Enable atomic flush (#4023)

Summary:
Adds a DB option `atomic_flush` to control whether to enable this feature. This PR is a subset of [PR 3752](https://github.com/facebook/rocksdb/pull/3752).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4023

Differential Revision: D8518381

Pulled By: riversand963

fbshipit-source-id: 1e3bb33e99bb102876a31b378d93b0138ff6634f

HISTORY.md

@@ -4,6 +4,7 @@
 * Introduced `Memoryllocator`, which lets the user specify custom allocator for memory in block cache.
 * Introduced `PerfContextByLevel` as part of `PerfContext` which allows storing perf context at each level. Also replaced `__thread` with `thread_local` keyword for perf_context.
 * With level_compaction_dynamic_level_bytes = true, level multiplier may be adjusted automatically when Level 0 to 1 compaction is lagged behind.
+* Introduced DB option `atomic_flush`. If true, RocksDB supports flushing multiple column families and atomically committing the result to MANIFEST. Useful when WAL is disabled.
 
 ### Bug Fixes
 * Fix corner case where a write group leader blocked due to write stall blocks other writers in queue with WriteOptions::no_slowdown set.

db/db_filesnapshot.cc

@@ -87,19 +87,28 @@ Status DBImpl::GetLiveFiles(std::vector<std::string>& ret,
   if (flush_memtable) {
     // flush all dirty data to disk.
     Status status;
-    for (auto cfd : *versions_->GetColumnFamilySet()) {
-      if (cfd->IsDropped()) {
-        continue;
-      }
-      cfd->Ref();
+    if (atomic_flush_) {
+      autovector<ColumnFamilyData*> cfds;
+      SelectColumnFamiliesForAtomicFlush(&cfds);
       mutex_.Unlock();
-      status = FlushMemTable(cfd, FlushOptions(), FlushReason::kGetLiveFiles);
-      TEST_SYNC_POINT("DBImpl::GetLiveFiles:1");
-      TEST_SYNC_POINT("DBImpl::GetLiveFiles:2");
+      status = AtomicFlushMemTables(cfds, FlushOptions(),
+                                    FlushReason::kGetLiveFiles);
       mutex_.Lock();
-      cfd->Unref();
-      if (!status.ok()) {
-        break;
+    } else {
+      for (auto cfd : *versions_->GetColumnFamilySet()) {
+        if (cfd->IsDropped()) {
+          continue;
+        }
+        cfd->Ref();
+        mutex_.Unlock();
+        status = FlushMemTable(cfd, FlushOptions(), FlushReason::kGetLiveFiles);
+        TEST_SYNC_POINT("DBImpl::GetLiveFiles:1");
+        TEST_SYNC_POINT("DBImpl::GetLiveFiles:2");
+        mutex_.Lock();
+        cfd->Unref();
+        if (!status.ok()) {
+          break;
+        }
       }
     }
     versions_->GetColumnFamilySet()->FreeDeadColumnFamilies();

db/db_flush_test.cc

@@ -25,6 +25,12 @@ class DBFlushDirectIOTest : public DBFlushTest,
   DBFlushDirectIOTest() : DBFlushTest() {}
 };
 
+class DBAtomicFlushTest : public DBFlushTest,
+                          public ::testing::WithParamInterface<bool> {
+ public:
+  DBAtomicFlushTest() : DBFlushTest() {}
+};
+
 // We had issue when two background threads trying to flush at the same time,
 // only one of them get committed. The test verifies the issue is fixed.
 TEST_F(DBFlushTest, FlushWhileWritingManifest) {
@@ -214,9 +220,82 @@ TEST_F(DBFlushTest, FlushError) {
   ASSERT_NE(s, Status::OK());
 }
 
+TEST_P(DBAtomicFlushTest, ManualAtomicFlush) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.atomic_flush = GetParam();
+  options.write_buffer_size = (static_cast<size_t>(64) << 20);
+
+  CreateAndReopenWithCF({"pikachu", "eevee"}, options);
+  size_t num_cfs = handles_.size();
+  ASSERT_EQ(3, num_cfs);
+  WriteOptions wopts;
+  wopts.disableWAL = true;
+  for (size_t i = 0; i != num_cfs; ++i) {
+    ASSERT_OK(Put(static_cast<int>(i) /*cf*/, "key", "value", wopts));
+  }
+  std::vector<int> cf_ids;
+  for (size_t i = 0; i != num_cfs; ++i) {
+    cf_ids.emplace_back(static_cast<int>(i));
+  }
+  ASSERT_OK(Flush(cf_ids));
+  for (size_t i = 0; i != num_cfs; ++i) {
+    auto cfh = static_cast<ColumnFamilyHandleImpl*>(handles_[i]);
+    ASSERT_EQ(0, cfh->cfd()->imm()->NumNotFlushed());
+    ASSERT_TRUE(cfh->cfd()->mem()->IsEmpty());
+  }
+}
+
+TEST_P(DBAtomicFlushTest, AtomicFlushTriggeredByMemTableFull) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.atomic_flush = GetParam();
+  // 4KB so that we can easily trigger auto flush.
+  options.write_buffer_size = 4096;
+
+  SyncPoint::GetInstance()->LoadDependency(
+      {{"DBImpl::BackgroundCallFlush:FlushFinish:0",
+        "DBAtomicFlushTest::AtomicFlushTriggeredByMemTableFull:BeforeCheck"}});
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  CreateAndReopenWithCF({"pikachu", "eevee"}, options);
+  size_t num_cfs = handles_.size();
+  ASSERT_EQ(3, num_cfs);
+  WriteOptions wopts;
+  wopts.disableWAL = true;
+  for (size_t i = 0; i != num_cfs; ++i) {
+    ASSERT_OK(Put(static_cast<int>(i) /*cf*/, "key", "value", wopts));
+  }
+  // Keep writing to one of them column families to trigger auto flush.
+  for (int i = 0; i != 4000; ++i) {
+    ASSERT_OK(Put(static_cast<int>(num_cfs) - 1 /*cf*/,
+                  "key" + std::to_string(i), "value" + std::to_string(i),
+                  wopts));
+  }
+
+  TEST_SYNC_POINT(
+      "DBAtomicFlushTest::AtomicFlushTriggeredByMemTableFull:BeforeCheck");
+  if (options.atomic_flush) {
+    for (size_t i = 0; i != num_cfs - 1; ++i) {
+      auto cfh = static_cast<ColumnFamilyHandleImpl*>(handles_[i]);
+      ASSERT_EQ(0, cfh->cfd()->imm()->NumNotFlushed());
+      ASSERT_TRUE(cfh->cfd()->mem()->IsEmpty());
+    }
+  } else {
+    for (size_t i = 0; i != num_cfs - 1; ++i) {
+      auto cfh = static_cast<ColumnFamilyHandleImpl*>(handles_[i]);
+      ASSERT_EQ(0, cfh->cfd()->imm()->NumNotFlushed());
+      ASSERT_FALSE(cfh->cfd()->mem()->IsEmpty());
+    }
+  }
+  SyncPoint::GetInstance()->DisableProcessing();
+}
+
 INSTANTIATE_TEST_CASE_P(DBFlushDirectIOTest, DBFlushDirectIOTest,
                         testing::Bool());
 
