Spark技术内幕：Client，Master和Worker 通信源码解析

http://blog.csdn.net/anzhsoft/article/details/30802603

Spark的Cluster Manager可以有几种部署模式：

1. Standlone
2. Mesos
3. YARN
4. EC2
5. Local
   

在向集群提交计算任务后，系统的运算模型就是Driver Program定义的SparkContext向APP Master提交，有APP Master进行计算资源的调度并最终完成计算。具体阐述可以阅读《Spark：大数据的电花火石！ 》。

那么Standalone模式下，Client，Master和Worker是如何进行通信，注册并开启服务的呢？

1. node之间的RPC - akka

模块间通信有很多成熟的实现，现在很多成熟的Framework已经早已经让我们摆脱原始的Socket编程了。简单归类，可以归纳为基于消息的传递和基于资源共享的同步机制。

基于消息的传递的机制应用比较广泛的有Message Queue。Message Queue, 是一种应用程序对应用程序的通信方法。应用程序通过读写出入队列的消息（针对应用程序的数据）来通信，而无需专用连接来链接它们。消 息传递指的是程序之间通过在消息中发送数据进行通信，而不是通过直接调用彼此来通信，直接调用通常是用于诸如远程过程调用的技术。排队指的是应用程序通过 队列来通信。队列的使用除去了接收和发送应用程序同时执行的要求。其中较为成熟的MQ产品有IBM WEBSPHERE MQ和RabbitMQ（AMQP的开源实现，现在由Pivotal维护）。

还有不得不提的是ZeroMQ，一个致力于进入Linux内核的基于Socket的编程框架。官方的说法： “ZeroMQ是一个简单好用的传输层，像框架一样的一个socket library，它使得Socket编程更加简单、简洁和性能更高。是一个消息处理队列库，可在多个线程、内核和主机盒之间弹性伸缩。ZMQ的明确目标是“成为标准网络协议栈的一部分，之后进入Linux内核”。

Spark在很多模块之间的通信选择是Scala原生支持的akka，一个用 Scala 编写的库，用于简化编写容错的、高可伸缩性的 Java 和 Scala 的 Actor 模型应用。akka有以下5个特性：

1. 易于构建并行和分布式应用 （Simple Concurrency & Distribution）:  Akka在设计时采用了异步通讯和分布式架构，并对上层进行抽象，如Actors、Futures ，STM等。
2. 可靠性（Resilient by Design）: 系统具备自愈能力，在本地/远程都有监护。
3. 高性能（High Performance）：在单机中每秒可发送50,000,000个消息。内存占用小，1GB内存中可保存2,500,000个actors。
4. 弹性，无中心（Elastic — Decentralized）：自适应的负责均衡，路由，分区，配置
5. 可扩展（Extensible）：可以使用Akka 扩展包进行扩展。
   

在Spark中的Client，Master和Worker实际上都是一个actor，拿Client来说：

import akka.actor._
import akka.pattern.ask
import akka.remote.{AssociationErrorEvent, DisassociatedEvent, RemotingLifecycleEvent}

private class ClientActor(driverArgs: ClientArguments, conf: SparkConf) extends Actor with Logging {
  var masterActor: ActorSelection = _
  val timeout = AkkaUtils.askTimeout(conf)

  override def preStart() = {
    masterActor = context.actorSelection(Master.toAkkaUrl(driverArgs.master))

    context.system.eventStream.subscribe(self, classOf[RemotingLifecycleEvent])

    println(s"Sending ${driverArgs.cmd} command to ${driverArgs.master}")

    driverArgs.cmd match {
      case "launch" =>
        ...
        masterActor ! RequestSubmitDriver(driverDescription)

      case "kill" =>
        val driverId = driverArgs.driverId
        val killFuture = masterActor ! RequestKillDriver(driverId)
    }
  }

  override def receive = {

    case SubmitDriverResponse(success, driverId, message) =>
      println(message)
      if (success) pollAndReportStatus(driverId.get) else System.exit(-1)

    case KillDriverResponse(driverId, success, message) =>
      println(message)
      if (success) pollAndReportStatus(driverId) else System.exit(-1)

    case DisassociatedEvent(_, remoteAddress, _) =>
      println(s"Error connecting to master ${driverArgs.master} ($remoteAddress), exiting.")
      System.exit(-1)

    case AssociationErrorEvent(cause, _, remoteAddress, _) =>
      println(s"Error connecting to master ${driverArgs.master} ($remoteAddress), exiting.")
      println(s"Cause was: $cause")
      System.exit(-1)
  }
}

