rocksdb flush
RocksDB的Flush代码走读。
什么时候会触发Flush
Write的所有leader 都会进入到这个函数主要在这个DBImpl::PreprocessWrite函数进行控制，包括：
- SwitchWAL (需要切换WAL)
- HandleWriteBufferFull (需要切换memtable)
- ScheduleFlushes (需要flush)
- DelayWrite (需要delay或者stop)
Status DBImpl::PreprocessWrite(const WriteOptions& write_options,
bool* need_log_sync,
WriteContext* write_context) {
...
assert(!single_column_family_mode_ ||
versions_->GetColumnFamilySet()->NumberOfColumnFamilies() == 1);
if (UNLIKELY(status.ok() && !single_column_family_mode_ &&
total_log_size_ > GetMaxTotalWalSize())) {
WaitForPendingWrites();
status = SwitchWAL(write_context);
}
if (UNLIKELY(status.ok() && write_buffer_manager_->ShouldFlush())) {
// Before a new memtable is added in SwitchMemtable(),
// write_buffer_manager_->ShouldFlush() will keep returning true. If another
// thread is writing to another DB with the same write buffer, they may also
// be flushed. We may end up with flushing much more DBs than needed. It's
// suboptimal but still correct.
WaitForPendingWrites();
status = HandleWriteBufferFull(write_context);
}
if (UNLIKELY(status.ok() && !trim_history_scheduler_.Empty())) {
status = TrimMemtableHistory(write_context);
}
if (UNLIKELY(status.ok() && !flush_scheduler_.Empty())) {
WaitForPendingWrites();
status = ScheduleFlushes(write_context);
}
PERF_TIMER_STOP(write_scheduling_flushes_compactions_time);
PERF_TIMER_GUARD(write_pre_and_post_process_time);
if (UNLIKELY(status.ok() && (write_controller_.IsStopped() ||
write_controller_.NeedsDelay()))) {
PERF_TIMER_STOP(write_pre_and_post_process_time);
PERF_TIMER_GUARD(write_delay_time);
// We don't know size of curent batch so that we always use the size
// for previous one. It might create a fairness issue that expiration
// might happen for smaller writes but larger writes can go through.
// Can optimize it if it is an issue.
status = DelayWrite(last_batch_group_size_, write_options);
PERF_TIMER_START(write_pre_and_post_process_time);
}
if (status.ok() && *need_log_sync) {
...
} else {
...
}
return status;
}
- SwitchWAL
- HandleWriteBufferFull
- ScheduleFlushes
这三个函数很相似都是先SwitchMemtable，再将flush加入调度队列
需要注意的是切换WAL之后需要SwitchMemtable，参见之前的图
for (const auto cfd : cfds) {
...
status = SwitchMemtable(cfd, write_context);
...
}
...
if (status.ok()) {
...
FlushRequest flush_req;
GenerateFlushRequest(cfds, &flush_req);
// 把flush的请求加入队列等待调度
SchedulePendingFlush(flush_req, FlushReason::kWriteBufferManager);
// 触发flush或者compaction
MaybeScheduleFlushOrCompaction();
}
return status;
}
Flush Job
FLushJob::Run
Status FlushJob::Run(LogsWithPrepTracker* prep_tracker,
FileMetaData* file_meta) {
TEST_SYNC_POINT("FlushJob::Start");
db_mutex_->AssertHeld();
assert(pick_memtable_called);
AutoThreadOperationStageUpdater stage_run(
ThreadStatus::STAGE_FLUSH_RUN);
if (mems_.empty()) {
ROCKS_LOG_BUFFER(log_buffer_, "[%s] Nothing in memtable to flush",
cfd_->GetName().c_str());
return Status::OK();
}
...
// This will release and re-acquire the mutex.
