1. 程式人生 > >Yarn方式下的ScheduleBackend是怎麼實現的?

Yarn方式下的ScheduleBackend是怎麼實現的?

Yarn方式下的ScheduleBackend是用的啥?

在SparkContext中建立ScheduleBackend時,會根據指定的”master“引數的字首決定建立哪種ScheduleBackend,對於"yarn://host:port"這樣的URL來說,如果是cluster模式,就是建立YarnClusterSchedulerBackend,如果是client模式,就是建立YarnClientSchedulerBackend。

我們還是先看看YarnClusterSchedulerBackend的程式碼結構把。

YarnClusterSchedulerBackend繼承了YarnSchedulerBackend,沒有太多的發揮程式碼,我們直接看YarnSchedulerBackend把。估計client模式下也差不多。

YarnSchedulerBackend又繼承了CoarseGrainedSchedulerBackend,我們看看不同點在哪裡。

覆寫了doRequestTotalExecutors和doKillExecutors方法,一個申請Executor,一個殺死Executor。

override def doRequestTotalExecutors(requestedTotal: Int): Future[Boolean] = {
    yarnSchedulerEndpointRef.ask[Boolean](prepareRequestExecutors(requestedTotal))
  }  
  override def doKillExecutors(executorIds: Seq[String]): Future[Boolean] = {
    yarnSchedulerEndpointRef.ask[Boolean](KillExecutors(executorIds))
  }

yarnSchedulerEndpointRef就是同一個檔案裡的endpoint端,看看具體的執行程式碼是什麼:

      case r: RequestExecutors =>
        amEndpoint match {
          case Some(am) =>
            am.ask[Boolean](r).andThen {
              case Success(b) => context.reply(b)
              case Failure(NonFatal(e)) =>
                logError(s"Sending $r to AM was unsuccessful", e)
                context.sendFailure(e)
            }(ThreadUtils.sameThread)         
        }
      case k: KillExecutors =>
        amEndpoint match {
          case Some(am) =>
            am.ask[Boolean](k).andThen {
              case Success(b) => context.reply(b)
              case Failure(NonFatal(e)) =>
                logError(s"Sending $k to AM was unsuccessful", e)
                context.sendFailure(e)
            }(ThreadUtils.sameThread)          
        }

我們看到它又將訊息轉給了amEndpoint,就是轉給了yarn工程裡的ApplicationManager。又要跳到ApplicationManager去看看裡面的實現邏輯了,真是一波三折啊。

ApplicationManager裡是怎麼處理RequestExecutors和KillExecutors兩個訊息的呢?

      case r: RequestExecutors =>
        Option(allocator) match {
          case Some(a) =>
            if (a.requestTotalExecutorsWithPreferredLocalities(r.requestedTotal,
              r.localityAwareTasks, r.hostToLocalTaskCount, r.nodeBlacklist)) {
              resetAllocatorInterval()
            }
            context.reply(true)
        }
      case KillExecutors(executorIds) =>
        Option(allocator) match {
          case Some(a) => executorIds.foreach(a.killExecutor)
        }
        context.reply(true)

呼叫allocator的killExecutor和requestTotalExecutorsWithPreferredLocalities方法。allocator又是啥?這裡是不是類有的太多了啊。。

allocator = client.createAllocator(
      yarnConf,
      _sparkConf,
      appAttemptId,
      driverUrl,
      driverRef,
      securityMgr,
      localResources)

是client的createAllocator方法創建出來的,client是啥?是YarnRMClient,我們就要先看看YarnRMClient了,看名字就大概能猜到,YarnRMClient就是來向Yarn機器申請Executor和殺死Executor的。

createAllocator方法返回下面的YarnAllocator:

 return new YarnAllocator(driverUrl, driverRef, conf, sparkConf, amClient, appAttemptId, securityMgr,
      localResources, SparkRackResolver.get(conf))

來到YarnAllocator。

YarnAllocator的killExecutor方法很好理解,就是釋放Yarn中的Container:

 def killExecutor(executorId: String): Unit = synchronized {
    executorIdToContainer.get(executorId) match {
      case Some(container) if !releasedContainers.contains(container.getId) =>
        internalReleaseContainer(container)
        runningExecutors.remove(executorId)
      case _ => logWarning(s"Attempted to kill unknown executor $executorId!")
    }
  }

申請Executor其實最終是在runAllocatedContainers方法中實現的。

核心程式碼看一下把,完整的可以看原始碼:

    if (runningExecutors.size() < targetNumExecutors) {
        numExecutorsStarting.incrementAndGet()
        if (launchContainers) {
          launcherPool.execute(() => {
            try {
              new ExecutorRunnable(
                Some(container),
                conf,
                sparkConf,
                driverUrl,
                executorId,
                executorHostname,
                executorMemory,
                executorCores,
                appAttemptId.getApplicationId.toString,
                securityMgr,
                localResources
              ).run()
              updateInternalState()
            } catch {              
            }
          })
        } 

申請targetNumExecutors個ExecutorRunner,這樣就和Standalone的申請Executor對應起來了。好了,整個過程就是這樣了。

最終就會在Yarn叢集中申請了所需數目的Container,並且在Container中啟動ExecutorRunner,來向Driver彙報成績。

這裡的ExecutorRunner就是YarnCoarseGrainedExecutorBackend執行緒,在ExecutorRunner類中可以看到。

多看幾遍原始碼吧,當你真正看懂了,你會感覺