本文为精度Spark-core的源码的第一节，主要内容包括Spark Deployment的简介和Standalone模式下启动集群的详细流程精读。

注：本专题的文章皆使用Spark-1.6.3版本的源码为参考，如果Spark-2.1.0版本有重大改进的地方也会进行说明。

Spark Deployment

Spark 的部署主要有四种方式：local、standalone、yarn、mesos

图片来源：Spark-Essentials-SSW2016-TE1.pdf

其中local和standalone模式主要用于测试学习，实际生产环境下国内一般都是使用yarn，这是历史原因造成的（考虑到集群中同时有Hadoop）；而国外一般都是使用mesos，而且个人认为mesos也是一种趋势，关于yarn和mesos的部分，以后会单独进行分析，下面我们详细解读standalone模式下的集群启动的具体流程。

Standalone mode下集群启动源码精读

我们就从start-all.sh开始，主要代码如下：

# Load the Spark configuration
. "${SPARK_HOME}/sbin/spark-config.sh"

# Start Master
"${SPARK_HOME}/sbin"/start-master.sh $TACHYON_STR

# Start Workers
"${SPARK_HOME}/sbin"/start-slaves.sh $TACHYON_STR

注释说的很明确了，我们继续追踪start-master.sh

CLASS="org.apache.spark.deploy.master.Master"
...
"${SPARK_HOME}/sbin"/spark-daemon.sh start $CLASS 1 \
  --ip $SPARK_MASTER_IP --port $SPARK_MASTER_PORT --webui-port $SPARK_MASTER_WEBUI_PORT \
  $ORIGINAL_ARGS
...

可以看出，是执行了spark-daemon.sh的start方法，即通过动态加载的方式将org.apache.spark.deploy.master.Master作为一个daemon（守护线程）来运行，所以我们直接分析Master的源码：

private[deploy] object Master extends Logging {
  val SYSTEM_NAME = "sparkMaster"
  val ENDPOINT_NAME = "Master"

  def main(argStrings: Array[String]) {
    //注册log
    SignalLogger.register(log)
    //实例化SparkConf，会加载`spark.*`格式的配置信息
    val conf = new SparkConf
    //使用MasterArguments对传入的参数argStrings和默认加载的conf进行封装，并执行一些初始化操作
    val args = new MasterArguments(argStrings, conf)
    val (rpcEnv, _, _) = startRpcEnvAndEndpoint(args.host, args.port, args.webUiPort, conf)
    rpcEnv.awaitTermination()
  }

  /**
   * Start the Master and return a three tuple of:
   *   (1) The Master RpcEnv
   *   (2) The web UI bound port
   *   (3) The REST server bound port, if any
   */
  def startRpcEnvAndEndpoint(
      host: String,
      port: Int,
      webUiPort: Int,
      conf: SparkConf): (RpcEnv, Int, Option[Int]) = {
    val securityMgr = new SecurityManager(conf)
    val rpcEnv = RpcEnv.create(SYSTEM_NAME, host, port, conf, securityMgr)
    val masterEndpoint = rpcEnv.setupEndpoint(ENDPOINT_NAME,
      new Master(rpcEnv, rpcEnv.address, webUiPort, securityMgr, conf))
    val portsResponse = masterEndpoint.askWithRetry[BoundPortsResponse](BoundPortsRequest)
    (rpcEnv, portsResponse.webUIPort, portsResponse.restPort)
  }
}

首先注册log，实例化SparkConf并加载spark.*格式的配置信息，然后使用MasterArguments对传入的参数argStrings和默认加载的conf进行封装，并执行一些初始化操作，主要是加载配置信息，这里不做详细说明，我们接着往下看。

下面才是真正意义上的Start Master，startRpcEnvAndEndpoint函数中首先实例化了SecurityManager（Spark中负责安全的类），然后创建了RpcEnv（Spark的Rpc通信有三个抽象：RpcEnv、RpcEndpoint、RpcEndpointRef，这样做屏蔽了底层的实现，方便用户进行扩展，Spark-1.6.3底层的默认实现方式是Netty，而Spark-2.x已经将Akka的依赖移除），接着实例化Master，实际上就是实例化了一个RpcEndpoint（因为Master实现了ThreadSafeRpcEndpoint接口，而ThreadSafeRpcEndpoint又继承了RpcEndpoint），实例化完成后通过RpcEnv的setupEndpoint向RpcEnv进行注册，注册的时候执行了Master的onStart方法，最后返回了一个RpcEndpointRef（实际上是NettyRpcEndpointRef），通过获得的RpcEndpointRef向Master（Endpoint）发送了一条BoundPortsRequest消息，Master通过receiveAndReply方法接受到该消息（实际上是通过NettyRpcEnv中的Dispatcher进行消息的分配），模式匹配到是BoundPortsRequest类型的消息，然后执行reply方法进行回复，源码如下：

case BoundPortsRequest => {
      context.reply(BoundPortsResponse(address.port, webUi.boundPort, restServerBoundPort))
    }

