Flink 任务提交与执行流程

Flink 任务提交与执行流程 JobGrap的接受与运行

上文我们讲解了客户端将用户代码最终转化为JobGrap之后，通过Dispatcher的网关将JobGrap提交给Dispatcher。之后Dispatcher通过JobManagerRunnerFactory工厂类创建JobManagerRunner实例，最终调用JobManagerRunner实例启动JobManager服务。JobManager服务的底层主要通过JobMaster实现的，负责整个作业的生命周期和Task调度工作。

JobGrap提交整体流程

flink客户端通过ClusterClient调用Dispatcher.submit(JobGrap)的RPC方法，将创建好的JobGrap提交给Dispatcher。集群运行时调用Dispatcher.internalSubmitJob()方法执行JobGraph，然后调用waitForTerminatingJobManager()方法检查该JobGraph是否已经创建过，并调用persistAndRunJob()方法，对JobGraph持久化并执行。之后调用runJob(JobGraph)方法，通过JobManagerRunnerFactory创建JobManagerRunner，并调用JobManagerRunner.start()方法启动JobManagerRunner。由于JobManagerRunner实现了高可用，因此，需要leaderElectionService.start(this)选主，给予当前JobManagerRunner领导权;当前JobManagerRunner获取到leader权限后，会回调JobManagerRunner.grantLeaderShip(UUID)方法，然后调用startJobMaster(UUID)方法，通过JobMasterService.start(JobMasterId)启动JobMaster，此时JobMaster的RPC服务启动完成并对外提供服务JobMaster启动完成后，会异步调用JobMaster.startJobExecution()方法开始JobGraph的调度运行。作业如果成功，会将Acknowledge信息范湖给Dispatcher服务。Dispatcher将Acknowledge信息返回给ClusterClient，该消息最终返回给客户端。

Dispatcher会对JobGraph进行持久化并创建对应的JobManagerRunner，通过JobManagerRunner对象启动JobMaster，执行任务

public abstract class Dispatcher extends PermanentlyFencedRpcEndpoint implements DispatcherGateway {

	//Step1:CLusterClient会调用Dispatcher的RPC方法，上传JobGraph
	public CompletableFuture submitJob(JobGraph jobGraph, Time timeout) {
		log.info("Received JobGraph submission {} ({}).", jobGraph.getJobID(), jobGraph.getName());

		try {
			if (isDuplicateJob(jobGraph.getJobID())) {
				return FutureUtils.completedExceptionally(
					new DuplicateJobSubmissionException(jobGraph.getJobID()));
			} else if (isPartialResourceConfigured(jobGraph)) {
				return FutureUtils.completedExceptionally(
					new JobSubmissionException(jobGraph.getJobID(), "Currently jobs is not supported if parts of the vertices have " +
							"resources configured. The limitation will be removed in future versions."));
			} else {
				//进入具体JobGraph提交的逻辑
				return internalSubmitJob(jobGraph);
			}
		} catch (FlinkException e) {
			return FutureUtils.completedExceptionally(e);
		}
	}
	//Step2
	private CompletableFuture internalSubmitJob(JobGraph jobGraph) {
		log.info("Submitting job {} ({}).", jobGraph.getJobID(), jobGraph.getName());
		//会检查该JobGraph是否提交过，并异步persistAndRunJob方法持久化并执行JobGraph
		final CompletableFuture persistAndRunFuture = waitForTerminatingJobManager(jobGraph.getJobID(), jobGraph, this::persistAndRunJob)
			.thenApply(ignored -> Acknowledge.get());

		//...
		//最终会返回一个ACK的Future对象
	}

	//Step3:持久化并执行JobGraph
	private CompletableFuture persistAndRunJob(JobGraph jobGraph) throws Exception {
		//持久化JobGraph，如果是Zookeeper实现的高可用，则通过ZooKeeperJobGraphStore组件将JobGraph写到ZK上
		//异常情况下通过ZooKeeperJobGraphStore组件恢复作业
		jobGraphWriter.putJobGraph(jobGraph);
		//执行对应的JobGraph
		final CompletableFuture runJobFuture = runJob(jobGraph);
		//对runJobFuture 执行完毕后的 *** 作并处理异常
		return runJobFuture.whenComplete(BiConsumerWithException.unchecked((Object ignored, Throwable throwable) -> {
			if (throwable != null) {
				jobGraphWriter.removeJobGraph(jobGraph.getJobID());
			}
		}));
	}
	
