@nxtedition/rocksdb 5.2.21 → 5.2.28

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

@@ -0,0 +1,92 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+
+#pragma once
+#include <algorithm>
+
+#include "db/version_edit.h"
+
+namespace ROCKSDB_NAMESPACE {
+// We boost files that are closer to TTL limit. This boosting could be
+// through FileMetaData.compensated_file_size but this compensated size
+// is widely used as something similar to file size so dramatically boost
+// the value might cause unintended consequences.
+//
+// This boosting algorithm can go very fancy, but here we use a simple
+// formula which can satisify:
+// (1) Different levels are triggered slightly differently to avoid
+//     too many cascading cases
+// (2) Files in the same level get boosting more when TTL gets closer.
+//
+// Don't do any boosting before TTL has past by half. This is to make
+// sure lower write amp for most of the case. And all levels should be
+// fully boosted when total TTL compaction threshold triggers.
+// Differientiate boosting ranges of each level by 1/2. This will make
+// range for each level exponentially increasing. We could do it by
+// having them to be equal, or go even fancier. We can adjust it after
+// we observe the behavior in production.
+// The threshold starting boosting:
+// +------------------------------------------------------------------ +
+// ^                            ^   ^     ^       ^                 ^
+// Age 0                        ... |     |    second last level    thresold
+//                                  |     |
+//                                  |  third last level
+//                                  |
+//                            forth last level
+//
+// We arbitrarily set with 0 when a file is aged boost_age_start and
+// grow linearly. The ratio is arbitrarily set so that when the next level
+// starts to boost, the previous level's boosting amount is 16.
+class FileTtlBooster {
+ public:
+  FileTtlBooster(uint64_t current_time, uint64_t ttl, int num_non_empty_levels,
+                 int level)
+      : current_time_(current_time) {
+    if (ttl == 0 || level == 0 || level >= num_non_empty_levels - 1) {
+      enabled_ = false;
+      boost_age_start_ = 0;
+      boost_step_ = 1;
+    } else {
+      enabled_ = true;
+      uint64_t all_boost_start_age = ttl / 2;
+      uint64_t all_boost_age_range = (ttl / 32) * 31 - all_boost_start_age;
+      uint64_t boost_age_range =
+          all_boost_age_range >> (num_non_empty_levels - level - 1);
+      boost_age_start_ = all_boost_start_age + boost_age_range;
+      const uint64_t kBoostRatio = 16;
+      // prevent 0 value to avoid divide 0 error.
+      boost_step_ = std::max(boost_age_range / kBoostRatio, uint64_t{1});
+    }
+  }
+
+  uint64_t GetBoostScore(FileMetaData* f) {
+    if (!enabled_) {
+      return 1;
+    }
+    uint64_t oldest_ancester_time = f->TryGetOldestAncesterTime();
+    if (oldest_ancester_time >= current_time_) {
+      return 1;
+    }
+    uint64_t age = current_time_ - oldest_ancester_time;
+    if (age > boost_age_start_) {
+      // Use integer just for convenience.
+      // We could make all file_to_order double if we want.
+      // Technically this can overflow if users override timing and
+      // give a very high current time. Ignore the case for simplicity.
+      // Boosting is addition to current value, so +1. This will effectively
+      // make boosting to kick in after the first boost_step_ is reached.
+      return (age - boost_age_start_) / boost_step_ + 1;
+    }
+    return 1;
+  }
+
+ private:
+  bool enabled_;
+  uint64_t current_time_;
+  uint64_t boost_age_start_;
+  uint64_t boost_step_;
+};
+}  // namespace ROCKSDB_NAMESPACE

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

@@ -8,7 +8,24 @@
 
 #include <algorithm>
 
+#include "rocksdb/utilities/customizable_util.h"
+#include "rocksdb/utilities/object_registry.h"
+#include "rocksdb/utilities/options_type.h"
+
 namespace ROCKSDB_NAMESPACE {
+static std::unordered_map<std::string, OptionTypeInfo>
+    sst_fixed_prefix_type_info = {
+#ifndef ROCKSDB_LITE
+        {"length",
+         {0, OptionType::kSizeT, OptionVerificationType::kNormal,
+          OptionTypeFlags::kNone}},
+#endif  // ROCKSDB_LITE
+};
+
+SstPartitionerFixedPrefixFactory::SstPartitionerFixedPrefixFactory(size_t len)
+    : len_(len) {
+  RegisterOptions("Length", &len_, &sst_fixed_prefix_type_info);
+}
 