至此Master启动完成，Rpc部分可以参考另一篇文章：Spark RPC 到底是个什么鬼？，下面贴出Master实例化部分和onStart方法的源码及中文注释：

Master实例化部分：

  //默认的情况下，取消的task不会从工作的队列中移除直到延迟时间完成，所以创建一个守护线程来“手动”移除它
  private val forwardMessageThread =
    ThreadUtils.newDaemonSingleThreadScheduledExecutor("master-forward-message-thread")

  //用于执行重建UI代码的守护线程
  private val rebuildUIThread =
    ThreadUtils.newDaemonSingleThreadExecutor("master-rebuild-ui-thread")
    
  //通过rebuildUIThread获得重建UI的执行上下文
  private val rebuildUIContext = ExecutionContext.fromExecutor(rebuildUIThread)

  //获取hadoop的配置文件
  private val hadoopConf = SparkHadoopUtil.get.newConfiguration(conf)

  //时间格式，用于构建application ID
  private def createDateFormat = new SimpleDateFormat("yyyyMMddHHmmss") // For application IDs

  //如果Master在60s内没有收到Worker发送的heartbeat信息就认为这个Worker timeout
  private val WORKER_TIMEOUT_MS = conf.getLong("spark.worker.timeout", 60) * 1000
  //webUI中显示的完成的application的最大个数，超过200个就移除掉(200/10,1)=20个完成的applications
  private val RETAINED_APPLICATIONS = conf.getInt("spark.deploy.retainedApplications", 200)
  //webUI中显示的完成的drivers的最大个数，超过200个就移除掉(200/10,1)=20个完成的drivers
  private val RETAINED_DRIVERS = conf.getInt("spark.deploy.retainedDrivers", 200)
  //如果Master在(REAPER_ITERATIONS + 1) * WORKER_TIMEOUT_MS)秒内仍然没有收到Worker发送的heartbeat信息，就删除这个Worker
  private val REAPER_ITERATIONS = conf.getInt("spark.dead.worker.persistence", 15)
  //recoveryMode：NONE、ZOOKEEPER、FILESYSTEM、CUSTOM，默认是NONE
  private val RECOVERY_MODE = conf.get("spark.deploy.recoveryMode", "NONE")
  //Executor失败的最大重试次数
  private val MAX_EXECUTOR_RETRIES = conf.getInt("spark.deploy.maxExecutorRetries", 10)

  //下面是各种“数据结构”，不再一一说明
  val workers = new HashSet[WorkerInfo]
  val idToApp = new HashMap[String, ApplicationInfo]
  val waitingApps = new ArrayBuffer[ApplicationInfo]
  val apps = new HashSet[ApplicationInfo]

  private val idToWorker = new HashMap[String, WorkerInfo]
  private val addressToWorker = new HashMap[RpcAddress, WorkerInfo]

  private val endpointToApp = new HashMap[RpcEndpointRef, ApplicationInfo]
  private val addressToApp = new HashMap[RpcAddress, ApplicationInfo]
  private val completedApps = new ArrayBuffer[ApplicationInfo]
  private var nextAppNumber = 0
  // Using ConcurrentHashMap so that master-rebuild-ui-thread can add a UI after asyncRebuildUI
  private val appIdToUI = new ConcurrentHashMap[String, SparkUI]

  private val drivers = new HashSet[DriverInfo]
  private val completedDrivers = new ArrayBuffer[DriverInfo]
  // Drivers currently spooled for scheduling
  private val waitingDrivers = new ArrayBuffer[DriverInfo]
  private var nextDriverNumber = 0

  Utils.checkHost(address.host, "Expected hostname")

  //下面是Metrics系统相关的代码
  private val masterMetricsSystem = MetricsSystem.createMetricsSystem("master", conf, securityMgr)
  private val applicationMetricsSystem = MetricsSystem.createMetricsSystem("applications", conf,
    securityMgr)
  private val masterSource = new MasterSource(this)

  // After onStart, webUi will be set
  private var webUi: MasterWebUI = null

  private val masterPublicAddress = {
    val envVar = conf.getenv("SPARK_PUBLIC_DNS")
    if (envVar != null) envVar else address.host
  }

  private val masterUrl = address.toSparkURL
  private var masterWebUiUrl: String = _

