Flink推测机制

1、配置

execution.batch.speculative.enabled：false，推测机制开关，必须在AdaptiveBatchScheduler模式下使用

execution.batch.speculative.max-concurrent-executions：2，同时最多几次执行

execution.batch.speculative.block-slow-node-duration：1分钟，慢速节点会如黑名单，控制在黑名单中的时长

slow-task-detector.check-interval：1秒，慢任务检查间隔

slow-task-detector.execution-time.baseline-lower-bound：1分钟，慢任务检测基线的下限

slow-task-detector.execution-time.baseline-ratio：0.75，开始检测慢任务基线的任务完成率，即有75%任务完成后，开始计算剩下的任务是否为慢任务

slow-task-detector.execution-time.baseline-multiplier：1.5，慢任务基线乘数

2、SpeculativeScheduler

推测机制在AdaptiveBatchScheduler模式下使用，在AdaptiveBatchSchedulerFactory当中，创建调度器时，如果开启了推测机制，会创建SpeculativeScheduler

if (enableSpeculativeExecution) {
    return new SpeculativeScheduler(
            log,
            jobGraph,
            ioExecutor,
            jobMasterConfiguration,

2.1、启动

调度器启动时有三个操作：1、注册指标；2、父类通用的启动流程，会有算子的一些初始化；3、启动慢任务检测任务

protected void startSchedulingInternal() {
    registerMetrics(jobManagerJobMetricGroup);

    super.startSchedulingInternal();
    slowTaskDetector.start(getExecutionGraph(), this, getMainThreadExecutor());
}

2.2、SlowTaskDetector

SlowTaskDetector负责检测慢任务，实现类是ExecutionTimeBasedSlowTaskDetector，基于schedule进行检测

this.scheduledDetectionFuture =
        mainThreadExecutor.schedule(
                () -> {
                    listener.notifySlowTasks(findSlowTasks(executionGraph));
                    scheduleTask(executionGraph, listener, mainThreadExecutor);
                },
                checkIntervalMillis,
                TimeUnit.MILLISECONDS);

核心是findSlowTasks，首先是获取需要校验的拓扑集

private List<ExecutionJobVertex> getJobVerticesToCheck(final ExecutionGraph executionGraph) {
    return IterableUtils.toStream(executionGraph.getVerticesTopologically())
            .filter(ExecutionJobVertex::isInitialized)
            .filter(ejv -> ejv.getAggregateState() != ExecutionState.FINISHED)
            .filter(ejv -> getFinishedRatio(ejv) >= baselineRatio)
            .collect(Collectors.toList());
}

getFinishedRatio就是获取完成任务数超过基线比率的，就是拓扑集中完成任务数和总任务数的比值

private double getFinishedRatio(final ExecutionJobVertex executionJobVertex) {
    checkState(executionJobVertex.getTaskVertices().length > 0);
    long finishedCount =
            Arrays.stream(executionJobVertex.getTaskVertices())
                    .filter(ev -> ev.getExecutionState() == ExecutionState.FINISHED)
                    .count();
    return (double) finishedCount / executionJobVertex.getTaskVertices().length;
}

接下来是获取基线和在基线基础上计算慢速任务的，接口是getBaseline和findExecutionsExceedingBaseline，本质就是执行时间和基线的对比，注意这里不仅用到了时间，还用到了输入字节数，所以慢任务的检测可能是基于吞吐来的

private ExecutionTimeWithInputBytes getBaseline(
        final ExecutionJobVertex executionJobVertex, final long currentTimeMillis) {
    final ExecutionTimeWithInputBytes weightedExecutionTimeMedian =
            calculateFinishedTaskExecutionTimeMedian(executionJobVertex, currentTimeMillis);
    long multipliedBaseline =
            (long) (weightedExecutionTimeMedian.getExecutionTime() * baselineMultiplier);

    return new ExecutionTimeWithInputBytes(
            multipliedBaseline, weightedExecutionTimeMedian.getInputBytes());
}

return Double.compare(
        (double) executionTime / Math.max(inputBytes, Double.MIN_VALUE),
        (double) other.getExecutionTime()
                / Math.max(other.getInputBytes(), Double.MIN_VALUE));

2.3、notifySlowTasks

获取慢速任务以后，SlowTaskDetector会触发监听器，监听器的处理实现在SpeculativeScheduler的notifySlowTasks接口

首先把节点加入黑名单

// add slow nodes to blocklist before scheduling new speculative executions
blockSlowNodes(slowTasks, currentTimestamp);

这边会检测任务是否支持推测，默认是支持

if (!executionVertex.isSupportsConcurrentExecutionAttempts()) {
    continue;
}

基于时间戳，对慢任务新建Execution

final Collection<Execution> attempts =
        IntStream.range(0, newSpeculativeExecutionsToDeploy)
                .mapToObj(
                        i ->
                                executionVertex.createNewSpeculativeExecution(
                                        currentTimestamp))
                .collect(Collectors.toList());

之后会进行一系列的配置，加入监控

setupSubtaskGatewayForAttempts(executionVertex, attempts);
verticesToDeploy.add(executionVertexId);
newSpeculativeExecutions.addAll(attempts);

最后发起调度

executionDeployer.allocateSlotsAndDeploy(
        newSpeculativeExecutions,
        executionVertexVersioner.getExecutionVertexVersions(verticesToDeploy));

3、任务结束

任务结束主要核心在DefaultExecutionGraph的jobFinished，判断在上层ExecutionJobVertex.executionVertexFinished，这里是通过任务并行度来判断的，所有子任务完成则认为job完成

void executionVertexFinished() {
    checkState(isInitialized());
    numExecutionVertexFinished++;
    if (numExecutionVertexFinished == parallelismInfo.getParallelism()) {
        getGraph().jobVertexFinished();
    }
}

这个的调用是由Execution触发的，也就是每个子任务完成会去调用一次

if (transitionState(current, FINISHED)) {
    try {
        finishPartitionsAndUpdateConsumers();
        updateAccumulatorsAndMetrics(userAccumulators, metrics);
        releaseAssignedResource(null);
        vertex.getExecutionGraphAccessor().deregisterExecution(this);
    } finally {
        vertex.executionFinished(this);
    }
    return;
}

最终一个jobVertex（对应Job的一个任务，任务根据并行度有子任务）完成的时候会通知所有子任务完成

public void jobVertexFinished() {
    assertRunningInJobMasterMainThread();
    final int numFinished = ++numFinishedJobVertices;
    if (numFinished == numJobVerticesTotal) {
        FutureUtils.assertNoException(
                waitForAllExecutionsTermination().thenAccept(ignored -> jobFinished()));
    }
}