Spark Streaming之运行原理

最新推荐文章于 2023-03-20 14:11:36 发布

莫言静好、

最新推荐文章于 2023-03-20 14:11:36 发布

阅读量1.2k

点赞数

分类专栏：大数据/spark 文章标签： spark spark streaming 运行原理

本文链接：https://blog.csdn.net/zhanglh046/article/details/78505053

版权

大数据/spark 专栏收录该内容

23 篇文章 9 订阅

订阅专栏

一启动流处理引擎

1.1初始化StreamingContext

首先需要初始化StreamingContext，在初始化的过程中会对DStreamGraph、JobScheduler等进行初始化，DStreamGraph类似于RDD的有向无环图，包含DStream之间相互依赖的有向无环图；JobScheduler定时查看DStreamGraph，然后根据流入的数据生成运行作业

1.2 创建InputDStream

根据你采用不同的数据源，可能生成的输入数据流不一样

1.3 启动JobScheduler

创建完成InputDStream之后，调用StreamingContext的start方法启动应用程序，并且需要启动JobScheduler，启动JobScheduler的时候会实例化ReceiverTracker和JobGenerator.

def start(): Unit = synchronized {
// JobShceduler已经启动则退出
if (eventLoop != null) return

logDebug("Starting JobScheduler")
eventLoop = new EventLoop[JobSchedulerEvent]("JobScheduler") {
    overrideprotected def onReceive(event: JobSchedulerEvent): Unit = processEvent(event)

    overrideprotected def onError(e: Throwable): Unit = reportError("Error in job scheduler", e)
}
eventLoop.start()

for {
    // 获取InputDStream
    inputDStream<- ssc.graph.getInputStreams
    rateController<- inputDStream.rateController
} ssc.addStreamingListener(rateController)

listenerBus.start()
// 构建ReceiverTracker和InputInfoTracker
receiverTracker= new ReceiverTracker(ssc)
inputInfoTracker= new InputInfoTracker(ssc)

val executorAllocClient: ExecutorAllocationClient= ssc.sparkContext.schedulerBackend match {
    case b: ExecutorAllocationClient => b.asInstanceOf[ExecutorAllocationClient]
    case _ => null
}

executorAllocationManager= ExecutorAllocationManager.createIfEnabled(
    executorAllocClient,
    receiverTracker,
    ssc.conf,
    ssc.graph.batchDuration.milliseconds,
    clock)
executorAllocationManager.foreach(ssc.addStreamingListener)
// 启动ReceiverTracker
receiverTracker.start()
// 启动JobGenerator
jobGenerator.start()
executorAllocationManager.foreach(_.start())
logInfo("Started JobScheduler")
}

1.4 启动JobGenerator

启动JobGenerator需要判断是否第一次运行，如果不是第一次运行需要进行上次检查点的恢复，如果是第一次运行则调用startFirstTime方法，在该方法中初始化了定时器的开启时间，并启动了DStreamGraph和定时器timer

private def startFirstTime() {
  val startTime = new Time(timer.getStartTime())
  graph.start(startTime - graph.batchDuration)
  timer.start(startTime.milliseconds)
  logInfo("Started JobGenerator at " + startTime)
}

timer的getStartTime方法会计算出来下一个周期到期时间，计算公式：当前时间 / 间隔时间

二接收机存储流数据

2.1 启动ReceiverTracker

启动ReceiverTracker的时候，如果输入数据流不为空，则调用launchReceivers方法，然后他就会向ReceiverTrackerEndpoint发送StartAllReceivers方法，启动所有Receivers

private def launchReceivers(): Unit = {
  val receivers = receiverInputStreams.map { nis =>
    val rcvr = nis.getReceiver()
    rcvr.setReceiverId(nis.id)
    rcvr
  }
  runDummySparkJob()
  // 发送启动所有receiver的消息
  endpoint.send(StartAllReceivers(receivers))
}

case StartAllReceivers(receivers) =>
  // 根据receiver分发策略,获取与之对应的receiver和executor调度信息
  val scheduledLocations = schedulingPolicy.scheduleReceivers(receivers, getExecutors)
  // 遍历receivers，为根据receiver获取候选的executor,更新被调度receiver的位置信息，即executor信息
  // 开启receiver
  for (receiver <- receivers) {
    val executors = scheduledLocations(receiver.streamId)
    updateReceiverScheduledExecutors(receiver.streamId, executors)
    // 保存流数据接收器Receiver首选位置
    receiverPreferredLocations(receiver.streamId) = receiver.preferredLocation
    // 启动每一个Receiver
    startReceiver(receiver, executors)
  }

