spark core源碼分析3 Master HA

博客地址: http://blog.csdn.net/yueqian_zhu/

這一節講解master 選舉以及之後的處理流程

上一節說到在Master啓動過程中，首先調用了 Akka actor的preStart方法。

override def preStart() {
  logInfo("Starting Spark master at " + masterUrl)
  logInfo(s"Running Spark version ${org.apache.spark.SPARK_VERSION}")
  // Listen for remote client disconnection events, since they don't go through Akka's watch()
  context.system.eventStream.subscribe(self, classOf[RemotingLifecycleEvent])
  webUi.bind()
  masterWebUiUrl = "http://" + masterPublicAddress + ":" + webUi.boundPort
  context.system.scheduler.schedule(0 millis, WORKER_TIMEOUT millis, self, CheckForWorkerTimeOut)

  masterMetricsSystem.registerSource(masterSource)
  masterMetricsSystem.start()
  applicationMetricsSystem.start()
  // Attach the master and app metrics servlet handler to the web ui after the metrics systems are
  // started.
  masterMetricsSystem.getServletHandlers.foreach(webUi.attachHandler)
  applicationMetricsSystem.getServletHandlers.foreach(webUi.attachHandler)

  //HA的流程從這裏開始
  //這裏可以選擇Master的元數據信息保存在哪裏，我們以ZK爲例講解
  //這裏用Apache Curator作爲zk的client，它包裝了zk client 複雜的api
  val (persistenceEngine_, leaderElectionAgent_) = RECOVERY_MODE match {
    case "ZOOKEEPER" =>
      logInfo("Persisting recovery state to ZooKeeper")
      val zkFactory =
        new ZooKeeperRecoveryModeFactory(conf, SerializationExtension(context.system))
      (zkFactory.createPersistenceEngine(), zkFactory.createLeaderElectionAgent(this))
    case "FILESYSTEM" =>
      val fsFactory =
        new FileSystemRecoveryModeFactory(conf, SerializationExtension(context.system))
      (fsFactory.createPersistenceEngine(), fsFactory.createLeaderElectionAgent(this))
    case "CUSTOM" =>
      val clazz = Class.forName(conf.get("spark.deploy.recoveryMode.factory"))
      val factory = clazz.getConstructor(classOf[SparkConf], classOf[Serialization])
        .newInstance(conf, SerializationExtension(context.system))
        .asInstanceOf[StandaloneRecoveryModeFactory]
      (factory.createPersistenceEngine(), factory.createLeaderElectionAgent(this))
    case _ =>
      (new BlackHolePersistenceEngine(), new MonarchyLeaderAgent(this))
  }
  persistenceEngine = persistenceEngine_
  leaderElectionAgent = leaderElectionAgent_
}

上面的persistenceEngine_封裝了在zk中讀寫元數據信息，以及序列化反序列化的接口

leaderElectionAgent_封裝了master的選舉過程，見下面代碼註釋中的解釋

private[master] class ZooKeeperLeaderElectionAgent(val masterActor: LeaderElectable,
    conf: SparkConf) extends LeaderLatchListener with LeaderElectionAgent with Logging  {

  //依賴zk中的一個節點來判斷選主
  val WORKING_DIR = conf.get("spark.deploy.zookeeper.dir", "/spark") + "/leader_election"

  private var zk: CuratorFramework = _
  private var leaderLatch: LeaderLatch = _
  private var status = LeadershipStatus.NOT_LEADER

  //構造這個對象之後就調用了start方法
  start()
  //leaderLatch.start()一旦調用，LeaderLatch會和其它使用相同latch path的其它LeaderLatch交涉，然後隨機的選擇其中一個作爲leader
  private def start() {
    logInfo("Starting ZooKeeper LeaderElection agent")
    zk = SparkCuratorUtil.newClient(conf)
    leaderLatch = new LeaderLatch(zk, WORKING_DIR)
    leaderLatch.addListener(this)
    leaderLatch.start()
  }

  override def stop() {
    leaderLatch.close()
    zk.close()
  }

  //當一個master被選爲主時，isLeader方法被回調，說明在這一輪選舉中勝出
  override def isLeader() {
    synchronized {
      // could have lost leadership by now.
      if (!leaderLatch.hasLeadership) {
        return
      }

      logInfo("We have gained leadership")
      updateLeadershipStatus(true)
    }
  }

