简介

在RocksDB中，每次写入它都会先写WAL,然后再写入MemTable,这次我们就来分析这两个逻辑具体是如何实现的. 首先需要明确的是在RocksDB中，WAL的写入是单线程顺序串行写入的，而MemTable则是可以并发多线程写入的。

而在RocksDB 5.5中引进了一个选项enable_pipelined_write，这个选项的目的就是将WAL和MemTable的写入pipeline化， 也就是说当一个线程写完毕WAL之后，此时在WAL的write队列中等待的其他的write则会开始继续写入WAL, 而当前线程将会继续 写入MemTable.此时就将不同的Writer的写入WAL和写入MemTable并发执行了.

实现

我们这里只来分析pipeline的实现，核心函数就是DBImpl::PipelinedWriteImpl.

  • 每一个DB(DBImpl)都有一个write_thread_(class WriteThread).
  • 每次调用Write的时候会先写入WAL, 此时新建一个WriteThread::Writer对象，并将这个对象加入到一个Group中(调用JoinBatchGroup)

```

  1. WriteThread::Writer w(write_options, my_batch, callback, log_ref,
  2. disable_memtable);
  3. write_thread_.JoinBatchGroup(&w); ``` - 然后我们来看JoinBatchGroup,这个函数主要是用来讲所有的写入WAL加入到一个Group中.这里可以看到当当前的Writer 对象是leader(比如第一个进入的对象)的时候将会直接返回，否则将会等待知道更新为对应的状态．

```

  1. void WriteThread::JoinBatchGroup(Writer* w) {
  2. ...................................
  3. bool linked_as_leader = LinkOne(w, &newest_writer_);
  4. if (linked_as_leader) {
  5. SetState(w, STATE_GROUP_LEADER);

  6. }

  7. TEST_SYNC_POINT_CALLBACK("WriteThread::JoinBatchGroup:Wait", w);

  8. if (!linked_as_leader) {

  9. /**
    • Wait util:
    • 1) An existing leader pick us as the new leader when it finishes
    • 2) An existing leader pick us as its follewer and
    • 2.1) finishes the memtable writes on our behalf
    • 2.2) Or tell us to finish the memtable writes in pralallel
    • 3) (pipelined write) An existing leader pick us as its follower and
    • finish book-keeping and WAL write for us, enqueue us as pending
    • memtable writer, and
    • 3.1) we become memtable writer group leader, or
    • 3.2) an existing memtable writer group leader tell us to finish memtable
    • writes in parallel.
  21. */
  22. AwaitState(w, STATE_GROUP_LEADER | STATE_MEMTABLE_WRITER_LEADER |
  23. STATE_PARALLEL_MEMTABLE_WRITER | STATE_COMPLETED,
  24. &jbg_ctx);
  25. TEST_SYNC_POINT_CALLBACK("WriteThread::JoinBatchGroup:DoneWaiting", w);
  26. }
  27. } ``` - 然后我们来看LinkOne函数，这个函数主要用来讲当前的Writer对象加入到group中，这里可以看到由于 写入是并发的因此对应的newest_writer_(保存最新的写入对象)需要原子操作来更新.

```

  1. bool WriteThread::LinkOne(Writer w, std::atomic newest_writer) {
  2. assert(newest_writer != nullptr);
  3. assert(w->state == STATE_INIT);
  4. Writer* writers = newest_writer->load(std::memory_order_relaxed);
  5. while (true) {
  6. w->link_older = writers;
  7. if (newest_writer->compare_exchange_weak(writers, w)) {
  8. return (writers == nullptr);
  9. }
  10. }
  11. } ``` - 当从JoinBatchGroup返回之后，当当前的Writer对象为leader的话，则将会把此leader下的所有的write都 链接到一个WriteGroup中(调用EnterAsBatchGroupLeader函数),　并开始写入WAL,这里要注意非leader的write将会直接 进入memtable的写入，这是因为非leader的write都将会被当前它所从属的leader来打包(group)写入，后面我们会看到实现.