  //当前Master的状态：STANDBY, ALIVE, RECOVERING, COMPLETING_RECOVERY
  private var state = RecoveryState.STANDBY

  private var persistenceEngine: PersistenceEngine = _

  private var leaderElectionAgent: LeaderElectionAgent = _

  private var recoveryCompletionTask: ScheduledFuture[_] = _

  private var checkForWorkerTimeOutTask: ScheduledFuture[_] = _

  // As a temporary workaround before better ways of configuring memory, we allow users to set
  // a flag that will perform round-robin scheduling across the nodes (spreading out each app
  // among all the nodes) instead of trying to consolidate each app onto a small # of nodes.
  // 避免将application的运行限制在固定的几个节点上
  private val spreadOutApps = conf.getBoolean("spark.deploy.spreadOut", true)

  // Default maxCores for applications that don't specify it (i.e. pass Int.MaxValue)
  private val defaultCores = conf.getInt("spark.deploy.defaultCores", Int.MaxValue)
  if (defaultCores < 1) {
    throw new SparkException("spark.deploy.defaultCores must be positive")
  }

  // Alternative application submission gateway that is stable across Spark versions
  // 用来接受application提交的restServer
  private val restServerEnabled = conf.getBoolean("spark.master.rest.enabled", true)
  private var restServer: Option[StandaloneRestServer] = None
  private var restServerBoundPort: Option[Int] = None

onStart方法：

override def onStart(): Unit = {
    //打日志
    logInfo("Starting Spark master at " + masterUrl)
    logInfo(s"Running Spark version ${org.apache.spark.SPARK_VERSION}")
    //实例化standalone模式下的MasterWebUI并绑定到HTTP Server
    webUi = new MasterWebUI(this, webUiPort)
    webUi.bind()
    //可以通过这个Url地址看到Master的信息
    masterWebUiUrl = "http://" + masterPublicAddress + ":" + webUi.boundPort
    
    //以固定的时间间隔检查并移除time-out的worker
    checkForWorkerTimeOutTask = forwardMessageThread.scheduleAtFixedRate(new Runnable {
      override def run(): Unit = Utils.tryLogNonFatalError {
        self.send(CheckForWorkerTimeOut)
      }
    }, 0, WORKER_TIMEOUT_MS, TimeUnit.MILLISECONDS)

    //实例化并启动restServer用于接受application的提交
    if (restServerEnabled) {
      val port = conf.getInt("spark.master.rest.port", 6066)
      restServer = Some(new StandaloneRestServer(address.host, port, conf, self, masterUrl))
    }
    restServerBoundPort = restServer.map(_.start())

    //启动MetricsSystem
    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)

    //序列化器
    val serializer = new JavaSerializer(conf)
    
    //恢复机制，包括持久化引擎和选举机制
    val (persistenceEngine_, leaderElectionAgent_) = RECOVERY_MODE match {
      case "ZOOKEEPER" =>
        logInfo("Persisting recovery state to ZooKeeper")
        val zkFactory =
          new ZooKeeperRecoveryModeFactory(conf, serializer)
        (zkFactory.createPersistenceEngine(), zkFactory.createLeaderElectionAgent(this))
      case "FILESYSTEM" =>
        val fsFactory =
          new FileSystemRecoveryModeFactory(conf, serializer)
        (fsFactory.createPersistenceEngine(), fsFactory.createLeaderElectionAgent(this))
      case "CUSTOM" =>
        val clazz = Utils.classForName(conf.get("spark.deploy.recoveryMode.factory"))
        val factory = clazz.getConstructor(classOf[SparkConf], classOf[Serializer])
          .newInstance(conf, serializer)
          .asInstanceOf[StandaloneRecoveryModeFactory]
        (factory.createPersistenceEngine(), factory.createLeaderElectionAgent(this))
      case _ =>
        (new BlackHolePersistenceEngine(), new MonarchyLeaderAgent(this))
    }
    persistenceEngine = persistenceEngine_
    leaderElectionAgent = leaderElectionAgent_
  }

下面介绍Worker的启动

start-slaves.sh:

# Launch the slaves
"${SPARK_HOME}/sbin/slaves.sh" cd "${SPARK_HOME}" \; "${SPARK_HOME}/sbin/start-slave.sh" "spark://$SPARK_MASTER_IP:$SPARK_MASTER_PORT"

start-slave.sh:

CLASS="org.apache.spark.deploy.worker.Worker"
...
  "${SPARK_HOME}/sbin"/spark-daemon.sh start $CLASS $WORKER_NUM \
     --webui-port "$WEBUI_PORT" $PORT_FLAG $PORT_NUM $MASTER "$@"

