20、Task原理剖析与源码分析

一、Task原理

1、图解

二、源码分析

1、

###org.apache.spark.executor/Executor.scala

/**
    * 从TaskRunner开始，来看Task的运行的工作原理
    */
  class TaskRunner(
      execBackend: ExecutorBackend,
      val taskId: Long,
      val attemptNumber: Int,
      taskName: String,
      serializedTask: ByteBuffer)
    extends Runnable {

    @volatile private var killed = false
    @volatile var task: Task[Any] = _
    @volatile var attemptedTask: Option[Task[Any]] = None
    @volatile var startGCTime: Long = _

    def kill(interruptThread: Boolean) {
      logInfo(s"Executor is trying to kill $taskName (TID $taskId)")
      killed = true
      if (task != null) {
        task.kill(interruptThread)
      }
    }

    override def run() {
      val deserializeStartTime = System.currentTimeMillis()
      Thread.currentThread.setContextClassLoader(replClassLoader)
      val ser = env.closureSerializer.newInstance()
      logInfo(s"Running $taskName (TID $taskId)")
      execBackend.statusUpdate(taskId, TaskState.RUNNING, EMPTY_BYTE_BUFFER)
      var taskStart: Long = 0
      startGCTime = gcTime

      try {
        // 对序列化的task数据，进行反序列化
        val (taskFiles, taskJars, taskBytes) = Task.deserializeWithDependencies(serializedTask)
        // 然后，通过网络通信，将需要的文件、资源、jar拷贝过来
        updateDependencies(taskFiles, taskJars)
        // 最后，通过正式的反序列化操作，将整个task的数据集反序列化回来
        // 这里用到了java的ClassLoader，因为java的ClassLoader可以干很多事情，比如，用反射的方式来动态加载一个类，创建这个类的对象，
        // 还有比如，可以用于对指定上下文的相关资源，进行加载和读取
        task = ser.deserialize[Task[Any]](taskBytes, Thread.currentThread.getContextClassLoader)

        // If this task has been killed before we deserialized it, let's quit now. Otherwise,
        // continue executing the task.
        if (killed) {
          // Throw an exception rather than returning, because returning within a try{} block
          // causes a NonLocalReturnControl exception to be thrown. The NonLocalReturnControl
          // exception will be caught by the catch block, leading to an incorrect ExceptionFailure
          // for the task.
          throw new TaskKilledException
        }

        attemptedTask = Some(task)
        logDebug("Task " + taskId + "'s epoch is " + task.epoch)
        env.mapOutputTracker.updateEpoch(task.epoch)

        // Run the actual task and measure its runtime.
        // 计算出task开始的时间
        taskStart = System.currentTimeMillis()
        // 执行task，用的是task的run()方法
        // 这里的value，对于ShuffleMapTask来说，其实就是MapStatus,封装了ShuffleMapTask计算的数据，输出的位置
        // 后面还是一个ShuffleMapTask，那么就会去联系MapOutputTracker，来获取上一个ShuffleMapTasks的输出位置，然后通过网络拉取数据
        // ResultTask，也是一样的
        val value = task.run(taskAttemptId = taskId, attemptNumber = attemptNumber)
        // 计算出task结束的时间
        val taskFinish = System.currentTimeMillis()

        // If the task has been killed, let's fail it.
        if (task.killed) {
          throw new TaskKilledException
        }
        // 这个，其实就是针对MapStatus进行了各种序列化和封装，因为后面要发送给Driver(通过网络)
        //
        val resultSer = env.serializer.newInstance()
        val beforeSerialization = System.currentTimeMillis()
        val valueBytes = resultSer.serialize(value)
        val afterSerialization = System.currentTimeMillis()

        // 计算出task相关的一些metrics，就是统计信息，包括运行了多长时间、反序列化消耗了多长时间、java虚拟机gc耗费了多长时间
        // 结果的序列化耗费了多长时间，这些东西，其实会在我们的SparkUI上显示
        for (m <- task.metrics) {
          m.setExecutorDeserializeTime(taskStart - deserializeStartTime)
          m.setExecutorRunTime(taskFinish - taskStart)
          m.setJvmGCTime(gcTime - startGCTime)
          m.setResultSerializationTime(afterSerialization - beforeSerialization)
        }

        val accumUpdates = Accumulators.values

        val directResult = new DirectTaskResult(valueBytes, accumUpdates, task.metrics.orNull)
        val serializedDirectResult = ser.serialize(directResult)
        val resultSize = serializedDirectResult.limit