  //當一個master被選爲備時，notLeader方法被回調，說明在這一輪選舉中落敗
  override def notLeader() {
    synchronized {
      // could have gained leadership by now.
      if (leaderLatch.hasLeadership) {
        return
      }

      logInfo("We have lost leadership")
      updateLeadershipStatus(false)
    }
  }
  private def updateLeadershipStatus(isLeader: Boolean) {
    //當一個master之前狀態爲備，目前被選爲主
    if (isLeader && status == LeadershipStatus.NOT_LEADER) {
      status = LeadershipStatus.LEADER
      masterActor.electedLeader()//調用master類的electedLeader方法	
      //當一個master之前狀態爲主，目前被選爲備
    } else if (!isLeader && status == LeadershipStatus.LEADER) {
      status = LeadershipStatus.NOT_LEADER
      masterActor.revokedLeadership()//調用master類的revokedLeadership方法	
    }
  }

  private object LeadershipStatus extends Enumeration {
    type LeadershipStatus = Value
    val LEADER, NOT_LEADER = Value
  }
}

繼續查看master中的邏輯

override def receiveWithLogging: PartialFunction[Any, Unit] = {
  case ElectedLeader => {
    //既然之前是備，現在想變成主，就需要讀取zk中的必要的信息來構造元數據
    val (storedApps, storedDrivers, storedWorkers) = persistenceEngine.readPersistedData()
    state = if (storedApps.isEmpty && storedDrivers.isEmpty && storedWorkers.isEmpty) {
      RecoveryState.ALIVE//如果沒有任何元數據需要構造，則直接置爲alive狀態
    } else {
      RecoveryState.RECOVERING//不然需要置爲恢復中
    }
    logInfo("I have been elected leader! New state: " + state)
    if (state == RecoveryState.RECOVERING) {
      beginRecovery(storedApps, storedDrivers, storedWorkers)//見下面介紹
      recoveryCompletionTask = context.system.scheduler.scheduleOnce(WORKER_TIMEOUT millis, self,
        CompleteRecovery)
    }
  }

  case CompleteRecovery => completeRecovery()

  //之前是主，現在被置爲備了，不需要額外操作，退出即可
  case RevokedLeadership => {
    logError("Leadership has been revoked -- master shutting down.")
    System.exit(0)
  }

開始恢復

private def beginRecovery(storedApps: Seq[ApplicationInfo], storedDrivers: Seq[DriverInfo],
    storedWorkers: Seq[WorkerInfo]) {
  for (app <- storedApps) {
    logInfo("Trying to recover app: " + app.id)
    try {
      registerApplication(app)//將讀到的app加載到內存
      app.state = ApplicationState.UNKNOWN//狀態置爲unknown
      app.driver ! MasterChanged(masterUrl, masterWebUiUrl)//向driver發送MasterChanged消息
    } catch {
      case e: Exception => logInfo("App " + app.id + " had exception on reconnect")
    }
  }

  for (driver <- storedDrivers) {
    // Here we just read in the list of drivers. Any drivers associated with now-lost workers
    // will be re-launched when we detect that the worker is missing.
    drivers += driver//將讀到的driver加載到內存
  }

  for (worker <- storedWorkers) {
    logInfo("Trying to recover worker: " + worker.id)
    try {
      registerWorker(worker)//將讀到的worker信息加載到內存
      worker.state = WorkerState.UNKNOWN//同樣狀態需要置爲unknown，需要等到worker發送消息過來之後才能認爲該worker是可用的
      worker.actor ! MasterChanged(masterUrl, masterWebUiUrl)//向worker發送MasterChanged消息
    } catch {
      case e: Exception => logInfo("Worker " + worker.id + " had exception on reconnect")
    }
  }
}