和Master的启动类似，我们直接看Worker文件，仍然从main方法开始：

def main(argStrings: Array[String]) {
    SignalLogger.register(log)
    val conf = new SparkConf
    val args = new WorkerArguments(argStrings, conf)
    val rpcEnv = startRpcEnvAndEndpoint(args.host, args.port, args.webUiPort, args.cores,
      args.memory, args.masters, args.workDir, conf = conf)
    rpcEnv.awaitTermination()
  }
  
def startRpcEnvAndEndpoint(
      host: String,
      port: Int,
      webUiPort: Int,
      cores: Int,
      memory: Int,
      masterUrls: Array[String],
      workDir: String,
      workerNumber: Option[Int] = None,
      conf: SparkConf = new SparkConf): RpcEnv = {

    // The LocalSparkCluster runs multiple local sparkWorkerX RPC Environments
    val systemName = SYSTEM_NAME + workerNumber.map(_.toString).getOrElse("")
    val securityMgr = new SecurityManager(conf)
    val rpcEnv = RpcEnv.create(systemName, host, port, conf, securityMgr)
    val masterAddresses = masterUrls.map(RpcAddress.fromSparkURL(_))
    rpcEnv.setupEndpoint(ENDPOINT_NAME, new Worker(rpcEnv, webUiPort, cores, memory,
      masterAddresses, systemName, ENDPOINT_NAME, workDir, conf, securityMgr))
    rpcEnv
  }

可以看到前面和Master类似，只不过Worker有可能是多个，所以需要根据workerNumber构造一个systemName，用来创建不同的RpcEnv，然后实例化Worker（即实例化Endpoint），实例化的时候需要传入masterAddresses（注意此处可能有多个Master），以便以后向Master注册，同时由于要向对应的RpcEnv注册，注册的时候同样要执行Worker的onStart方法，我会将Worker实例化和onStart的源码放到后面，这里我们先来看一下Worker向Master注册的代码（onStart方法中调用registerWithMaster）：

private def registerWithMaster() {
    // onDisconnected may be triggered multiple times, so don't attempt registration
    // if there are outstanding registration attempts scheduled.
    registrationRetryTimer match {
      case None =>
        registered = false
        registerMasterFutures = tryRegisterAllMasters()
        connectionAttemptCount = 0
        registrationRetryTimer = Some(forwordMessageScheduler.scheduleAtFixedRate(
          new Runnable {
            override def run(): Unit = Utils.tryLogNonFatalError {
              Option(self).foreach(_.send(ReregisterWithMaster))
            }
          },
          INITIAL_REGISTRATION_RETRY_INTERVAL_SECONDS,
          INITIAL_REGISTRATION_RETRY_INTERVAL_SECONDS,
          TimeUnit.SECONDS))
      case Some(_) =>
        logInfo("Not spawning another attempt to register with the master, since there is an" +
          " attempt scheduled already.")
    }
  }

可以看到内部调用了tryRegisterAllMasters方法：

private def tryRegisterAllMasters(): Array[JFuture[_]] = {
    masterRpcAddresses.map { masterAddress =>
      registerMasterThreadPool.submit(new Runnable {
        override def run(): Unit = {
          try {
            logInfo("Connecting to master " + masterAddress + "...")
            val masterEndpoint =
              rpcEnv.setupEndpointRef(Master.SYSTEM_NAME, masterAddress, Master.ENDPOINT_NAME)
            registerWithMaster(masterEndpoint)
          } catch {
            case ie: InterruptedException => // Cancelled
            case NonFatal(e) => logWarning(s"Failed to connect to master $masterAddress", e)
          }
        }
      })
    }
  }

通过一个名为registerMasterThreadPool的线程池（最大线程数为Worker的个数）来运行run方法中的内容：首先通过setupEndpointRef方法获得其中一个Master的一个引用（RpcEndpointRef），然后执行registerWithMaster(masterEndpoint)方法，刚才得到的Master的引用作为参数传入，下面进入registerWithMaster方法：（注意此处的registerWithMaster方法是有一个RpcEndpointRef作为参数的，和刚开始的那个不一样）

private def registerWithMaster(masterEndpoint: RpcEndpointRef): Unit = {
    masterEndpoint.ask[RegisterWorkerResponse](RegisterWorker(
      workerId, host, port, self, cores, memory, webUi.boundPort, publicAddress))
      .onComplete {
        // This is a very fast action so we can use "ThreadUtils.sameThread"
        case Success(msg) =>
          Utils.tryLogNonFatalError {
            handleRegisterResponse(msg)
          }
        case Failure(e) =>
          logError(s"Cannot register with master: ${masterEndpoint.address}", e)
          System.exit(1)
      }(ThreadUtils.sameThread)
  }