	//Step4：创建并启动对应的JobManagerRunner，并在startJobManagerRunner()中调用JobManagerRunner.start()方法启动JobManagerRunner
	private CompletableFuture runJob(JobGraph jobGraph) {
		Preconditions.checkState(!jobManagerRunnerFutures.containsKey(jobGraph.getJobID()));
		//JobManagerRunnerFactory创建对应的jobManagerRunner
		final CompletableFuture jobManagerRunnerFuture = createJobManagerRunner(jobGraph);
		//将JobManagerRunner的异步对象添加到集合中，用于避免重复创建Job
		jobManagerRunnerFutures.put(jobGraph.getJobID(), jobManagerRunnerFuture);
		//异步启动JobManagerRunner，并处理异常
		return jobManagerRunnerFuture
			.thenApply(FunctionUtils.uncheckedFunction(this::startJobManagerRunner))
			.thenApply(FunctionUtils.nullFn())
			.whenCompleteAsync(
				(ignored, throwable) -> {
					if (throwable != null) {
						jobManagerRunnerFutures.remove(jobGraph.getJobID());
					}
				},
				getMainThreadExecutor());
	}
}

JobManagerRunner根据具体HA策略进行选主，然后启动JobMaster，通过JobMaster调度并执行Job

public class JobManagerRunnerImpl implements LeaderContender, OnCompletionActions, JobManagerRunner {
	//Step1:HA，选主，选取成功后，回调grantLeadership()方法
	@Override
	public void start() throws Exception {
		try {
			leaderElectionService.start(this);
		} catch (Exception e) {
			log.error("Could not start the JobManager because the leader election service did not start.", e);
			throw new Exception("Could not start the leader election service.", e);
		}
	}

	//Step2：选主成功后，检验Job状态，然后调用startJobMaster()方法启动JobMaster
	@Override
	public void grantLeadership(final UUID leaderSessionID) {
		synchronized (lock) {
			if (shutdown) {
				log.info("JobManagerRunner already shutdown.");
				return;
			}

			leadershipOperation = leadershipOperation.thenCompose(
				(ignored) -> {
					synchronized (lock) {
						return verifyJobSchedulingStatusAndStartJobManager(leaderSessionID);
					}
				});

			handleException(leadershipOperation, "Could not start the job manager.");
		}
	}

	//Step3:启动JobMaster，调度并执行Job
	private CompletionStage startJobMaster(UUID leaderSessionId) {
		log.info("JobManager runner for job {} ({}) was granted leadership with session id {} at {}.",
			jobGraph.getName(), jobGraph.getJobID(), leaderSessionId, jobMasterService.getAddress());

		try {
			//注册Job信息
			runningJobsRegistry.setJobRunning(jobGraph.getJobID());
		} catch (IOException e) {
			return FutureUtils.completedExceptionally(
				new FlinkException(
					String.format("Failed to set the job %s to running in the running jobs registry.", jobGraph.getJobID()),
					e));
		}

		final CompletableFuture startFuture;
		try {
			//启动JobMaster，并调度执行整个JobGraph
			startFuture = jobMasterService.start(new JobMasterId(leaderSessionId));
		} catch (Exception e) {
			return FutureUtils.completedExceptionally(new FlinkException("Failed to start the JobMaster.", e));
		}
		//异步确认当前JobManagerRunner具有LeaderShip
		final CompletableFuture currentLeaderGatewayFuture = leaderGatewayFuture;
		return startFuture.thenAcceptAsync(
			(Acknowledge ack) -> /confirm/iLeaderSessionIdIfStillLeader(
				leaderSessionId,
				jobMasterService.getAddress(),
				currentLeaderGatewayFuture),
			executor);
	}
}

JobMaster启动后建立和RM的心跳，并启动SlotPool服务，创建调度器(创建过程中将JobGrap装换为ExecutionGraph)，通过调度器调度Job

public class JobMaster extends FencedRpcEndpoint implements JobMasterGateway, JobMasterService {

	//Step1
	public CompletableFuture start(final JobMasterId newJobMasterId) throws Exception {
		// 启动JobMaster的RPC服务
		start();
		//启动JobMaster并通过JobMaster执行分配的Job
		return callAsyncWithoutFencing(() -> startJobExecution(newJobMasterId), RpcUtils.INF_TIMEOUT);
	}
	