+INSTANTIATE_TEST_CASE_P(DBAtomicFlushTest, DBAtomicFlushTest, testing::Bool());
+
 }  // namespace rocksdb
 
 int main(int argc, char** argv) {

db/db_impl.cc

@@ -220,6 +220,7 @@ DBImpl::DBImpl(const DBOptions& options, const std::string& dbname,
       preserve_deletes_(options.preserve_deletes),
       closed_(false),
       error_handler_(this, immutable_db_options_, &mutex_),
+      atomic_flush_(options.atomic_flush),
       atomic_flush_commit_in_progress_(false) {
   // !batch_per_trx_ implies seq_per_batch_ because it is only unset for
   // WriteUnprepared, which should use seq_per_batch_.
@@ -305,7 +306,30 @@ Status DBImpl::ResumeImpl() {
   // We cannot guarantee consistency of the WAL. So force flush Memtables of
   // all the column families
   if (s.ok()) {
-    s = FlushAllCFs(FlushReason::kErrorRecovery);
+    FlushOptions flush_opts;
+    // We allow flush to stall write since we are trying to resume from error.
+    flush_opts.allow_write_stall = true;
+    if (atomic_flush_) {
+      autovector<ColumnFamilyData*> cfds;
+      SelectColumnFamiliesForAtomicFlush(&cfds);
+      mutex_.Unlock();
+      s = AtomicFlushMemTables(cfds, flush_opts, FlushReason::kErrorRecovery);
+      mutex_.Lock();
+    } else {
+      for (auto cfd : *versions_->GetColumnFamilySet()) {
+        if (cfd->IsDropped()) {
+          continue;
+        }
+        cfd->Ref();
+        mutex_.Unlock();
+        s = FlushMemTable(cfd, flush_opts, FlushReason::kErrorRecovery);
+        mutex_.Lock();
+        cfd->Unref();
+        if (!s.ok()) {
+          break;
+        }
+      }
+    }
     if (!s.ok()) {
       ROCKS_LOG_INFO(immutable_db_options_.info_log,
                      "DB resume requested but failed due to Flush failure [%s]",
@@ -377,13 +401,21 @@ void DBImpl::CancelAllBackgroundWork(bool wait) {
   if (!shutting_down_.load(std::memory_order_acquire) &&
       has_unpersisted_data_.load(std::memory_order_relaxed) &&
       !mutable_db_options_.avoid_flush_during_shutdown) {
-    for (auto cfd : *versions_->GetColumnFamilySet()) {
-      if (!cfd->IsDropped() && cfd->initialized() && !cfd->mem()->IsEmpty()) {
-        cfd->Ref();
-        mutex_.Unlock();
-        FlushMemTable(cfd, FlushOptions(), FlushReason::kShutDown);
-        mutex_.Lock();
-        cfd->Unref();
+    if (atomic_flush_) {
+      autovector<ColumnFamilyData*> cfds;
+      SelectColumnFamiliesForAtomicFlush(&cfds);
+      mutex_.Unlock();
+      AtomicFlushMemTables(cfds, FlushOptions(), FlushReason::kShutDown);
+      mutex_.Lock();
+    } else {
+      for (auto cfd : *versions_->GetColumnFamilySet()) {
+        if (!cfd->IsDropped() && cfd->initialized() && !cfd->mem()->IsEmpty()) {
+          cfd->Ref();
+          mutex_.Unlock();
+          FlushMemTable(cfd, FlushOptions(), FlushReason::kShutDown);
+          mutex_.Lock();
+          cfd->Unref();
+        }
       }
     }
     versions_->GetColumnFamilySet()->FreeDeadColumnFamilies();
@@ -3099,11 +3131,21 @@ Status DBImpl::IngestExternalFile(
       TEST_SYNC_POINT_CALLBACK("DBImpl::IngestExternalFile:NeedFlush",
                                &need_flush);
       if (status.ok() && need_flush) {
-        mutex_.Unlock();
-        status = FlushMemTable(cfd, FlushOptions(),
-                               FlushReason::kExternalFileIngestion,
-                               true /* writes_stopped */);
-        mutex_.Lock();
+        if (atomic_flush_) {
+          mutex_.Unlock();
+          autovector<ColumnFamilyData*> cfds;
+          SelectColumnFamiliesForAtomicFlush(&cfds);
+          status = AtomicFlushMemTables(cfds, FlushOptions(),
+                                        FlushReason::kExternalFileIngestion,
+                                        true /* writes_stopped */);
+          mutex_.Lock();
+        } else {
+          mutex_.Unlock();
+          status = FlushMemTable(cfd, FlushOptions(),
+                                 FlushReason::kExternalFileIngestion,
+                                 true /* writes_stopped */);
+          mutex_.Lock();
+        }
       }
     }
 

db/db_impl.h

@@ -228,6 +228,9 @@ class DBImpl : public DB {
   using DB::Flush;
   virtual Status Flush(const FlushOptions& options,
                        ColumnFamilyHandle* column_family) override;
+  virtual Status Flush(
+      const FlushOptions& options,
+      const std::vector<ColumnFamilyHandle*>& column_families) override;
   virtual Status FlushWAL(bool sync) override;
   bool TEST_WALBufferIsEmpty();
   virtual Status SyncWAL() override;
@@ -965,10 +968,17 @@ class DBImpl : public DB {
 
   Status SwitchMemtable(ColumnFamilyData* cfd, WriteContext* context);
 
+  void SelectColumnFamiliesForAtomicFlush(autovector<ColumnFamilyData*>* cfds);
+
   // Force current memtable contents to be flushed.
   Status FlushMemTable(ColumnFamilyData* cfd, const FlushOptions& options,
                        FlushReason flush_reason, bool writes_stopped = false);
 
+  Status AtomicFlushMemTables(
+      const autovector<ColumnFamilyData*>& column_family_datas,
+      const FlushOptions& options, FlushReason flush_reason,
+      bool writes_stopped = false);
+
   // Wait until flushing this column family won't stall writes
   Status WaitUntilFlushWouldNotStallWrites(ColumnFamilyData* cfd,
                                            bool* flush_needed);
@@ -977,14 +987,22 @@ class DBImpl : public DB {
   // If flush_memtable_id is non-null, wait until the memtable with the ID
   // gets flush. Otherwise, wait until the column family don't have any
   // memtable pending flush.
+  // resuming_from_bg_err indicates whether the caller is attempting to resume
+  // from background error.
   Status WaitForFlushMemTable(ColumnFamilyData* cfd,
-                              const uint64_t* flush_memtable_id = nullptr) {
-    return WaitForFlushMemTables({cfd}, {flush_memtable_id});
+                              const uint64_t* flush_memtable_id = nullptr,
+                              bool resuming_from_bg_err = false) {
+    return WaitForFlushMemTables({cfd}, {flush_memtable_id},
+                                 resuming_from_bg_err);
   }
   // Wait for memtables to be flushed for multiple column families.
   Status WaitForFlushMemTables(
       const autovector<ColumnFamilyData*>& cfds,
-      const autovector<const uint64_t*>& flush_memtable_ids);
+      const autovector<const uint64_t*>& flush_memtable_ids,
+      bool resuming_from_bg_err);
+
+  // REQUIRES: mutex locked and in write thread.
+  void AssignAtomicFlushSeq(const autovector<ColumnFamilyData*>& cfds);
 