 PartitionerResult SstPartitionerFixedPrefix::ShouldPartition(
     const PartitionerRequest& request) {
@@ -41,4 +58,33 @@ std::shared_ptr<SstPartitionerFactory> NewSstPartitionerFixedPrefixFactory(
   return std::make_shared<SstPartitionerFixedPrefixFactory>(prefix_len);
 }
 
+#ifndef ROCKSDB_LITE
+namespace {
+static int RegisterSstPartitionerFactories(ObjectLibrary& library,
+                                           const std::string& /*arg*/) {
+  library.AddFactory<SstPartitionerFactory>(
+      SstPartitionerFixedPrefixFactory::kClassName(),
+      [](const std::string& /*uri*/,
+         std::unique_ptr<SstPartitionerFactory>* guard,
+         std::string* /* errmsg */) {
+        guard->reset(new SstPartitionerFixedPrefixFactory(0));
+        return guard->get();
+      });
+  return 1;
+}
+}  // namespace
+#endif  // ROCKSDB_LITE
+
+Status SstPartitionerFactory::CreateFromString(
+    const ConfigOptions& options, const std::string& value,
+    std::shared_ptr<SstPartitionerFactory>* result) {
+#ifndef ROCKSDB_LITE
+  static std::once_flag once;
+  std::call_once(once, [&]() {
+    RegisterSstPartitionerFactories(*(ObjectLibrary::Default().get()), "");
+  });
+#endif  // ROCKSDB_LITE
+  return LoadSharedObject<SstPartitionerFactory>(options, value, nullptr,
+                                                 result);
+}
 }  // namespace ROCKSDB_NAMESPACE

package/deps/rocksdb/rocksdb/db/comparator_db_test.cc

@@ -317,7 +317,7 @@ class ComparatorDBTest
 INSTANTIATE_TEST_CASE_P(FormatDef, ComparatorDBTest,
                         testing::Values(test::kDefaultFormatVersion));
 INSTANTIATE_TEST_CASE_P(FormatLatest, ComparatorDBTest,
-                        testing::Values(test::kLatestFormatVersion));
+                        testing::Values(kLatestFormatVersion));
 
 TEST_P(ComparatorDBTest, Bytewise) {
   for (int rand_seed = 301; rand_seed < 306; rand_seed++) {

package/deps/rocksdb/rocksdb/db/convenience.cc

@@ -44,7 +44,7 @@ Status VerifySstFileChecksum(const Options& options,
   std::unique_ptr<FSRandomAccessFile> file;
   uint64_t file_size;
   InternalKeyComparator internal_comparator(options.comparator);
-  ImmutableCFOptions ioptions(options);
+  ImmutableOptions ioptions(options);
 
   Status s = ioptions.fs->NewRandomAccessFile(file_path,
                                               FileOptions(env_options),
@@ -56,10 +56,13 @@ Status VerifySstFileChecksum(const Options& options,
   }
   std::unique_ptr<TableReader> table_reader;
   std::unique_ptr<RandomAccessFileReader> file_reader(
-      new RandomAccessFileReader(std::move(file), file_path));
+      new RandomAccessFileReader(
+          std::move(file), file_path, ioptions.clock, nullptr /* io_tracer */,
+          nullptr /* stats */, 0 /* hist_type */, nullptr /* file_read_hist */,
+          ioptions.rate_limiter.get()));
   const bool kImmortal = true;
   s = ioptions.table_factory->NewTableReader(
-      TableReaderOptions(ioptions, options.prefix_extractor.get(), env_options,
+      TableReaderOptions(ioptions, options.prefix_extractor, env_options,
                          internal_comparator, false /* skip_filters */,
                          !kImmortal, false /* force_direct_prefetch */,
                          -1 /* level */),

