Flink源码阅读之基于Flink1.10的任务执行流程

在基于Flink1.10的任务提交流程中分析了任务的提交流程。本文基于前文基础上进行job执行流程的源码分析。

前文这里已经已经说明，执行流程就从resetAndStartScheduler（）开始

private Acknowledge startJobExecution(JobMasterId newJobMasterId) throws Exception {

        //.......

        startJobMasterServices();
        resetAndStartScheduler();//这里开始正式的任务调度过程了
        //......
    }

通过SchedulerNG来调度job的运行

private void resetAndStartScheduler() throws Exception {
//......
final SchedulerNG newScheduler = createScheduler(newJobManagerJobMetricGroup);//通过SchedulerNGFactory来实例化SchedulerNG ，里面包含了ExecutionGraph的生成
//......
schedulerAssignedFuture.thenRun(this::startScheduling);
}

SchedulerNGFactory有两个实现：DefaultSchedulerFactory和LegacySchedulerFactory分别创建DefaultScheduler和LegacyScheduler实例，这两者都继承SchedulerBase，实例化时都会调用SchedulerBase的构造方法，其中会构造ExecutionGraph

this.executionGraph = createAndRestoreExecutionGraph(jobManagerJobMetricGroup, checkNotNull(shuffleMaster), checkNotNull(partitionTracker));

private ExecutionGraph createAndRestoreExecutionGraph(
        JobManagerJobMetricGroup currentJobManagerJobMetricGroup,
        ShuffleMaster shuffleMaster,
        JobMasterPartitionTracker partitionTracker) throws Exception {
ExecutionGraph newExecutionGraph = createExecutionGraph(currentJobManagerJobMetricGroup, shuffleMaster, partitionTracker);
//.....
}

private ExecutionGraph createExecutionGraph(
        JobManagerJobMetricGroup currentJobManagerJobMetricGroup,
        ShuffleMaster shuffleMaster,
        final JobMasterPartitionTracker partitionTracker) throws JobExecutionException, JobException {

        final FailoverStrategy.Factory failoverStrategy = legacyScheduling ?
            FailoverStrategyLoader.loadFailoverStrategy(jobMasterConfiguration, log) :
            new NoOpFailoverStrategy.Factory();
        //调用ExecutionGraphBuilder.buildGraph来构造ExecutionGraph
        return ExecutionGraphBuilder.buildGraph(
            null,
            jobGraph,
            jobMasterConfiguration,
            futureExecutor,
            ioExecutor,
            slotProvider,
            userCodeLoader,
            checkpointRecoveryFactory,
            rpcTimeout,
            restartStrategy,
            currentJobManagerJobMetricGroup,
            blobWriter,
            slotRequestTimeout,
            log,
            shuffleMaster,
            partitionTracker,
            failoverStrategy);
    }

然后通过startScheduling进行调度

private void startScheduling() {
        checkState(jobStatusListener == null);
        // register self as job status change listener
        jobStatusListener = new JobManagerJobStatusListener();//注册状态监听器，在触发checkpoint时会用到
        schedulerNG.registerJobStatusListener(jobStatusListener);

        schedulerNG.startScheduling();//开始调度
    }
public final void startScheduling() {
        mainThreadExecutor.assertRunningInMainThread();
        registerJobMetrics();
        startSchedulingInternal();
    }
protected void startSchedulingInternal() {
        final ExecutionGraph executionGraph = getExecutionGraph();
        try {
            executionGraph.scheduleForExecution();//通过executiongraph调度执行
        }
        catch (Throwable t) {
            executionGraph.failGlobal(t);
        }
    }

public void scheduleForExecution() throws JobException {

        assertRunningInJobMasterMainThread();

        if (isLegacyScheduling()) {
            LOG.info("Job recovers via failover strategy: {}", failoverStrategy.getStrategyName());
        }

        final long currentGlobalModVersion = globalModVersion;

        if (transitionState(JobStatus.CREATED, JobStatus.RUNNING)) {
            //调度工具类，把流和批调度整合到一起
            final CompletableFuture newSchedulingFuture = SchedulingUtils.schedule(
                scheduleMode,
                getAllExecutionVertices(),
                this);

            if (state == JobStatus.RUNNING && currentGlobalModVersion == globalModVersion) {
                schedulingFuture = newSchedulingFuture;
                newSchedulingFuture.whenComplete(
                    (Void ignored, Throwable throwable) -> {
                        if (throwable != null) {
                            final Throwable strippedThrowable = ExceptionUtils.stripCompletionException(throwable);