最后创建ReceiverSupervisor，并启动，在启动的时候，由它启动Receiver

2.2Receiver启动并接收数据

Receiver启动会调用各个具体子类的onstart方法，这里面就会接收数据，以kafka为例，则会根据提供配置创建连接,获取消息流，构造一个线程池，为每一个topic分区分配一个线程处理数据

def onStart() {
  // 获取kafka连接参数
  val props = new Properties()
  kafkaParams.foreach(param => props.put(param._1, param._2))

  val zkConnect = kafkaParams("zookeeper.connect")
  // Create the connection to the cluster
  logInfo("Connecting to Zookeeper: " + zkConnect)
  // 构造消费者配置文件
  val consumerConfig = new ConsumerConfig(props)
  // 根据消费者配置文件创建消费者连接
  consumerConnector = Consumer.create(consumerConfig)
  logInfo("Connected to " + zkConnect)
  // 构造keyDecoder和valueDecoder
  val keyDecoder = classTag[U].runtimeClass.getConstructor(classOf[VerifiableProperties])
    .newInstance(consumerConfig.props)
    .asInstanceOf[Decoder[K]]
  val valueDecoder = classTag[T].runtimeClass.getConstructor(classOf[VerifiableProperties])
    .newInstance(consumerConfig.props)
    .asInstanceOf[Decoder[V]]

  // Create threads for each topic/message Stream we are listening
  // 创建消息流
  val topicMessageStreams = consumerConnector.createMessageStreams(
    topics, keyDecoder, valueDecoder)
  // 构造线程池
  val executorPool =
    ThreadUtils.newDaemonFixedThreadPool(topics.values.sum, "KafkaMessageHandler")
  try {
    // 开始处理每一个分区的数据
    topicMessageStreams.values.foreach { streams =>
      streams.foreach { stream => executorPool.submit(new MessageHandler(stream)) }
    }
  } finally {
    executorPool.shutdown() // Just causes threads to terminate after work is done
  }
}

2.3 启动BlockGenerator生成block

在ReceiverSupervisorImpl的onstart方法中调用BlockGenerator的start启动BlockGenerator

override protected def onStart() {
  registeredBlockGenerators.asScala.foreach { _.start() }
}

启动时候会先更新自身状态为Active，然后启动2个线程：

blockIntervalTimer：定义开始一个新batch，然后准备把之前的batch作为一个block

blockPushingThread：把数据块 push到block manager

def start(): Unit = synchronized {
  if (state == Initialized) {
    // 更改状态
    state = Active
    // 开启一个定时器，定期的把缓存中的数据封装成数据块
    blockIntervalTimer.start()
    // 开始一个线程，不断将封装好的数据封装成数据块
    blockPushingThread.start()
    logInfo("Started BlockGenerator")
  } else {
    throw new SparkException(
      s"Cannot start BlockGenerator as its not in the Initialized state [state = $state]")
  }
}

private def updateCurrentBuffer(time: Long): Unit = {
  try {
    var newBlock: Block = null
    synchronized {
      // 判断当前放数据的buffer是否为空，如果不为空
      if (currentBuffer.nonEmpty) {
        // 则赋给一个新的block buffer，然后再把置为currentBuffer空
        val newBlockBuffer = currentBuffer
        currentBuffer = new ArrayBuffer[Any]
        // 构建一个blockId
        val blockId = StreamBlockId(receiverId, time - blockIntervalMs)
        listener.onGenerateBlock(blockId)
        // 构建block
        newBlock = new Block(blockId, newBlockBuffer)
      }
    }
    // 新的block不为空，则放入push队列，如果该队列满了则由其他线程push到block manager
    if (newBlock != null) {
      blocksForPushing.put(newBlock)  // put is blocking when queue is full
    }
  } catch {
    case ie: InterruptedException =>
      logInfo("Block updating timer thread was interrupted")
    case e: Exception =>
      reportError("Error in block updating thread", e)
  }
}

