(版本定制)第9课：Spark Streaming源码解读之Receiver在Driver的精妙实现全生命周期彻底研究和思考

标签：spark streaming receiver解析

本期内容：

    1、Receiver启动方式的设想

    2、Receiver启动源码彻底分析

一：Receiver启动方式的设想 
1. Spark Streaming通过Receiver持续不断的从外部数据源接收数据，并把数据汇报给Driver端，由此每个Batch Durations就可以根据汇报的数据生成不同的Job。 
2. Receiver是在Spark Streaming应用程序启动时启动的，那么我们找Receiver在哪里启动就应该去找Spark Streaming的启动。 
3. Receivers和InputDStreams是一一对应的，默认情况下一般只有一个Receiver.

如何启动Receiver？ 
1. 从Spark Core的角度来看，Receiver的启动Spark Core并不知道，就相当于Linux的内核之上所有的都是应用程序，因此Receiver是通过Job的方式启动的

2. 一般情况下，只有一个Receiver，但是可以创建不同的数据来源的InputDStream.

final private[streaming] class DStreamGraph extends Serializable with Logging {

  private val inputStreams = new ArrayBuffer[InputDStream[_]]() //数组
  private val outputStreams = new ArrayBuffer[DStream[_]]()

3.  启动Receiver的时候,启动一个Job，这个Job里面有RDD的transformations操作和action的操作，这个Job只有一个partition.这个partition的特殊是里面只有一个成员，
这个成员就是启动的Receiver.
4.  这样做的问题：
a)  如果有多个InputDStream，那就要启动多个Receiver，每个Receiver也就相当于分片partition，那我们启动Receiver的时候理想的情况下是在不同的机器上启动Receiver，
但是Spark Core的角度来看就是应用程序，感觉不到Receiver的特殊性，所以就会按照正常的Job启动的方式来处理，极有可能在一个Executor上启动多个Receiver.
这样的话就可能导致负载不均衡。
b)  有可能启动Receiver失败，只要集群存在Receiver就不应该失败。
c)  运行过程中，就默认的而言如果是一个partition的话，那启动的时候就是一个Task，但是此Task也很可能失败，因此以Task启动的Receiver也会挂掉。

由此，可以得出，对于Receiver失败的话，后果是非常严重的，那么Spark Streaming如何防止这些事的呢，下面就寻找Receiver的创建

这里先给出答案，后面源码会详细分析： 
a) Spark使用一个Job启动一个Receiver.最大程度的保证了负载均衡。 
b) Spark Streaming指定每个Receiver运行在哪些Executor上。 
c) 如果Receiver启动失败，此时并不是Job失败，在内部会重新启动Receiver.

接下来我们通过代码一步一步解析Receiver是如何启动的

1、首先我们在编写具体的应用程序的时候，都会调用StreamingContext的start方法，其实这就是job启动的源头，我们先来看下start方法的源码：

def start(): Unit = synchronized {
  state match {
    case INITIALIZED =>
      startSite.set(DStream.getCreationSite())
      StreamingContext.ACTIVATION_LOCK.synchronized {
        StreamingContext.assertNoOtherContextIsActive()
        try {
          validate()

          // Start the streaming scheduler in a new thread, so that thread local properties
          // like call sites and job groups can be reset without affecting those of the
          // current thread.
          ThreadUtils.runInNewThread("streaming-start") {
            sparkContext.setCallSite(startSite.get)
            sparkContext.clearJobGroup()
            sparkContext.setLocalProperty(SparkContext.SPARK_JOB_INTERRUPT_ON_CANCEL, "false")
            scheduler.start() //启动JobScheduler的start方法，启动子线程，一方面为了本地初始化工作，另外一方面是不要阻塞主线程。
          }
          state = StreamingContextState.ACTIVE
        } catch {
          case NonFatal(e) =>
            logError("Error starting the context, marking it as stopped", e)
            scheduler.stop(false)
            state = StreamingContextState.STOPPED
            throw e
        }
        StreamingContext.setActiveContext(this)
      }
      shutdownHookRef = ShutdownHookManager.addShutdownHook(
        StreamingContext.SHUTDOWN_HOOK_PRIORITY)(stopOnShutdown)
      // Registering Streaming Metrics at the start of the StreamingContext
      assert(env.metricsSystem != null)
      env.metricsSystem.registerSource(streamingSource)
      uiTab.foreach(_.attach())
      logInfo("StreamingContext started")
    case ACTIVE =>
      logWarning("StreamingContext has already been started")
    case STOPPED =>
      throw new IllegalStateException("StreamingContext has already been stopped")
  }
}

