roman_日积跬步-终至千里

roman_日积跬步-终至千里

关注
发信
关注(28)粉丝(399)

【Flink状态管理（八）】Checkpoint：CheckpointBarrier对齐后Checkpoint的完成、通知与对学习状态管理源码的思考

上文介绍了CheckpointBarrier的对齐操作，当CheckpointBarrier完成对齐操作后，接下来就是通过notifyCheckpoint()方法触发StreamTask节点的Checkpoint操作。

一. 调用StreamTask执行Checkpoint操作

如下代码，notifyCheckpoint()方法主要包含如下逻辑。

> 1. 判断toNotifyOnCheckpoint不为空。
> 2. 创建CheckpointMetaData和CheckpointMetrics实例，CheckpointMetaData用于存储
> Checkpoint的元信息，CheckpointMetrics用于记录和监控Checkpoint监控指标。
> 3. 触发StreamTask中算子的Checkpoint操作。
protected void notifyCheckpoint(CheckpointBarrier checkpointBarrier, 
                                long bufferedBytes, 
                                long alignmentDurationNanos) throws Exception {
   if (toNotifyOnCheckpoint != null) {
      // 创建CheckpointMetaData对象用于存储Meta信息
      CheckpointMetaData checkpointMetaData =
         new CheckpointMetaData(checkpointBarrier.getId(), 
                                checkpointBarrier.getTimestamp());
            // 创建CheckpointMetrics对象用于记录监控指标
      CheckpointMetrics checkpointMetrics = new CheckpointMetrics()
         .setBytesBufferedInAlignment(bufferedBytes)
         .setAlignmentDurationNanos(alignmentDurationNanos);
      // 调用toNotifyOnCheckpoint.triggerCheckpointOnBarrier()方法触发Checkpoint
        操作
      toNotifyOnCheckpoint.triggerCheckpointOnBarrier(
         checkpointMetaData,
         checkpointBarrier.getCheckpointOptions(),
         checkpointMetrics);
   }
}

注意：StreamTask是唯一实现了Checkpoint方法的子类，即只有StreamTask才能触发当前Task实例中的Checkpoint操作。

 

接下来具体看Checkpoint执行细节

1. 执行Checkpoint总体代码流程

不管是哪种方式触发Checkpoint，最终都是调用StreamTask.performCheckpoint()方法实现StreamTask实例中状态数据的持久化操作。

 

在StreamTask.performCheckpoint()方法中，首先判断当前的Task是否运行正常，然后使用actionExecutor线程池执行Checkpoint操作，Checkpoint的实际执行过程如下。

  1. task挂掉情况处理：
private boolean performCheckpoint(
      CheckpointMetaData checkpointMetaData,
      CheckpointOptions checkpointOptions,
      CheckpointMetrics checkpointMetrics,
      boolean advanceToEndOfTime) throws Exception {
   LOG.debug("Starting checkpoint ({}) {} on task {}",
             checkpointMetaData.getCheckpointId(), 
             checkpointOptions.getCheckpointType(), 
             getName());
   final long checkpointId = checkpointMetaData.getCheckpointId();
   if (isRunning) {
      // 使用actionExecutor执行Checkpoint逻辑
      actionExecutor.runThrowing(() -> {
         if (checkpointOptions.getCheckpointType().isSynchronous()) {
             setSynchronousSavepointId(checkpointId);
             if (advanceToEndOfTime) {
                 advanceToEndOfEventTime();
            }
         }
         //Checkpoint操作的准备工作
         operatorChain.prepareSnapshotPreBarrier(checkpointId);
         //将checkpoint barrier发送到下游的stream中
         operatorChain.broadcastCheckpointBarrier(
               checkpointId,
               checkpointMetaData.getTimestamp(),
               checkpointOptions);
         //对算子中的状态进行快照操作，此步骤是异步操作，
         //不影响streaming拓扑中数据的正常处理
         checkpointState(checkpointMetaData, checkpointOptions, 
            checkpointMetrics);
      });
      return true;
   } else {
      // 如果Task处于其他状态，则向下游广播CancelCheckpointMarker消息
      actionExecutor.runThrowing(() -> {
         final CancelCheckpointMarker message = 
             new CancelCheckpointMarker(checkpointMetaData.getCheckpointId());
         recordWriter.broadcastEvent(message);
      });
      return false;
   }
}