```

  1. size_t WriteThread::EnterAsBatchGroupLeader(Writer* leader,
  2. WriteGroup* write_group) {
  3. assert(leader->link_older == nullptr);
  4. assert(leader->batch != nullptr);
  5. assert(write_group != nullptr);
  6. ................................................

  7. Writer* newest_writer = newest_writer_.load(std::memory_order_acquire);

  8. // This is safe regardless of any db mutex status of the caller. Previous

  9. // calls to ExitAsGroupLeader either didn't call CreateMissingNewerLinks
  10. // (they emptied the list and then we added ourself as leader) or had to
  11. // explicitly wake us up (the list was non-empty when we added ourself,
  12. // so we have already received our MarkJoined).

  13. CreateMissingNewerLinks(newest_writer);

  14. // Tricky. Iteration start (leader) is exclusive and finish

  15. // (newest_writer) is inclusive. Iteration goes from old to new.
  16. Writer* w = leader;
  17. while (w != newest_writer) {
  18. w = w->link_newer;
  19. .........................................
  20. w->write_group = write_group;
  21. size += batch_size;
  22. write_group->last_writer = w;
  23. write_group->size++;
  24. }
  25. ..............................
  26. } ``` - 这里注意到遍历是通过link_newer进行的，之所以这样做是相当于在写入WAL之前，对于当前leader的Write 做一次snapshot(通过CreateMissingNewerLinks函数).

```

  1. void WriteThread::CreateMissingNewerLinks(Writer* head) {
  2. while (true) {
  3. Writer* next = head->link_older;
  4. if (next == nullptr || next->link_newer != nullptr) {
  5. assert(next == nullptr || next->link_newer == head);
  6. break;
  7. }
  8. next->link_newer = head;
  9. head = next;
  10. }
  11. } ``` - 上述操作进行完毕之后，进入写WAL操作,最终会把这个write_group打包成一个writeBatch(通过MergeBatch函数)进行写入.

1. if (w.ShouldWriteToWAL()) {
2. ...............................
3. w.status = WriteToWAL(wal_write_group, log_writer, log_used,
4. need_log_sync, need_log_dir_sync, current_sequence);
5. }
  • 当当前的leader将它自己与它的follow写入之后，此时它将需要写入memtable,那么此时之前还阻塞的Writer，分为两种情况 第一种是已经被当前的leader打包写入到WAL，这些writer(包括leader自己)需要将他们链接到memtable writer list.还有一种情况，那就是还没有写入WAL的，此时这类writer则需要选择一个leader然后继续写入WAL.

```

  1. void WriteThread::ExitAsBatchGroupLeader(WriteGroup& write_group,
  2. Status status) {
  3. Writer* leader = write_group.leader;
  4. Writer* last_writer = write_group.last_writer;
  5. assert(leader->link_older == nullptr);

  6. .....................................

  7. if (enable_pipelined_write_) {

  8. // Notify writers don't write to memtable to exit.
  9. ......................................
  10. // Link the ramaining of the group to memtable writer list.
  11. if (write_group.size > 0) {
  12. if (LinkGroup(write_group, &newest_memtable_writer_)) {
  13. // The leader can now be different from current writer.
  14. SetState(write_group.leader, STATE_MEMTABLE_WRITER_LEADER);
  15. }
  16. }
  17. // Reset newest_writer_ and wake up the next leader.
  18. Writer* newest_writer = last_writer;
  19. if (!newest_writer_.compare_exchange_strong(newest_writer, nullptr)) {
  20. Writer* next_leader = newest_writer;
  21. while (next_leader->link_older != last_writer) {
  22. next_leader = next_leader->link_older;
  23. assert(next_leader != nullptr);
  24. }
  25. next_leader->link_older = nullptr;
  26. SetState(next_leader, STATE_GROUP_LEADER);
  27. }
  28. AwaitState(leader, STATE_MEMTABLE_WRITER_LEADER |
  29. STATE_PARALLEL_MEMTABLE_WRITER | STATE_COMPLETED,
  30. &eabgl_ctx);
  31. } else {
  32. .....................................
  33. }
  34. } ``` - 接下来我们来看写入memtable的操作，这里逻辑类似写入WAL,如果是leader的话，则依旧会创建一个group(WriteGroup),然后遍历需要写入memtable的writer,将他们都加入到group中(EnterAsMemTableWriter),然后则设置并发执行的大小，以及设置对应状态(LaunchParallelMemTableWriters).这里注意每次setstate就将会唤醒之前阻塞的Writer.

```

  1. void WriteThread::LaunchParallelMemTableWriters(WriteGroup* write_group) {
  2. assert(write_group != nullptr);
  3. write_group->running.store(write_group->size);
  4. for (auto w : *write_group) {
  5. SetState(w, STATE_PARALLEL_MEMTABLE_WRITER);
  6. }
  7. } ``` - 这里要注意，在构造memtable的group的时候，我们不需要创建link_newer，因为之前在写入WAL的时候，我们已经构造好link_newer,那么此时我们使用构造好的group也就是表示这个group中包含的都是已经写入到WAL的操作.