2、上面调用start方法的时候，会调用JobScheduler的start()方法，在该方法里面，receiverTracker启动了，源码如下：

def start(): Unit = synchronized {
  if (eventLoop != null) return // scheduler has already been started

  logDebug("Starting JobScheduler")
  eventLoop = new EventLoop[JobSchedulerEvent]("JobScheduler") {
    override protected def onReceive(event: JobSchedulerEvent): Unit = processEvent(event)
    override protected def onError(e: Throwable): Unit = reportError("Error in job scheduler", e)
  }
  eventLoop.start()
  // attach rate controllers of input streams to receive batch completion updates
  for {
    inputDStream <- ssc.graph.getInputStreams
    rateController <- inputDStream.rateController
  } ssc.addStreamingListener(rateController)

  listenerBus.start(ssc.sparkContext)
  receiverTracker = new ReceiverTracker(ssc)
  inputInfoTracker = new InputInfoTracker(ssc)
  receiverTracker.start() //启动Receiver
  jobGenerator.start()
  logInfo("Started JobScheduler")
}

3、我们接着看下receiverTracker的start()方法，在start方法里启动了RPC消息通信体，为啥呢？因为receiverTracker会监控整个集群中的Receiver，Receiver转过来要向ReceiverTrackerEndpoint汇报自己的状态，接收的数据，包括生命周期等信息

/** Start the endpoint and receiver execution thread. */
def start(): Unit = synchronized {
  if (isTrackerStarted) {
    throw new SparkException("ReceiverTracker already started")
  }

  if (!receiverInputStreams.isEmpty) { //Receiver的启动是依据数据流的
    endpoint = ssc.env.rpcEnv.setupEndpoint(
      "ReceiverTracker", new ReceiverTrackerEndpoint(ssc.env.rpcEnv)) //汇报状态信息
    if (!skipReceiverLaunch) launchReceivers() //发起Receiver
    logInfo("ReceiverTracker started")
    trackerState = Started
  }
}

4、基于ReceiverInputDStream(是在Driver端)来获得具体的Receivers实例，然后再把他们分不到Worker节点上。一个ReceiverInputDStream只产生一个Receiver

/**
 * Get the receivers from the ReceiverInputDStreams, distributes them to the
 * worker nodes as a parallel collection, and runs them.
 */
private def launchReceivers(): Unit = {
  val receivers = receiverInputStreams.map(nis => {
    //一个输入数据来源只产生一个Receiver
    val rcvr = nis.getReceiver()
    rcvr.setReceiverId(nis.id)
    rcvr
  })

  runDummySparkJob() //启动虚拟Job来分配Receiver到不同的executor上

  logInfo("Starting " + receivers.length + " receivers")
  endpoint.send(StartAllReceivers(receivers))
}

private def runDummySparkJob(): Unit = {
  if (!ssc.sparkContext.isLocal) {
    ssc.sparkContext.makeRDD(1 to 50, 50).map(x => (x, 1)).reduceByKey(_ + _, 20).collect()
  }
  assert(getExecutors.nonEmpty)
}

ReceiverInputDStream中的getReceiver()方法获得receiver对象然后将它发送到worker节点上实例化receiver,然后去接收数据。 
此方法必须要在子类中实现。

/**
 * Gets the receiver object that will be sent to the worker nodes
 * to receive data. This method needs to defined by any specific implementation
 * of a ReceiverInputDStream.
 */
def getReceiver(): Receiver[T]

ReceiverInputDStream是抽象类，所以getReceiver方法必须要在继承的子类中实现

private[streaming]
class SocketInputDStream[T: ClassTag](
    ssc_ : StreamingContext,
    host: String,
    port: Int,
    bytesToObjects: InputStream => Iterator[T],
    storageLevel: StorageLevel
  ) extends ReceiverInputDStream[T](ssc_) {

  def getReceiver(): Receiver[T] = {
    new SocketReceiver(host, port, bytesToObjects, storageLevel) //调用SocketReceiver
  }
}

private[streaming]
class SocketReceiver[T: ClassTag](
    host: String,
    port: Int,
    bytesToObjects: InputStream => Iterator[T],
    storageLevel: StorageLevel
  ) extends Receiver[T](storageLevel) with Logging {

  def onStart() {
    // Start the thread that receives data over a connection
    new Thread("Socket Receiver") {
      setDaemon(true)
      override def run() { receive() } //启动线程，调用Receiver方法
    }.start()
  }