// 核心方法 生成L0的SST
Status s = WriteLevel0Table();
...
if (!s.ok()) {
cfd_->imm()->RollbackMemtableFlush(mems_, meta_.fd.GetNumber());
} else if (write_manifest_) {
TEST_SYNC_POINT("FlushJob::InstallResults");
// Replace immutable memtable with the generated Table
IOStatus tmp_io_s;
// 在MANIFEST中增加flush以及新生成的SST文件信息
s = cfd_->imm()->TryInstallMemtableFlushResults(
cfd_, mutable_cf_options_, mems_, prep_tracker, versions_, db_mutex_,
meta_.fd.GetNumber(), &job_context_->memtables_to_free, db_directory_,
log_buffer_, &committed_flush_jobs_info_, &tmp_io_s);
if (!tmp_io_s.ok()) {
io_status_ = tmp_io_s;
}
}
if (s.ok() && file_meta != nullptr) {
*file_meta = meta_;
}
RecordFlushIOStats();
...
return s;
}
FlushJob::WriteLevel0Table()
Status FlushJob::WriteLevel0Table() {
AutoThreadOperationStageUpdater stage_updater(
ThreadStatus::STAGE_FLUSH_WRITE_L0);
db_mutex_->AssertHeld();
const uint64_t start_micros = db_options_.env->NowMicros();
const uint64_t start_cpu_micros = db_options_.env->NowCPUNanos() / 1000;
Status s;
std::vector<BlobFileAddition> blob_file_additions;
{
auto write_hint = cfd_->CalculateSSTWriteHint(0);
db_mutex_->Unlock();
if (log_buffer_) {
log_buffer_->FlushBufferToLog();
}
// memtables and range_del_iters store internal iterators over each data
// memtable and its associated range deletion memtable, respectively, at
// corresponding indexes.
std::vector<InternalIterator*> memtables;
std::vector<std::unique_ptr<FragmentedRangeTombstoneIterator>>
range_del_iters;
ReadOptions ro;
ro.total_order_seek = true;
Arena arena;
uint64_t total_num_entries = 0, total_num_deletes = 0;
uint64_t total_data_size = 0;
size_t total_memory_usage = 0;
// 根据要进行flush的各个memtable中的range delete，构建FragmentedRangeTombstoneIterator
for (MemTable* m : mems_) {
...
memtables.push_back(m->NewIterator(ro, &arena));
auto* range_del_iter =
m->NewRangeTombstoneIterator(ro, kMaxSequenceNumber);
if (range_del_iter != nullptr) {
range_del_iters.emplace_back(range_del_iter);
}
total_num_entries += m->num_entries();
total_num_deletes += m->num_deletes();
total_data_size += m->get_data_size();
total_memory_usage += m->ApproximateMemoryUsage();
}
...
{
ScopedArenaIterator iter(
NewMergingIterator(&cfd_->internal_comparator(), &memtables[0],
static_cast<int>(memtables.size()), &arena));
...
uint64_t oldest_key_time =
mems_.front()->ApproximateOldestKeyTime();
// It's not clear whether oldest_key_time is always available. In case
// it is not available, use current_time.
uint64_t oldest_ancester_time = std::min(current_time, oldest_key_time);
TEST_SYNC_POINT_CALLBACK(
"FlushJob::WriteLevel0Table:oldest_ancester_time",
&oldest_ancester_time);
meta_.oldest_ancester_time = oldest_ancester_time;
meta_.file_creation_time = current_time;
uint64_t creation_time = (cfd_->ioptions()->compaction_style ==
CompactionStyle::kCompactionStyleFIFO)
? current_time
: meta_.oldest_ancester_time;
IOStatus io_s;
const std::string* const full_history_ts_low =
(full_history_ts_low_.empty()) ? nullptr : &full_history_ts_low_;
// 核心方法 创建sst文件
s = BuildTable(
dbname_, versions_, db_options_, *cfd_->ioptions(),
mutable_cf_options_, file_options_, cfd_->table_cache(), iter.get(),
std::move(range_del_iters), &meta_, &blob_file_additions,
cfd_->internal_comparator(), cfd_->int_tbl_prop_collector_factories(),
cfd_->GetID(), cfd_->GetName(), existing_snapshots_,
earliest_write_conflict_snapshot_, snapshot_checker_,
output_compression_, mutable_cf_options_.sample_for_compression,
mutable_cf_options_.compression_opts,
mutable_cf_options_.paranoid_file_checks, cfd_->internal_stats(),
TableFileCreationReason::kFlush, &io_s, io_tracer_, event_logger_,
job_context_->job_id, Env::IO_HIGH, &table_properties_, 0 /* level */,
creation_time, oldest_key_time, write_hint, current_time, db_id_,
db_session_id_, full_history_ts_low);
if (!io_s.ok()) {
io_status_ = io_s;
}
LogFlush(db_options_.info_log);
}
...
if (s.ok() && output_file_directory_ != nullptr && sync_output_directory_) {
s = output_file_directory_->Fsync(IOOptions(), nullptr);
}
TEST_SYNC_POINT_CALLBACK("FlushJob::WriteLevel0Table", &mems_);
db_mutex_->Lock();
}
base_->Unref();
// Note that if file_size is zero, the file has been deleted and
// should not be added to the manifest.
const bool has_output = meta_.fd.GetFileSize() > 0;
if (s.ok() && has_output) {
// if we have more than 1 background thread, then we cannot
// insert files directly into higher levels because some other
// threads could be concurrently producing compacted files for
// that key range.