   // REQUIRES: mutex locked
   Status SwitchWAL(WriteContext* write_context);
@@ -1049,6 +1067,9 @@ class DBImpl : public DB {
   // column families in this request, this flush is considered complete.
   typedef std::vector<std::pair<ColumnFamilyData*, uint64_t>> FlushRequest;
 
+  void GenerateFlushRequest(const autovector<ColumnFamilyData*>& cfds,
+                            FlushRequest* req);
+
   void SchedulePendingFlush(const FlushRequest& req, FlushReason flush_reason);
 
   void SchedulePendingCompaction(ColumnFamilyData* cfd);
@@ -1109,8 +1130,6 @@ class DBImpl : public DB {
 
   Status CloseHelper();
 
-  Status FlushAllCFs(FlushReason flush_reason);
-
   void WaitForBackgroundWork();
 
   // table_cache_ provides its own synchronization
@@ -1584,6 +1603,9 @@ class DBImpl : public DB {
 
   ErrorHandler error_handler_;
 
+  // True if DB enables atomic flush.
+  bool atomic_flush_;
+
   // True if the DB is committing atomic flush.
   // TODO (yanqin) the current impl assumes that the entire DB belongs to
   // a single atomic flush group. In the future we need to add a new class

db/db_impl_compaction_flush.cc

@@ -211,6 +211,10 @@ Status DBImpl::FlushMemTableToOutputFile(
 Status DBImpl::FlushMemTablesToOutputFiles(
     const autovector<BGFlushArg>& bg_flush_args, bool* made_progress,
     JobContext* job_context, LogBuffer* log_buffer) {
+  if (atomic_flush_) {
+    return AtomicFlushMemTablesToOutputFiles(bg_flush_args, made_progress,
+                                             job_context, log_buffer);
+  }
   Status s;
   for (auto& arg : bg_flush_args) {
     ColumnFamilyData* cfd = arg.cfd_;
@@ -318,11 +322,23 @@ Status DBImpl::AtomicFlushMemTablesToOutputFiles(
     s = SyncClosedLogs(job_context);
   }
 
+  // exec_status stores the execution status of flush_jobs as
+  // <bool /* executed */, Status /* status code */>
+  autovector<std::pair<bool, Status>> exec_status;
+  for (int i = 0; i != num_cfs; ++i) {
+    // Initially all jobs are not executed, with status OK.
+    std::pair<bool, Status> elem(false, Status::OK());
+    exec_status.emplace_back(elem);
+  }
+
   if (s.ok()) {
     // TODO (yanqin): parallelize jobs with threads.
     for (int i = 0; i != num_cfs; ++i) {
-      s = jobs[i].Run(&logs_with_prep_tracker_, &file_meta[i]);
-      if (!s.ok()) {
+      exec_status[i].second =
+          jobs[i].Run(&logs_with_prep_tracker_, &file_meta[i]);
+      exec_status[i].first = true;
+      if (!exec_status[i].second.ok()) {
+        s = exec_status[i].second;
         break;
       }
     }
@@ -401,12 +417,19 @@ Status DBImpl::AtomicFlushMemTablesToOutputFiles(
   }
 
   if (!s.ok()) {
+    // Have to cancel the flush jobs that have NOT executed because we need to
+    // unref the versions.
     for (int i = 0; i != num_cfs; ++i) {
-      auto& mems = jobs[i].GetMemTables();
-      cfds[i]->imm()->RollbackMemtableFlush(mems, file_meta[i].fd.GetNumber());
-      jobs[i].Cancel();
+      if (!exec_status[i].first) {
+        jobs[i].Cancel();
+      }
     }
     if (!s.IsShutdownInProgress()) {
+      for (int i = 0; i != num_cfs; ++i) {
+        auto& mems = jobs[i].GetMemTables();
+        cfds[i]->imm()->RollbackMemtableFlush(mems,
+                                              file_meta[i].fd.GetNumber());
+      }
       Status new_bg_error = s;
       error_handler_.SetBGError(new_bg_error, BackgroundErrorReason::kFlush);
     }
@@ -543,8 +566,15 @@ Status DBImpl::CompactRange(const CompactRangeOptions& options,
   if (flush_needed) {
     FlushOptions fo;
     fo.allow_write_stall = options.allow_write_stall;
-    s = FlushMemTable(cfd, fo, FlushReason::kManualCompaction,
-      false /* writes_stopped*/);
+    if (atomic_flush_) {
+      autovector<ColumnFamilyData*> cfds;
+      SelectColumnFamiliesForAtomicFlush(&cfds);
+      s = AtomicFlushMemTables(cfds, fo, FlushReason::kManualCompaction,
+                               false /* writes_stopped */);
+    } else {
+      s = FlushMemTable(cfd, fo, FlushReason::kManualCompaction,
+                        false /* writes_stopped*/);
+    }
     if (!s.ok()) {
       LogFlush(immutable_db_options_.info_log);
       return s;
@@ -1165,72 +1195,59 @@ Status DBImpl::Flush(const FlushOptions& flush_options,
   auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
   ROCKS_LOG_INFO(immutable_db_options_.info_log, "[%s] Manual flush start.",
                  cfh->GetName().c_str());
-  Status s =
-      FlushMemTable(cfh->cfd(), flush_options, FlushReason::kManualFlush);
+  Status s;
+  if (atomic_flush_) {
+    s = AtomicFlushMemTables({cfh->cfd()}, flush_options,
+                             FlushReason::kManualFlush);
+  } else {
+    s = FlushMemTable(cfh->cfd(), flush_options, FlushReason::kManualFlush);
+  }
+
   ROCKS_LOG_INFO(immutable_db_options_.info_log,
                  "[%s] Manual flush finished, status: %s\n",
                  cfh->GetName().c_str(), s.ToString().c_str());
   return s;
 }
 