在receive()方法中启动socket接收数据

  /** Create a socket connection and receive data until receiver is stopped */
  def receive() {
    var socket: Socket = null
    try {
      logInfo("Connecting to " + host + ":" + port)
      socket = new Socket(host, port) //根据我们应用程序传入的host和post创建socket对象
      logInfo("Connected to " + host + ":" + port)
      val iterator = bytesToObjects(socket.getInputStream()) //接收数据
      while(!isStopped && iterator.hasNext) {
        store(iterator.next) //接收后的数据进行存储
      }
      if (!isStopped()) {
        restart("Socket data stream had no more data")
      } else {
        logInfo("Stopped receiving")
      }
    } catch {
      case e: java.net.ConnectException =>
        restart("Error connecting to " + host + ":" + port, e)
      case NonFatal(e) =>
        logWarning("Error receiving data", e)
        restart("Error receiving data", e)
    } finally {
      if (socket != null) {
        socket.close()
        logInfo("Closed socket to " + host + ":" + port)
      }
    }
  }
}

6、ReceiverTrackerEndpoint源码如下：

override def receive: PartialFunction[Any, Unit] = {
  // Local messages
  case StartAllReceivers(receivers) =>
    val scheduledLocations = schedulingPolicy.scheduleReceivers(receivers, getExecutors) // receivers就是要启动的receiver，getExecutors获得集群中的Executors的列表
    for (receiver <- receivers) {
      val executors = scheduledLocations(receiver.streamId)
      updateReceiverScheduledExecutors(receiver.streamId, executors)
      receiverPreferredLocations(receiver.streamId) = receiver.preferredLocation
      startReceiver(receiver, executors) //循环receivers，每次将一个receiver传入过去。
    }
  case RestartReceiver(receiver) =>
    // Old scheduled executors minus the ones that are not active any more
    val oldScheduledExecutors = getStoredScheduledExecutors(receiver.streamId)
    val scheduledLocations = if (oldScheduledExecutors.nonEmpty) {
        // Try global scheduling again
        oldScheduledExecutors
      } else {
        val oldReceiverInfo = receiverTrackingInfos(receiver.streamId)
        // Clear "scheduledLocations" to indicate we are going to do local scheduling
        val newReceiverInfo = oldReceiverInfo.copy(
          state = ReceiverState.INACTIVE, scheduledLocations = None)
        receiverTrackingInfos(receiver.streamId) = newReceiverInfo
        schedulingPolicy.rescheduleReceiver(
          receiver.streamId,
          receiver.preferredLocation,
          receiverTrackingInfos,
          getExecutors)
      }
    // Assume there is one receiver restarting at one time, so we don‘t need to update
    // receiverTrackingInfos
    startReceiver(receiver, scheduledLocations)
  case c: CleanupOldBlocks =>
    receiverTrackingInfos.values.flatMap(_.endpoint).foreach(_.send(c))
  case UpdateReceiverRateLimit(streamUID, newRate) =>
    for (info <- receiverTrackingInfos.get(streamUID); eP <- info.endpoint) {
      eP.send(UpdateRateLimit(newRate))
    }
  // Remote messages
  case ReportError(streamId, message, error) =>
    reportError(streamId, message, error)
}

从注释中可以看到，Spark Streaming指定receiver在那些Executors运行，而不是基于Spark Core中的Task来指定。
通过StartAllReceivers将消息发送给ReceiverTrackerEndpoint

在for循环中为每个receiver分配相应的executor。并调用startReceiver方法：

Receiver是以job的方式启动的！！！ 这里你可能会有疑惑，没有RDD和来的Job呢？首先，在startReceiver方法中，会将Receiver封装成RDD

receiverRDD: RDD[Receiver[_]] =
  (scheduledLocations.isEmpty) {
    ssc..makeRDD((receiver))
  } {
    preferredLocations = scheduledLocations.map(_.toString).distinct
    ssc..makeRDD((receiver -> preferredLocations))
  }