                            if (!(strippedThrowable instanceof CancellationException)) {
                                // only fail if the scheduling future was not canceled
                                failGlobal(strippedThrowable);
                            }
                        }
                    });
            } else {
                newSchedulingFuture.cancel(false);
            }
        }
        else {
            throw new IllegalStateException("Job may only be scheduled from state " + JobStatus.CREATED);
        }
    }
    
public static CompletableFuture schedule(
            ScheduleMode scheduleMode,
            final Iterable vertices,
            final ExecutionGraph executionGraph) {

        switch (scheduleMode) {
            case LAZY_FROM_SOURCES:
            case LAZY_FROM_SOURCES_WITH_BATCH_SLOT_REQUEST:
                return scheduleLazy(vertices, executionGraph);//批调度

            case EAGER:
                return scheduleEager(vertices, executionGraph);//流调度

            default:
                throw new IllegalStateException(String.format("Schedule mode %s is invalid.", scheduleMode));
        }
    }

流程序必须所有提前将所有task都启动成功，批程序不需要提前启动所有task，上游task计算完后可以把结果落盘，下游task在启动运行。

/**
     * Schedule vertices eagerly. That means all vertices will be scheduled at once.
     *
     * @param vertices Topologically sorted vertices to schedule.
     * @param executionGraph The graph the given vertices belong to.
     */
    public static CompletableFuture scheduleEager(
            final Iterable vertices,
            final ExecutionGraph executionGraph) {
        //......
        // allocate the slots (obtain all their futures)
        //分配slots
        for (ExecutionVertex ev : vertices) {
            // these calls are not blocking, they only return futures
            CompletableFuture allocationFuture = ev.getCurrentExecutionAttempt().allocateResourcesForExecution(
                slotProviderStrategy,
                LocationPreferenceConstraint.ALL,
                allPreviousAllocationIds);

            allAllocationFutures.add(allocationFuture);
        }
// this future is complete once all slot futures are complete.
        // the future fails once one slot future fails.
        final ConjunctFuture> allAllocationsFuture = FutureUtils.combineAll(allAllocationFutures);
        //slot都分配成功后
return allAllocationsFuture.thenAccept(
            (Collection executionsToDeploy) -> {
                for (Execution execution : executionsToDeploy) {
                    try {
                        execution.deploy();//执行
                    } catch (Throwable t) {
                        throw new CompletionException(
                            new FlinkException(
                                String.format("Could not deploy execution %s.", execution),
                                t));
                    }
                }
            })      
}

把申请的slot通过rpc调用分配给对应的TaskManager执行

/**
     * Deploys the execution to the previously assigned resource.
     *
     * @throws JobException if the execution cannot be deployed to the assigned resource
     */
    public void deploy() throws JobException {
    //......
    //一些校验
    final TaskManagerGateway taskManagerGateway = slot.getTaskManagerGateway();
    //通过RPC调用吧executor提交给taskmanager执行
    CompletableFuture.supplyAsync(() -> taskManagerGateway.submitTask(deployment, rpcTimeout), executor)
                .thenCompose(Function.identity())
                .whenCompleteAsync(
                    (ack, failure) -> {
                        // only respond to the failure case
                        if (failure != null) {
                            if (failure instanceof TimeoutException) {
                                String taskname = vertex.getTaskNameWithSubtaskIndex() + " (" + attemptId + ')';

                                markFailed(new Exception(
                                    "Cannot deploy task " + taskname + " - TaskManager (" + getAssignedResourceLocation()
                                        + ") not responding after a rpcTimeout of " + rpcTimeout, failure));
                            } else {
                                markFailed(failure);
                            }
                        }
                    },
                    jobMasterMainThreadExecutor);
}

提交task的过程，最终被提交到TaskExecutor中，创建Task线程并启动。

public CompletableFuture submitTask(TaskDeploymentDescriptor tdd, Time timeout) {
        return taskExecutorGateway.submitTask(tdd, jobMasterId, timeout);
    }
public CompletableFuture submitTask(
            TaskDeploymentDescriptor tdd,
            JobMasterId jobMasterId,
            Time timeout) {
        //......校验+一些处理
        Task task = new Task(
                jobInformation,
                taskInformation,
                tdd.getExecutionAttemptId(),
                tdd.getAllocationId(),
                tdd.getSubtaskIndex(),
                tdd.getAttemptNumber(),
                tdd.getProducedPartitions(),
                tdd.getInputGates(),
                tdd.getTargetSlotNumber(),
                memoryManager,
                taskExecutorServices.getIOManager(),
                taskExecutorServices.getShuffleEnvironment(),
                taskExecutorServices.getKvStateService(),
                taskExecutorServices.getBroadcastVariableManager(),
                taskExecutorServices.getTaskEventDispatcher(),
                taskStateManager,
                taskManagerActions,
                inputSplitProvider,
                checkpointResponder,
                aggregateManager,
                blobCacheService,
                libraryCache,
                fileCache,
                taskManagerConfiguration,
                taskMetricGroup,
                resultPartitionConsumableNotifier,
                partitionStateChecker,
                getRpcService().getExecutor());
                