// Add file to L0
edit_->AddFile(0 /* level */, meta_.fd.GetNumber(), meta_.fd.GetPathId(),
meta_.fd.GetFileSize(), meta_.smallest, meta_.largest,
meta_.fd.smallest_seqno, meta_.fd.largest_seqno,
meta_.marked_for_compaction, meta_.oldest_blob_file_number,
meta_.oldest_ancester_time, meta_.file_creation_time,
meta_.file_checksum, meta_.file_checksum_func_name);
edit_->SetBlobFileAdditions(std::move(blob_file_additions));
}
#ifndef ROCKSDB_LITE
// Piggyback FlushJobInfo on the first first flushed memtable.
mems_[0]->SetFlushJobInfo(GetFlushJobInfo());
#endif // !ROCKSDB_LITE
...
RecordFlushIOStats();
return s;
}
BuildTable
这参数还能更多点吗- -b
Status BuildTable(
const std::string& dbname, VersionSet* versions,
const ImmutableDBOptions& db_options, const ImmutableCFOptions& ioptions,
const MutableCFOptions& mutable_cf_options, const FileOptions& file_options,
TableCache* table_cache, InternalIterator* iter,
std::vector<std::unique_ptr<FragmentedRangeTombstoneIterator>>
range_del_iters,
FileMetaData* meta, std::vector<BlobFileAddition>* blob_file_additions,
const InternalKeyComparator& internal_comparator,
const std::vector<std::unique_ptr<IntTblPropCollectorFactory>>*
int_tbl_prop_collector_factories,
uint32_t column_family_id, const std::string& column_family_name,
std::vector<SequenceNumber> snapshots,
SequenceNumber earliest_write_conflict_snapshot,
SnapshotChecker* snapshot_checker, const CompressionType compression,
uint64_t sample_for_compression, const CompressionOptions& compression_opts,
bool paranoid_file_checks, InternalStats* internal_stats,
TableFileCreationReason reason, IOStatus* io_status,
const std::shared_ptr<IOTracer>& io_tracer, EventLogger* event_logger,
int job_id, const Env::IOPriority io_priority,
TableProperties* table_properties, int level, const uint64_t creation_time,
const uint64_t oldest_key_time, Env::WriteLifeTimeHint write_hint,
const uint64_t file_creation_time, const std::string& db_id,
const std::string& db_session_id, const std::string* full_history_ts_low) {
assert((column_family_id ==
TablePropertiesCollectorFactory::Context::kUnknownColumnFamily) ==
column_family_name.empty());
// Reports the IOStats for flush for every following bytes.
const size_t kReportFlushIOStatsEvery = 1048576;
OutputValidator output_validator(
internal_comparator,
/*enable_order_check=*/
mutable_cf_options.check_flush_compaction_key_order,
/*enable_hash=*/paranoid_file_checks);
Status s;
meta->fd.file_size = 0;
iter->SeekToFirst();
// 处理delete-range
std::unique_ptr<CompactionRangeDelAggregator> range_del_agg(
new CompactionRangeDelAggregator(&internal_comparator, snapshots));
for (auto& range_del_iter : range_del_iters) {
range_del_agg->AddTombstones(std::move(range_del_iter));
}
std::string fname = TableFileName(ioptions.cf_paths, meta->fd.GetNumber(),
meta->fd.GetPathId());
std::vector<std::string> blob_file_paths;
std::string file_checksum = kUnknownFileChecksum;
std::string file_checksum_func_name = kUnknownFileChecksumFuncName;
#ifndef ROCKSDB_LITE
EventHelpers::NotifyTableFileCreationStarted(