2.4 数据存储

Receiver会进行数据的存储，如果数据量很少，则攒多条数据成数据块在进行块存储；如果数据量很大，则直接进行存储，对于需要攒多条数据成数据块的操作在Receiver.store方法里面调用ReceiverSupervisor的pushSingle方法处理。在pushSingle中把数据先保存在内存中，这些内存数据被BlockGenerator的定时器线程blockIntervalTimer加入队列并调用ReceiverSupervisor的pushArrayBuffer方法进行处理。

他们其实都是调用的是pushAndReportBlock，该方法会调用ReceiveBlockHandler的storeBlock方法保存数据并根据配置进行预写日志;另外存储数据块并向driver报告：

def pushAndReportBlock(
    receivedBlock: ReceivedBlock,
    metadataOption: Option[Any],
    blockIdOption: Option[StreamBlockId]
  ) {
  // 获取一个blockId
  val blockId = blockIdOption.getOrElse(nextBlockId)
  val time = System.currentTimeMillis
  // 存储block
  val blockStoreResult = receivedBlockHandler.storeBlock(blockId, receivedBlock)
  logDebug(s"Pushed block $blockId in ${(System.currentTimeMillis - time)} ms")
  // 结果数量
  val numRecords = blockStoreResult.numRecords
  // 构建ReceivedBlockInfo
  val blockInfo = ReceivedBlockInfo(streamId, numRecords, metadataOption, blockStoreResult)
  // 向ReceiverTrackerEndpoint发送AddBlock消息
  trackerEndpoint.askWithRetry[Boolean](AddBlock(blockInfo))
  logDebug(s"Reported block $blockId")
}

三数据处理

我们知道DStream在进行action操作时，会触发job。我们以saveAsTextFiles方法为例：

def saveAsTextFiles(prefix: String, suffix: String = ""): Unit = ssc.withScope {
// 封装了一个保存函数，内部其实调用的RDD的saveAsTextFile
val saveFunc = (rdd: RDD[T], time: Time) => {
val file = rddToFileName(prefix, suffix, time)
rdd.saveAsTextFile(file)
}
// 调用foreachRDD方法遍历RDD
this.foreachRDD(saveFunc, displayInnerRDDOps = false)
}

foreachRDD：它会向DStreamGraph注册，根据返回的当前的DStream然后创建ForEachDStream

private def foreachRDD(
    foreachFunc: (RDD[T], Time) => Unit,
    displayInnerRDDOps: Boolean): Unit = {
// 它会向DStreamGraph注册，根据返回的当前的DStream然后创建ForEachDStream
new ForEachDStream(this,
    context.sparkContext.clean(foreachFunc, false), displayInnerRDDOps).register()
}

private[streaming] def register(): DStream[T] = {
// 向DStreamGraph添加输出流
ssc.graph.addOutputStream(this)
this
}

JobGenerator初始化的时候会构造一个timer定时器：

private val timer = new RecurringTimer(clock, ssc.graph.batchDuration.milliseconds,
longTime => eventLoop.post(GenerateJobs(new Time(longTime))), "JobGenerator")

它会启动一个后台线程，不断去调用triggerActionForNextInterval方法，该方法就会不断调用processsEvent方法，并且传递GenerateJobs事件

private def processEvent(event: JobGeneratorEvent) {
logDebug("Got event " + event)
event match {
    case GenerateJobs(time) => generateJobs(time)
    case ClearMetadata(time) => clearMetadata(time)
    case DoCheckpoint(time, clearCheckpointDataLater) =>
      doCheckpoint(time, clearCheckpointDataLater)
    case ClearCheckpointData(time) => clearCheckpointData(time)
}
}