        // directSend = sending directly back to the driver
        val serializedResult = {
          if (maxResultSize > 0 && resultSize > maxResultSize) {
            logWarning(s"Finished $taskName (TID $taskId). Result is larger than maxResultSize " +
              s"(${Utils.bytesToString(resultSize)} > ${Utils.bytesToString(maxResultSize)}), " +
              s"dropping it.")
            ser.serialize(new IndirectTaskResult[Any](TaskResultBlockId(taskId), resultSize))
          } else if (resultSize >= akkaFrameSize - AkkaUtils.reservedSizeBytes) {
            val blockId = TaskResultBlockId(taskId)
            env.blockManager.putBytes(
              blockId, serializedDirectResult, StorageLevel.MEMORY_AND_DISK_SER)
            logInfo(
              s"Finished $taskName (TID $taskId). $resultSize bytes result sent via BlockManager)")
            ser.serialize(new IndirectTaskResult[Any](blockId, resultSize))
          } else {
            logInfo(s"Finished $taskName (TID $taskId). $resultSize bytes result sent to driver")
            serializedDirectResult
          }
        }

        // 其实就是调用了Executor所在的CoarseGrainedExecutorBackend的statusUpdate()方法
        execBackend.statusUpdate(taskId, TaskState.FINISHED, serializedResult)

      }
}

###org.apache.spark.executor/Executor.scala

private def updateDependencies(newFiles: HashMap[String, Long], newJars: HashMap[String, Long]) {
    // 获取hadoop配置文件
    lazy val hadoopConf = SparkHadoopUtil.get.newConfiguration(conf)
    // 这里，使用java的synchronized进行了多线程并发访问的同步
    // 因为task实际上是以java线程的方式，在一个CoarseGrainedExecutorBackend进程内并发运行的
    // 如果在执行业务逻辑的时候，要访问一些共享的资源，那么就可能会出现多线程并发访问安全问题
    // 所以，spark在这里选择进行了多线程并发访问的同步(synchronized)，因为在这里面访问了诸如currentFiles等等这些共享资源

    synchronized {
      // Fetch missing dependencies
      // 遍历要拉取的文件
      // 通过Utils的fetchFile()方法，通过网络通信，从远程拉取文件
      for ((name, timestamp) <- newFiles if currentFiles.getOrElse(name, -1L) < timestamp) {
        logInfo("Fetching " + name + " with timestamp " + timestamp)
        // Fetch file with useCache mode, close cache for local mode.
        Utils.fetchFile(name, new File(SparkFiles.getRootDirectory), conf,
          env.securityManager, hadoopConf, timestamp, useCache = !isLocal)
        currentFiles(name) = timestamp
      }
      // 遍历要拉取的jar
      for ((name, timestamp) <- newJars) {
        val localName = name.split("/").last
        // 判断一下时间戳，要求jar当前时间戳必须小于目标时间戳
        // 通过Utils的fetchFile()，拉取jar文件
        val currentTimeStamp = currentJars.get(name)
          .orElse(currentJars.get(localName))
          .getOrElse(-1L)
        if (currentTimeStamp < timestamp) {
          logInfo("Fetching " + name + " with timestamp " + timestamp)
          // Fetch file with useCache mode, close cache for local mode.
          Utils.fetchFile(name, new File(SparkFiles.getRootDirectory), conf,
            env.securityManager, hadoopConf, timestamp, useCache = !isLocal)
          currentJars(name) = timestamp
          // Add it to our class loader
          val url = new File(SparkFiles.getRootDirectory, localName).toURI.toURL
          if (!urlClassLoader.getURLs.contains(url)) {
            logInfo("Adding " + url + " to class loader")
            urlClassLoader.addURL(url)
          }
        }
      }
    }
  }