内部使用masterEndpoint（Master的RpcEndpointRef）的ask方法向Master发送一条RegisterWorker的消息，并使用onComplete方法接受Master的处理结果，下面我们先来看一下消息到达Master端进行怎样的处理：

override def receiveAndReply(context: RpcCallContext): PartialFunction[Any, Unit] = {
    case RegisterWorker(
        id, workerHost, workerPort, workerRef, cores, memory, workerUiPort, publicAddress) => {
      logInfo("Registering worker %s:%d with %d cores, %s RAM".format(
        workerHost, workerPort, cores, Utils.megabytesToString(memory)))
      if (state == RecoveryState.STANDBY) {
        context.reply(MasterInStandby)
      } else if (idToWorker.contains(id)) {
        context.reply(RegisterWorkerFailed("Duplicate worker ID"))
      } else {
        val worker = new WorkerInfo(id, workerHost, workerPort, cores, memory,
          workerRef, workerUiPort, publicAddress)
        if (registerWorker(worker)) {
          persistenceEngine.addWorker(worker)
          context.reply(RegisteredWorker(self, masterWebUiUrl))
          schedule()
        } else {
          val workerAddress = worker.endpoint.address
          logWarning("Worker registration failed. Attempted to re-register worker at same " +
            "address: " + workerAddress)
          context.reply(RegisterWorkerFailed("Attempted to re-register worker at same address: "
            + workerAddress))
        }
      }
    }

首先receiveAndReply方法匹配到Worker发过来的RegisterWorker消息，然后执行具体的操作：打了一个日志，判断Master现在的状态，如果是STANDBY就reply一个MasterInStandby的消息，如果idToWorker中已经存在该Worker的ID就回复重复的worker ID的失败信息，如果都不是，将获得的Worker信息用WorkerInfo进行封装，然后执行registerWorker(worker)操作注册该Worker，如果成功就向persistenceEngine中添加该Worker并reply给Worker RegisteredWorker(self, masterWebUiUrl)消息并执行schedule方法，如果注册失败就reply RegisterWorkerFailed消息，下面我们具体看一下Master端是如何注册Worker的，即registerWorker(worker)方法：

private def registerWorker(worker: WorkerInfo): Boolean = {
    // There may be one or more refs to dead workers on this same node (w/ different ID's),
    // remove them.
    workers.filter { w =>
      (w.host == worker.host && w.port == worker.port) && (w.state == WorkerState.DEAD)
    }.foreach { w =>
      workers -= w
    }

    val workerAddress = worker.endpoint.address
    if (addressToWorker.contains(workerAddress)) {
      val oldWorker = addressToWorker(workerAddress)
      if (oldWorker.state == WorkerState.UNKNOWN) {
        // A worker registering from UNKNOWN implies that the worker was restarted during recovery.
        // The old worker must thus be dead, so we will remove it and accept the new worker.
        removeWorker(oldWorker)
      } else {
        logInfo("Attempted to re-register worker at same address: " + workerAddress)
        return false
      }
    }

    workers += worker
    idToWorker(worker.id) = worker
    addressToWorker(workerAddress) = worker
    true
  }

首先判断是否有和该Worker的host和port相同且状态为DEAD的Worker，如果有就remove掉，然后获得该Worker的RpcAddress，然后根据RpcAddress判断addressToWorker中是否有相同地址的记录，如果有记录且老的Worker的状态为UNKNOWN就remove掉老的Worker，如果没有记录就打日志并返回false（导致上一步reply：RegisterWorkerFailed）然后分别在workers、idToWorker、addressToWorker中添加该Worker，最后返回true，导致上一步向Worker reply注册成功的消息：context.reply(RegisteredWorker(self, masterWebUiUrl))，并执行schedule()，即向等待的applications分配当前可用的资源（每当新的application加入或者有资源变化时都会调用该方法），这个方法我会用单独的一片文章详细分析，现在我们先来看Worker端是如何进行回复的，回到上面的registerWithMaster方法（有参数的），我们直接看成功后执行的handleRegisterResponse(msg)这个方法：