	//Step2
	private Acknowledge startJobExecution(JobMasterId newJobMasterId) throws Exception {
		//检测当前RPC主线程是否正常运行，如未正常运行则抛出异常
		validateRunsInMainThread();
		
		checkNotNull(newJobMasterId, "The new JobMasterId must not be null.");
		//如果Fence Token已包含新的JobMasterId，则表明Job已通过该JobMasterId启动，返回Ack
		if (Objects.equals(getFencingToken(), newJobMasterId)) {
			log.info("Already started the job execution with JobMasterId {}.", newJobMasterId);

			return Acknowledge.get();
		}
		//为JobMasterId设置新的Fencing Token，用于RPC通信
		setNewFencingToken(newJobMasterId);
		//建立TM与RM之间的心态，并启动SlotPool
		startJobMasterServices();

		log.info("Starting execution of job {} ({}) under job master id {}.", jobGraph.getName(), jobGraph.getJobID(), newJobMasterId);
		//为Job分配调度器，调度执行Task
		resetAndStartScheduler();

		return Acknowledge.get();
	}
	
	//Step3：创建调度器，并调度执行Job
	private void resetAndStartScheduler() throws Exception {
		validateRunsInMainThread();
		//用于调度器的异步分配 *** 作
		final CompletableFuture schedulerAssignedFuture;

		if (schedulerNG.requestJobStatus() == JobStatus.CREATED) {
		//Job为新建状态，将MainThreadExecutor分配给schedulerNG调度器
			schedulerAssignedFuture = CompletableFuture.completedFuture(null);
			schedulerNG.setMainThreadExecutor(getMainThreadExecutor());
		} else {
			//重启任务时，需要重置ExecutionGraph，并创建新的调度器给当前Job
			suspendAndClearSchedulerFields(new FlinkException("ExecutionGraph is being reset in order to be rescheduled."));
			final JobManagerJobMetricGroup newJobManagerJobMetricGroup = jobMetricGroupFactory.create(jobGraph);
			//创建调度器，并将JobGraph转换为ExecutionGraph
			final SchedulerNG newScheduler = createScheduler(newJobManagerJobMetricGroup);

			schedulerAssignedFuture = schedulerNG.getTerminationFuture().handle(
				(ignored, throwable) -> {
					newScheduler.setMainThreadExecutor(getMainThreadExecutor());
					assignScheduler(newScheduler, newJobManagerJobMetricGroup);
					return null;
				}
			);
		}
		//调度器调度执行Job，分别执行ExecutionGraph中的节点
		schedulerAssignedFuture.thenRun(this::startScheduling);
	}

	//通过调度器，调度执行Job
	private void startScheduling() {
		checkState(jobStatusListener == null);
		// register self as job status change listener
		jobStatusListener = new JobManagerJobStatusListener();
		schedulerNG.registerJobStatusListener(jobStatusListener);

		schedulerNG.startScheduling();
	}
}

ExecutionGrap的调度执行

public class DefaultScheduler extends Schedulerbase implements SchedulerOperations {
	//Step1
	@Override
	public final void startScheduling() {
		mainThreadExecutor.assertRunningInMainThread();
		registerJobMetrics();
		startSchedulingInternal();
	}

	//Step2
	@Override
	protected void startSchedulingInternal() {
		log.info("Starting scheduling with scheduling strategy [{}]", schedulingStrategy.getClass().getName());
		prepareExecutionGraphForNgScheduling();
		//根据具体调度策略调度Job
		//eager：用于流计算，一次申请所有资源，资源不足则失败
		//lazy：用于批计算，从SourceTask分阶段调度
		schedulingStrategy.startScheduling();
	}
}

public class EagerSchedulingStrategy implements SchedulingStrategy {
	@Override
	public void startScheduling() {
		allocateSlotsAndDeploy(SchedulingStrategyUtils.getAllVertexIdsFromTopology(schedulingTopology));
	}

	private void allocateSlotsAndDeploy(final Set verticesToDeploy) {
		//将需要部署执行的ExecutionVertexID集合转换成ExecutionVertexDeploymentOption集合
		final List executionVertexDeploymentOptions =
			SchedulingStrategyUtils.createExecutionVertexDeploymentOptionsInTopologicalOrder(
				schedulingTopology,
				verticesToDeploy,
				id -> deploymentOption);
		//通过DefaultScheduler.allocateSlotsAndDeploy(),部署ExecutionVertex
		schedulerOperations.allocateSlotsAndDeploy(executionVertexDeploymentOptions);
	}
}