 

1.1. StreamTask.checkpointState()

接下来我们看StreamTask.checkpointState()方法的具体实现，如下代码。

private void checkpointState(
      CheckpointMetaData checkpointMetaData,
      CheckpointOptions checkpointOptions,
      CheckpointMetrics checkpointMetrics) throws Exception {
     // 创建CheckpointStreamFactory实例
   CheckpointStreamFactory storage = checkpointStorage.resolveCheckpointStorag
      eLocation(
         checkpointMetaData.getCheckpointId(),
         checkpointOptions.getTargetLocation());
     // 创建CheckpointingOperation实例
   CheckpointingOperation checkpointingOperation = new CheckpointingOperation(
      this,
      checkpointMetaData,
      checkpointOptions,
      storage,
      checkpointMetrics);
   // 执行Checkpoint操作
   checkpointingOperation.executeCheckpointing();
}

 

1.2. executeCheckpointing

如代码所示，CheckpointingOperation.executeCheckpointing()方法主要包含如下逻辑。

public void executeCheckpointing() throws Exception {
   //通过算子创建执行快照操作的OperatorSnapshotFutures对象
   for (StreamOperator<?> op : allOperators) {
      checkpointStreamOperator(op);
   }
   // 此处省略部分代码
   startAsyncPartNano = System.nanoTime();
   checkpointMetrics.setSyncDurationMillis(
      (startAsyncPartNano - startSyncPartNano) / 1_000_000);
   AsyncCheckpointRunnable asyncCheckpointRunnable = new 
      AsyncCheckpointRunnable(
      owner,
      operatorSnapshotsInProgress,
      checkpointMetaData,
      checkpointMetrics,
      startAsyncPartNano);
   // 注册Closeable操作
   owner.cancelables.registerCloseable(asyncCheckpointRunnable);
   // 执行asyncCheckpointRunnable
         owner.asyncOperationsThreadPool.execute(asyncCheckpointRunnable);
 }

 

1.3. 将算子中的状态快照操作封装在OperatorSnapshotFutures中

如下代码，AbstractStreamOperator.snapshotState()方法将当前算子的状态快照操作封装在OperatorSnapshotFutures对象中，然后通过asyncOperationsThreadPool线程池异步触发所有的OperatorSnapshotFutures操作，方法主要步骤如下。

  1. 向snapshotInProgress中指定KeyedStateRawFuture和OperatorStateRawFuture，专门用于处理原生状态数据的快照操作。
  1. 返回创建的snapshotInProgress异步Future对象，snapshotInProgress中封装了当前算子需要执行的所有快照操作。
public final OperatorSnapshotFutures snapshotState(long checkpointId, 
                                                   long timestamp, 
                                                   CheckpointOptions 
                                                   checkpointOptions,
                                                   CheckpointStreamFactory factory
                                                   ) throws Exception {
      // 获取KeyGroupRange
   KeyGroupRange keyGroupRange = null != keyedStateBackend ?
         keyedStateBackend.getKeyGroupRange() : KeyGroupRange.EMPTY_KEY_GROUP_
            RANGE;
      // 创建OperatorSnapshotFutures处理对象
   OperatorSnapshotFutures snapshotInProgress = new OperatorSnapshotFutures();
      // 创建snapshotContext上下文对象
   StateSnapshotContextSynchronousImpl snapshotContext = 
   new StateSnapshotContextSynchronousImpl(
      checkpointId,
      timestamp,
      factory,
      keyGroupRange,
      getContainingTask().getCancelables());
   try {
      snapshotState(snapshotContext);
      // 设定KeyedStateRawFuture和OperatorStateRawFuture
      snapshotInProgress
      .setKeyedStateRawFuture(snapshotContext.getKeyedStateStreamFuture());
      snapshotInProgress
      .setOperatorStateRawFuture(snapshotContext.getOperatorStateStreamFuture());
            // 如果operatorStateBackend不为空，设定OperatorStateManagedFuture
      if (null != operatorStateBackend) {
         snapshotInProgress.setOperatorStateManagedFuture(
            operatorStateBackend
            .snapshot(checkpointId, timestamp, factory, checkpointOptions));
      }
      // 如果keyedStateBackend不为空，设定KeyedStateManagedFuture
      if (null != keyedStateBackend) {
         snapshotInProgress.setKeyedStateManagedFuture(
            keyedStateBackend
            .snapshot(checkpointId, timestamp, factory, checkpointOptions));
      }
   } catch (Exception snapshotException) {
    // 此处省略部分代码
   }
   return snapshotInProgress;
}

