine if
   // a key is safe to output to the penultimate level (details see
   // `Compaction::WithinPenultimateLevelOutputRange()`.
@@ -568,8 +565,8 @@ class Compaction {
   // Key range for penultimate level output
   // includes timestamp if user-defined timestamp is enabled.
   // penultimate_output_range_type_ shows the range type
-  Slice penultimate_level_smallest_user_key_;
-  Slice penultimate_level_largest_user_key_;
+  InternalKey penultimate_level_smallest_;
+  InternalKey penultimate_level_largest_;
   PenultimateOutputRangeType penultimate_output_range_type_ =
       PenultimateOutputRangeType::kNotSupported;
 };
@@ -598,6 +595,6 @@ struct PerKeyPlacementContext {
 #endif /* !NDEBUG */
 
 // Return sum of sizes of all files in `files`.
-extern uint64_t TotalFileSize(const std::vector<FileMetaData*>& files);
+uint64_t TotalFileSize(const std::vector<FileMetaData*>& files);
 
 }  // namespace ROCKSDB_NAMESPACE

package/deps/rocksdb/rocksdb/db/compaction/compaction_iteration_stats.h

@@ -44,6 +44,13 @@ struct CompactionIterationStats {
   uint64_t total_blob_bytes_read = 0;
   uint64_t num_blobs_relocated = 0;
   uint64_t total_blob_bytes_relocated = 0;
+
+  // TimedPut diagnostics
+  // Total number of kTypeValuePreferredSeqno records encountered.
+  uint64_t num_input_timed_put_records = 0;
+  // Number of kTypeValuePreferredSeqno records we ended up swapping in
+  // preferred seqno.
+  uint64_t num_timed_put_swap_preferred_seqno = 0;
 };
 
 }  // namespace ROCKSDB_NAMESPACE

package/deps/rocksdb/rocksdb/db/compaction/compaction_iterator.cc

@@ -491,6 +491,8 @@ void CompactionIterator::NextFromInput() {
     if (ikey_.type == kTypeDeletion || ikey_.type == kTypeSingleDeletion ||
         ikey_.type == kTypeDeletionWithTimestamp) {
       iter_stats_.num_input_deletion_records++;
+    } else if (ikey_.type == kTypeValuePreferredSeqno) {
+      iter_stats_.num_input_timed_put_records++;
     }
     iter_stats_.total_input_raw_key_bytes += key_.size();
     iter_stats_.total_input_raw_value_bytes += value_.size();
@@ -618,7 +620,8 @@ void CompactionIterator::NextFromInput() {
       // not compact out.  We will keep this Put, but can drop it's data.
       // (See Optimization 3, below.)
       if (ikey_.type != kTypeValue && ikey_.type != kTypeBlobIndex &&
-          ikey_.type != kTypeWideColumnEntity) {
+          ikey_.type != kTypeWideColumnEntity &&
+          ikey_.type != kTypeValuePreferredSeqno) {
         ROCKS_LOG_FATAL(info_log_, "Unexpected key %s for compaction output",
                         ikey_.DebugString(allow_data_in_errors_, true).c_str());
         assert(false);
@@ -632,7 +635,8 @@ void CompactionIterator::NextFromInput() {
         assert(false);
       }
 
-      if (ikey_.type == kTypeBlobIndex || ikey_.type == kTypeWideColumnEntity) {
+      if (ikey_.type == kTypeBlobIndex || ikey_.type == kTypeWideColumnEntity ||
+          ikey_.type == kTypeValuePreferredSeqno) {
         ikey_.type = kTypeValue;
         current_key_.UpdateInternalKey(ikey_.sequence, ikey_.type);
       }
@@ -798,7 +802,8 @@ void CompactionIterator::NextFromInput() {
             // happened
             if (next_ikey.type != kTypeValue &&
                 next_ikey.type != kTypeBlobIndex &&
-                next_ikey.type != kTypeWideColumnEntity) {
+                next_ikey.type != kTypeWideColumnEntity &&
+                next_ikey.type != kTypeValuePreferredSeqno) {
               ++iter_stats_.num_single_del_mismatch;
             }
 
