【深入浅出 Yarn 架构与实现】4-5 RM 行为探究

本节开始，将对 ResourceManager 中一些常见行为进行分析探究，看某些具体关键的行为，在 RM 中是如何流转的。本节将深入源码探究「启动 ApplicationMaster」的具体流程。

一、整体流程

本小节介绍从应用程序提交到启动 ApplicationMaster 的整个过程，期间涉及 Client、RMService、 RMAppManager、RMApplmpl、RMAppAttemptImpl、RMNode、ResourceScheduler 等几个主要组件。当客户端调用 RPC 函数 ApplicationClientProtocol#submitApplication 后， ResourceManager 端的处理过程如下图所示。

二、具体流程分析

接下来跟随上面的流程图，我们深入源码具体分析每一步都是如何执行的：
最开始由客户端发起任务提交 submitApplication()，经过 ClientRMService 和 RMAppManager 发送 RMAppEventType.START 事件，之后交由 RMAppImpl 处理。

  protected void submitApplication(
      ApplicationSubmissionContext submissionContext, long submitTime,
      String user) throws YarnException {
    ApplicationId applicationId = submissionContext.getApplicationId();

    RMAppImpl application =
        createAndPopulateNewRMApp(submissionContext, submitTime, user, false);
    Credentials credentials = null;
    try {
      credentials = parseCredentials(submissionContext);
      if (UserGroupInformation.isSecurityEnabled()) {
        this.rmContext.getDelegationTokenRenewer()
            .addApplicationAsync(applicationId, credentials,
                submissionContext.getCancelTokensWhenComplete(),
                application.getUser());
      } else {
        // Dispatcher is not yet started at this time, so these START events
        // enqueued should be guaranteed to be first processed when dispatcher
        // gets started.
        // 这里发送 RMAppEventType.START 事件
        this.rmContext.getDispatcher().getEventHandler()
            .handle(new RMAppEvent(applicationId, RMAppEventType.START));
      }

RMAppImpl 这东西是个状态机，收到事件之后会自己转换状态并且处理相应的逻辑。
（状态机还不熟悉的同学，可翻到我前面的文章进行学习《2-4 Yarn 基础库 - 状态机库》）

截取一部分状态转换代码：

  private static final StateMachineFactory<RMAppImpl,
                                           RMAppState,
                                           RMAppEventType,
                                           RMAppEvent> stateMachineFactory
                               = new StateMachineFactory<RMAppImpl,
                                           RMAppState,
                                           RMAppEventType,
                                           RMAppEvent>(RMAppState.NEW)


     // Transitions from NEW state
    .addTransition(RMAppState.NEW, RMAppState.NEW,
        RMAppEventType.NODE_UPDATE, new RMAppNodeUpdateTransition())
     // 收到 RMAppEventType.START 事件
    .addTransition(RMAppState.NEW, RMAppState.NEW_SAVING,
        RMAppEventType.START, new RMAppNewlySavingTransition())
    .addTransition(RMAppState.NEW, EnumSet.of(RMAppState.SUBMITTED,
            RMAppState.ACCEPTED, RMAppState.FINISHED, RMAppState.FAILED,
            RMAppState.KILLED, RMAppState.FINAL_SAVING),
        RMAppEventType.RECOVER, new RMAppRecoveredTransition())
    .addTransition(RMAppState.NEW, RMAppState.KILLED, RMAppEventType.KILL,
        new AppKilledTransition())
    .addTransition(RMAppState.NEW, RMAppState.FINAL_SAVING,
        RMAppEventType.APP_REJECTED,
        new FinalSavingTransition(new AppRejectedTransition(),
          RMAppState.FAILED))

一）RMAppImpl - START

收到 RMAppEventType.START 事件之后，会执行 RMAppNewlySavingTransition()。

  private static final class RMAppNewlySavingTransition extends RMAppTransition {
    @Override
    public void transition(RMAppImpl app, RMAppEvent event) {

      // If recovery is enabled then store the application information in a
      // non-blocking call so make sure that RM has stored the information
      // needed to restart the AM after RM restart without further client
      // communication
      LOG.info("Storing application with id " + app.applicationId);
      app.rmContext.getStateStore().storeNewApplication(app);
    }
  }

跟下去会发现它发出 RMStateStoreEventType.STORE_APP 事件，去 RMStateStore 中找一下对应的事件处理。发现也是个状态机：