-Status DBImpl::FlushAllCFs(FlushReason flush_reason) {
+Status DBImpl::Flush(const FlushOptions& flush_options,
+                     const std::vector<ColumnFamilyHandle*>& column_families) {
   Status s;
-  WriteContext context;
-  WriteThread::Writer w;
-
-  mutex_.AssertHeld();
-  write_thread_.EnterUnbatched(&w, &mutex_);
-
-  FlushRequest flush_req;
-  for (auto cfd : *versions_->GetColumnFamilySet()) {
-    if (cfd->imm()->NumNotFlushed() == 0 && cfd->mem()->IsEmpty() &&
-        cached_recoverable_state_empty_.load()) {
-      // Nothing to flush
-      continue;
+  if (!atomic_flush_) {
+    for (auto cfh : column_families) {
+      s = Flush(flush_options, cfh);
+      if (!s.ok()) {
+        break;
+      }
     }
-
-    // SwitchMemtable() will release and reacquire mutex during execution
-    s = SwitchMemtable(cfd, &context);
-    if (!s.ok()) {
-      break;
+  } else {
+    ROCKS_LOG_INFO(immutable_db_options_.info_log,
+                   "Manual atomic flush start.\n"
+                   "=====Column families:=====");
+    for (auto cfh : column_families) {
+      auto cfhi = static_cast<ColumnFamilyHandleImpl*>(cfh);
+      ROCKS_LOG_INFO(immutable_db_options_.info_log, "%s",
+                     cfhi->GetName().c_str());
     }
-
-    cfd->imm()->FlushRequested();
-
-    flush_req.emplace_back(cfd, cfd->imm()->GetLatestMemTableID());
-  }
-
-  // schedule flush
-  if (s.ok() && !flush_req.empty()) {
-    SchedulePendingFlush(flush_req, flush_reason);
-    MaybeScheduleFlushOrCompaction();
-  }
-
-  write_thread_.ExitUnbatched(&w);
-
-  if (s.ok()) {
-    for (auto& flush : flush_req) {
-      auto cfd = flush.first;
-      auto flush_memtable_id = flush.second;
-      while (cfd->imm()->NumNotFlushed() > 0 &&
-             cfd->imm()->GetEarliestMemTableID() <= flush_memtable_id) {
-        if (!error_handler_.GetRecoveryError().ok()) {
-          break;
-        }
-        if (cfd->IsDropped()) {
-          // FlushJob cannot flush a dropped CF, if we did not break here
-          // we will loop forever since cfd->imm()->NumNotFlushed() will never
-          // drop to zero
-          continue;
-        }
-        cfd->Ref();
-        bg_cv_.Wait();
-        cfd->Unref();
-      }
+    ROCKS_LOG_INFO(immutable_db_options_.info_log,
+                   "=====End of column families list=====");
+    autovector<ColumnFamilyData*> cfds;
+    std::for_each(column_families.begin(), column_families.end(),
+                  [&cfds](ColumnFamilyHandle* elem) {
+                    auto cfh = static_cast<ColumnFamilyHandleImpl*>(elem);
+                    cfds.emplace_back(cfh->cfd());
+                  });
+    s = AtomicFlushMemTables(cfds, flush_options, FlushReason::kManualFlush);
+    ROCKS_LOG_INFO(immutable_db_options_.info_log,
+                   "Manual atomic flush finished, status: %s\n",
+                   "=====Column families:=====", s.ToString().c_str());
+    for (auto cfh : column_families) {
+      auto cfhi = static_cast<ColumnFamilyHandleImpl*>(cfh);
+      ROCKS_LOG_INFO(immutable_db_options_.info_log, "%s",
+                     cfhi->GetName().c_str());
     }
+    ROCKS_LOG_INFO(immutable_db_options_.info_log,
+                   "=====End of column families list=====");
   }
-
-  flush_req.clear();
   return s;
 }
 
@@ -1364,6 +1381,19 @@ Status DBImpl::RunManualCompaction(ColumnFamilyData* cfd, int input_level,
   return manual.status;
 }
 
+void DBImpl::GenerateFlushRequest(const autovector<ColumnFamilyData*>& cfds,
+                                  FlushRequest* req) {
+  assert(req != nullptr);
+  for (const auto cfd : cfds) {
+    if (nullptr == cfd) {
+      // cfd may be null, see DBImpl::ScheduleFlushes
+      continue;
+    }
+    uint64_t max_memtable_id = cfd->imm()->GetLatestMemTableID();
+    req->emplace_back(cfd, max_memtable_id);
+  }
+}
+
 Status DBImpl::FlushMemTable(ColumnFamilyData* cfd,
                              const FlushOptions& flush_options,
                              FlushReason flush_reason, bool writes_stopped) {
@@ -1415,12 +1445,89 @@ Status DBImpl::FlushMemTable(ColumnFamilyData* cfd,
       cfds.push_back(iter.first);
       flush_memtable_ids.push_back(&(iter.second));
     }
-    s = WaitForFlushMemTables(cfds, flush_memtable_ids);
+    s = WaitForFlushMemTables(cfds, flush_memtable_ids,
+                              (flush_reason == FlushReason::kErrorRecovery));
   }
   TEST_SYNC_POINT("FlushMemTableFinished");
   return s;
 }
 
+// Flush all elments in 'column_family_datas'
+// and atomically record the result to the MANIFEST.
+Status DBImpl::AtomicFlushMemTables(
+    const autovector<ColumnFamilyData*>& column_family_datas,
+    const FlushOptions& flush_options, FlushReason flush_reason,
+    bool writes_stopped) {
+  Status s;
+  if (!flush_options.allow_write_stall) {
+    int num_cfs_to_flush = 0;
+    for (auto cfd : column_family_datas) {
+      bool flush_needed = true;
+      s = WaitUntilFlushWouldNotStallWrites(cfd, &flush_needed);
+      if (!s.ok()) {
+        return s;
+      } else if (flush_needed) {
+        ++num_cfs_to_flush;
+      }
+    }
+    if (0 == num_cfs_to_flush) {
+      return s;
+    }
+  }
+  FlushRequest flush_req;
+  autovector<ColumnFamilyData*> cfds;
+  {
+    WriteContext context;
+    InstrumentedMutexLock guard_lock(&mutex_);
+
+    WriteThread::Writer w;
+    if (!writes_stopped) {
+      write_thread_.EnterUnbatched(&w, &mutex_);
+    }
+
+    for (auto cfd : column_family_datas) {
+      if (cfd->IsDropped()) {
+        continue;
+      }
+      if (cfd->imm()->NumNotFlushed() != 0 || !cfd->mem()->IsEmpty() ||
+          !cached_recoverable_state_empty_.load()) {
+        cfds.emplace_back(cfd);
+      }
+    }
+    for (auto cfd : cfds) {
+      cfd->Ref();
+      s = SwitchMemtable(cfd, &context);
+      cfd->Unref();
+      if (!s.ok()) {
+        break;
+      }
+    }
+    if (s.ok()) {
+      AssignAtomicFlushSeq(cfds);
+      for (auto cfd : cfds) {
+        cfd->imm()->FlushRequested();
+      }
+      GenerateFlushRequest(cfds, &flush_req);
+      SchedulePendingFlush(flush_req, flush_reason);
+      MaybeScheduleFlushOrCompaction();
+    }
+
+    if (!writes_stopped) {
+      write_thread_.ExitUnbatched(&w);
+    }
+  }
+
+  if (s.ok() && flush_options.wait) {
+    autovector<const uint64_t*> flush_memtable_ids;
+    for (auto& iter : flush_req) {
+      flush_memtable_ids.push_back(&(iter.second));
+    }
+    s = WaitForFlushMemTables(cfds, flush_memtable_ids,
+                              (flush_reason == FlushReason::kErrorRecovery));
+  }
+  return s;
+}
+
 // Calling FlushMemTable(), whether from DB::Flush() or from Backup Engine, can
 // cause write stall, for example if one memtable is being flushed already.
 // This method tries to avoid write stall (similar to CompactRange() behavior)
@@ -1492,16 +1599,25 @@ Status DBImpl::WaitUntilFlushWouldNotStallWrites(ColumnFamilyData* cfd,
 //  2) if flush_memtable_ids[i] is null, then all memtables in THIS column
 //     family have to be flushed.
 // Finish waiting when ALL column families finish flushing memtables.
+// resuming_from_bg_err indicates whether the caller is trying to resume from
+// background error or in normal processing.
 Status DBImpl::WaitForFlushMemTables(
     const autovector<ColumnFamilyData*>& cfds,
-    const autovector<const uint64_t*>& flush_memtable_ids) {
+    const autovector<const uint64_t*>& flush_memtable_ids,
+    bool resuming_from_bg_err) {
   int num = static_cast<int>(cfds.size());
   // Wait until the compaction completes
   InstrumentedMutexLock l(&mutex_);
-  while (!error_handler_.IsDBStopped()) {
+  // If the caller is trying to resume from bg error, then
+  // error_handler_.IsDBStopped() is true.
+  while (resuming_from_bg_err || !error_handler_.IsDBStopped()) {
     if (shutting_down_.load(std::memory_order_acquire)) {
       return Status::ShutdownInProgress();
     }
+    // If an error has occurred during resumption, then no need to wait.
+    if (!error_handler_.GetRecoveryError().ok()) {
+      break;
+    }
     // Number of column families that have been dropped.
     int num_dropped = 0;
     // Number of column families that have finished flush.
@@ -1527,7 +1643,9 @@ Status DBImpl::WaitForFlushMemTables(
     bg_cv_.Wait();
   }
   Status s;
-  if (error_handler_.IsDBStopped()) {
+  // If not resuming from bg error, and an error has caused the DB to stop,
+  // then report the bg error to caller.
+  if (!resuming_from_bg_err && error_handler_.IsDBStopped()) {
     s = error_handler_.GetBGError();
   }
   return s;
@@ -1867,6 +1985,7 @@ void DBImpl::BackgroundCallFlush() {
       mutex_.Lock();
     }
 