这里可以看出，原生状态和管理状态的RunnableFuture对象会有所不同

 

1.4. 算子状态进行快照

我们知道所有的状态快照操作都会被封装到OperatorStateManagedFuture对象中，最终通过AsyncCheckpointRunnable线程触发执行。

下面我们看AsyncCheckpointRunnable线程的定义。如代码所示，AsyncCheckpointRunnable.run()方法主要逻辑如下。

  1. 执行所有状态快照线程操作
  1. 从finalizedSnapshots中获取JobManagerOwnedState和TaskLocalState，分别存储在jobManagerTaskOperatorSubtaskStates和localTaskOperatorSubtaskStates集合中。
  2. 调用checkpointMetrics对象记录Checkpoint执行的时间并汇总到Metric监控系统中。
  3. 如果AsyncCheckpointState为COMPLETED状态，则调用reportCompletedSnapshotStates()方法向JobManager汇报Checkpoint的执行结果。
  4. 如果出现其他异常情况，则调用handleExecutionException()方法进行处理。
public void run() {
   FileSystemSafetyNet.initializeSafetyNetForThread();
   try {
      // 创建TaskStateSnapshot
      TaskStateSnapshot jobManagerTaskOperatorSubtaskStates =
         new TaskStateSnapshot(operatorSnapshotsInProgress.size());
      TaskStateSnapshot localTaskOperatorSubtaskStates =
         new TaskStateSnapshot(operatorSnapshotsInProgress.size());
      for (Map.Entry<OperatorID, OperatorSnapshotFutures> entry : 
           operatorSnapshotsInProgress.entrySet()) {
         OperatorID operatorID = entry.getKey();
         OperatorSnapshotFutures snapshotInProgress = entry.getValue();
         // 创建OperatorSnapshotFinalizer对象
         OperatorSnapshotFinalizer finalizedSnapshots =
            new OperatorSnapshotFinalizer(snapshotInProgress);
         jobManagerTaskOperatorSubtaskStates.putSubtaskStateByOperatorID(
            operatorID,
            finalizedSnapshots.getJobManagerOwnedState());
         localTaskOperatorSubtaskStates.putSubtaskStateByOperatorID(
            operatorID,
            finalizedSnapshots.getTaskLocalState());
      }
      final long asyncEndNanos = System.nanoTime();
      final long asyncDurationMillis = (asyncEndNanos - asyncStartNanos) / 1_000_000L;
      checkpointMetrics.setAsyncDurationMillis(asyncDurationMillis);
      if (asyncCheckpointState.compareAndSet(
          CheckpointingOperation.AsyncCheckpointState.RUNNING,
         CheckpointingOperation.AsyncCheckpointState.COMPLETED)) {
         reportCompletedSnapshotStates(
            jobManagerTaskOperatorSubtaskStates,
            localTaskOperatorSubtaskStates,
            asyncDurationMillis);
      } else {
         LOG.debug("{} - asynchronous part of checkpoint {} could not be 
            completed because it was closed before.",
            owner.getName(),
            checkpointMetaData.getCheckpointId());
      }
   } catch (Exception e) {
      handleExecutionException(e);
   } finally {
      owner.cancelables.unregisterCloseable(this);
      FileSystemSafetyNet.closeSafetyNetAndGuardedResourcesForThread();
   }
}