private def handleRegisterResponse(msg: RegisterWorkerResponse): Unit = synchronized {
    msg match {
      case RegisteredWorker(masterRef, masterWebUiUrl) =>
        logInfo("Successfully registered with master " + masterRef.address.toSparkURL)
        registered = true
        changeMaster(masterRef, masterWebUiUrl)
        forwordMessageScheduler.scheduleAtFixedRate(new Runnable {
          override def run(): Unit = Utils.tryLogNonFatalError {
            self.send(SendHeartbeat)
          }
        }, 0, HEARTBEAT_MILLIS, TimeUnit.MILLISECONDS)
        if (CLEANUP_ENABLED) {
          logInfo(
            s"Worker cleanup enabled; old application directories will be deleted in: $workDir")
          forwordMessageScheduler.scheduleAtFixedRate(new Runnable {
            override def run(): Unit = Utils.tryLogNonFatalError {
              self.send(WorkDirCleanup)
            }
          }, CLEANUP_INTERVAL_MILLIS, CLEANUP_INTERVAL_MILLIS, TimeUnit.MILLISECONDS)
        }

      case RegisterWorkerFailed(message) =>
        if (!registered) {
          logError("Worker registration failed: " + message)
          System.exit(1)
        }

      case MasterInStandby =>
        // Ignore. Master not yet ready.
    }
  }

依然是模式匹配的方式：

• 如果接受到的是RegisteredWorker，会执行changeMaster方法，取消最后一次的重试，然后向自己的RpcEnv发送SendHeartBeat消息，使用receive方法接受到该消息后会通过sendToMaster方法向Master发送心跳，最后判断CLEANUP_ENABLED如果开启就向自己的RpcEnv发送WorkDirCleanup消息，接受到消息后将老的application的目录清除
• 如果接受到的是RegisterWorkerFailed就表明注册失败

changeMaster发送：

private def changeMaster(masterRef: RpcEndpointRef, uiUrl: String) {
    // activeMasterUrl it's a valid Spark url since we receive it from master.
    activeMasterUrl = masterRef.address.toSparkURL
    activeMasterWebUiUrl = uiUrl
    master = Some(masterRef)
    connected = true
    // Cancel any outstanding re-registration attempts because we found a new master
    cancelLastRegistrationRetry()
  }

cancelLastRegistrationRetry:

private def cancelLastRegistrationRetry(): Unit = {
    if (registerMasterFutures != null) {
      registerMasterFutures.foreach(_.cancel(true))
      registerMasterFutures = null
    }
    registrationRetryTimer.foreach(_.cancel(true))
    registrationRetryTimer = None
  }

如果Worker注册失败同样会通过registrationRetryTimer进行重试：

registrationRetryTimer = Some(forwordMessageScheduler.scheduleAtFixedRate(
          new Runnable {
            override def run(): Unit = Utils.tryLogNonFatalError {
              Option(self).foreach(_.send(ReregisterWithMaster))
            }
          },
          INITIAL_REGISTRATION_RETRY_INTERVAL_SECONDS,
          INITIAL_REGISTRATION_RETRY_INTERVAL_SECONDS,
          TimeUnit.SECONDS))

可以看到向自己发送重新注册的消息：ReregisterWithMaster，receive接收到后会执行reregisterWithMaster()方法：

private def reregisterWithMaster(): Unit = {
    Utils.tryOrExit {
      //重试次数加1
      connectionAttemptCount += 1
      if (registered) {
        //如果已经注册了，就取消重试
        cancelLastRegistrationRetry()
      } else if (connectionAttemptCount <= TOTAL_REGISTRATION_RETRIES) {  //判断是否超过最大重试次数
        logInfo(s"Retrying connection to master (attempt # $connectionAttemptCount)")
        /**
         * Re-register with the active master this worker has been communicating with. If there
         * is none, then it means this worker is still bootstrapping and hasn't established a
         * connection with a master yet, in which case we should re-register with all masters.
         *
         * It is important to re-register only with the active master during failures. Otherwise,
         * if the worker unconditionally attempts to re-register with all masters, the following
         * race condition may arise and cause a "duplicate worker" error detailed in SPARK-4592:
         *
         *   (1) Master A fails and Worker attempts to reconnect to all masters
         *   (2) Master B takes over and notifies Worker
         *   (3) Worker responds by registering with Master B
         *   (4) Meanwhile, Worker's previous reconnection attempt reaches Master B,
         *       causing the same Worker to register with Master B twice
         *
         * Instead, if we only register with the known active master, we can assume that the
         * old master must have died because another master has taken over. Note that this is
         * still not safe if the old master recovers within this interval, but this is a much
         * less likely scenario.
         */
        master match {
          case Some(masterRef) =>
            // registered == false && master != None means we lost the connection to master, so
            // masterRef cannot be used and we need to recreate it again. Note: we must not set
            // master to None due to the above comments.
            // 这里说的很清楚，如果注册失败了，但是master != None说明我们失去了和master的连接，所以需要重新创建一个masterRef
            // 先取消原来阻塞的用来等待消息回复的线程
            if (registerMasterFutures != null) {
              registerMasterFutures.foreach(_.cancel(true))
            }
            