DefaultScheduler最终会调用deployTaskSafe()方法部署指定ExecutionVertexID的ExecutionVertex。

public class DefaultScheduler extends Schedulerbase implements SchedulerOperations {
	private void deployTaskSafe(final ExecutionVertexID executionVertexId) {
		try {
			//获取ExecutionVertex节点
			final ExecutionVertex executionVertex = getExecutionVertex(executionVertexId);
			//通过executionVertexOperations部署获取到的executionVertex 
			executionVertexOperations.deploy(executionVertex);
		} catch (Throwable e) {
			handleTaskDeploymentFailure(executionVertexId, e);
		}
	}
}

executionVertexOperations的deploy方法，最终实际调用的是ExecutionVertex.deploy()方法，将ExecutionVertex节点的Task部署和运行到执行的TaskExecution中。
每个ExecutionVertex节点都有一个Execution对象，通过调用Execution.deploy()方法完成Execution的部署和运行。

public class ExecutionVertex implements AccessExecutionVertex, Archiveable {
	public void deploy() throws JobException {
		currentExecution.deploy();
	}
}

每个Execution 对象都分配了一个LogicSlot，通过LogicSlot可以拿到对于TM的网关对象，然后RPC调用TaskExecutor的submitTask方法，提交任务

public class Execution implements AccessExecution, Archiveable, LogicalSlot.Payload {
	public void deploy() throws JobException {
		assertRunningInJobMasterMainThread();
		//该Execution 以及分配的Slot资源
		final LogicalSlot slot  = assignedResource;

		checkNotNull(slot, "In order to deploy the execution we first have to assign a resource via tryAssignResource.");

		// 判断当前Slot对于的TM是否存活
		if (!slot.isAlive()) {
			throw new JobException("Target slot (TaskManager) for deployment is no longer alive.");
		}

		// 判断当前Execution的状态
		ExecutionState previous = this.state;
		if (previous == SCHEDULED || previous == CREATED) {
			if (!transitionState(previous, DEPLOYING)) {
				
				throw new IllegalStateException("Cannot deploy task: Concurrent deployment call race.");
			}
		}
		else {
			// 当前Execution已经被取消或执行
			throw new IllegalStateException("The vertex must be in CREATED or SCHEDULED state to be deployed. Found state " + previous);
		}
		//判断当前Execution分配的SLot资源的有效性
		if (this != slot.getPayload()) {
			throw new IllegalStateException(
				String.format("The execution %s has not been assigned to the assigned slot.", this));
		}

		try {

			// 当前Execution的状态不是DEPLOYING，则释放资源
			if (this.state != DEPLOYING) {
				slot.releaseSlot(new FlinkException("Actual state of execution " + this + " (" + state + ") does not match expected state DEPLOYING."));
				return;
			}

			if (LOG.isInfoEnabled()) {
				LOG.info(String.format("Deploying %s (attempt #%d) to %s", vertex.getTaskNameWithSubtaskIndex(),
						attemptNumber, getAssignedResourceLocation()));
			}
			//创建TaskDeploymentDescriptor 用于将Task部署到TM上
			final TaskDeploymentDescriptor deployment = TaskDeploymentDescriptorFactory
				.fromExecutionVertex(vertex, attemptNumber)
				.createDeploymentDescriptor(
					slot.getAllocationId(),
					slot.getPhysicalSlotNumber(),
					taskRestore,
					producedPartitions.values());

			// 将taskRestore 置空，以进行GC回收
			taskRestore = null;
			// 获取Slot对于的TM的网关
			final TaskManagerGateway taskManagerGateway = slot.getTaskManagerGateway();

			final ComponentMainThreadExecutor jobMasterMainThreadExecutor =
				vertex.getExecutionGraph().getJobMasterMainThreadExecutor();

			// 通过网关RPC调用方TM的submitTask方法，将Task部署到TM上分
			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);

		}
		catch (Throwable t) {
			markFailed(t);

			if (isLegacyScheduling()) {
				ExceptionUtils.rethrow(t);
			}
		}
	}
}

完成上述步骤后，TM接受到JM提交的TaskDeploymentDescriptor 信息，完成Task线程的构建并启动运行。当Job所有的Task实例全部启动后，系统就可以正常处理接入的数据。