package/deps/rocksdb/rocksdb/db/corruption_test.cc

@@ -19,7 +19,6 @@
 #include "db/db_test_util.h"
 #include "db/log_format.h"
 #include "db/version_set.h"
-#include "env/composite_env_wrapper.h"
 #include "file/filename.h"
 #include "port/stack_trace.h"
 #include "rocksdb/cache.h"
@@ -27,6 +26,7 @@
 #include "rocksdb/db.h"
 #include "rocksdb/env.h"
 #include "rocksdb/table.h"
+#include "rocksdb/utilities/transaction_db.h"
 #include "rocksdb/write_batch.h"
 #include "table/block_based/block_based_table_builder.h"
 #include "table/meta_blocks.h"
@@ -40,11 +40,35 @@
 namespace ROCKSDB_NAMESPACE {
 
 static constexpr int kValueSize = 1000;
-
+namespace {
+// A wrapper that allows injection of errors.
+class ErrorEnv : public EnvWrapper {
+ public:
+  bool writable_file_error_;
+  int num_writable_file_errors_;
+
+  explicit ErrorEnv(Env* _target)
+      : EnvWrapper(_target),
+        writable_file_error_(false),
+        num_writable_file_errors_(0) {}
+  const char* Name() const override { return "ErrorEnv"; }
+
+  virtual Status NewWritableFile(const std::string& fname,
+                                 std::unique_ptr<WritableFile>* result,
+                                 const EnvOptions& soptions) override {
+    result->reset();
+    if (writable_file_error_) {
+      ++num_writable_file_errors_;
+      return Status::IOError(fname, "fake error");
+    }
+    return target()->NewWritableFile(fname, result, soptions);
+  }
+};
+}  // namespace
 class CorruptionTest : public testing::Test {
  public:
   std::shared_ptr<Env> env_guard_;
-  test::ErrorEnv* env_;
+  ErrorEnv* env_;
   std::string dbname_;
   std::shared_ptr<Cache> tiny_cache_;
   Options options_;
@@ -56,15 +80,10 @@ class CorruptionTest : public testing::Test {
     // bug in recovery code. Keep it 4 for now to make the test passes.
     tiny_cache_ = NewLRUCache(100, 4);
     Env* base_env = Env::Default();
-#ifndef ROCKSDB_LITE
-    const char* test_env_uri = getenv("TEST_ENV_URI");
-    if (test_env_uri) {
-      Status s = Env::LoadEnv(test_env_uri, &base_env, &env_guard_);
-      EXPECT_OK(s);
-      EXPECT_NE(Env::Default(), base_env);
-    }
-#endif  //! ROCKSDB_LITE
-    env_ = new test::ErrorEnv(base_env);
+    EXPECT_OK(
+        test::CreateEnvFromSystem(ConfigOptions(), &base_env, &env_guard_));
+    EXPECT_NE(base_env, nullptr);
+    env_ = new ErrorEnv(base_env);
     options_.wal_recovery_mode = WALRecoveryMode::kTolerateCorruptedTailRecords;
     options_.env = env_;
     dbname_ = test::PerThreadDBPath(env_, "corruption_test");
@@ -257,6 +276,42 @@ class CorruptionTest : public testing::Test {
     }
     return Slice(*storage);
   }
+
+  void GetSortedWalFiles(std::vector<uint64_t>& file_nums) {
+    std::vector<std::string> tmp_files;
+    ASSERT_OK(env_->GetChildren(dbname_, &tmp_files));
+    FileType type = kWalFile;
+    for (const auto& file : tmp_files) {
+      uint64_t number = 0;
+      if (ParseFileName(file, &number, &type) && type == kWalFile) {
+        file_nums.push_back(number);
+      }
+    }
+    std::sort(file_nums.begin(), file_nums.end());
+  }
+
+  void CorruptFileWithTruncation(FileType file, uint64_t number,
+                                 uint64_t bytes_to_truncate = 0) {
+    std::string path;
+    switch (file) {
+      case FileType::kWalFile:
+        path = LogFileName(dbname_, number);
+        break;
+      // TODO: Add other file types as this method is being used for those file
+      // types.
+      default:
+        return;
+    }
+    uint64_t old_size = 0;
+    ASSERT_OK(env_->GetFileSize(path, &old_size));
+    assert(old_size > bytes_to_truncate);
+    uint64_t new_size = old_size - bytes_to_truncate;
+    // If bytes_to_truncate == 0, it will do full truncation.
+    if (bytes_to_truncate == 0) {
+      new_size = old_size;
+    }
+    ASSERT_OK(test::TruncateFile(env_, path, new_size));
+  }
 };
 