###org.apache.spark.scheduler/Task.scala

  final def run(taskAttemptId: Long, attemptNumber: Int): T = {
    // 创建一个TaskContext，就是task的执行上下文，里面记录了task执行的一些全局性的数据，比如task重试了几次
    // 比如task属于哪个stage，task要处理的是rdd的哪个partition等等
    context = new TaskContextImpl(stageId = stageId, partitionId = partitionId,
      taskAttemptId = taskAttemptId, attemptNumber = attemptNumber, runningLocally = false)
    TaskContextHelper.setTaskContext(context)
    context.taskMetrics.setHostname(Utils.localHostName())
    taskThread = Thread.currentThread()
    if (_killed) {
      kill(interruptThread = false)
    }
    try {
      // 调用抽象方法，runTask()
      runTask(context)
    } finally {
      context.markTaskCompleted()
      TaskContextHelper.unset()
    }
  }

###org.apache.spark.scheduler/Task.scala

  // 调用到了抽象方法，那就意味着这个类，只是一个模板类，或者抽象父类，仅仅封装了一些子类通用的数据和操作
  // 而关键的操作，全部都要依赖于子类的实现，task的子类，有ShuffleMapTask、ResultTask
  // 要运行子类的runTask()方法，才能执行我们自己定义的算子和逻辑
  def runTask(context: TaskContext): T

2、

接下来分别看下ShuffleMapTask和ResultTask的runTask()方法：

一个ShuffleMapTask会将一个RDD的元素，切分为多个bucket，基于一个在ShuffleDependency中指定的partitioner，默认就是HashPartition；

###org.apache.spark.scheduler/ShuffleMapTask.scala

/**
    * ShuffleMapTask的runTask()方法有MapStatus返回值
    */
  override def runTask(context: TaskContext): MapStatus = {
    // Deserialize the RDD using the broadcast variable.
    // 对task要处理的rdd相关的数据，做一些反序列化操作
    // 这里有一个问题，如何拿到这个要处理的RDD
    // 多个task运行在多个Executor中，都是并行运行，或者并发运行的，可能都不在一个地方，但是一个stage的task，其实要处理的rdd是一样，所以task如何拿到自己要处理的rdd数据？
    // 这里会通过broadcast variable 直接拿到
    val ser = SparkEnv.get.closureSerializer.newInstance()
    val (rdd, dep) = ser.deserialize[(RDD[_], ShuffleDependency[_, _, _])](
      ByteBuffer.wrap(taskBinary.value), Thread.currentThread.getContextClassLoader)

    metrics = Some(context.taskMetrics)
    var writer: ShuffleWriter[Any, Any] = null
    try {
      // 获取ShuffleManager
      val manager = SparkEnv.get.shuffleManager
      // 从ShuffleManager中获取ShuffleWriter
      writer = manager.getWriter[Any, Any](dep.shuffleHandle, partitionId, context)
      // 首先调用了，rdd的iterator()方法，并且传入了，当前task要处理哪个partition
      // 所以核心的逻辑，就在rdd的iterator()方法中，在这里，实现了针对rdd的某个partition，执行我们自己定义的算子，或者是函数
      // 执行完了我们自己定义的算子、或者函数，就相当于是，针对rdd的partition执行了处理，会有返回的数据
      // 返回的数据，都是通过ShuffleWriter，经过HashPartitioner进行分区之后，写入自己对应的分区bucket
      writer.write(rdd.iterator(partition, context).asInstanceOf[Iterator[_ <: Product2[Any, Any]]])
      // 最后，返回结果MapStatus，MapStatus里面封装了ShuffleMapTask计算后的数据，数据存储在哪里，其实就是BlockManager的相关信息
      // BlockManager是Spark底层的内存，数据，磁盘数据管理的组件
      return writer.stop(success = true).get
    } catch {
      case e: Exception =>
        try {
          if (writer != null) {
            writer.stop(success = false)
          }
        } catch {
          case e: Exception =>
            log.debug("Could not stop writer", e)
        }
        throw e
    }
  }

###org.apache.spark.rdd/RDD.scalal

  final def iterator(split: Partition, context: TaskContext): Iterator[T] = {
    if (storageLevel != StorageLevel.NONE) {
      // cacheManager相关东西
      SparkEnv.get.cacheManager.getOrCompute(this, split, context, storageLevel)
    } else {
      // 进行rdd partition的计算
      computeOrReadCheckpoint(split, context)
    }
  }

