【深入淺出 Yarn 架構與實現】4-5 RM 行為探究 - 啟動 ApplicationMaster

本節開始，將對 ResourceManager 中一些常見行為進行分析探究，看某些具體關鍵的行為，在 RM 中是如何流轉的。本節將深入源碼探究「啟動 ApplicationMaster」的具體流程。

一、整體流程

本小節介紹從應用程式提交到啟動 ApplicationMaster 的整個過程，期間涉及 Client、RMService、 RMAppManager、RMApplmpl、RMAppAttemptImpl、RMNode、ResourceScheduler 等幾個主要組件。當客戶端調用 RPC 函數 ApplicationClientProtocol#submitApplication 後， ResourceManager 端的處理過程如下圖所示。

二、具體流程分析

接下來跟隨上面的流程圖，我們深入源碼具體分析每一步都是如何執行的：
最開始由客戶端發起任務提交 submitApplication()，經過 ClientRMService 和 RMAppManager 發送 RMAppEventType.START 事件，之後交由 RMAppImpl 處理。

  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.
        // 這裡發送 RMAppEventType.START 事件
        this.rmContext.getDispatcher().getEventHandler()
            .handle(new RMAppEvent(applicationId, RMAppEventType.START));
      }

RMAppImpl 這東西是個狀態機，收到事件之後會自己轉換狀態並且處理相應的邏輯。
（狀態機還不熟悉的同學，可翻到我前面的文章進行學習《2-4 Yarn 基礎庫 - 狀態機庫》）

截取一部分狀態轉換代碼：

  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 事件之後，會執行 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);
    }
  }

跟下去會發現它發出 RMStateStoreEventType.STORE_APP 事件，去 RMStateStore 中找一下對應的事件處理。發現也是個狀態機：

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

跟著 StoreAppTransition 看看做了啥（發送 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);
        // 這裡發送了 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

我們再回到 RMAppImpl，找到對應的狀態轉移邏輯。

    // 剛剛我們的狀態是 NEW_SAVING，收到了 APP_NEW_SAVED 事件，執行 AddApplicationToSchedulerTransition() 後，轉換為 SUBMITTED 狀態
    .addTransition(RMAppState.NEW_SAVING, RMAppState.SUBMITTED,
        RMAppEventType.APP_NEW_SAVED, new AddApplicationToSchedulerTransition())

在 AddApplicationToSchedulerTransition() 中會發送 SchedulerEventType.APP_ADDED 事件。之後 RMAppImpl 轉換為 RMAppState.SUBMITTED 狀態。
SchedulerEventType.APP_ADDED 會被多個事件處理器捕獲處理：
1）ResourceSchedulerWrapper 事件處理器，僅記錄

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

2）各個 AbstractYarnScheduler 的實現類。以 CapacityScheduler 為例：
執行 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() 中會提交 Application 併發送 RMAppEventType.APP_ACCEPTED 事件。

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

三）RMAppImpl - APP_ACCEPTED（重點）

繼續回到 RMAppImpl，執行 StartAppAttemptTransition()，創建 newAttempt，發送事件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 中會捕獲這個事件，執行 AttemptStartedTransition()，其中會發送 SchedulerEventType.APP_ATTEMPT_ADDED 事件，由 AbstractYarnScheduler 實現類處理

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

如在 CapacityScheduler 中由 addApplicationAttempt 處理，會提交 ApplicationAttempt，併發送 RMAppAttemptEventType.ATTEMPT_ADDED 事件

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

RMAppAttemptImpl 收到 event 後繼續處理，在 ScheduleTransition 會 allocate am container 資源。

      .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 將資源返回給它之前，會向 RMContainerlmpl 發送一個 RMContainerEventType.ACQUIRED 事件。
在 RMContainerImpl 接到 RMContainerEventType.START，發送 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 後續狀態機，執行 AMContainerAllocatedTransition，在其中又一次為 am allocate，和上一個狀態中 allocate 僅參數不同，沒搞懂為啥。這裡如果發現 allocate container 資源還是 0，會退回上一步，狀態還是 RMAppAttemptState.SCHEDULED 等待再次獲取資源。如果正常獲取到了資源，就會轉為 RMAppAttemptState.ALLOCATED_SAVING 狀態。

      .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);

日誌記錄完成後，RMStateStore 向 RMAppAttemptImpl 發送 RMAppAttemptEventType.ATTEMPT_NEW_SAVED 事件。
RMAppAttemptImpl 後續向 ApplicationMasterLauncher 發 送 AMLauncherEventType.LAUNCH 事件（實際執行是在 AMLauncher 中），並將狀態從 ALLOCATED_SAVING 轉移為 ALLOCATED。

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

ApplicationMasterLauncher 收到 AMLauncherEventType.LAUNCH 事件後，會將該事件放到事件隊列中，等待 AMLauncher 線程池中的線程處理該事件。它將與對應的 NodeManager 通信，啟動 ApplicationMaster，一旦成功啟動後，將向 RMAppAttemptImpl 發送 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 事件後，會向 AMLivelinessMonitor 註冊，以監控運行狀態。RMAppAttemptImpl 狀態從 ALLOCATED 轉移為 LAUNCHED。

之後，NodeManager 通過心跳機制彙報 ApplicationMaster 所在 Container 已經成功啟動，收到該信息後，ResourceScheduler 將發送一個 RMContainerEventType.LAUNCHED 事件，RMContainerImpl 收到該事件後，會從 ContainerAllocationExpirer 監控列表中移除。

啟動的 ApplicationMaster 通過RPC 函數 ApplicationMasterProtocol#registerApplicationMaster 向 ResourceManager 註冊，ResourceManager 中的 ApplicationMasterService 服務接收到該請求後，發送 RMAppAttemptEventType.REGISTERED 事件。

// ApplicationMasterService#registerApplicationMaster

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

RMAppAttemptImpl 收到該事件後，首先保存該 ApplicationMaster 的基本信息（比如所在 host、啟用的 RPC 埠號等），然後向 RMApplmpl 發送一個 RMAppEventType.ATTEMPT_REGISTERED 事件。RMAppAttemptImpl 狀態從 LAUNCHED 轉移為 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 事件後，將狀態從 ACCEPTED 轉換為 RUNNING。

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

到這裡，啟動 ApplicationMaster 的整體流程分析完畢！

三、總結

本篇文章分析了從應用程式提交到啟動 ApplicationMaster 的整個過程，分析具體過程看的可能會有些繁瑣。但只要抓住核心本質，就很容易捋清楚。重點就是事件處理和狀態機，這兩個部件理解清楚，就很容易看明白程式的流轉。
實際邏輯無非就是幾個服務之間互相發送對應的事件，接收到事件後會執行啟動服務、記錄日誌、監控狀態，然後再發送個新的事件。
本身不難，但需要耐下心來一點點去梳理。