Flink rocksdb如何做checkpoint

最近在做实时数仓的过程中遇到了一些问题，于是有了这篇博客。

我们知道当设置 backend 为 RocksDBBackend 时，mapState.put 操作最终转化为 rockdb.put 操作，如：

public void put(UK userKey, UV userValue) throws IOException, RocksDBException {

		byte[] rawKeyBytes = serializeCurrentKeyWithGroupAndNamespacePlusUserKey(userKey, userKeySerializer);
		byte[] rawValueBytes = serializeValueNullSensitive(userValue, userValueSerializer);

		backend.db.put(columnFamily, writeOptions, rawKeyBytes, rawValueBytes);

数据是存起来了，但是当进行 checkpoint 的时候， rocksdb 又做了什么呢？

这就要从 RocksDBKeyedStateBackend 说起了了，这个类很好的说明了 checkpoint 与 rocksdb 还有 hdfs 的关系

//当进行 checkpoint 的时候并且要对 keyed state 做 snapshot 时，会触发此方法
	public RunnableFuture> snapshot(
		final long checkpointId,
		final long timestamp,
		@Nonnull final CheckpointStreamFactory streamFactory,
		@Nonnull CheckpointOptions checkpointOptions) throws Exception {

		long startTime = System.currentTimeMillis();

		// flush everything into db before taking a snapshot
		writeBatchWrapper.flush();
		
    //是进行 savepointSnapshotStrategy 还是 checkpointSnapshotStrategy
    //实际上 savepointSnapshotStrategy 和 checkpointSnapshotStrategy 均是 RocksDBKeyedStateBackend 的属性，均是 RocksDBSnapshotStrategyBase。
    //我们知道 RocksDBStateBackend 有两种 snapshot 策略，一种是 full snapshot ，一种是 incremental snapshot。这两种策略分别对应 RocksFullSnapshotStrategy 和 RocksIncrementalSnapshotStrategy。
    //RocksDBSnapshotStrategyBase 为 RocksFullSnapshotStrategy 和 RocksIncrementalSnapshotStrategy 的父类
		RocksDBSnapshotStrategyBase chosenSnapshotStrategy =
			CheckpointType.SAVEPOINT == checkpointOptions.getCheckpointType() ?
				savepointSnapshotStrategy : checkpointSnapshotStrategy;

    //开始进行 snapshot，此处为多态，我们以 RocksIncrementalSnapshotStrategy 为主,交叉查看
		RunnableFuture> snapshotRunner =
			chosenSnapshotStrategy.snapshot(checkpointId, timestamp, streamFactory, checkpointOptions);

		chosenSnapshotStrategy.logSyncCompleted(streamFactory, startTime);

		return snapshotRunner;
	}

进入 RocksIncrementalSnapshotStrategy

@Nonnull
	@Override
	protected RunnableFuture> doSnapshot(
		long checkpointId,
		long checkpointTimestamp,
		@Nonnull CheckpointStreamFactory checkpointStreamFactory,
		@Nonnull CheckpointOptions checkpointOptions) throws Exception {

		final SnapshotDirectory snapshotDirectory = prepareLocalSnapshotDirectory(checkpointId);
		LOG.trace("Local RocksDB checkpoint goes to backup path {}.", snapshotDirectory);

		// RocksDBIncrementalRestoreOperation 中 kvStateInformation 赋值(RocksFullSnapshotStrategy 和 RocksIncrementalSnapshotStrategy 对应的 kvStateInformation 是一样的 )
		//		kvStateInformation.put(columnFamilyName, registeredColumn(RocksDBKeyedStateBackend.RocksDbKvStateInfo));
		final List stateMetaInfoSnapshots = new ArrayList<>(kvStateInformation.size());
    //对 meta data 做全量 snapshot ，并将结果赋值给 stateMetaInfoSnapshots
    final Set baseSstFiles = snapshotMetaData(checkpointId, stateMetaInfoSnapshots);

    //对 rocksdb 做 checkpoint 为 RocksDBIncrementalSnapshotOperation.uploadSstFiles 做准备
		takeDBNativeCheckpoint(snapshotDirectory);