+    TEST_SYNC_POINT("DBImpl::BackgroundCallFlush:FlushFinish:0");
     ReleaseFileNumberFromPendingOutputs(pending_outputs_inserted_elem);
 
     // If flush failed, we want to delete all temporary files that we might have

db/db_impl_debug.cc

@@ -123,7 +123,7 @@ Status DBImpl::TEST_WaitForFlushMemTable(ColumnFamilyHandle* column_family) {
     auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
     cfd = cfh->cfd();
   }
-  return WaitForFlushMemTable(cfd);
+  return WaitForFlushMemTable(cfd, nullptr, false);
 }
 
 Status DBImpl::TEST_WaitForCompact(bool wait_unscheduled) {

db/db_impl_write.cc

@@ -1014,6 +1014,28 @@ Status DBImpl::WriteRecoverableState() {
   return Status::OK();
 }
 
+void DBImpl::SelectColumnFamiliesForAtomicFlush(
+    autovector<ColumnFamilyData*>* cfds) {
+  for (ColumnFamilyData* cfd : *versions_->GetColumnFamilySet()) {
+    if (cfd->IsDropped()) {
+      continue;
+    }
+    if (cfd->imm()->NumNotFlushed() != 0 || !cfd->mem()->IsEmpty() ||
+        !cached_recoverable_state_empty_.load()) {
+      cfds->push_back(cfd);
+    }
+  }
+}
+
+// Assign sequence number for atomic flush.
+void DBImpl::AssignAtomicFlushSeq(const autovector<ColumnFamilyData*>& cfds) {
+  assert(atomic_flush_);
+  auto seq = versions_->LastSequence();
+  for (auto cfd : cfds) {
+    cfd->imm()->AssignAtomicFlushSeq(seq);
+  }
+}
+
 Status DBImpl::SwitchWAL(WriteContext* write_context) {
   mutex_.AssertHeld();
   assert(write_context != nullptr);
@@ -1062,21 +1084,36 @@ Status DBImpl::SwitchWAL(WriteContext* write_context) {
                  oldest_alive_log, total_log_size_.load(), GetMaxTotalWalSize());
   // no need to refcount because drop is happening in write thread, so can't
   // happen while we're in the write thread
-  FlushRequest flush_req;
-  for (auto cfd : *versions_->GetColumnFamilySet()) {
-    if (cfd->IsDropped()) {
-      continue;
-    }
-    if (cfd->OldestLogToKeep() <= oldest_alive_log) {
-      status = SwitchMemtable(cfd, write_context);
-      if (!status.ok()) {
-        break;
+  autovector<ColumnFamilyData*> cfds;
+  if (atomic_flush_) {
+    SelectColumnFamiliesForAtomicFlush(&cfds);
+  } else {
+    for (auto cfd : *versions_->GetColumnFamilySet()) {
+      if (cfd->IsDropped()) {
+        continue;
       }
-      flush_req.emplace_back(cfd, cfd->imm()->GetLatestMemTableID());
-      cfd->imm()->FlushRequested();
+      if (cfd->OldestLogToKeep() <= oldest_alive_log) {
+        cfds.push_back(cfd);
+      }
+    }
+  }
+  for (const auto cfd : cfds) {
+    cfd->Ref();
+    status = SwitchMemtable(cfd, write_context);
+    cfd->Unref();
+    if (!status.ok()) {
+      break;
     }
   }
   if (status.ok()) {
+    if (atomic_flush_) {
+      AssignAtomicFlushSeq(cfds);
+    }
+    for (auto cfd : cfds) {
+      cfd->imm()->FlushRequested();
+    }
+    FlushRequest flush_req;
+    GenerateFlushRequest(cfds, &flush_req);
     SchedulePendingFlush(flush_req, FlushReason::kWriteBufferManager);
     MaybeScheduleFlushOrCompaction();
   }
@@ -1101,29 +1138,32 @@ Status DBImpl::HandleWriteBufferFull(WriteContext* write_context) {
       write_buffer_manager_->buffer_size());
   // no need to refcount because drop is happening in write thread, so can't
   // happen while we're in the write thread
-  ColumnFamilyData* cfd_picked = nullptr;
-  SequenceNumber seq_num_for_cf_picked = kMaxSequenceNumber;
+  autovector<ColumnFamilyData*> cfds;
+  if (atomic_flush_) {
+    SelectColumnFamiliesForAtomicFlush(&cfds);
+  } else {
+    ColumnFamilyData* cfd_picked = nullptr;
+    SequenceNumber seq_num_for_cf_picked = kMaxSequenceNumber;
 
-  for (auto cfd : *versions_->GetColumnFamilySet()) {
-    if (cfd->IsDropped()) {
-      continue;
-    }
-    if (!cfd->mem()->IsEmpty()) {
-      // We only consider active mem table, hoping immutable memtable is
-      // already in the process of flushing.
-      uint64_t seq = cfd->mem()->GetCreationSeq();
-      if (cfd_picked == nullptr || seq < seq_num_for_cf_picked) {
-        cfd_picked = cfd;
-        seq_num_for_cf_picked = seq;
+    for (auto cfd : *versions_->GetColumnFamilySet()) {
+      if (cfd->IsDropped()) {
+        continue;
       }
+      if (!cfd->mem()->IsEmpty()) {
+        // We only consider active mem table, hoping immutable memtable is
+        // already in the process of flushing.
+        uint64_t seq = cfd->mem()->GetCreationSeq();
+        if (cfd_picked == nullptr || seq < seq_num_for_cf_picked) {
+          cfd_picked = cfd;
+          seq_num_for_cf_picked = seq;
+        }
+      }
+    }
+    if (cfd_picked != nullptr) {
+      cfds.push_back(cfd_picked);
     }
   }
 