ioptions.listeners, dbname, column_family_name, fname, job_id, reason);
#endif // !ROCKSDB_LITE
Env* env = db_options.env;
assert(env);
FileSystem* fs = db_options.fs.get();
assert(fs);
TableProperties tp;
if (iter->Valid() || !range_del_agg->IsEmpty()) {
TableBuilder* builder;
std::unique_ptr<WritableFileWriter> file_writer;
{
std::unique_ptr<FSWritableFile> file;
#ifndef NDEBUG
bool use_direct_writes = file_options.use_direct_writes;
TEST_SYNC_POINT_CALLBACK("BuildTable:create_file", &use_direct_writes);
#endif // !NDEBUG
// 创建文件
IOStatus io_s = NewWritableFile(fs, fname, &file, file_options);
assert(s.ok());
s = io_s;
if (io_status->ok()) {
*io_status = io_s;
}
if (!s.ok()) {
EventHelpers::LogAndNotifyTableFileCreationFinished(
event_logger, ioptions.listeners, dbname, column_family_name, fname,
job_id, meta->fd, kInvalidBlobFileNumber, tp, reason, s,
file_checksum, file_checksum_func_name);
return s;
}
file->SetIOPriority(io_priority);
file->SetWriteLifeTimeHint(write_hint);
file_writer.reset(new WritableFileWriter(
std::move(file), fname, file_options, env, io_tracer,
ioptions.statistics, ioptions.listeners,
ioptions.file_checksum_gen_factory));
builder = NewTableBuilder(
ioptions, mutable_cf_options, internal_comparator,
int_tbl_prop_collector_factories, column_family_id,
column_family_name, file_writer.get(), compression,
sample_for_compression, compression_opts, level,
false /* skip_filters */, creation_time, oldest_key_time,
0 /*target_file_size*/, file_creation_time, db_id, db_session_id);
}
MergeHelper merge(env, internal_comparator.user_comparator(),
ioptions.merge_operator, nullptr, ioptions.info_log,
true /* internal key corruption is not ok */,
snapshots.empty() ? 0 : snapshots.back(),
snapshot_checker);
std::unique_ptr<BlobFileBuilder> blob_file_builder(
(mutable_cf_options.enable_blob_files && blob_file_additions)
? new BlobFileBuilder(versions, env, fs, &ioptions,
&mutable_cf_options, &file_options, job_id,
column_family_id, column_family_name,
io_priority, write_hint, &blob_file_paths,
blob_file_additions)
: nullptr);
// CompactionIterator用来处理memtable中的重复数据 只保留最新的
CompactionIterator c_iter(
iter, internal_comparator.user_comparator(), &merge, kMaxSequenceNumber,
&snapshots, earliest_write_conflict_snapshot, snapshot_checker, env,
ShouldReportDetailedTime(env, ioptions.statistics),
true /* internal key corruption is not ok */, range_del_agg.get(),
blob_file_builder.get(), ioptions.allow_data_in_errors,
/*compaction=*/nullptr,
/*compaction_filter=*/nullptr, /*shutting_down=*/nullptr,
/*preserve_deletes_seqnum=*/0, /*manual_compaction_paused=*/nullptr,
db_options.info_log, full_history_ts_low);
c_iter.SeekToFirst();
// 将key/value交给TableBuilder处理
for (; c_iter.Valid(); c_iter.Next()) {
const Slice& key = c_iter.key();
const Slice& value = c_iter.value();
const ParsedInternalKey& ikey = c_iter.ikey();
// Generate a rolling 64-bit hash of the key and values
s = output_validator.Add(key, value);
if (!s.ok()) {
break;
}
builder->Add(key, value);
meta->UpdateBoundaries(key, value, ikey.sequence, ikey.type);
// TODO(noetzli): Update stats after flush, too.