 TEST_F(CorruptionTest, Recovery) {
@@ -539,19 +594,20 @@ TEST_F(CorruptionTest, RangeDeletionCorrupted) {
   ASSERT_EQ(static_cast<size_t>(1), metadata.size());
   std::string filename = dbname_ + metadata[0].name;
 
-  std::unique_ptr<RandomAccessFile> file;
-  ASSERT_OK(options_.env->NewRandomAccessFile(filename, &file, EnvOptions()));
-  std::unique_ptr<RandomAccessFileReader> file_reader(
-      new RandomAccessFileReader(NewLegacyRandomAccessFileWrapper(file),
-                                 filename));
+  FileOptions file_opts;
+  const auto& fs = options_.env->GetFileSystem();
+  std::unique_ptr<RandomAccessFileReader> file_reader;
+  ASSERT_OK(RandomAccessFileReader::Create(fs, filename, file_opts,
+                                           &file_reader, nullptr));
 
   uint64_t file_size;
-  ASSERT_OK(options_.env->GetFileSize(filename, &file_size));
+  ASSERT_OK(
+      fs->GetFileSize(filename, file_opts.io_options, &file_size, nullptr));
 
   BlockHandle range_del_handle;
-  ASSERT_OK(FindMetaBlock(
+  ASSERT_OK(FindMetaBlockInFile(
       file_reader.get(), file_size, kBlockBasedTableMagicNumber,
-      ImmutableCFOptions(options_), kRangeDelBlock, &range_del_handle));
+      ImmutableOptions(options_), kRangeDelBlockName, &range_del_handle));
 
   ASSERT_OK(TryReopen());
   ASSERT_OK(test::CorruptFile(env_, filename,
@@ -882,7 +938,7 @@ TEST_F(CorruptionTest, VerifyWholeTableChecksum) {
   SyncPoint::GetInstance()->ClearAllCallBacks();
   int count{0};
   SyncPoint::GetInstance()->SetCallBack(
-      "DBImpl::VerifySstFileChecksum:mismatch", [&](void* arg) {
+      "DBImpl::VerifyFullFileChecksum:mismatch", [&](void* arg) {
         auto* s = reinterpret_cast<Status*>(arg);
         ASSERT_NE(s, nullptr);
         ++count;
@@ -893,15 +949,314 @@ TEST_F(CorruptionTest, VerifyWholeTableChecksum) {
   ASSERT_EQ(1, count);
 }
 