            // 然后创建新的masterRef，然后重新注册
            val masterAddress = masterRef.address
            registerMasterFutures = Array(registerMasterThreadPool.submit(new Runnable {
              override def run(): Unit = {
                try {
                  logInfo("Connecting to master " + masterAddress + "...")
                  val masterEndpoint =
                    rpcEnv.setupEndpointRef(Master.SYSTEM_NAME, masterAddress, Master.ENDPOINT_NAME)
                  registerWithMaster(masterEndpoint)
                } catch {
                  case ie: InterruptedException => // Cancelled
                  case NonFatal(e) => logWarning(s"Failed to connect to master $masterAddress", e)
                }
              }
            }))
          case None =>
            // 如果没有masterRef，先取消原来阻塞的用来等待消息回复的线程
            if (registerMasterFutures != null) {
              registerMasterFutures.foreach(_.cancel(true))
            }
            
            // 然后执行最初的注册，即tryRegisterAllMasters
            // We are retrying the initial registration
            registerMasterFutures = tryRegisterAllMasters()
        }
        // We have exceeded the initial registration retry threshold
        // All retries from now on should use a higher interval
        // 如果超过刚开始设置的重试注册次数，取消之前的重试，开启新的注册，并改变重试次数和时间间隔
        // 刚开始的重试默认为6次，时间间隔在5到15秒之间，接下来的10次重试时间间隔在30到90秒之间
        if (connectionAttemptCount == INITIAL_REGISTRATION_RETRIES) {
          registrationRetryTimer.foreach(_.cancel(true))
          registrationRetryTimer = Some(
            forwordMessageScheduler.scheduleAtFixedRate(new Runnable {
              override def run(): Unit = Utils.tryLogNonFatalError {
                self.send(ReregisterWithMaster)
              }
            }, PROLONGED_REGISTRATION_RETRY_INTERVAL_SECONDS,
              PROLONGED_REGISTRATION_RETRY_INTERVAL_SECONDS,
              TimeUnit.SECONDS))
        }
      } else {
        logError("All masters are unresponsive! Giving up.")
        System.exit(1)
      }
    }
  }

至此Worker的启动和注册完成，即start-all.sh执行完成。

下面是Worker的初始化部分和onStart方法的源码及注释（重要部分）：

初始化部分：

  private val host = rpcEnv.address.host
  private val port = rpcEnv.address.port

  Utils.checkHost(host, "Expected hostname")
  assert (port > 0)

  // A scheduled executor used to send messages at the specified time.
  private val forwordMessageScheduler =
    ThreadUtils.newDaemonSingleThreadScheduledExecutor("worker-forward-message-scheduler")

  // A separated thread to clean up the workDir. Used to provide the implicit parameter of `Future`
  // methods.
  private val cleanupThreadExecutor = ExecutionContext.fromExecutorService(
    ThreadUtils.newDaemonSingleThreadExecutor("worker-cleanup-thread"))

  // For worker and executor IDs
  private def createDateFormat = new SimpleDateFormat("yyyyMMddHHmmss")
  // 发送心跳的时间间隔：timeout的时间 / 4
  // Send a heartbeat every (heartbeat timeout) / 4 milliseconds
  private val HEARTBEAT_MILLIS = conf.getLong("spark.worker.timeout", 60) * 1000 / 4

  // 重试的模型及其次数设置
  // Model retries to connect to the master, after Hadoop's model.
  // The first six attempts to reconnect are in shorter intervals (between 5 and 15 seconds)
  // Afterwards, the next 10 attempts are between 30 and 90 seconds.
  // A bit of randomness is introduced so that not all of the workers attempt to reconnect at
  // the same time.
  private val INITIAL_REGISTRATION_RETRIES = 6
  private val TOTAL_REGISTRATION_RETRIES = INITIAL_REGISTRATION_RETRIES + 10
  private val FUZZ_MULTIPLIER_INTERVAL_LOWER_BOUND = 0.500
  private val REGISTRATION_RETRY_FUZZ_MULTIPLIER = {
    val randomNumberGenerator = new Random(UUID.randomUUID.getMostSignificantBits)
    randomNumberGenerator.nextDouble + FUZZ_MULTIPLIER_INTERVAL_LOWER_BOUND
  }
  private val INITIAL_REGISTRATION_RETRY_INTERVAL_SECONDS = (math.round(10 *
    REGISTRATION_RETRY_FUZZ_MULTIPLIER))
  private val PROLONGED_REGISTRATION_RETRY_INTERVAL_SECONDS = (math.round(60
    * REGISTRATION_RETRY_FUZZ_MULTIPLIER))