-  autovector<ColumnFamilyData*> cfds;
-  if (cfd_picked != nullptr) {
-    cfds.push_back(cfd_picked);
-  }
-  FlushRequest flush_req;
   for (const auto cfd : cfds) {
     cfd->Ref();
     status = SwitchMemtable(cfd, write_context);
@@ -1131,11 +1171,16 @@ Status DBImpl::HandleWriteBufferFull(WriteContext* write_context) {
     if (!status.ok()) {
       break;
     }
-    uint64_t flush_memtable_id = cfd->imm()->GetLatestMemTableID();
-    cfd->imm()->FlushRequested();
-    flush_req.emplace_back(cfd, flush_memtable_id);
   }
   if (status.ok()) {
+    if (atomic_flush_) {
+      AssignAtomicFlushSeq(cfds);
+    }
+    for (const auto cfd : cfds) {
+      cfd->imm()->FlushRequested();
+    }
+    FlushRequest flush_req;
+    GenerateFlushRequest(cfds, &flush_req);
     SchedulePendingFlush(flush_req, FlushReason::kWriteBufferFull);
     MaybeScheduleFlushOrCompaction();
   }
@@ -1258,25 +1303,36 @@ Status DBImpl::ThrottleLowPriWritesIfNeeded(const WriteOptions& write_options,
 }
 
 Status DBImpl::ScheduleFlushes(WriteContext* context) {
-  ColumnFamilyData* cfd;
-  FlushRequest flush_req;
+  autovector<ColumnFamilyData*> cfds;
+  if (atomic_flush_) {
+    SelectColumnFamiliesForAtomicFlush(&cfds);
+    for (auto cfd : cfds) {
+      cfd->Ref();
+    }
+    flush_scheduler_.Clear();
+  } else {
+    ColumnFamilyData* tmp_cfd;
+    while ((tmp_cfd = flush_scheduler_.TakeNextColumnFamily()) != nullptr) {
+      cfds.push_back(tmp_cfd);
+    }
+  }
   Status status;
-  while ((cfd = flush_scheduler_.TakeNextColumnFamily()) != nullptr) {
+  for (auto& cfd : cfds) {
     status = SwitchMemtable(cfd, context);
-    bool should_schedule = true;
     if (cfd->Unref()) {
       delete cfd;
-      should_schedule = false;
+      cfd = nullptr;
     }
     if (!status.ok()) {
       break;
     }
-    if (should_schedule) {
-      uint64_t flush_memtable_id = cfd->imm()->GetLatestMemTableID();
-      flush_req.emplace_back(cfd, flush_memtable_id);
-    }
   }
   if (status.ok()) {
+    if (atomic_flush_) {
+      AssignAtomicFlushSeq(cfds);
+    }
+    FlushRequest flush_req;
+    GenerateFlushRequest(cfds, &flush_req);
     SchedulePendingFlush(flush_req, FlushReason::kWriteBufferFull);
     MaybeScheduleFlushOrCompaction();
   }

db/db_test.cc

@@ -2659,6 +2659,11 @@ class ModelDB : public DB {
     Status ret;
     return ret;
   }
+  virtual Status Flush(
+      const rocksdb::FlushOptions& /*options*/,
+      const std::vector<ColumnFamilyHandle*>& /*column_families*/) override {
+    return Status::OK();
+  }
 
   virtual Status SyncWAL() override { return Status::OK(); }
 

db/db_test_util.cc

@@ -646,6 +646,13 @@ Status DBTestBase::Flush(int cf) {
   }
 }
 
+Status DBTestBase::Flush(const std::vector<int>& cf_ids) {
+  std::vector<ColumnFamilyHandle*> cfhs;
+  std::for_each(cf_ids.begin(), cf_ids.end(),
+                [&cfhs, this](int id) { cfhs.emplace_back(handles_[id]); });
+  return db_->Flush(FlushOptions(), cfhs);
+}
+
 Status DBTestBase::Put(const Slice& k, const Slice& v, WriteOptions wo) {
   if (kMergePut == option_config_) {
     return db_->Merge(wo, k, v);

db/db_test_util.h

@@ -799,6 +799,8 @@ class DBTestBase : public testing::Test {
 
   Status Flush(int cf = 0);
 
+  Status Flush(const std::vector<int>& cf_ids);
+
   Status Put(const Slice& k, const Slice& v, WriteOptions wo = WriteOptions());
 
   Status Put(int cf, const Slice& k, const Slice& v,

db/memtable_list.h

@@ -283,6 +283,21 @@ class MemTableList {
     return memlist.front()->GetID();
   }
 
+  void AssignAtomicFlushSeq(const SequenceNumber& seq) {
+    const auto& memlist = current_->memlist_;
+    // Scan the memtable list from new to old
+    for (auto it = memlist.begin(); it != memlist.end(); ++it) {
+      MemTable* mem = *it;
+      if (mem->atomic_flush_seqno_ == kMaxSequenceNumber) {
+        mem->atomic_flush_seqno_ = seq;
+      } else {
+        // Earlier memtables must have been assigned a atomic flush seq, no
+        // need to continue scan.
+        break;
+      }
+    }
+  }
+
  private:
   // DB mutex held
   void InstallNewVersion();

include/rocksdb/db.h

@@ -900,11 +900,22 @@ class DB {
   virtual DBOptions GetDBOptions() const = 0;
 
   // Flush all mem-table data.
+  // Flush a single column family, even when atomic flush is enabled. To flush
+  // multiple column families, use Flush(options, column_families).
   virtual Status Flush(const FlushOptions& options,
                        ColumnFamilyHandle* column_family) = 0;
   virtual Status Flush(const FlushOptions& options) {
     return Flush(options, DefaultColumnFamily());
   }
+  // Flushes multiple column families.
+  // If atomic flush is not enabled, Flush(options, column_families) is
+  // equivalent to calling Flush(options, column_family) multiple times.
+  // If atomic flush is enabled, Flush(options, column_families) will flush all
+  // column families specified in 'column_families' up to the latest sequence
+  // number at the time when flush is requested.
+  virtual Status Flush(
+      const FlushOptions& options,
+      const std::vector<ColumnFamilyHandle*>& column_families) = 0;
 
   // Flush the WAL memory buffer to the file. If sync is true, it calls SyncWAL
   // afterwards.

include/rocksdb/options.h

@@ -943,6 +943,20 @@ struct DBOptions {
   // relies on manual invocation of FlushWAL to write the WAL buffer to its
   // file.
   bool manual_wal_flush = false;
+
+  // If true, RocksDB supports flushing multiple column families and committing
+  // their results atomically to MANIFEST. Note that it is not
+  // necessary to set atomic_flush to true if WAL is always enabled since WAL
+  // allows the database to be restored to the last persistent state in WAL.
+  // This option is useful when there are column families with writes NOT
+  // protected by WAL.
+  // For manual flush, application has to specify which column families to
+  // flush atomically in DB::Flush.
+  // For auto-triggered flush, RocksDB atomically flushes ALL column families.
+  //
+  // Currently, any WAL-enabled writes after atomic flush may be replayed
+  // independently if the process crashes later and tries to recover.
+  bool atomic_flush = false;
 };
 