-}  // namespace ROCKSDB_NAMESPACE
+class CrashDuringRecoveryWithCorruptionTest
+    : public CorruptionTest,
+      public testing::WithParamInterface<std::tuple<bool, bool>> {
+ public:
+  explicit CrashDuringRecoveryWithCorruptionTest()
+      : CorruptionTest(),
+        avoid_flush_during_recovery_(std::get<0>(GetParam())),
+        track_and_verify_wals_in_manifest_(std::get<1>(GetParam())) {}
+
+ protected:
+  const bool avoid_flush_during_recovery_;
+  const bool track_and_verify_wals_in_manifest_;
+};
+
+INSTANTIATE_TEST_CASE_P(CorruptionTest, CrashDuringRecoveryWithCorruptionTest,
+                        ::testing::Values(std::make_tuple(true, false),
+                                          std::make_tuple(false, false),
+                                          std::make_tuple(true, true),
+                                          std::make_tuple(false, true)));
+
+// In case of non-TransactionDB with avoid_flush_during_recovery = true, RocksDB
+// won't flush the data from WAL to L0 for all column families if possible. As a
+// result, not all column families can increase their log_numbers, and
+// min_log_number_to_keep won't change.
+// It may prematurely persist a new MANIFEST even before we can declare the DB
+// is in consistent state after recovery (this is when the new WAL is synced)
+// and advances log_numbers for some column families.
+//
+// If there is power failure before we sync the new WAL, we will end up in
+// a situation in which after persisting the MANIFEST, RocksDB will see some
+// column families' log_numbers larger than the corrupted wal, and
+// "Column family inconsistency: SST file contains data beyond the point of
+// corruption" error will be hit, causing recovery to fail.
+//
+// After adding the fix, corrupted WALs whose numbers are larger than the
+// corrupted wal and smaller than the new WAL are moved to a separate folder.
+// Only after new WAL is synced, RocksDB persist a new MANIFEST with column
+// families to ensure RocksDB is in consistent state.
+// RocksDB writes an empty WriteBatch as a sentinel to the new WAL which is
+// synced immediately afterwards. The sequence number of the sentinel
+// WriteBatch will be the next sequence number immediately after the largest
+// sequence number recovered from previous WALs and MANIFEST because of which DB
+// will be in consistent state.
+TEST_P(CrashDuringRecoveryWithCorruptionTest, CrashDuringRecovery) {
+  CloseDb();
+  Options options;
+  options.track_and_verify_wals_in_manifest =
+      track_and_verify_wals_in_manifest_;
+  options.wal_recovery_mode = WALRecoveryMode::kPointInTimeRecovery;
+  options.avoid_flush_during_recovery = false;
+  options.env = env_;
+  ASSERT_OK(DestroyDB(dbname_, options));
+  options.create_if_missing = true;
+  options.max_write_buffer_number = 3;
+
+  Reopen(&options);
+  Status s;
+  const std::string test_cf_name = "test_cf";
+  ColumnFamilyHandle* cfh = nullptr;
+  s = db_->CreateColumnFamily(options, test_cf_name, &cfh);
+  ASSERT_OK(s);
+  delete cfh;
+  CloseDb();
+
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  cf_descs.emplace_back(kDefaultColumnFamilyName, options);
+  cf_descs.emplace_back(test_cf_name, options);
+  std::vector<ColumnFamilyHandle*> handles;
+
+  // 1. Open and populate the DB. Write and flush default_cf several times to
+  // advance wal number so that some column families have advanced log_number
+  // while other don't.
+  {
+    ASSERT_OK(DB::Open(options, dbname_, cf_descs, &handles, &db_));
+    auto* dbimpl = static_cast_with_check<DBImpl>(db_);
+    assert(dbimpl);
+
+    // Write one key to test_cf.
+    ASSERT_OK(db_->Put(WriteOptions(), handles[1], "old_key", "dontcare"));
+    // Write to default_cf and flush this cf several times to advance wal
+    // number.
+    for (int i = 0; i < 2; ++i) {
+      ASSERT_OK(db_->Put(WriteOptions(), "key" + std::to_string(i), "value"));
+      ASSERT_OK(dbimpl->TEST_SwitchMemtable());
+    }
+    ASSERT_OK(db_->Put(WriteOptions(), handles[1], "dontcare", "dontcare"));
+
+    for (auto* h : handles) {
+      delete h;
+    }
+    handles.clear();
+    CloseDb();
+  }
+
+  // 2. Corrupt second last wal file to emulate power reset which caused the DB
+  // to lose the un-synced WAL.
+  {
+    std::vector<uint64_t> file_nums;
+    GetSortedWalFiles(file_nums);
+    size_t size = file_nums.size();
+    uint64_t log_num = file_nums[size - 2];
+    CorruptFileWithTruncation(FileType::kWalFile, log_num,
+                              /*bytes_to_truncate=*/8);
+  }
 