###org.apache.spark.rdd/MapPartitionsRDD.scala

  // 这里，就是针对rdd中的某个partition执行我们给这个rdd定义的算子和函数
  // 这里的f，可以理解为我们自己定义的算子和函数，但是是Spark内部进行了封装的，还实现了一些其他的逻辑
  // 执行到了这里，就是在针对RDD的partition，执行自定义的计算操作，并返回新的rdd的partition数据
  override def compute(split: Partition, context: TaskContext) =
    f(context, split.index, firstParent[T].iterator(split, context))

###org.apache.spark.scheduler/ResultTask.scala

  override def runTask(context: TaskContext): U = {
    // Deserialize the RDD and the func using the broadcast variables.
    // 进行了基本的反序列化
    val ser = SparkEnv.get.closureSerializer.newInstance()
    val (rdd, func) = ser.deserialize[(RDD[T], (TaskContext, Iterator[T]) => U)](
      ByteBuffer.wrap(taskBinary.value), Thread.currentThread.getContextClassLoader)
    metrics = Some(context.taskMetrics)
    // 执行通过rdd的iterator，执行我们定义的算子和函数
    func(context, rdd.iterator(partition, context))
  }

###org.apache.spark.executor/Executor.scala

// 其实就是调用了Executor所在的CoarseGrainedExecutorBackend的statusUpdate()方法

execBackend.statusUpdate(taskId, TaskState.FINISHED, serializedResult)

###org.apache.spark.executor/CoarseGrainedExecutorBackend.scalal

  override def statusUpdate(taskId: Long, state: TaskState, data: ByteBuffer) {
    // 向CoarseGrainedSchedulerBackend发送一个StatusUpdate消息
    driver ! StatusUpdate(executorId, taskId, state, data)
  }

###org.apache.spark.scheduler.cluster/CoarseGrainedSchedulerBackend.scalla

      // 处理task执行结束的事件
      case StatusUpdate(executorId, taskId, state, data) =>
        scheduler.statusUpdate(taskId, state, data.value)
        if (TaskState.isFinished(state)) {
          executorDataMap.get(executorId) match {
            case Some(executorInfo) =>
              executorInfo.freeCores += scheduler.CPUS_PER_TASK
              makeOffers(executorId)
            case None =>
              // Ignoring the update since we don't know about the executor.
              logWarning(s"Ignored task status update ($taskId state $state) " +
                "from unknown executor $sender with ID $executorId")
          }
        }

###org.apache.spark.scheduler/TaskSchedulerlmpl.scala

  def statusUpdate(tid: Long, state: TaskState, serializedData: ByteBuffer) {
    var failedExecutor: Option[String] = None
    synchronized {
      try {
        // 判断如果task是lost了，实际上，可能会经常发现task lost了，这就是因为各种各样的原因，执行失败了
        if (state == TaskState.LOST && taskIdToExecutorId.contains(tid)) {
          // We lost this entire executor, so remember that it's gone
          // 移除Executor，将它加入失败队列
          val execId = taskIdToExecutorId(tid)
          if (activeExecutorIds.contains(execId)) {
            removeExecutor(execId)
            failedExecutor = Some(execId)
          }
        }
        // 获取对应的taskSet
        taskIdToTaskSetId.get(tid) match {
          case Some(taskSetId) =>
            // 如果task结束了，从内存缓存中移除
            if (TaskState.isFinished(state)) {
              taskIdToTaskSetId.remove(tid)
              taskIdToExecutorId.remove(tid)
            }
            // 如果正常结束，也做相应的处理
            activeTaskSets.get(taskSetId).foreach { taskSet =>
              if (state == TaskState.FINISHED) {
                taskSet.removeRunningTask(tid)
                taskResultGetter.enqueueSuccessfulTask(taskSet, tid, serializedData)
              } else if (Set(TaskState.FAILED, TaskState.KILLED, TaskState.LOST).contains(state)) {
                taskSet.removeRunningTask(tid)
                taskResultGetter.enqueueFailedTask(taskSet, tid, state, serializedData)
              }
            }
          case None =>
            logError(
              ("Ignoring update with state %s for TID %s because its task set is gone (this is " +
               "likely the result of receiving duplicate task finished status updates)")
              .format(state, tid))
        }
      } catch {
        case e: Exception => logError("Exception in statusUpdate", e)
      }
    }
    // Update the DAGScheduler without holding a lock on this, since that can deadlock
    if (failedExecutor.isDefined) {
      dagScheduler.executorLost(failedExecutor.get)
      backend.reviveOffers()
    }
  }