@@ -968,6 +973,62 @@ void CompactionIterator::NextFromInput() {
         validity_info_.SetValid(ValidContext::kKeepDel);
         at_next_ = true;
       }
+    } else if (ikey_.type == kTypeValuePreferredSeqno &&
+               DefinitelyInSnapshot(ikey_.sequence, earliest_snapshot_) &&
+               (bottommost_level_ ||
+                (compaction_ != nullptr &&
+                 compaction_->KeyNotExistsBeyondOutputLevel(ikey_.user_key,
+                                                            &level_ptrs_)))) {
+      // This section that attempts to swap preferred sequence number will not
+      // be invoked if this is a CompactionIterator created for flush, since
+      // `compaction_` will be nullptr and it's not bottommost either.
+      //
+      // The entries with the same user key and smaller sequence numbers are
+      // all in this earliest snapshot range to be iterated. Since those entries
+      // will be hidden by this entry [rule A], it's safe to swap in the
+      // preferred seqno now.
+      //
+      // It's otherwise not safe to swap in the preferred seqno since it's
+      // possible for entries in earlier snapshots to have sequence number that
+      // is smaller than this entry's sequence number but bigger than this
+      // entry's preferred sequence number. Swapping in the preferred sequence
+      // number will break the internal key ordering invariant for this key.
+      //
+      // A special case involving range deletion is handled separately below.
+      auto [unpacked_value, preferred_seqno] =
+          ParsePackedValueWithSeqno(value_);
+      assert(preferred_seqno < ikey_.sequence || ikey_.sequence == 0);
+      if (range_del_agg_->ShouldDelete(
+              key_, RangeDelPositioningMode::kForwardTraversal)) {
+        ++iter_stats_.num_record_drop_hidden;
+        ++iter_stats_.num_record_drop_range_del;
+        AdvanceInputIter();
+      } else {
+        InternalKey ikey_after_swap(ikey_.user_key,
+                                    std::min(preferred_seqno, ikey_.sequence),
+                                    kTypeValue);
+        Slice ikey_after_swap_slice(*ikey_after_swap.rep());
+        if (range_del_agg_->ShouldDelete(
+                ikey_after_swap_slice,
+                RangeDelPositioningMode::kForwardTraversal)) {
+          // A range tombstone that doesn't cover this kTypeValuePreferredSeqno
+          // entry will end up covering the entry, so it's not safe to swap
+          // preferred sequence number. In this case, we output the entry as is.
+          validity_info_.SetValid(ValidContext::kNewUserKey);
+        } else {
+          if (ikey_.sequence != 0) {
+            iter_stats_.num_timed_put_swap_preferred_seqno++;
+            saved_seq_for_penul_check_ = ikey_.sequence;
+            ikey_.sequence = preferred_seqno;
+          }
+          ikey_.type = kTypeValue;
+          current_key_.UpdateInternalKey(ikey_.sequence, ikey_.type);
+          key_ = current_key_.GetInternalKey();
+          ikey_.user_key = current_key_.GetUserKey();
+          value_ = unpacked_value;
+          validity_info_.SetValid(ValidContext::kSwapPreferredSeqno);
+        }
+      }
     } else if (ikey_.type == kTypeMerge) {
       if (!merge_helper_->HasOperator()) {
         status_ = Status::InvalidArgument(
@@ -1219,6 +1280,21 @@ void CompactionIterator::DecideOutputLevel() {
   }
 #endif  // NDEBUG
 