            if (taskAdded) {
                task.startTaskThread();
                //......
                }
}
    public void run() {
        try {
            doRun();
        } finally {
            terminationFuture.complete(executionState);
        }
    }
private void doRun() {
    //......
    //实例化invokable对象
    invokable = loadAndInstantiateInvokable(userCodeClassLoader, nameOfInvokableClass, env);
    invokable.invoke();//执行
    //......
}

看看比较常用的StreamTask的构造方法

protected StreamTask(
            Environment environment,
            @Nullable TimerService timerService,
            Thread.UncaughtExceptionHandler uncaughtExceptionHandler,
            StreamTaskActionExecutor.SynchronizedStreamTaskActionExecutor actionExecutor,
            TaskMailbox mailbox) {

        super(environment);

        this.timerService = timerService;
        this.uncaughtExceptionHandler = Preconditions.checkNotNull(uncaughtExceptionHandler);
        this.configuration = new StreamConfig(getTaskConfiguration());
        this.accumulatorMap = getEnvironment().getAccumulatorRegistry().getUserMap();
        this.recordWriter = createRecordWriterDelegate(configuration, environment);
        this.actionExecutor = Preconditions.checkNotNull(actionExecutor);
        this.mailboxProcessor = new MailboxProcessor(this::processInput, mailbox, actionExecutor);//调用processInput
        this.asyncExceptionHandler = new StreamTaskAsyncExceptionHandler(environment);
    }
protected void processInput(MailboxDefaultAction.Controller controller) throws Exception {
        InputStatus status = inputProcessor.processInput();
        if (status == InputStatus.MORE_AVAILABLE && recordWriter.isAvailable()) {
            return;
        }
        if (status == InputStatus.END_OF_INPUT) {
            controller.allActionsCompleted();
            return;
        }
        CompletableFuture jointFuture = getInputOutputJointFuture(status);
        MailboxDefaultAction.Suspension suspendedDefaultAction = controller.suspendDefaultAction();
        jointFuture.thenRun(suspendedDefaultAction::resume);
    }

processInput处理输入数据，StreamOneInputProcessor#processInput

public InputStatus processInput() throws Exception {
        InputStatus status = input.emitNext(output);

        if (status == InputStatus.END_OF_INPUT) {
            synchronized (lock) {
                operatorChain.endHeadOperatorInput(1);
            }
        }

        return status;
    }

public InputStatus emitNext(DataOutput output) throws Exception {

        while (true) {
            // get the stream element from the deserializer
            if (currentRecordDeserializer != null) {
                DeserializationResult result = currentRecordDeserializer.getNextRecord(deserializationDelegate);
                if (result.isBufferConsumed()) {
                    currentRecordDeserializer.getCurrentBuffer().recycleBuffer();
                    currentRecordDeserializer = null;
                }

                if (result.isFullRecord()) {
                    processElement(deserializationDelegate.getInstance(), output);//处理数据
                    return InputStatus.MORE_AVAILABLE;
                }
            }

            Optional bufferOrEvent = checkpointedInputGate.pollNext();
            if (bufferOrEvent.isPresent()) {
                processBufferOrEvent(bufferOrEvent.get());
            } else {
                if (checkpointedInputGate.isFinished()) {
                    checkState(checkpointedInputGate.getAvailableFuture().isDone(), "Finished BarrierHandler should be available");
                    if (!checkpointedInputGate.isEmpty()) {
                        throw new IllegalStateException("Trailing data in checkpoint barrier handler.");
                    }
                    return InputStatus.END_OF_INPUT;
                }
                return InputStatus.NOTHING_AVAILABLE;
            }
        }
    }

private void processElement(StreamElement recordOrMark, DataOutput output) throws Exception {
        if (recordOrMark.isRecord()){//是正常数据，走该分支
            output.emitRecord(recordOrMark.asRecord());
        } else if (recordOrMark.isWatermark()) {//水印
            statusWatermarkValve.inputWatermark(recordOrMark.asWatermark(), lastChannel);
        } else if (recordOrMark.isLatencyMarker()) {
            output.emitLatencyMarker(recordOrMark.asLatencyMarker());
        } else if (recordOrMark.isStreamStatus()) {
            statusWatermarkValve.inputStreamStatus(recordOrMark.asStreamStatus(), lastChannel);
        } else {
            throw new UnsupportedOperationException("Unknown type of StreamElement");
        }
    }