封装成RDD后，将RDD提交到集群中运行

future = ssc.sparkContext.submitJob[Receiver[_]](
  receiverRDDstartReceiverFunc()(__) => ())

task被发送到executor中，从RDD中取出“Receiver”然后对它执行startReceiverFunc：

// Function to start the receiver on the worker node
val startReceiverFunc: Iterator[Receiver[_]] => Unit =
  (iterator: Iterator[Receiver[_]]) => {
    if (!iterator.hasNext) {
      throw new SparkException(
        "Could not start receiver as object not found.")
    }
    if (TaskContext.get().attemptNumber() == 0) {
      val receiver = iterator.next()
      assert(iterator.hasNext == false)
      val supervisor = new ReceiverSupervisorImpl( //Receiver注册
        receiver, SparkEnv.get, serializableHadoopConf.value, checkpointDirOption)
      supervisor.start() //启动Receiver
      supervisor.awaitTermination()
    } else {
      // It‘s restarted by TaskScheduler, but we want to reschedule it again. So exit it.
    }
  }

在函数中创建了一个ReceiverSupervisorImpl对象。它用来管理具体的Receiver。

首先它会将Receiver注册到ReceiverTracker中
override protected def onReceiverStart(): Boolean = {
 val msg = RegisterReceiver(
   streamId, receiver.getClass.getSimpleName, host, executorId, endpoint)
 trackerEndpoint.askWithRetry[Boolean](msg)
}

如果注册成功，通过supervisor.start（）来启动Receiver

/** Start the supervisor */
def start() {
  onStart()
  startReceiver() //启动Receiver
}

// We will keep restarting the receiver job until ReceiverTracker is stopped
future.onComplete {
  case Success(_) =>
    if (!shouldStartReceiver) {
      onReceiverJobFinish(receiverId)
    } else {
      logInfo(s"Restarting Receiver $receiverId")
      self.send(RestartReceiver(receiver))
    }
  case Failure(e) =>
    if (!shouldStartReceiver) {
      onReceiverJobFinish(receiverId)
    } else {
      logError("Receiver has been stopped. Try to restart it.", e)
      logInfo(s"Restarting Receiver $receiverId")
      self.send(RestartReceiver(receiver))
    }
}(submitJobThreadPool)
logInfo(s"Receiver ${receiver.streamId} started")

回到receiverTracker的startReceiver方法中，只要Receiver对应的Job结束了(无论是正常还是异常结束)，而ReceiverTracker还没有停止。
它将会向ReceiverTrackerEndpoint发送一个ReStartReceiver的方法。

// We will keep restarting the receiver job until ReceiverTracker is stopped
future.onComplete {
  case Success(_) =>
    if (!shouldStartReceiver) {
      onReceiverJobFinish(receiverId)
    } else {
      logInfo(s"Restarting Receiver $receiverId")
      self.send(RestartReceiver(receiver))
    }
  case Failure(e) =>
    if (!shouldStartReceiver) {
      onReceiverJobFinish(receiverId)
    } else {
      logError("Receiver has been stopped. Try to restart it.", e)
      logInfo(s"Restarting Receiver $receiverId")
      self.send(RestartReceiver(receiver))
    }
}(submitJobThreadPool)
logInfo(s"Receiver ${receiver.streamId} started")

重新为Receiver选择一个executor，并再次运行Receiver。直到ReceiverTracker启动为止。

Spark使用submit Job的方式启动Receiver,而在应用程序执行的时候会有很多Receiver，这个时候是启动一个Receiver呢，还是把所有的Receiver通过这一个Job启动？ 
在ReceiverTracker的receive方法中startReceiver方法第一个参数就是receiver,从实现的可以看出for循环不断取出receiver,然后调用startReceiver。由此就可以得出一个Job只启动一个Receiver. 
如果Receiver启动失败，此时并不会认为是作业失败，会重新发消息给ReceiverTrackerEndpoint重新启动Receiver，这样也就确保了Receivers一定会被启动，这样就不会像Task启动Receiver的话如果失败受重试次数的影响。

简单的流程图：

备注：

1、DT大数据梦工厂微信公众号DT_Spark 
2、IMF晚8点大数据实战YY直播频道号：68917580
3、新浪微博: http://www.weibo.com/ilovepains

本文出自 “DT_Spark大数据梦工厂” 博客，请务必保留此出处http://18610086859.blog.51cto.com/11484530/1775759

(版本定制)第9课：Spark Streaming源码解读之Receiver在Driver的精妙实现全生命周期彻底研究和思考

标签：spark streaming receiver解析

原文地址：http://18610086859.blog.51cto.com/11484530/1775759