+  // saved_seq_for_penul_check_ is populated in `NextFromInput` when the
+  // entry's sequence number is non zero and validity context for output this
+  // entry is kSwapPreferredSeqno for use in `DecideOutputLevel`. It should be
+  // cleared out here unconditionally. Otherwise, it may end up getting consumed
+  // incorrectly by a different entry.
+  SequenceNumber seq_for_range_check =
+      (saved_seq_for_penul_check_.has_value() &&
+       saved_seq_for_penul_check_.value() != kMaxSequenceNumber)
+          ? saved_seq_for_penul_check_.value()
+          : ikey_.sequence;
+  saved_seq_for_penul_check_ = std::nullopt;
+  ParsedInternalKey ikey_for_range_check = ikey_;
+  if (seq_for_range_check != ikey_.sequence) {
+    ikey_for_range_check.sequence = seq_for_range_check;
+  }
   if (output_to_penultimate_level_) {
     // If it's decided to output to the penultimate level, but unsafe to do so,
     // still output to the last level. For example, moving the data from a lower
@@ -1227,7 +1303,7 @@ void CompactionIterator::DecideOutputLevel() {
     // not from this compaction.
     // TODO: add statistic for declined output_to_penultimate_level
     bool safe_to_penultimate_level =
-        compaction_->WithinPenultimateLevelOutputRange(ikey_.user_key);
+        compaction_->WithinPenultimateLevelOutputRange(ikey_for_range_check);
     if (!safe_to_penultimate_level) {
       output_to_penultimate_level_ = false;
       // It could happen when disable/enable `last_level_temperature` while
@@ -1239,7 +1315,7 @@ void CompactionIterator::DecideOutputLevel() {
       // snapshot is released before enabling `last_level_temperature` feature
       // We will migrate the feature to `last_level_temperature` and maybe make
       // it not dynamically changeable.
-      if (ikey_.sequence > earliest_snapshot_) {
+      if (seq_for_range_check > earliest_snapshot_) {
         status_ = Status::Corruption(
             "Unsafe to store Seq later than snapshot in the last level if "
             "per_key_placement is enabled");
@@ -1256,10 +1332,13 @@ void CompactionIterator::PrepareOutput() {
       } else if (ikey_.type == kTypeBlobIndex) {
         GarbageCollectBlobIfNeeded();
       }
-    }
 
-    if (compaction_ != nullptr && compaction_->SupportsPerKeyPlacement()) {
-      DecideOutputLevel();
+      // For range del sentinel, we don't use it to cut files for bottommost
+      // compaction. So it should not make a difference which output level we
+      // decide.
+      if (compaction_ != nullptr && compaction_->SupportsPerKeyPlacement()) {
+        DecideOutputLevel();
+      }
     }
 
     // Zeroing out the sequence number leads to better compression.

package/deps/rocksdb/rocksdb/db/compaction/compaction_iterator.h

@@ -119,7 +119,8 @@ class CompactionIterator {
     virtual bool SupportsPerKeyPlacement() const = 0;
 
     // `key` includes timestamp if user-defined timestamp is enabled.
-    virtual bool WithinPenultimateLevelOutputRange(const Slice& key) const = 0;
+    virtual bool WithinPenultimateLevelOutputRange(
+        const ParsedInternalKey&) const = 0;
   };
 
   class RealCompaction : public CompactionProxy {
@@ -186,8 +187,9 @@ class CompactionIterator {
     // Check if key is within penultimate level output range, to see if it's
     // safe to output to the penultimate level for per_key_placement feature.
     // `key` includes timestamp if user-defined timestamp is enabled.
-    bool WithinPenultimateLevelOutputRange(const Slice& key) const override {
-      return compaction_->WithinPenultimateLevelOutputRange(key);
+    bool WithinPenultimateLevelOutputRange(
+        const ParsedInternalKey& ikey) const override {
+      return compaction_->WithinPenultimateLevelOutputRange(ikey);
     }
 
    private:
@@ -408,6 +410,7 @@ class CompactionIterator {
     kKeepDel = 9,
     kNewUserKey = 10,
     kRangeDeletion = 11,
+    kSwapPreferredSeqno = 12,
   };
 
   struct ValidityInfo {
@@ -434,6 +437,17 @@ class CompactionIterator {
   // iterator output (or current key in the underlying iterator during
   // NextFromInput()).
   ParsedInternalKey ikey_;
+
+  // When a kTypeValuePreferredSeqno entry's preferred seqno is safely swapped
+  // in in this compaction, this field saves its original sequence number for
+  // range checking whether it's safe to be placed on the penultimate level.
+  // This is to ensure when such an entry happens to be the right boundary of
+  // penultimate safe range, it won't get excluded because with the preferred
+  // seqno swapped in, it's now larger than the right boundary (itself before
+  // the swap). This is safe to do, because preferred seqno is swapped in only
+  // when no entries with the same user key exist on lower levels and this entry
+  // is already visible in the earliest snapshot.
+  std::optional<SequenceNumber> saved_seq_for_penul_check_ = kMaxSequenceNumber;
   // Stores whether ikey_.user_key is valid. If set to false, the user key is
   // not compared against the current key in the underlying iterator.
   bool has_current_user_key_ = false;