.addTransition(RMStateStoreState.ACTIVE,
    EnumSet.of(RMStateStoreState.ACTIVE, RMStateStoreState.FENCED),
    RMStateStoreEventType.STORE_APP, new StoreAppTransition())

跟着 StoreAppTransition 看看做了啥（发送 RMAppEventType.APP_NEW_SAVED 事件）

  private static class StoreAppTransition
      implements MultipleArcTransition<RMStateStore, RMStateStoreEvent,
          RMStateStoreState> {
    @Override
    public RMStateStoreState transition(RMStateStore store,
        RMStateStoreEvent event) {
      if (!(event instanceof RMStateStoreAppEvent)) {
        // should never happen
        LOG.error("Illegal event type: " + event.getClass());
        return RMStateStoreState.ACTIVE;
      }
      boolean isFenced = false;
      ApplicationStateData appState =
          ((RMStateStoreAppEvent) event).getAppState();
      ApplicationId appId =
          appState.getApplicationSubmissionContext().getApplicationId();
      LOG.info("Storing info for app: " + appId);
      try {
        store.storeApplicationStateInternal(appId, appState);
        // 这里发送了 RMAppEventType.APP_NEW_SAVED 事件
        store.notifyApplication(new RMAppEvent(appId,
               RMAppEventType.APP_NEW_SAVED));
      } catch (Exception e) {
        LOG.error("Error storing app: " + appId, e);
        isFenced = store.notifyStoreOperationFailedInternal(e);
      }
      return finalState(isFenced);
    };
  }

二）RMAppImpl - APP_NEW_SAVED

我们再回到 RMAppImpl，找到对应的状态转移逻辑。

    // 刚刚我们的状态是 NEW_SAVING，收到了 APP_NEW_SAVED 事件，执行 AddApplicationToSchedulerTransition() 后，转换为 SUBMITTED 状态
    .addTransition(RMAppState.NEW_SAVING, RMAppState.SUBMITTED,
        RMAppEventType.APP_NEW_SAVED, new AddApplicationToSchedulerTransition())

在 AddApplicationToSchedulerTransition() 中会发送 SchedulerEventType.APP_ADDED 事件。之后 RMAppImpl 转换为 RMAppState.SUBMITTED 状态。
SchedulerEventType.APP_ADDED 会被多个事件处理器捕获处理：
1）ResourceSchedulerWrapper 事件处理器，仅记录

      } else if (schedulerEvent.getType() == SchedulerEventType.APP_ADDED
          && schedulerEvent instanceof AppAddedSchedulerEvent) {
        AppAddedSchedulerEvent appAddEvent =
                (AppAddedSchedulerEvent) schedulerEvent;
        String queueName = appAddEvent.getQueue();
        appQueueMap.put(appAddEvent.getApplicationId(), queueName);
      }

2）各个 AbstractYarnScheduler 的实现类。以 CapacityScheduler 为例：
执行 addApplication()

    case APP_ADDED:
    {
      AppAddedSchedulerEvent appAddedEvent = (AppAddedSchedulerEvent) event;
      String queueName = resolveReservationQueueName(appAddedEvent.getQueue(),
          appAddedEvent.getApplicationId(), appAddedEvent.getReservationID(),
          appAddedEvent.getIsAppRecovering());
      if (queueName != null) {
        if (!appAddedEvent.getIsAppRecovering()) {
          addApplication(appAddedEvent.getApplicationId(), queueName,
              appAddedEvent.getUser(), appAddedEvent.getApplicatonPriority());
        } else {
          addApplicationOnRecovery(appAddedEvent.getApplicationId(), queueName,
              appAddedEvent.getUser(), appAddedEvent.getApplicatonPriority());
        }
      }
    }

在 addApplication() 中会提交 Application 并发送 RMAppEventType.APP_ACCEPTED 事件。

	queue.submitApplication(applicationId, user, queueName);
    rmContext.getDispatcher().getEventHandler()
        .handle(new RMAppEvent(applicationId, RMAppEventType.APP_ACCEPTED));

三）RMAppImpl - APP_ACCEPTED（重点）

继续回到 RMAppImpl，执行 StartAppAttemptTransition()，创建 newAttempt，发送事件RMAppAttemptEventType.START

