Spark技术内幕：Master的故障恢复

Spark技术内幕：Master基于ZooKeeper的High Availability（HA）源码实现  详细阐述了使用ZK实现的Master的HA，那么Master是如何快速故障恢复的呢？

处于Standby状态的Master在接收到org.apache.spark.deploy.master.ZooKeeperLeaderElectionAgent发送的ElectedLeader消息后，就开始通过ZK中保存的Application，Driver和Worker的元数据信息进行故障恢复了，它的状态也从RecoveryState.STANDBY变为RecoveryState.RECOVERING了。当然了，如果没有任何需要恢复的数据，Master的状态就直接变为RecoveryState.ALIVE，开始对外服务了。

一方面Master通过

beginRecovery(storedApps, storedDrivers, storedWorkers) 

recoveryCompletionTask = context.system.scheduler.scheduleOnce(WORKER_TIMEOUT millis, self,
          CompleteRecovery)

首先看一下如何通过ZooKeeperLeaderElectionAgent提供的接口恢复数据。

  override def readPersistedData(): (Seq[ApplicationInfo], Seq[DriverInfo], Seq[WorkerInfo]) = {
    val sortedFiles = zk.getChildren().forPath(WORKING_DIR).toList.sorted // 获取所有的文件
    val appFiles = sortedFiles.filter(_.startsWith("app_")) //获取Application的序列化文件
    val apps = appFiles.map(deserializeFromFile[ApplicationInfo]).flatten //将Application的元数据反序列化
    val driverFiles = sortedFiles.filter(_.startsWith("driver_")) //获取Driver的序列化文件
    val drivers = driverFiles.map(deserializeFromFile[DriverInfo]).flatten //将Driver的元数据反序列化
    val workerFiles = sortedFiles.filter(_.startsWith("worker_")) // 获取Worker的序列化文件
    val workers = workerFiles.map(deserializeFromFile[WorkerInfo]).flatten // 将Worker的元数据反序列化
    (apps, drivers, workers)
  }

恢复Application的步骤：

1. 置待恢复的Application的状态为UNKNOWN，向AppClient发送MasterChanged的消息
2. AppClient收到后改变其保存的Master的信息，包括URL和Master actor的信息，回复MasterChangeAcknowledged(appId)
3. Master收到后通过appId后将Application的状态置为WAITING
4. 检查如果所有的worker和Application的状态都不是UNKNOWN，那么恢复结束，调用completeRecovery()

恢复Worker的步骤：

1. 重新注册Worker（实际上是更新Master本地维护的数据结构），置状态为UNKNOWN
2. 向Worker发送MasterChanged的消息
3. Worker收到消息后，向Master回复 消息WorkerSchedulerStateResponse，并通过该消息上报executor和driver的信息。
4. Master收到消息后，会置该Worker的状态为ALIVE，并且会检查该Worker上报的信息是否与自己从ZK中获取的数据一致，包括executor和driver。一致的executor和driver将被恢复。对于Driver，其状态被置为RUNNING。
5. 检查如果所有的worker和Application的状态都不是UNKNOWN，那么恢复结束，调用completeRecovery()

  def beginRecovery(storedApps: Seq[ApplicationInfo], storedDrivers: Seq[DriverInfo],
      storedWorkers: Seq[WorkerInfo]) {
    for (app <- storedApps) { // 逐个恢复Application
      logInfo("Trying to recover app: " + app.id)
      try {
        registerApplication(app)
        app.state = ApplicationState.UNKNOWN
        app.driver ! MasterChanged(masterUrl, masterWebUiUrl) //向AppClient发送Master变化的消息，AppClient会回复MasterChangeAcknowledged
      } 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 // 在Worker恢复后，Worker会主动上报运行其上的executors和drivers从而使得Master恢复executor和driver的信息。
    }

    for (worker <- storedWorkers) { //逐个恢复Worker
      logInfo("Trying to recover worker: " + worker.id)
      try {
        registerWorker(worker) //重新注册Worker
        worker.state = WorkerState.UNKNOWN
        worker.actor ! MasterChanged(masterUrl, masterWebUiUrl) //向Worker发送Master变化的消息，Worker会回复WorkerSchedulerStateResponse
      } catch {
        case e: Exception => logInfo("Worker " + worker.id + " had exception on reconnect")
      }
    }
  }

  //调用时机
  // 1. 在恢复开始后的60s会被强制调用
  // 2. 在每次收到AppClient和Worker的消息回复后会检查如果Application和worker的状态都不为UNKNOWN，则调用
  def completeRecovery() {
    // Ensure "only-once" recovery semantics using a short synchronization period.
    synchronized {
      if (state != RecoveryState.RECOVERING) { return }
      state = RecoveryState.COMPLETING_RECOVERY
    }

    // Kill off any workers and apps that didn't respond to us. 删除在60s内没有回应的app和worker
    workers.filter(_.state == WorkerState.UNKNOWN).foreach(removeWorker)
    apps.filter(_.state == ApplicationState.UNKNOWN).foreach(finishApplication)

    // Reschedule drivers which were not claimed by any workers
    drivers.filter(_.worker.isEmpty).foreach { d => // 如果driver的worker为空，则relaunch。
      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
    schedule() 
    logInfo("Recovery complete - resuming operations!")
  }

但是对于一个拥有几千个节点的集群来说，60s设置的是否合理？毕竟现在没有使用Standalone模式部署几千个节点的吧？因此硬编码60s看上去也十分合理，毕竟都是逻辑很简单的调用，如果一些节点60S没有返回，那么下线这部分机器也是合理的。

通过设置spark.worker.timeout，可以自定义超时时间。