```

  1. void WriteThread::EnterAsMemTableWriter(Writer* leader,
  2. WriteGroup* write_group) {

  3. ....................................

  4. if (!allow_concurrent_memtable_write_ || !leader->batch->HasMerge()) {

  5. ....................................................

  6. }

  7. write_group->last_writer = last_writer;

  8. write_group->last_sequence =
  9. last_writer->sequence + WriteBatchInternal::Count(last_writer->batch) - 1;
  10. } ``` - 最后开始执行写入MemTable的操作，之前在写入WAL的时候被阻塞的所有Writer此时都会进入下面这个逻辑，此时也就意味着 并发写入MemTable．

```

  1. if (w.state == WriteThread::STATE_PARALLEL_MEMTABLE_WRITER) {
  2. ．．．．．．．．．．．．．．．．．．．．．．．．．
  3. w.status = WriteBatchInternal::InsertInto(
  4. &w, w.sequence, &column_family_memtables, &flush_scheduler_,
  5. write_options.ignore_missing_column_families, 0 /log_number/, this,
  6. true /concurrent_memtable_writes/);
  7. if (write_thread_.CompleteParallelMemTableWriter(&w)) {
  8. MemTableInsertStatusCheck(w.status);
  9. versions_->SetLastSequence(w.write_group->last_sequence);
  10. write_thread_.ExitAsMemTableWriter(&w, *w.write_group);
  11. }
  12. } ``` - 最后当当前group的所有Writer都写入MemTable之后，则将会调用ExitAsMemTableWriter来进行收尾工作.如果有新的memtable writer list需要处理，那么则唤醒对应的Writer,然后设置已经处理完毕的Writer的状态.

```

  1. void WriteThread::ExitAsMemTableWriter(Writer /self*/,
  2. WriteGroup& write_group) {
  3. Writer* leader = write_group.leader;

  4. Writer* last_writer = write_group.last_writer;

  5. Writer* newest_writer = last_writer;

  6. if (!newest_memtable_writer_.compare_exchange_strong(newest_writer,
  7. nullptr)) {
  8. CreateMissingNewerLinks(newest_writer);
  9. Writer* next_leader = last_writer->link_newer;
  10. assert(next_leader != nullptr);
  11. next_leader->link_older = nullptr;
  12. SetState(next_leader, STATE_MEMTABLE_WRITER_LEADER);
  13. }
  14. Writer* w = leader;
  15. while (true) {
  16. if (!write_group.status.ok()) {
  17. w->status = write_group.status;
  18. }
  19. Writer* next = w->link_newer;
  20. if (w != leader) {
  21. SetState(w, STATE_COMPLETED);
  22. }
  23. if (w == last_writer) {
  24. break;
  25. }
  26. w = next;
  27. }
  28. // Note that leader has to exit last, since it owns the write group.
  29. SetState(leader, STATE_COMPLETED);
  30. } ```

总结

我们可以看到在RocksDB中，WAL的写入始终是串行写入，而MemTable可以多线程并发写入，也就是说在系统压力到一定阶段的时候， 写入WAL肯定会成为瓶颈.

原文：http://mysql.taobao.org/monthly/2018/07/04/