-#ifdef ROCKSDB_UNITTESTS_WITH_CUSTOM_OBJECTS_FROM_STATIC_LIBS
-extern "C" {
-void RegisterCustomObjects(int argc, char** argv);
+  // 3. After first crash reopen the DB which contains corrupted WAL. Default
+  // family has higher log number than corrupted wal number.
+  //
+  // Case1: If avoid_flush_during_recovery = true, RocksDB won't flush the data
+  // from WAL to L0 for all column families (test_cf_name in this case). As a
+  // result, not all column families can increase their log_numbers, and
+  // min_log_number_to_keep won't change.
+  //
+  // Case2: If avoid_flush_during_recovery = false, all column families have
+  // flushed their data from WAL to L0 during recovery, and none of them will
+  // ever need to read the WALs again.
+  {
+    options.avoid_flush_during_recovery = avoid_flush_during_recovery_;
+    s = DB::Open(options, dbname_, cf_descs, &handles, &db_);
+    ASSERT_OK(s);
+    for (auto* h : handles) {
+      delete h;
+    }
+    handles.clear();
+    CloseDb();
+  }
+
+  // 4. Corrupt max_wal_num to emulate second power reset which caused the
+  // DB to again lose the un-synced WAL.
+  {
+    std::vector<uint64_t> file_nums;
+    GetSortedWalFiles(file_nums);
+    size_t size = file_nums.size();
+    uint64_t log_num = file_nums[size - 1];
+    CorruptFileWithTruncation(FileType::kWalFile, log_num);
+  }
+
+  // 5. After second crash reopen the db with second corruption. Default family
+  // has higher log number than corrupted wal number.
+  //
+  // Case1: If avoid_flush_during_recovery = true, we persist a new
+  // MANIFEST with advanced log_numbers for some column families only after
+  // syncing the WAL. So during second crash, RocksDB will skip the corrupted
+  // WAL files as they have been moved to different folder. Since newly synced
+  // WAL file's sequence number (sentinel WriteBatch) will be the next
+  // sequence number immediately after the largest sequence number recovered
+  // from previous WALs and MANIFEST, db will be in consistent state and opens
+  // successfully.
+  //
+  // Case2: If avoid_flush_during_recovery = false, the corrupted WAL is below
+  // this number. So during a second crash after persisting the new MANIFEST,
+  // RocksDB will skip the corrupted WAL(s) because they are all below this
+  // bound. Therefore, we won't hit the "column family inconsistency" error
+  // message.
+  {
+    options.avoid_flush_during_recovery = avoid_flush_during_recovery_;
+    ASSERT_OK(DB::Open(options, dbname_, cf_descs, &handles, &db_));
+    for (auto* h : handles) {
+      delete h;
+    }
+    handles.clear();
+    CloseDb();
+  }
 }
-#else
-void RegisterCustomObjects(int /*argc*/, char** /*argv*/) {}
-#endif  // !ROCKSDB_UNITTESTS_WITH_CUSTOM_OBJECTS_FROM_STATIC_LIBS
+
+// In case of TransactionDB, it enables two-phase-commit. The prepare section of
+// an uncommitted transaction always need to be kept. Even if we perform flush
+// during recovery, we may still need to hold an old WAL. The
+// min_log_number_to_keep won't change, and "Column family inconsistency: SST
+// file contains data beyond the point of corruption" error will be hit, causing
+// recovery to fail.
+//
+// After adding the fix, corrupted WALs whose numbers are larger than the
+// corrupted wal and smaller than the new WAL are moved to a separate folder.
+// Only after new WAL is synced, RocksDB persist a new MANIFEST with column
+// families to ensure RocksDB is in consistent state.
+// RocksDB writes an empty WriteBatch as a sentinel to the new WAL which is
+// synced immediately afterwards. The sequence number of the sentinel
+// WriteBatch will be the next sequence number immediately after the largest
+// sequence number recovered from previous WALs and MANIFEST because of which DB
+// will be in consistent state.
+TEST_P(CrashDuringRecoveryWithCorruptionTest, TxnDbCrashDuringRecovery) {
+  CloseDb();
+  Options options;
+  options.wal_recovery_mode = WALRecoveryMode::kPointInTimeRecovery;
+  options.track_and_verify_wals_in_manifest =
+      track_and_verify_wals_in_manifest_;
+  options.avoid_flush_during_recovery = false;
+  options.env = env_;
+  ASSERT_OK(DestroyDB(dbname_, options));
+  options.create_if_missing = true;