JobGenerator# generateJobs

调用DStreamGraph的generateJobs方法产生job,然后利用JobScheduler开始提交job集合

private def generateJobs(time: Time) {
// checkpoint所有那些标记为checkpointing状态的RDDs以确保他们的血缘
// 关系会定期删除，否则血缘关系太长会造成栈溢出
ssc.sparkContext.setLocalProperty(RDD.CHECKPOINT_ALL_MARKED_ANCESTORS, "true")
Try {
    // 根据时间分配block到batch,一个batch可能你有多个block
    jobScheduler.receiverTracker.allocateBlocksToBatch(time) // allocate received blocks to batch
    // DStreamDgraph根据时间产生job集合，使用分配的数据块
    graph.generateJobs(time) // generate jobs using allocated block
} match {
    case Success(jobs) =>
      // 如果成功，则提交jobset
      val streamIdToInputInfos = jobScheduler.inputInfoTracker.getInfo(time)
      jobScheduler.submitJobSet(JobSet(time, jobs, streamIdToInputInfos))
    case Failure(e) =>
      jobScheduler.reportError("Error generating jobs fortime " + time, e)
      PythonDStream.stopStreamingContextIfPythonProcessIsDead(e)
}
// 进行checkpoint
eventLoop.post(DoCheckpoint(time, clearCheckpointDataLater = false))
}

# DStreamGraph的generateJobs根据时间产生job集

def generateJobs(time: Time): Seq[Job] = {
logDebug("Generating jobs for time " + time)
// 根据DStreamGraph的输出流创建job集合
val jobs = this.synchronized {
    outputStreams.flatMap { outputStream =>
      // 调用DStream的generateJob方法产生job
      val jobOption = outputStream.generateJob(time)
      jobOption.foreach(_.setCallSite(outputStream.creationSite))
      jobOption
    }
}
logDebug("Generated " + jobs.length + " jobs for time " + time)
jobs
}

# 然后调用DStream的generateJobs产生job

private[streaming] def generateJob(time: Time): Option[Job] = {
  getOrCompute(time) match {
    case Some(rdd) =>
      val jobFunc = () => {
        val emptyFunc = { (iterator: Iterator[T]) => {} }
        context.sparkContext.runJob(rdd, emptyFunc)
      }
      Some(new Job(time, jobFunc))
    case None => None
  }
}

# 最后提交job集合

提交job集合，遍历每一个job，创建JobHandler，然后JobHandler是一个线程类，在其run方法中会向JobScheduler发送JobStarted事件，从而开始处理job

private class JobHandler(job: Job) extends Runnable with Logging {
  import JobScheduler._

  def run() {
    val oldProps = ssc.sparkContext.getLocalProperties
    try {
      ssc.sparkContext.setLocalProperties(SerializationUtils.clone(ssc.savedProperties.get()))
      val formattedTime = UIUtils.formatBatchTime(
        job.time.milliseconds, ssc.graph.batchDuration.milliseconds, showYYYYMMSS = false)
      val batchUrl = s"/streaming/batch/?id=${job.time.milliseconds}"
      val batchLinkText = s"[output operation ${job.outputOpId}, batch time ${formattedTime}]"

      ssc.sc.setJobDescription(
        s"""Streaming job from <a href="$batchUrl">$batchLinkText</a>""")
      ssc.sc.setLocalProperty(BATCH_TIME_PROPERTY_KEY, job.time.milliseconds.toString)
      ssc.sc.setLocalProperty(OUTPUT_OP_ID_PROPERTY_KEY, job.outputOpId.toString)
      ssc.sparkContext.setLocalProperty(RDD.CHECKPOINT_ALL_MARKED_ANCESTORS, "true")
      var _eventLoop = eventLoop
      if (_eventLoop != null) {
        _eventLoop.post(JobStarted(job, clock.getTimeMillis()))
        PairRDDFunctions.disableOutputSpecValidation.withValue(true) {
          job.run() //真正开始处理job
        }
        _eventLoop = eventLoop
        if (_eventLoop != null) {
          _eventLoop.post(JobCompleted(job, clock.getTimeMillis()))
        }
      } else {
      }
    } finally {
      ssc.sparkContext.setLocalProperties(oldProps)
    }
  }
}

private def handleJobStart(job: Job, startTime: Long) {
  // 根据时间获取jobSet
  val jobSet = jobSets.get(job.time)
  // 判断是否已经开始运行
  val isFirstJobOfJobSet = !jobSet.hasStarted
  // 更新jobset开始时间
  jobSet.handleJobStart(job)
  if (isFirstJobOfJobSet) {
    listenerBus.post(StreamingListenerBatchStarted(jobSet.toBatchInfo))
  }
  job.setStartTime(startTime)
  listenerBus.post(StreamingListenerOutputOperationStarted(job.toOutputOperationInfo))
  logInfo("Starting job " + job.id + " from job set of time " + jobSet.time)
}