Spark的standalone源码分析（二）

本文主要描述Spark的standalone模式启动时候，master和work的状态transfer，并简要分析相关的代码；先上一幅状态图

1. Master启动

"$bin"/spark-daemon.sh start spark.deploy.master.Master --ip $SPARK_MASTER_IP --port $SPARK_MASTER_PORT

 --webui-port $SPARK_MASTER_WEBUI_PORT

由上述启动脚本可见，Master的启动主函数位于spark.deploy.master.Master：

private[spark] object Master {
  def main(argStrings: Array[String]) {
    val args = new MasterArguments(argStrings)
    val (actorSystem, boundPort) = AkkaUtils.createActorSystem("spark", args.ip, args.port)
    val actor = actorSystem.actorOf(
      Props(new Master(args.ip, boundPort, args.webUiPort)), name = "Master")
    actorSystem.awaitTermination()
  }
}

首先：

1. 解析参数，获得master-ip，master-port等必要参数；

2. 创建MasterActor，每一个Actor实例初始化的时候会执行preStart方法，Master对象的preStart方法会开启webui，并且为了监听connect到master的work节点的状态，MasterActor订阅监听RemoteClientLifeCycleEvent，监听所有的outbound-related events；当worker节点出现disconnected或者shutdown的时候，receive中接受到这样的event，会remove该work节点，以及在该work节点上提交的Job；

  override def preStart() {
    logInfo("Starting Spark master at spark://" + ip + ":" + port)
    // Listen for remote client disconnection events, since they don't go through Akka's watch()
    context.system.eventStream.subscribe(self, classOf[RemoteClientLifeCycleEvent])
    startWebUi()
  }

    case RemoteClientDisconnected(transport, address) => {
      // The disconnected client could've been either a worker or a job; remove whichever it was
      addressToWorker.get(address).foreach(removeWorker)
      addressToJob.get(address).foreach(removeJob)
    }

    case RemoteClientShutdown(transport, address) => {
      // The disconnected client could've been either a worker or a job; remove whichever it was
      addressToWorker.get(address).foreach(removeWorker)
      addressToJob.get(address).foreach(removeJob)
    }

2. Worker的启动

"$bin"/spark-daemon.sh start spark.deploy.worker.Worker $1

Work启动的主函数位于spark.deploy.worker.Worker，其会初始化WorkActor，在connect to master的方法中执行如下操作：

1. 判断masterURL是否合法；

2. 生成MasterActor的ActorRef，并向MasterActor提交RegisterWorker的消息，MasterActor接到消息之后会将该Worker节点加入自己维护的一个works的hashset；

 def connectToMaster() {
    masterUrl match {
      case MASTER_REGEX(masterHost, masterPort) => {
        logInfo("Connecting to master spark://" + masterHost + ":" + masterPort)
        val akkaUrl = "akka://spark@%s:%s/user/Master".format(masterHost, masterPort)
        try {
          master = context.actorFor(akkaUrl)
          master ! RegisterWorker(workerId, ip, port, cores, memory, webUiPort, publicAddress)
          context.system.eventStream.subscribe(self, classOf[RemoteClientLifeCycleEvent])
          context.watch(master) // Doesn't work with remote actors, but useful for testing
        } catch {
          case e: Exception =>
            logError("Failed to connect to master", e)
            System.exit(1)
        }
      }

      case _ =>
        logError("Invalid master URL: " + masterUrl)
        System.exit(1)
    }
  }

3. 在提交注册消息以后，与MasterActor一样，WorkerActor也订阅监听RemoteClientLifeCycleEvent，并在receive方法里面，对相应的event调用masterDisconnected，kill掉该节点的executors，并退出：

    case Terminated(_) | RemoteClientDisconnected(_, _) | RemoteClientShutdown(_, _) =>
      masterDisconnected()

补充，AKKA的Remote System Events

在AKKA Concurrency一书中，对AKKA的remote node做了描述，即在AKKA中，节点与节点之间是peer的概念，对于彼此来说，它们既是clients又是servers，如图所示：

当以plane node为主的话，outbound的都是client event，inbound的都是server event；对于Airport node亦然；