    .addTransition(RMAppState.SUBMITTED, RMAppState.ACCEPTED,
        RMAppEventType.APP_ACCEPTED, new StartAppAttemptTransition())

  private static final class StartAppAttemptTransition extends RMAppTransition {
    @Override
    public void transition(RMAppImpl app, RMAppEvent event) {
      app.createAndStartNewAttempt(false);
    };
  }

  private void
      createAndStartNewAttempt(boolean transferStateFromPreviousAttempt) {
    createNewAttempt();
    handler.handle(new RMAppStartAttemptEvent(currentAttempt.getAppAttemptId(),
      transferStateFromPreviousAttempt));
  }

在 RMAppAttemptImpl 中会捕获这个事件，执行 AttemptStartedTransition()，其中会发送 SchedulerEventType.APP_ATTEMPT_ADDED 事件，由 AbstractYarnScheduler 实现类处理

      .addTransition(RMAppAttemptState.NEW, RMAppAttemptState.SUBMITTED,
          RMAppAttemptEventType.START, new AttemptStartedTransition())

如在 CapacityScheduler 中由 addApplicationAttempt 处理，会提交 ApplicationAttempt，并发送 RMAppAttemptEventType.ATTEMPT_ADDED 事件

private synchronized void addApplicationAttempt() {
    // 提交 attempt
	queue.submitApplicationAttempt(attempt, application.getUser());
    // 发送 RMAppAttemptEventType.ATTEMPT_ADDED 事件
	rmContext.getDispatcher().getEventHandler().handle(
    		new RMAppAttemptEvent(applicationAttemptId,
            RMAppAttemptEventType.ATTEMPT_ADDED));
}

RMAppAttemptImpl 收到 event 后继续处理，在 ScheduleTransition 会 allocate am container 资源。

      .addTransition(RMAppAttemptState.SUBMITTED, 
          EnumSet.of(RMAppAttemptState.LAUNCHED_UNMANAGED_SAVING,
                     RMAppAttemptState.SCHEDULED),
          RMAppAttemptEventType.ATTEMPT_ADDED,
          new ScheduleTransition())

        // AM resource has been checked when submission
        Allocation amContainerAllocation =
            appAttempt.scheduler.allocate(
                appAttempt.applicationAttemptId,
                Collections.singletonList(appAttempt.amReq),
                EMPTY_CONTAINER_RELEASE_LIST,
                amBlacklist.getBlacklistAdditions(),
                amBlacklist.getBlacklistRemovals(), null, null);

ResourceScheduler 将资源返回给它之前，会向 RMContainerlmpl 发送一个 RMContainerEventType.ACQUIRED 事件。
在 RMContainerImpl 接到 RMContainerEventType.START，发送 RMAppAttemptEventType.CONTAINER_ALLOCATED 事件。

    .addTransition(RMContainerState.NEW, RMContainerState.ALLOCATED,
        RMContainerEventType.START, new ContainerStartedTransition())

  private static final class ContainerStartedTransition extends
      BaseTransition {

    @Override
    public void transition(RMContainerImpl container, RMContainerEvent event) {
      container.eventHandler.handle(new RMAppAttemptEvent(
          container.appAttemptId, RMAppAttemptEventType.CONTAINER_ALLOCATED));
    }
  }

又回到RMAppAttemptImpl 后续状态机，执行 AMContainerAllocatedTransition，在其中又一次为 am allocate，和上一个状态中 allocate 仅参数不同，没搞懂为啥。这里如果发现 allocate container 资源还是 0，会退回上一步，状态还是 RMAppAttemptState.SCHEDULED 等待再次获取资源。如果正常获取到了资源，就会转为 RMAppAttemptState.ALLOCATED_SAVING 状态。

      .addTransition(RMAppAttemptState.SCHEDULED,
          EnumSet.of(RMAppAttemptState.ALLOCATED_SAVING,
            RMAppAttemptState.SCHEDULED),
          RMAppAttemptEventType.CONTAINER_ALLOCATED,
          new AMContainerAllocatedTransition())

      Allocation amContainerAllocation =
          appAttempt.scheduler.allocate(appAttempt.applicationAttemptId,
            EMPTY_CONTAINER_REQUEST_LIST, EMPTY_CONTAINER_RELEASE_LIST, null,
            null, null, null);