至此，算子状态数据快照的逻辑基本完成，算子中的托管状态主要借助KeyedStateBackend和OperatorStateBackend管理。

KeyedStateBackend和OperatorStateBackend都实现了SnapshotStrategy接口，提供了状态快照的方法。SnapshotStrategy根据不同类型存储后端，主要有HeapSnapshotStrategy和RocksDBSnapshotStrategy两种类型。

 

1.5. 状态数据快照持久化

这里我们以HeapSnapshotStrategy为例，介绍在StateBackend中对状态数据进行状态快照持久化操作的步骤。如代码所示，

HeapSnapshotStrategy.processSnapshotMetaInfoForAllStates()方法中定义了对KeyedState以及OperatorState的状态处理逻辑。

private void processSnapshotMetaInfoForAllStates(
   List metaInfoSnapshots,
   Map<StateUID, StateSnapshot> cowStateStableSnapshots,
   Map<StateUID, Integer> stateNamesToId,
   Map<String, ? extends StateSnapshotRestore> registeredStates,
   StateMetaInfoSnapshot.BackendStateType stateType) {
   for (Map.Entry<String, ? extends StateSnapshotRestore> kvState :
        registeredStates.entrySet()) {
      final StateUID stateUid = StateUID.of(kvState.getKey(), stateType);
      stateNamesToId.put(stateUid, stateNamesToId.size());
      StateSnapshotRestore state = kvState.getValue();
      if (null != state) {
         final StateSnapshot stateSnapshot = state.stateSnapshot();
         metaInfoSnapshots.add(stateSnapshot.getMetaInfoSnapshot());
         cowStateStableSnapshots.put(stateUid, stateSnapshot);
      }
   }
}

 

二. CheckpointCoordinator管理Checkpoint

1. Checkpoint执行完毕后的确认过程

当StreamTask中所有的算子完成状态数据的快照操作后，Task实例会立即将TaskStateSnapshot消息发送到管理节点的CheckpointCoordinator中，并在CheckpointCoordinator中完成后续的操作。如图所示，Checkpoint执行完毕后的确认过程如下。

【Flink状态管理（八）】Checkpoint：CheckpointBarrier对齐后Checkpoint的完成、通知与对学习状态管理源码的思考-LMLPHP

  1. 消息传递
  1. 管理PendingCheckpoint
  1. 添加CompletedCheckpoint：
  1. 通知Checkpoint操作结束。
  1. 通知同步

 

2. 触发并完成Checkpoint操作

CheckpointCoordinator组件接收到Task实例的Ack消息（快照完成了？）后，会触发并完成Checkpoint操作。如代码PendingCheckpoint.finalizeCheckpoint()方法的具体实现如下。

1）向sharedStateRegistry中注册operatorStates。
2）结束pendingCheckpoint中的Checkpoint操作并生成CompletedCheckpoint。
3）将completedCheckpoint添加到completedCheckpointStore中，
4）从pendingCheckpoint中移除checkpointId对应的PendingCheckpoint，
并触发队列中的Checkpoint请求。
5）向所有的ExecutionVertex节点发送CheckpointComplete消息，
通知Task实例本次Checkpoint操作完成。