  //CLEANUP相关的设置
  private val CLEANUP_ENABLED = conf.getBoolean("spark.worker.cleanup.enabled", false)
  // How often worker will clean up old app folders
  private val CLEANUP_INTERVAL_MILLIS =
    conf.getLong("spark.worker.cleanup.interval", 60 * 30) * 1000
  // TTL for app folders/data;  after TTL expires it will be cleaned up
  private val APP_DATA_RETENTION_SECONDS =
    conf.getLong("spark.worker.cleanup.appDataTtl", 7 * 24 * 3600)

  private val testing: Boolean = sys.props.contains("spark.testing")
  //对master的引用
  private var master: Option[RpcEndpointRef] = None
  private var activeMasterUrl: String = ""
  private[worker] var activeMasterWebUiUrl : String = ""
  private val workerUri = rpcEnv.uriOf(systemName, rpcEnv.address, endpointName)
  private var registered = false
  private var connected = false
  private val workerId = generateWorkerId()
  private val sparkHome =
    if (testing) {
      assert(sys.props.contains("spark.test.home"), "spark.test.home is not set!")
      new File(sys.props("spark.test.home"))
    } else {
      new File(sys.env.get("SPARK_HOME").getOrElse("."))
    }

  var workDir: File = null
  val finishedExecutors = new LinkedHashMap[String, ExecutorRunner]
  val drivers = new HashMap[String, DriverRunner]
  val executors = new HashMap[String, ExecutorRunner]
  val finishedDrivers = new LinkedHashMap[String, DriverRunner]
  val appDirectories = new HashMap[String, Seq[String]]
  val finishedApps = new HashSet[String]

  val retainedExecutors = conf.getInt("spark.worker.ui.retainedExecutors",
    WorkerWebUI.DEFAULT_RETAINED_EXECUTORS)
  val retainedDrivers = conf.getInt("spark.worker.ui.retainedDrivers",
    WorkerWebUI.DEFAULT_RETAINED_DRIVERS)

  // The shuffle service is not actually started unless configured.
  private val shuffleService = new ExternalShuffleService(conf, securityMgr)

  private val publicAddress = {
    val envVar = conf.getenv("SPARK_PUBLIC_DNS")
    if (envVar != null) envVar else host
  }
  private var webUi: WorkerWebUI = null

  private var connectionAttemptCount = 0

  private val metricsSystem = MetricsSystem.createMetricsSystem("worker", conf, securityMgr)
  private val workerSource = new WorkerSource(this)

  private var registerMasterFutures: Array[JFuture[_]] = null
  private var registrationRetryTimer: Option[JScheduledFuture[_]] = None

  // 用来和Master注册使用的线程池，默认线程的最大个数为Worker的个数
  // A thread pool for registering with masters. Because registering with a master is a blocking
  // action, this thread pool must be able to create "masterRpcAddresses.size" threads at the same
  // time so that we can register with all masters.
  private val registerMasterThreadPool = ThreadUtils.newDaemonCachedThreadPool(
    "worker-register-master-threadpool",
    masterRpcAddresses.size // Make sure we can register with all masters at the same time
  )

  var coresUsed = 0
  var memoryUsed = 0

onStart()方法：

override def onStart() {
    assert(!registered)
    logInfo("Starting Spark worker %s:%d with %d cores, %s RAM".format(
      host, port, cores, Utils.megabytesToString(memory)))
    logInfo(s"Running Spark version ${org.apache.spark.SPARK_VERSION}")
    logInfo("Spark home: " + sparkHome)
    // 创建Work的目录
    createWorkDir()
    // 开启 external shuffle service
    shuffleService.startIfEnabled()
    webUi = new WorkerWebUI(this, workDir, webUiPort)
    webUi.bind()
    // 向Master注册自己
    registerWithMaster()

    // metrics系统
    metricsSystem.registerSource(workerSource)
    metricsSystem.start()
    // Attach the worker metrics servlet handler to the web ui after the metrics system is started.
    metricsSystem.getServletHandlers.foreach(webUi.attachHandler)
  }

本文简单介绍了Spark的几种部署模式，并详细的分析了start-all.sh所执行源码（Master的启动和注册、Worker的启动和向Master的注册）的具体流程，当然Master的schedule方法并没有详细说明，我们会单独用一篇文章进行详细的分析。

本文参考和拓展阅读：

Spark-1.6.3源码

Spark-2.1.0源码

本文为原创，欢迎转载，转载请注明出处、作者，谢谢！