日志记录完成后，RMStateStore 向 RMAppAttemptImpl 发送 RMAppAttemptEventType.ATTEMPT_NEW_SAVED 事件。
RMAppAttemptImpl 后续向 ApplicationMasterLauncher 发送 AMLauncherEventType.LAUNCH 事件（实际执行是在 AMLauncher 中），并将状态从 ALLOCATED_SAVING 转移为 ALLOCATED。

      .addTransition(RMAppAttemptState.ALLOCATED_SAVING, 
          RMAppAttemptState.ALLOCATED,
          RMAppAttemptEventType.ATTEMPT_NEW_SAVED, new AttemptStoredTransition())

ApplicationMasterLauncher 收到 AMLauncherEventType.LAUNCH 事件后，会将该事件放到事件队列中，等待 AMLauncher 线程池中的线程处理该事件。它将与对应的 NodeManager 通信，启动 ApplicationMaster，一旦成功启动后，将向 RMAppAttemptImpl 发送 RMAppAttemptEventType.LAUNCHED 事件。

  public void run() {
    switch (eventType) {
    case LAUNCH:
      try {
        LOG.info("Launching master" + application.getAppAttemptId());
        launch();
        handler.handle(new RMAppAttemptEvent(application.getAppAttemptId(),
            RMAppAttemptEventType.LAUNCHED));

RMAppAttemptImpl 收到 RMAppAttemptEventType.LAUNCHED 事件后，会向 AMLivelinessMonitor 注册，以监控运行状态。RMAppAttemptImpl 状态从 ALLOCATED 转移为 LAUNCHED。

之后，NodeManager 通过心跳机制汇报 ApplicationMaster 所在 Container 已经成功启动，收到该信息后，ResourceScheduler 将发送一个 RMContainerEventType.LAUNCHED 事件，RMContainerImpl 收到该事件后，会从 ContainerAllocationExpirer 监控列表中移除。

启动的 ApplicationMaster 通过RPC 函数 ApplicationMasterProtocol#registerApplicationMaster 向 ResourceManager 注册，ResourceManager 中的 ApplicationMasterService 服务接收到该请求后，发送 RMAppAttemptEventType.REGISTERED 事件。

// ApplicationMasterService#registerApplicationMaster

	LOG.info("AM registration " + applicationAttemptId);
      this.rmContext
        .getDispatcher()
        .getEventHandler()
        .handle(
          // 这里发送 RMAppAttemptEventType.REGISTERED 事件
          new RMAppAttemptRegistrationEvent(applicationAttemptId, request
            .getHost(), request.getRpcPort(), request.getTrackingUrl()));

RMAppAttemptImpl 收到该事件后，首先保存该 ApplicationMaster 的基本信息（比如所在 host、启用的 RPC 端口号等），然后向 RMApplmpl 发送一个 RMAppEventType.ATTEMPT_REGISTERED 事件。RMAppAttemptImpl 状态从 LAUNCHED 转移为 RUNNING。

      .addTransition(RMAppAttemptState.LAUNCHED, RMAppAttemptState.RUNNING,
          RMAppAttemptEventType.REGISTERED, REGISTERED_TRANSITION)

// AMRegisteredTransition
	appAttempt.eventHandler.handle(new RMAppEvent(appAttempt
          .getAppAttemptId().getApplicationId(),
          RMAppEventType.ATTEMPT_REGISTERED));

四）RMAppImpl - ATTEMPT_REGISTERED

RMAppImpl 收到 RMAppEventType.ATTEMPT_REGISTERED 事件后，将状态从 ACCEPTED 转换为 RUNNING。

    .addTransition(RMAppState.ACCEPTED, RMAppState.RUNNING,
        RMAppEventType.ATTEMPT_REGISTERED, new RMAppStateUpdateTransition(
            YarnApplicationState.RUNNING))

到这里，启动 ApplicationMaster 的整体流程分析完毕！

三、总结

本篇文章分析了从应用程序提交到启动 ApplicationMaster 的整个过程，分析具体过程看的可能会有些繁琐。但只要抓住核心本质，就很容易捋清楚。重点就是事件处理和状态机，这两个部件理解清楚，就很容易看明白程序的流转。
实际逻辑无非就是几个服务之间互相发送对应的事件，接收到事件后会执行启动服务、记录日志、监控状态，然后再发送个新的事件。
本身不难，但需要耐下心来一点点去梳理。