private void completePendingCheckpoint(PendingCheckpoint pendingCheckpoint) 
   throws CheckpointException {
   final long checkpointId = pendingCheckpoint.getCheckpointId();
   final CompletedCheckpoint completedCheckpoint;
   // 首先向sharedStateRegistry中注册operatorStates
   Map<OperatorID, OperatorState> operatorStates = 
      pendingCheckpoint.getOperatorStates();
   sharedStateRegistry.registerAll(operatorStates.values());
   // 对pendingCheckpoint中的Checkpoint做结束处理并生成CompletedCheckpoint
   try {
      try {
         completedCheckpoint = pendingCheckpoint.finalizeCheckpoint();
         failureManager.handleCheckpointSuccess(pendingCheckpoint.
            getCheckpointId());
      }
      catch (Exception e1) {
         // 如果出现异常则中止运行并抛出CheckpointExecution
         if (!pendingCheckpoint.isDiscarded()) {
             failPendingCheckpoint(pendingCheckpoint,
                                   CheckpointFailureReason.FINALIZE_CHECKPOINT_
                                        FAILURE, e1);
         }
         throw new CheckpointException("Could not finalize the pending 
                                       checkpoint " +
                                       checkpointId + '.',
                                       CheckpointFailureReason
                                       .FINALIZE_CHECKPOINT_FAILURE, e1);
      }
      // 当完成finalization后，PendingCheckpoint必须被丢弃
      Preconditions.checkState(pendingCheckpoint.isDiscarded() 
                               && completedCheckpoint != null);
      // 将completedCheckpoint添加到completedCheckpointStore中
      try {
         completedCheckpointStore.addCheckpoint(completedCheckpoint);
      } catch (Exception exception) {
         // 如果completed checkpoint存储出现异常则进行清理
         executor.execute(new Runnable() {
            @Override
            public void run() {
               try {
                  completedCheckpoint.discardOnFailedStoring();
               } catch (Throwable t) {
                  LOG.warn("Could not properly discard completed checkpoint {}.",
                           completedCheckpoint.getCheckpointID(), t);
               }
            }
         });
         throw new CheckpointException("Could not complete the pending 
                                       checkpoint " + 
                                       checkpointId + '.', 
                                       CheckpointFailureReason.
                                       FINALIZE_CHECKPOINT_FAILURE, exception);
      }
   } finally {
      // 最后从pendingCheckpoints中移除checkpointId对应的PendingCheckpoint
      pendingCheckpoints.remove(checkpointId);
      // 触发队列中的Checkpoint请求
      triggerQueuedRequests();
   }
   // 记录checkpointId
   rememberRecentCheckpointId(checkpointId);
   // 清除之前的Checkpoints
   dropSubsumedCheckpoints(checkpointId);
   // 计算和前面Checkpoint操作之间的最低延时
   lastCheckpointCompletionRelativeTime = clock.relativeTimeMillis();
   LOG.info("Completed checkpoint {} for job {} ({} bytes in {} ms).", 
            checkpointId, job,
            completedCheckpoint.getStateSize(), completedCheckpoint.getDuration());
   // 通知所有的ExecutionVertex节点Checkpoint操作完成
   final long timestamp = completedCheckpoint.getTimestamp();
   for (ExecutionVertex ev : tasksToCommitTo) {
      Execution ee = ev.getCurrentExecutionAttempt();
      if (ee != null) {
          ee.notifyCheckpointComplete(checkpointId, timestamp);
      }
   }
}

 

3. 通知CheckpointComplete给TaskExecutor

当TaskExecutor接收到来自CheckpointCoordinator的CheckpointComplete消息后，会调用Task.notifyCheckpointComplete()方法将消息传递到指定的Task实例中。Task线程会将CheckpointComplete消息通知给StreamTask中的算子。

如下代码，

/**
将notifyCheckpointComplete()转换成RunnableWithException线程并提交到Mailbox中运行，且在MailboxExecutor线程模型中获取和执行的优先级是最高的。
最终notifyCheckpointComplete()方法会在MailboxProcessor中运行。
**/

public Future<Void> notifyCheckpointCompleteAsync(long checkpointId) {
   return mailboxProcessor.getMailboxExecutor(TaskMailbox.MAX_PRIORITY).submit(
      () -> notifyCheckpointComplete(checkpointId),
      "checkpoint %d complete", checkpointId);
}