		// snapshot
		final RocksDBIncrementalSnapshotOperation snapshotOperation =
			new RocksDBIncrementalSnapshotOperation(
				checkpointId,
				checkpointStreamFactory,
				snapshotDirectory,
				baseSstFiles,
				stateMetaInfoSnapshots);
		
    // 执行增量快照
		return snapshotOperation.toAsyncSnapshotFutureTask(cancelStreamRegistry);
	}

我们来具体看一下 RocksDBIncrementalSnapshotOperation 是怎么执行的

@Override
		protected SnapshotResult callInternal() throws Exception {

			boolean completed = false;

			// Handle to the meta data file
			SnapshotResult metaStateHandle = null;
			// Handles to new sst files since the last completed checkpoint will go here
			final Map sstFiles = new HashMap<>();
			// Handles to the misc files in the current snapshot will go here
			final Map miscFiles = new HashMap<>();

			try {

				// 写 meta 到 hdfs
				metaStateHandle = materializeMetaData();

				// Sanity checks - they should never fail
				Preconditions.checkNotNull(metaStateHandle, "Metadata was not properly created.");
				Preconditions.checkNotNull(metaStateHandle.getJobManagerOwnedSnapshot(),
					"Metadata for job manager was not properly created.");

				// 将新产生的 sst file、misc file upload to checkpointFs
				uploadSstFiles(sstFiles, miscFiles);

				synchronized (materializedSstFiles) {
					materializedSstFiles.put(checkpointId, sstFiles.keySet());
				}

				final IncrementalRemoteKeyedStateHandle jmIncrementalKeyedStateHandle =
					new IncrementalRemoteKeyedStateHandle(
						backendUID,
						keyGroupRange,
						checkpointId,
						sstFiles,
						miscFiles,
						metaStateHandle.getJobManagerOwnedSnapshot());

				//PermanentSnapshotDirectory
				final DirectoryStateHandle directoryStateHandle = localBackupDirectory.completeSnapshotAndGetHandle();
				final SnapshotResult snapshotResult;
				if (directoryStateHandle != null && metaStateHandle.getTaskLocalSnapshot() != null) {

					//增量的 localSnapshot
					IncrementalLocalKeyedStateHandle localDirKeyedStateHandle =
						new IncrementalLocalKeyedStateHandle(
							backendUID,
							checkpointId,
							directoryStateHandle,
							keyGroupRange,
							metaStateHandle.getTaskLocalSnapshot(),
							sstFiles.keySet());

					// localSnapshot report to local state manager, jobManagerState(jmIncrementalKeyedStateHandle) report to job manager
					snapshotResult = SnapshotResult.withLocalState(jmIncrementalKeyedStateHandle, localDirKeyedStateHandle);
				} else {
					//jobManagerState(jmIncrementalKeyedStateHandle) report to job manager
					snapshotResult = SnapshotResult.of(jmIncrementalKeyedStateHandle);
				}

				completed = true;

				return snapshotResult;
			} finally {
				if (!completed) {
					final List statesToDiscard =
						new ArrayList<>(1 + miscFiles.size() + sstFiles.size());
					statesToDiscard.add(metaStateHandle);
					statesToDiscard.addAll(miscFiles.values());
					statesToDiscard.addAll(sstFiles.values());
					cleanupIncompleteSnapshot(statesToDiscard);
				}
			}
		}

上述就是 rocksdb 做 checkpoint 的全过程了，从中可以看出元数据是全量更新的，具体的 state data 是增量更新的。

那么元数据和 state data 具体是什么呢？让我们来感性的认识一下
先看 meta data：
meta data 主要就是一些元数据，像:rocksdb.column.family.name以及 state data的全目录。

下面来看一下 state data：
主要是 rocksdb 的一些配置信息以及当前的 db 目录，
主要就是状态中(比如 mapState )存储的一些东西的具体。

我们都知道 checkpoint 是异步的，那么拥有 key state 的 operator 进行 notifyCheckpointComplete 的呢？

首先是 AbstractUdfStreamOperator.notifyCheckpointComplete ------> 
AbstractStreamOperator.notifyCheckpointComplete --------> 
RocksDBKeyedStateBackend.notifyCheckpointComplete --------->
RocksIncrementalSnapshotStrategy.notifyCheckpointComplete

到该处为止，对应的 key state 状态的算子对应的增量 checkpoint 就做完了。