 // Options to control the behavior of a database (passed to DB::Open)

include/rocksdb/utilities/stackable_db.h

@@ -278,6 +278,11 @@ class StackableDB : public DB {
                        ColumnFamilyHandle* column_family) override {
     return db_->Flush(fopts, column_family);
   }
+  virtual Status Flush(
+      const FlushOptions& fopts,
+      const std::vector<ColumnFamilyHandle*>& column_families) override {
+    return db_->Flush(fopts, column_families);
+  }
 
   virtual Status SyncWAL() override {
     return db_->SyncWAL();

options/db_options.h

@@ -78,6 +78,7 @@ struct ImmutableDBOptions {
   bool preserve_deletes;
   bool two_write_queues;
   bool manual_wal_flush;
+  bool atomic_flush;
 };
 
 struct MutableDBOptions {

options/options_helper.cc

@@ -1554,7 +1554,11 @@ std::unordered_map<std::string, OptionTypeInfo>
           offsetof(struct ImmutableDBOptions, manual_wal_flush)}},
         {"seq_per_batch",
          {0, OptionType::kBoolean, OptionVerificationType::kDeprecated, false,
-          0}}};
+          0}},
+        {"atomic_flush",
+         {offsetof(struct DBOptions, atomic_flush), OptionType::kBoolean,
+          OptionVerificationType::kNormal, false,
+          offsetof(struct ImmutableDBOptions, atomic_flush)}}};
 
 std::unordered_map<std::string, BlockBasedTableOptions::IndexType>
     OptionsHelper::block_base_table_index_type_string_map = {

options/options_settable_test.cc

@@ -291,7 +291,8 @@ TEST_F(OptionsSettableTest, DBOptionsAllFieldsSettable) {
                              "concurrent_prepare=false;"
                              "two_write_queues=false;"
                              "manual_wal_flush=false;"
-                             "seq_per_batch=false;",
+                             "seq_per_batch=false;"
+                             "atomic_flush=false",
                              new_options));
 
   ASSERT_EQ(unset_bytes_base, NumUnsetBytes(new_options_ptr, sizeof(DBOptions),

tools/db_stress.cc

@@ -133,6 +133,8 @@ DEFINE_bool(test_batches_snapshots, false,
             "\t(b) No long validation at the end (more speed up)\n"
             "\t(c) Test snapshot and atomicity of batch writes");
 
+DEFINE_bool(atomic_flush, false, "If true, the test enables atomic flush\n");
+
 DEFINE_int32(threads, 32, "Number of concurrent threads to run.");
 
 DEFINE_int32(ttl, -1,
@@ -2218,6 +2220,8 @@ class StressTest {
     fprintf(stdout, "Format version            : %d\n", FLAGS_format_version);
     fprintf(stdout, "TransactionDB             : %s\n",
             FLAGS_use_txn ? "true" : "false");
+    fprintf(stdout, "Atomic flush              : %s\n",
+            FLAGS_atomic_flush ? "true" : "false");
     fprintf(stdout, "Column families           : %d\n", FLAGS_column_families);
     if (!FLAGS_test_batches_snapshots) {
       fprintf(stdout, "Clear CFs one in          : %d\n",
@@ -2363,6 +2367,7 @@ class StressTest {
           FLAGS_universal_max_merge_width;
       options_.compaction_options_universal.max_size_amplification_percent =
           FLAGS_universal_max_size_amplification_percent;
+      options_.atomic_flush = FLAGS_atomic_flush;
     } else {
 #ifdef ROCKSDB_LITE
       fprintf(stderr, "--options_file not supported in lite mode\n");
@@ -3327,6 +3332,252 @@ class BatchedOpsStressTest : public StressTest {
   virtual void VerifyDb(ThreadState* /* thread */) const {}
 };
 