继续具体看StreamTask.notifyCheckpointComplete()，如下代码：

1）获取当前Task中算子链的算子，并发送Checkpoint完成的消息。
2）获取TaskStateManager对象，向其通知Checkpoint完成消息，这里主要调用
TaskLocalStateStore清理本地无用的Checkpoint数据。
3）如果当前Checkpoint是同步的Savepoint操作，直接完成并终止当前Task实例，并调用
resetSynchronousSavepointId()方法将syncSavepointId重置为空。

private void notifyCheckpointComplete(long checkpointId) {
   try {
      boolean success = actionExecutor.call(() -> {
         if (isRunning) {
            LOG.debug("Notification of complete checkpoint for task {}", 
               getName());
            // 获取当前Task中operatorChain所有的Operator，并通知每个Operator 
               Checkpoint执行成功的消息
            for (StreamOperator<?> operator : operatorChain.getAllOperators()) {
               if (operator != null) {
                  operator.notifyCheckpointComplete(checkpointId);
               }
            }
            return true;
         } else {
            LOG.debug("Ignoring notification of complete checkpoint for 
               not-running task {}", getName());
            return true;
         }
      });
      // 获取TaskStateManager，并通知Checkpoint执行完成的消息
      getEnvironment().getTaskStateManager().notifyCheckpointComplete(checkpointId);
      // 如果是同步的Savepoint操作，则直接完成当前Task
      if (success && isSynchronousSavepointId(checkpointId)) {
         finishTask();
         // Reset to "notify" the internal synchronous savepoint mailbox loop.
         resetSynchronousSavepointId();
      }
   } catch (Exception e) {
      handleException(new RuntimeException("Error while confirming checkpoint", e));
   }
}

算子接收到Checkpoint完成消息后，会根据自身需要进行后续的处理，默认在AbstractStreamOperator基本实现类中会通知keyedStateBackend进行后续操作。

对于AbstractUdfStreamOperator实例，会判断当前userFunction是否实现了CheckpointListener，如果实现了，则向UserFucntion通知Checkpoint执行完成的信息

public void notifyCheckpointComplete(long checkpointId) throws Exception {
   super.notifyCheckpointComplete(checkpointId);
   if (userFunction instanceof CheckpointListener) {
      ((CheckpointListener) userFunction).notifyCheckpointComplete(checkpointId);
   }
}

 

三. 状态管理学习小结

通过学习状态管理的源码，我们可以再来思考下如下几个场景问题，是不是有一点“庖丁解牛”的意思！

flink中状态存在的意义是什么，涉及到哪些场景。

  1. 实时聚合：比如，计算过去一小时内的平均销售额。这时，你会需要使用到Flink的状态来存储过去一小时内的所有销售数据。
  2. 窗口操作：Flink SQL支持滚动窗口、滑动窗口、会话窗口等。这些窗口操作都需要Flink的状态来存储在窗口期限内的数据。
  3. 状态的持久化与任务恢复：实时任务挂掉之后，为了快速从上一个点恢复任务，可以使用savepoint和checkpoint。
  4. 多流join：Flink至少存储一个流中的数据，以便于在新的记录到来时进行匹配。

 

其次通过学习Flink状态管理相关源码，可以进一步了解状态管理的细节操作，为解决更加复杂的问题打下理论基础

  1. 深入理解任务运行过程中，各算子状态的流转机制；
  2. 快速定位问题：在遇到实际问题时，能够快速反应出是哪块逻辑出现了问题；
  3. 应对故障：状态管理和Flink容错机制相关，可以了解Flink发生故障时如何保证状态的一致性和可恢复性
  4. 二次开发：可以自定义状态后端，或者拓展优化已有的例如RocksDB状态后端等；
  5. 性能优化：了解了Flink是如何有效的处理和管理状态，就可以优化任务性能，减少资源消耗。

 

参考：《Flink设计与实现：核心原理与源码解析》–张利兵

02-21 18:26
Powered by LMLPHP ©2024 roman_日积跬步-终至千里 0.002059