/**
 * Executable utility for starting and terminating drivers inside of a standalone cluster.
 */
object Client {
  def main(args: Array[String]) {
    println("WARNING: This client is deprecated and will be removed in a future version of Spark.")
    println("Use ./bin/spark-submit with "--master spark://host:port"")

    val conf = new SparkConf()
    val driverArgs = new ClientArguments(args)

    if (!driverArgs.logLevel.isGreaterOrEqual(Level.WARN)) {
      conf.set("spark.akka.logLifecycleEvents", "true")
    }
    conf.set("spark.akka.askTimeout", "10")
    conf.set("akka.loglevel", driverArgs.logLevel.toString.replace("WARN", "WARNING"))
    Logger.getRootLogger.setLevel(driverArgs.logLevel)

    // TODO: See if we can initialize akka so return messages are sent back using the same TCP
    //       flow. Else, this (sadly) requires the DriverClient be routable from the Master.
    val (actorSystem, _) = AkkaUtils.createActorSystem(
      "driverClient", Utils.localHostName(), 0, conf, new SecurityManager(conf))

    actorSystem.actorOf(Props(classOf[ClientActor], driverArgs, conf))

    actorSystem.awaitTermination()
  }
}

其中第19行的含义就是向Master提交Driver的请求，

masterActor ! RequestSubmitDriver(driverDescription)

而Master将在receive里处理这个请求。当然了27行到44行的是处理Client Actor收到的消息。

可以看出，通过akka，可以非常简单高效的处理模块间的通信，这可以说是Spark RPC的一大特色。

2. Client，Master和Workerq启动通信详解

源码位置：spark-1.0.0coresrcmainscalaorgapachesparkdeploy。主要涉及的类：Client.scala, Master.scala和Worker.scala。这三大模块之间的通信框架如下图。

Standalone模式下存在的角色：

1. Client：负责提交作业到Master。

2. Master：接收Client提交的作业，管理Worker，并命令Worker启动Driver和Executor。

3. Worker：负责管理本节点的资源，定期向Master汇报心跳，接收Master的命令，比如启动Driver和Executor。

实际上，Master和Worker要处理的消息要比这多得多，本图只是反映了集群启动和向集群提交运算时候的主要消息处理。

接下来将分别走读这三大角色的源码。

2.1 Client源码解析

Client启动：

object Client {
  def main(args: Array[String]) {
    println("WARNING: This client is deprecated and will be removed in a future version of Spark.")
    println("Use ./bin/spark-submit with "--master spark://host:port"")

    val conf = new SparkConf()
    val driverArgs = new ClientArguments(args)

    if (!driverArgs.logLevel.isGreaterOrEqual(Level.WARN)) {
      conf.set("spark.akka.logLifecycleEvents", "true")
    }
    conf.set("spark.akka.askTimeout", "10")
    conf.set("akka.loglevel", driverArgs.logLevel.toString.replace("WARN", "WARNING"))
    Logger.getRootLogger.setLevel(driverArgs.logLevel)

    // TODO: See if we can initialize akka so return messages are sent back using the same TCP
    //       flow. Else, this (sadly) requires the DriverClient be routable from the Master.
    val (actorSystem, _) = AkkaUtils.createActorSystem(
      "driverClient", Utils.localHostName(), 0, conf, new SecurityManager(conf))
    // 使用ClientActor初始化actorSystem
    actorSystem.actorOf(Props(classOf[ClientActor], driverArgs, conf))
    //启动并等待actorSystem的结束
    actorSystem.awaitTermination()
  }
}

从行21可以看出，核心实现是由ClientActor实现的。Client的Actor是akka.Actor的一个扩展。对于Actor，从它对recevie的override就可以看出它需要处理的消息。

override def receive = {

  case SubmitDriverResponse(success, driverId, message) =>
    println(message)
    if (success) pollAndReportStatus(driverId.get) else System.exit(-1)

  case KillDriverResponse(driverId, success, message) =>
    println(message)
    if (success) pollAndReportStatus(driverId) else System.exit(-1)

  case DisassociatedEvent(_, remoteAddress, _) =>
    println(s"Error connecting to master ${driverArgs.master} ($remoteAddress), exiting.")
    System.exit(-1)

  case AssociationErrorEvent(cause, _, remoteAddress, _) =>
    println(s"Error connecting to master ${driverArgs.master} ($remoteAddress), exiting.")
    println(s"Cause was: $cause")
    System.exit(-1)
}