+  options.max_write_buffer_number = 3;
+  Reopen(&options);
+
+  // Create cf test_cf_name.
+  ColumnFamilyHandle* cfh = nullptr;
+  const std::string test_cf_name = "test_cf";
+  Status s = db_->CreateColumnFamily(options, test_cf_name, &cfh);
+  ASSERT_OK(s);
+  delete cfh;
+  CloseDb();
+
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  cf_descs.emplace_back(kDefaultColumnFamilyName, options);
+  cf_descs.emplace_back(test_cf_name, options);
+  std::vector<ColumnFamilyHandle*> handles;
+
+  TransactionDB* txn_db = nullptr;
+  TransactionDBOptions txn_db_opts;
+
+  // 1. Open and populate the DB. Write and flush default_cf several times to
+  // advance wal number so that some column families have advanced log_number
+  // while other don't.
+  {
+    options.avoid_flush_during_recovery = avoid_flush_during_recovery_;
+    ASSERT_OK(TransactionDB::Open(options, txn_db_opts, dbname_, cf_descs,
+                                  &handles, &txn_db));
+
+    auto* txn = txn_db->BeginTransaction(WriteOptions(), TransactionOptions());
+    // Put cf1
+    ASSERT_OK(txn->Put(handles[1], "foo", "value"));
+    ASSERT_OK(txn->SetName("txn0"));
+    ASSERT_OK(txn->Prepare());
+    delete txn;
+    txn = nullptr;
+
+    auto* dbimpl = static_cast_with_check<DBImpl>(txn_db->GetRootDB());
+    assert(dbimpl);
+
+    // Put and flush cf0
+    for (int i = 0; i < 2; ++i) {
+      ASSERT_OK(txn_db->Put(WriteOptions(), "dontcare", "value"));
+      ASSERT_OK(dbimpl->TEST_SwitchMemtable());
+    }
+
+    // Put cf1
+    txn = txn_db->BeginTransaction(WriteOptions(), TransactionOptions());
+    ASSERT_OK(txn->Put(handles[1], "foo1", "value"));
+    ASSERT_OK(txn->Commit());
+
+    delete txn;
+    txn = nullptr;
+    for (auto* h : handles) {
+      delete h;
+    }
+    handles.clear();
+    delete txn_db;
+  }
+
+  // 2. Corrupt second last wal to emulate power reset which caused the DB to
+  // lose the un-synced WAL.
+  {
+    std::vector<uint64_t> file_nums;
+    GetSortedWalFiles(file_nums);
+    size_t size = file_nums.size();
+    uint64_t log_num = file_nums[size - 2];
+    CorruptFileWithTruncation(FileType::kWalFile, log_num,
+                              /*bytes_to_truncate=*/8);
+  }
+
+  // 3. After first crash reopen the DB which contains corrupted WAL. Default
+  // family has higher log number than corrupted wal number. There may be old
+  // WAL files that it must not delete because they can contain data of
+  // uncommitted transactions. As a result, min_log_number_to_keep won't change.
+  {
+    options.avoid_flush_during_recovery = avoid_flush_during_recovery_;
+    ASSERT_OK(TransactionDB::Open(options, txn_db_opts, dbname_, cf_descs,
+                                  &handles, &txn_db));
+
+    for (auto* h : handles) {
+      delete h;
+    }
+    handles.clear();
+    delete txn_db;
+  }
+
+  // 4. Corrupt max_wal_num to emulate second power reset which caused the
+  // DB to again lose the un-synced WAL.
+  {
+    std::vector<uint64_t> file_nums;
+    GetSortedWalFiles(file_nums);
+    size_t size = file_nums.size();
+    uint64_t log_num = file_nums[size - 1];
+    CorruptFileWithTruncation(FileType::kWalFile, log_num);
+  }
+
+  // 5. After second crash reopen the db with second corruption. Default family
+  // has higher log number than corrupted wal number.
+  // We persist a new MANIFEST with advanced log_numbers for some column
+  // families only after syncing the WAL. So during second crash, RocksDB will
+  // skip the corrupted WAL files as they have been moved to different folder.
+  // Since newly synced WAL file's sequence number (sentinel WriteBatch) will be
+  // the next sequence number immediately after the largest sequence number
+  // recovered from previous WALs and MANIFEST, db will be in consistent state
+  // and opens successfully.
+  {
+    options.avoid_flush_during_recovery = false;
+
+    ASSERT_OK(TransactionDB::Open(options, txn_db_opts, dbname_, cf_descs,
+                                  &handles, &txn_db));
+    for (auto* h : handles) {
+      delete h;
+    }
+    delete txn_db;
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
 
 int main(int argc, char** argv) {
   ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();