@Override
        public void emitRecord(StreamRecord record) throws Exception {
            synchronized (lock) {
                numRecordsIn.inc();
                operator.setKeyContextElement1(record);
                operator.processElement(record);
            }
        }

processElement的具体实现就很多了，不同的算子有不同的实现。

在这里插入图片描述

以StreamFlatMap为例，最终会调用udf中的flatMap实现来处理数据。

public void processElement(StreamRecord element) throws Exception {
        collector.setTimestamp(element);
        userFunction.flatMap(element.getValue(), collector);
    }

至此job的调用过程基本清晰了。
总结：本文分析了flink job，从任务提交到集群后，是如何执行到用户定义的不同算子function实现的过程。
首先jobGraph提交到JM后会生成ExecutionGraph，然后向ResourceManager申请slot资源，申请成功后把slot分配给TaskManager执行具体的task任务。Task会运行不同的算子的实现。

留一个疑问，在Task#doRun方法中会调用invokable.invoke();具体实现有

在这里插入图片描述

以StreamTask为例

public final void invoke() throws Exception {
        try {
            beforeInvoke();//一些初始化操作

            // final check to exit early before starting to run
            if (canceled) {
                throw new CancelTaskException();
            }

            // let the task do its work
            isRunning = true;
            runMailboxLoop();

            // if this left the run() method cleanly despite the fact that this was canceled,
            // make sure the "clean shutdown" is not attempted
            if (canceled) {
                throw new CancelTaskException();
            }

            afterInvoke();
        }
        finally {
            cleanUpInvoke();
        }
    }
private void runMailboxLoop() throws Exception {
        try {
            mailboxProcessor.runMailboxLoop();
        }
        catch (Exception e) {
            Optional interruption = ExceptionUtils.findThrowable(e, InterruptedException.class);
            if (interruption.isPresent()) {
                if (!canceled) {
                    Thread.currentThread().interrupt();
                    throw interruption.get();
                }
            } else if (canceled) {
                LOG.warn("Error while canceling task.", e);
            }
            else {
                throw e;
            }
        }
    }

问题：
MailboxProcessor、TaskMailbox、Mail 这几个概念是什么意思，有什么作用？
还有为什么在AbstractInvokable的实例化中处理元素？
等我弄明白了再来更新。

4-19来填上面的坑
真正的执行逻辑还是在这一行

// run the invokable
invokable.invoke();

然后上面提到在实例化AbstractInvokable时，比如具体的StreamTask,在其构造函数中会把processInput传递给MailboxProcessor的mailboxDefaultAction

protected StreamTask(
            Environment environment,
            @Nullable TimerService timerService,
            Thread.UncaughtExceptionHandler uncaughtExceptionHandler,
            StreamTaskActionExecutor.SynchronizedStreamTaskActionExecutor actionExecutor,
            TaskMailbox mailbox) {

        super(environment);

        this.timerService = timerService;
        this.uncaughtExceptionHandler = Preconditions.checkNotNull(uncaughtExceptionHandler);
        this.configuration = new StreamConfig(getTaskConfiguration());
        this.accumulatorMap = getEnvironment().getAccumulatorRegistry().getUserMap();
        this.recordWriter = createRecordWriterDelegate(configuration, environment);
        this.actionExecutor = Preconditions.checkNotNull(actionExecutor);
        this.mailboxProcessor = new MailboxProcessor(this::processInput, mailbox, actionExecutor);
        this.asyncExceptionHandler = new StreamTaskAsyncExceptionHandler(environment);
    }

在StreamTask#invoke()->runMailboxLoop()->mailboxProcessor.runMailboxLoop();中会调用这个mailboxDefaultAction

public void runMailboxLoop() throws Exception {

        final TaskMailbox localMailbox = mailbox;

        Preconditions.checkState(
            localMailbox.isMailboxThread(),
            "Method must be executed by declared mailbox thread!");

        assert localMailbox.getState() == TaskMailbox.State.OPEN : "Mailbox must be opened!";

        final MailboxController defaultActionContext = new MailboxController(this);

        while (processMail(localMailbox)) {
            mailboxDefaultAction.runDefaultAction(defaultActionContext); //就是这里调用defaultAction,也就是processInput方法// lock is acquired inside default action as needed
        }
    }

后面的流程就如上面说的一样了