2.2 Master的源码分析

源码分析详见注释。

override def receive = {
   case ElectedLeader => {
     // 被选为Master，首先判断是否该Master原来为active，如果是那么进行Recovery。
   }
   case CompleteRecovery => completeRecovery() // 删除没有响应的worker和app，并且将所有没有worker的Driver分配worker
   case RevokedLeadership => {
     // Master将关闭。
   }
   case RegisterWorker(id, workerHost, workerPort, cores, memory, workerUiPort, publicAddress) =>
   {
     // 如果该Master不是active，不做任何操作，返回
     // 如果注册过该worker id，向sender返回错误
     sender ! RegisterWorkerFailed("Duplicate worker ID")
     // 注册worker，如果worker注册成功则返回成功的消息并且进行调度
     sender ! RegisteredWorker(masterUrl, masterWebUiUrl)
     schedule()
     // 如果worker注册失败，发送消息到sender
     sender ! RegisterWorkerFailed("Attempted to re-register worker at same address: " + workerAddress)
   }
   case RequestSubmitDriver(description) => {
       // 如果master不是active，返回错误
       sender ! SubmitDriverResponse(false, None, msg)
       // 否则创建driver，返回成功的消息
       sender ! SubmitDriverResponse(true, Some(driver.id), s"Driver successfully submitted as ${driver.id}")
     }
   }
   case RequestKillDriver(driverId) => {
     if (state != RecoveryState.ALIVE) {
       // 如果master不是active，返回错误
       val msg = s"Can only kill drivers in ALIVE state. Current state: $state."
       sender ! KillDriverResponse(driverId, success = false, msg)
     } else {
       logInfo("Asked to kill driver " + driverId)
       val driver = drivers.find(_.id == driverId)
       driver match {
         case Some(d) =>
             //如果driver仍然在等待队列，从等待队列删除并且更新driver状态为KILLED
           } else {
             // 通知worker kill driver id的driver。结果会由workder发消息给master ! DriverStateChanged
             d.worker.foreach { w => w.actor ! KillDriver(driverId) }
           }
           // 注意，此时driver不一定被kill，master只是通知了worker去kill driver。
           sender ! KillDriverResponse(driverId, success = true, msg)
         case None =>
           // driver已经被kill，直接返回结果
           sender ! KillDriverResponse(driverId, success = false, msg)
       }
     }
   }
   case RequestDriverStatus(driverId) => {
     // 查找请求的driver，如果找到则返回driver的状态
     (drivers ++ completedDrivers).find(_.id == driverId) match {
       case Some(driver) =>
         sender ! DriverStatusResponse(found = true, Some(driver.state),
           driver.worker.map(_.id), driver.worker.map(_.hostPort), driver.exception)
       case None =>
         sender ! DriverStatusResponse(found = false, None, None, None, None)
     }
   }
   case RegisterApplication(description) => {
       //如果是standby，那么忽略这个消息
       //否则注册application；返回结果并且开始调度
   }
   case ExecutorStateChanged(appId, execId, state, message, exitStatus) => {
     // 通过idToApp获得app，然后通过app获得executors，从而通过execId获得executor
     val execOption = idToApp.get(appId).flatMap(app => app.executors.get(execId))
     execOption match {
       case Some(exec) => {
         exec.state = state
         exec.application.driver ! ExecutorUpdated(execId, state, message, exitStatus)
         if (ExecutorState.isFinished(state)) {
           val appInfo = idToApp(appId)
           // Remove this executor from the worker and app
           logInfo("Removing executor " + exec.fullId + " because it is " + state)
           appInfo.removeExecutor(exec)
           exec.worker.removeExecutor(exec)
          }
     ｝
   }
   case DriverStateChanged(driverId, state, exception) => {
     //  如果Driver的state为ERROR | FINISHED | KILLED | FAILED， 删除它。
   }
   case Heartbeat(workerId) => {
     // 更新worker的时间戳 workerInfo.lastHeartbeat = System.currentTimeMillis()
   }
   case MasterChangeAcknowledged(appId) => {
     //  将appId的app的状态置为WAITING，为切换Master做准备。
     }
   case WorkerSchedulerStateResponse(workerId, executors, driverIds) => {
     // 通过workerId查找到worker，那么worker的state置为ALIVE，
     // 并且查找状态为idDefined的executors，并且将这些executors都加入到app中，
     // 然后保存这些app到worker中。可以理解为Worker在Master端的Recovery
     idToWorker.get(workerId) match {
       case Some(worker) =>
         logInfo("Worker has been re-registered: " + workerId)
         worker.state = WorkerState.ALIVE

         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)
         }
         // 将所有的driver设置为RUNNING然后加入到worker中。
         for (driverId <- driverIds) {
           drivers.find(_.id == driverId).foreach { driver =>
             driver.worker = Some(worker)
             driver.state = DriverState.RUNNING
             worker.drivers(driverId) = driver
           }
         }
     }
   }
   case DisassociatedEvent(_, address, _) => {
     // 这个请求是Worker或者是App发送的。删除address对应的Worker和App
     // 如果Recovery可以结束，那么结束Recovery
   }
   case RequestMasterState => {
     //向sender返回master的状态
     sender ! MasterStateResponse(host, port, workers.toArray, apps.toArray, completedApps.toArray, drivers.toArray, completedDrivers.toArray, state)
   }
   case CheckForWorkerTimeOut => {
     //删除超时的Worker
   }
   case RequestWebUIPort => {
     //向sender返回web ui的端口号
     sender ! WebUIPortResponse(webUi.boundPort)
   }
 }

2.3 Worker 源码解析

通过对Client和Master的源码解析，相信你也知道如何去分析Worker是如何和Master进行通信的了，没错，答案就在下面：

override def receive