if (io_priority == Env::IO_HIGH &&
IOSTATS(bytes_written) >= kReportFlushIOStatsEvery) {
ThreadStatusUtil::SetThreadOperationProperty(
ThreadStatus::FLUSH_BYTES_WRITTEN, IOSTATS(bytes_written));
}
}
if (!s.ok()) {
c_iter.status().PermitUncheckedError();
} else if (!c_iter.status().ok()) {
s = c_iter.status();
}
// TableBuilder处理delete-range
if (s.ok()) {
auto range_del_it = range_del_agg->NewIterator();
for (range_del_it->SeekToFirst(); range_del_it->Valid();
range_del_it->Next()) {
auto tombstone = range_del_it->Tombstone();
auto kv = tombstone.Serialize();
builder->Add(kv.first.Encode(), kv.second);
meta->UpdateBoundariesForRange(kv.first, tombstone.SerializeEndKey(),
tombstone.seq_, internal_comparator);
}
}
// 收尾工作
TEST_SYNC_POINT("BuildTable:BeforeFinishBuildTable");
const bool empty = builder->IsEmpty();
if (!s.ok() || empty) {
builder->Abandon();
} else {
s = builder->Finish();
}
if (io_status->ok()) {
*io_status = builder->io_status();
}
if (s.ok() && !empty) {
uint64_t file_size = builder->FileSize();
meta->fd.file_size = file_size;
meta->marked_for_compaction = builder->NeedCompact();
assert(meta->fd.GetFileSize() > 0);
tp = builder->GetTableProperties(); // refresh now that builder is finished
if (table_properties) {
*table_properties = tp;
}
}
delete builder;
// Finish and check for file errors
TEST_SYNC_POINT("BuildTable:BeforeSyncTable");
if (s.ok() && !empty) {
StopWatch sw(env, ioptions.statistics, TABLE_SYNC_MICROS);
*io_status = file_writer->Sync(ioptions.use_fsync);
}
TEST_SYNC_POINT("BuildTable:BeforeCloseTableFile");
if (s.ok() && io_status->ok() && !empty) {
*io_status = file_writer->Close();
}
if (s.ok() && io_status->ok() && !empty) {
// Add the checksum information to file metadata.
meta->file_checksum = file_writer->GetFileChecksum();
meta->file_checksum_func_name = file_writer->GetFileChecksumFuncName();
file_checksum = meta->file_checksum;
file_checksum_func_name = meta->file_checksum_func_name;
}
if (s.ok()) {
s = *io_status;
}
if (blob_file_builder) {
if (s.ok()) {
s = blob_file_builder->Finish();
}
blob_file_builder.reset();
}
// TODO Also check the IO status when create the Iterator.
if (s.ok() && !empty) {
// Verify that the table is usable
// We set for_compaction to false and don't OptimizeForCompactionTableRead
// here because this is a special case after we finish the table building
// No matter whether use_direct_io_for_flush_and_compaction is true,
// we will regrad this verification as user reads since the goal is
// to cache it here for further user reads
ReadOptions read_options;
std::unique_ptr<InternalIterator> it(table_cache->NewIterator(
read_options, file_options, internal_comparator, *meta,
nullptr /* range_del_agg */,
mutable_cf_options.prefix_extractor.get(), nullptr,
(internal_stats == nullptr) ? nullptr
: internal_stats->GetFileReadHist(0),
TableReaderCaller::kFlush, /*arena=*/nullptr,
/*skip_filter=*/false, level,
MaxFileSizeForL0MetaPin(mutable_cf_options),
/*smallest_compaction_key=*/nullptr,
/*largest_compaction_key*/ nullptr,
/*allow_unprepared_value*/ false));
s = it->status();
if (s.ok() && paranoid_file_checks) {
OutputValidator file_validator(internal_comparator,
/*enable_order_check=*/true,
/*enable_hash=*/true);
for (it->SeekToFirst(); it->Valid(); it->Next()) {
// Generate a rolling 64-bit hash of the key and values
file_validator.Add(it->key(), it->value()).PermitUncheckedError();
}
s = it->status();
if (s.ok() && !output_validator.CompareValidator(file_validator)) {
s = Status::Corruption("Paranoid checksums do not match");
}
}
}
}
// Check for input iterator errors
if (!iter->status().ok()) {
s = iter->status();
}
if (!s.ok() || meta->fd.GetFileSize() == 0) {
TEST_SYNC_POINT("BuildTable:BeforeDeleteFile");
constexpr IODebugContext* dbg = nullptr;
Status ignored = fs->DeleteFile(fname, IOOptions(), dbg);
ignored.PermitUncheckedError();
assert(blob_file_additions || blob_file_paths.empty());
if (blob_file_additions) {
for (const std::string& blob_file_path : blob_file_paths) {
ignored = fs->DeleteFile(blob_file_path, IOOptions(), dbg);
ignored.PermitUncheckedError();
}
}
}
if (meta->fd.GetFileSize() == 0) {
fname = "(nil)";
}
// Output to event logger and fire events.
EventHelpers::LogAndNotifyTableFileCreationFinished(
event_logger, ioptions.listeners, dbname, column_family_name, fname,
job_id, meta->fd, meta->oldest_blob_file_number, tp, reason, s,
file_checksum, file_checksum_func_name);
return s;
}