+class AtomicFlushStressTest : public StressTest {
+ public:
+  AtomicFlushStressTest() : batch_id_(0) {}
+
+  virtual ~AtomicFlushStressTest() {}
+
+  virtual Status TestPut(ThreadState* thread, WriteOptions& write_opts,
+                         const ReadOptions& /* read_opts */,
+                         const std::vector<int>& rand_column_families,
+                         const std::vector<int64_t>& rand_keys,
+                         char (&value)[100],
+                         std::unique_ptr<MutexLock>& /* lock */) {
+    std::string key_str = Key(rand_keys[0]);
+    Slice key = key_str;
+    uint64_t value_base = batch_id_.fetch_add(1);
+    size_t sz =
+        GenerateValue(static_cast<uint32_t>(value_base), value, sizeof(value));
+    Slice v(value, sz);
+    WriteBatch batch;
+    for (auto cf : rand_column_families) {
+      ColumnFamilyHandle* cfh = column_families_[cf];
+      if (FLAGS_use_merge) {
+        batch.Merge(cfh, key, v);
+      } else { /* !FLAGS_use_merge */
+        batch.Put(cfh, key, v);
+      }
+    }
+    Status s = db_->Write(write_opts, &batch);
+    if (!s.ok()) {
+      fprintf(stderr, "multi put or merge error: %s\n", s.ToString().c_str());
+      thread->stats.AddErrors(1);
+    } else {
+      size_t num = rand_column_families.size();
+      thread->stats.AddBytesForWrites(num, (sz + 1) * num);
+    }
+
+    return s;
+  }
+
+  virtual Status TestDelete(ThreadState* thread, WriteOptions& write_opts,
+                            const std::vector<int>& rand_column_families,
+                            const std::vector<int64_t>& rand_keys,
+                            std::unique_ptr<MutexLock>& /* lock */) {
+    std::string key_str = Key(rand_keys[0]);
+    Slice key = key_str;
+    WriteBatch batch;
+    for (auto cf : rand_column_families) {
+      ColumnFamilyHandle* cfh = column_families_[cf];
+      batch.Delete(cfh, key);
+    }
+    Status s = db_->Write(write_opts, &batch);
+    if (!s.ok()) {
+      fprintf(stderr, "multidel error: %s\n", s.ToString().c_str());
+      thread->stats.AddErrors(1);
+    } else {
+      thread->stats.AddDeletes(rand_column_families.size());
+    }
+    return s;
+  }
+
+  virtual Status TestDeleteRange(ThreadState* thread, WriteOptions& write_opts,
+                                 const std::vector<int>& rand_column_families,
+                                 const std::vector<int64_t>& rand_keys,
+                                 std::unique_ptr<MutexLock>& /* lock */) {
+    int64_t rand_key = rand_keys[0];
+    auto shared = thread->shared;
+    int64_t max_key = shared->GetMaxKey();
+    if (rand_key > max_key - FLAGS_range_deletion_width) {
+      rand_key =
+          thread->rand.Next() % (max_key - FLAGS_range_deletion_width + 1);
+    }
+    std::string key_str = Key(rand_key);
+    Slice key = key_str;
+    std::string end_key_str = Key(rand_key + FLAGS_range_deletion_width);
+    Slice end_key = end_key_str;
+    WriteBatch batch;
+    for (auto cf : rand_column_families) {
+      ColumnFamilyHandle* cfh = column_families_[rand_column_families[cf]];
+      batch.DeleteRange(cfh, key, end_key);
+    }
+    Status s = db_->Write(write_opts, &batch);
+    if (!s.ok()) {
+      fprintf(stderr, "multi del range error: %s\n", s.ToString().c_str());
+      thread->stats.AddErrors(1);
+    } else {
+      thread->stats.AddRangeDeletions(rand_column_families.size());
+    }
+    return s;
+  }
+
+  virtual void TestIngestExternalFile(
+      ThreadState* /* thread */,
+      const std::vector<int>& /* rand_column_families */,
+      const std::vector<int64_t>& /* rand_keys */,
+      std::unique_ptr<MutexLock>& /* lock */) {
+    assert(false);
+    fprintf(stderr,
+            "AtomicFlushStressTest does not support TestIngestExternalFile "
+            "because it's not possible to verify the result\n");
+    std::terminate();
+  }
+
+  virtual Status TestGet(ThreadState* thread, const ReadOptions& readoptions,
+                         const std::vector<int>& rand_column_families,
+                         const std::vector<int64_t>& rand_keys) {
+    std::string key_str = Key(rand_keys[0]);
+    Slice key = key_str;
+    auto cfh =
+        column_families_[rand_column_families[thread->rand.Next() %
+                                              rand_column_families.size()]];
+    std::string from_db;
+    Status s = db_->Get(readoptions, cfh, key, &from_db);
+    if (s.ok()) {
+      thread->stats.AddGets(1, 1);
+    } else if (s.IsNotFound()) {
+      thread->stats.AddGets(1, 0);
+    } else {
+      thread->stats.AddErrors(1);
+    }
+    return s;
+  }
+
+  virtual Status TestPrefixScan(ThreadState* thread,
+                                const ReadOptions& readoptions,
+                                const std::vector<int>& rand_column_families,
+                                const std::vector<int64_t>& rand_keys) {
+    std::string key_str = Key(rand_keys[0]);
+    Slice key = key_str;
+    Slice prefix = Slice(key.data(), FLAGS_prefix_size);
+
+    std::string upper_bound;
+    Slice ub_slice;
+    ReadOptions ro_copy = readoptions;
+    if (thread->rand.OneIn(2) && GetNextPrefix(prefix, &upper_bound)) {
+      ub_slice = Slice(upper_bound);
+      ro_copy.iterate_upper_bound = &ub_slice;
+    }
+    auto cfh =
+        column_families_[rand_column_families[thread->rand.Next() %
+                                              rand_column_families.size()]];
+    Iterator* iter = db_->NewIterator(ro_copy, cfh);
+    int64_t count = 0;
+    for (iter->Seek(prefix); iter->Valid() && iter->key().starts_with(prefix);
+         iter->Next()) {
+      ++count;
+    }
+    assert(count <= (static_cast<int64_t>(1) << ((8 - FLAGS_prefix_size) * 8)));
+    Status s = iter->status();
+    if (s.ok()) {
+      thread->stats.AddPrefixes(1, static_cast<int>(count));
+    } else {
+      thread->stats.AddErrors(1);
+    }
+    delete iter;
+    return s;
+  }
+
+  virtual void VerifyDb(ThreadState* thread) const {
+    ReadOptions options(FLAGS_verify_checksum, true);
+    // We must set total_order_seek to true because we are doing a SeekToFirst
+    // on a column family whose memtables may support (by default) prefix-based
+    // iterator. In this case, NewIterator with options.total_order_seek being
+    // false returns a prefix-based iterator. Calling SeekToFirst using this
+    // iterator causes the iterator to become invalid. That means we cannot
+    // iterate the memtable using this iterator any more, although the memtable
+    // contains the most up-to-date key-values.
+    options.total_order_seek = true;
+    assert(thread != nullptr);
+    auto shared = thread->shared;
+    std::vector<unique_ptr<Iterator> > iters(column_families_.size());
+    for (size_t i = 0; i != column_families_.size(); ++i) {
+      iters[i].reset(db_->NewIterator(options, column_families_[i]));
+    }
+    for (auto& iter : iters) {
+      iter->SeekToFirst();
+    }
+    size_t num = column_families_.size();
+    assert(num == iters.size());
+    do {
+      size_t valid_cnt = 0;
+      for (auto& iter : iters) {
+        if (iter->Valid()) {
+          ++valid_cnt;
+        }
+      }
+      if (valid_cnt == 0) {
+        break;
+      } else if (valid_cnt != iters.size()) {
+        fprintf(stderr, "Finished iterating the following column families:\n");
+        for (size_t i = 0; i != num; ++i) {
+          if (!iters[i]->Valid()) {
+            fprintf(stderr, "%s ", column_families_[i]->GetName().c_str());
+          }
+        }
+        fprintf(stderr,
+                "\nThe following column families have data that have not been "
+                "scanned:\n");
+        for (size_t i = 0; i != num; ++i) {
+          if (iters[i]->Valid()) {
+            fprintf(stderr, "%s ", column_families_[i]->GetName().c_str());
+          }
+        }
+        fprintf(stderr, "\n");
+      }
+      // If the program reaches here, then all column families' iterators are
+      // still valid.
+      Slice key;
+      Slice value;
+      for (size_t i = 0; i != num; ++i) {
+        if (i == 0) {
+          key = iters[i]->key();
+          value = iters[i]->value();
+        } else {
+          if (key.compare(iters[i]->key()) != 0) {
+            fprintf(stderr, "Verification failed\n");
+            fprintf(stderr, "cf%s: %s => %s\n",
+                    column_families_[0]->GetName().c_str(),
+                    key.ToString(true /* hex */).c_str(),
+                    value.ToString(/* hex */).c_str());
+            fprintf(stderr, "cf%s: %s => %s\n",
+                    column_families_[i]->GetName().c_str(),
+                    iters[i]->key().ToString(true /* hex */).c_str(),
+                    iters[i]->value().ToString(true /* hex */).c_str());
+            shared->SetVerificationFailure();
+          }
+        }
+      }
+      for (auto& iter : iters) {
+        iter->Next();
+      }
+    } while (true);
+  }
+
+  virtual std::vector<int> GenerateColumnFamilies(
+      const int /* num_column_families */, int /* rand_column_family */) const {
+    std::vector<int> ret;
+    int num = static_cast<int>(column_families_.size());
+    int k = 0;
+    std::generate_n(back_inserter(ret), num, [&k]() -> int { return k++; });
+    return ret;
+  }
+
+ private:
+  std::atomic<int64_t> batch_id_;
+};
+
 }  // namespace rocksdb
 
 int main(int argc, char** argv) {
@@ -3428,7 +3679,9 @@ int main(int argc, char** argv) {
   rocksdb_kill_prefix_blacklist = SplitString(FLAGS_kill_prefix_blacklist);
 
   std::unique_ptr<rocksdb::StressTest> stress;
-  if (FLAGS_test_batches_snapshots) {
+  if (FLAGS_atomic_flush) {
+    stress.reset(new rocksdb::AtomicFlushStressTest());
+  } else if (FLAGS_test_batches_snapshots) {
     stress.reset(new rocksdb::BatchedOpsStressTest());
   } else {
     stress.reset(new rocksdb::NonBatchedOpsStressTest());