看driver端收到MasterChanged消息會發生什麼？在AppClient.scala中

只有主master會發送MasterChanged消息，所以這裏的masterUrl肯定是新的主master的

case MasterChanged(masterUrl, masterWebUiUrl) =>
  logInfo("Master has changed, new master is at " + masterUrl)
  //收到這個消息之後，driver需要修改之前保存的master信息，用於之後向新的master通信
  changeMaster(masterUrl)
  alreadyDisconnected = false
  sender ! MasterChangeAcknowledged(appId)//向master反饋MasterChangeAcknowledged消息

master這時會收到所有app中driver發來的消息，我們看master收到MasterChangeAcknowledged消息的處理方式，參數爲appId

case MasterChangeAcknowledged(appId) => {
  idToApp.get(appId) match {
    case Some(app) =>
      logInfo("Application has been re-registered: " + appId)
      app.state = ApplicationState.WAITING  //收到消息後將app狀態置爲WAITING
    case None =>
      logWarning("Master change ack from unknown app: " + appId)
  }

  if (canCompleteRecovery) { completeRecovery() }  //這個只是優先判斷消息處理是否都結束了，這樣就不用等待worker_timeout的時間間隔再調用completeRecovery了
}

看worker端收到MasterChanged消息會發生什麼？在Worker.scala中

case MasterChanged(masterUrl, masterWebUiUrl) =>
  logInfo("Master has changed, new master is at " + masterUrl)
  changeMaster(masterUrl, masterWebUiUrl)//同上

//master不與Executor交互，所以需要worker來告訴master關於Executor的信息
val execs = executors.values.
    map(e => new ExecutorDescription(e.appId, e.execId, e.cores, e.state))
  sender ! WorkerSchedulerStateResponse(workerId, execs.toList, drivers.keys.toSeq)

繼續看master中的處理邏輯

case WorkerSchedulerStateResponse(workerId, executors, driverIds) => {
  idToWorker.get(workerId) match {
    case Some(worker) =>
      logInfo("Worker has been re-registered: " + workerId)
      worker.state = WorkerState.ALIVE //這時可以將之前worker狀態unknown修改爲ALIVE，代表該worker可用

      //將接受到的Executor信息更新到相關的app，worker中
      val validExecutors = executors.filter(exec => idToApp.get(exec.appId).isDefined)
      for (exec <- validExecutors) {
        val app = idToApp.get(exec.appId).get
        val execInfo = app.addExecutor(worker, exec.cores, Some(exec.execId))
        worker.addExecutor(execInfo)
        execInfo.copyState(exec)
      }

      //將master中driver信息更新，狀態置爲RUNNING
      for (driverId <- driverIds) {
        drivers.find(_.id == driverId).foreach { driver =>
          driver.worker = Some(worker)
          driver.state = DriverState.RUNNING
          worker.drivers(driverId) = driver
        }
      }
    case None =>
      logWarning("Scheduler state from unknown worker: " + workerId)
  }

  if (canCompleteRecovery) { completeRecovery() }  //同上
}

這一切都處理完畢之後，看master的completeRecovery，這個是在beginRecovery調用之後，在延遲worker_timeout時間之後調用，一般情況下，上面的消息來回發送處理應該都已經結束了

private def completeRecovery() {
  // Ensure "only-once" recovery semantics using a short synchronization period.
  synchronized {
    if (state != RecoveryState.RECOVERING) { return }
    state = RecoveryState.COMPLETING_RECOVERY//狀態置爲恢復完成
  }

  // Kill off any workers and apps that didn't respond to us.
  //清理在這個worker_timeout間隔過後還未處理成功的worker和app
  workers.filter(_.state == WorkerState.UNKNOWN).foreach(removeWorker)
  apps.filter(_.state == ApplicationState.UNKNOWN).foreach(finishApplication)

  // Reschedule drivers which were not claimed by any workers
  //在一番消息通信之後，本應該在driver中更新的worker信息不見了，則重啓driver或者刪除
  drivers.filter(_.worker.isEmpty).foreach { d =>
    logWarning(s"Driver ${d.id} was not found after master recovery")
    if (d.desc.supervise) {
      logWarning(s"Re-launching ${d.id}")
      relaunchDriver(d)
    } else {
      removeDriver(d.id, DriverState.ERROR, None)
      logWarning(s"Did not re-launch ${d.id} because it was not supervised")
    }
  }

  state = RecoveryState.ALIVE  //這時恢復狀態真正結束了
  schedule() //整個選主流程結束時候，重新調度一次
  logInfo("Recovery complete - resuming operations!")
}