Spark学习之7:Job触发及Stage划分

1. Job提交触发

流程图:
Spark学习之7:Job触发及Stage划分_第1张图片

作业提交流程由RDD的action操作触发,继而调用SparkContext.runJob。
在RDD的action操作后可能会调用多个SparkContext.runJob的重载函数,但最终会调用的runJob见1.1。

1.1. SparkContext.runJob

  def runJob[T, U: ClassTag](
      rdd: RDD[T],
      func: (TaskContext, Iterator[T]) => U,
      partitions: Seq[Int],
      allowLocal: Boolean,
      resultHandler: (Int, U) => Unit) {
    if (stopped) {
      throw new IllegalStateException("SparkContext has been shutdown")
    }
    val callSite = getCallSite
    val cleanedFunc = clean(func)
    logInfo("Starting job: " + callSite.shortForm)
    if (conf.getBoolean("spark.logLineage", false)) {
      logInfo("RDD's recursive dependencies:\n" + rdd.toDebugString)
    }
    dagScheduler.runJob(rdd, cleanedFunc, partitions, callSite, allowLocal,
      resultHandler, localProperties.get)
    progressBar.foreach(_.finishAll())
    rdd.doCheckpoint()
  }
参数说明:
(1)func,将在每个分区上执行的函数;
(2)partitions,分区索引号,从0开始;
(3)resultHandler,结果聚合函数;
在job执行完成后,将调用RDD.doCheckPoint检查是否需要做checkpoint。

1.2. DAGScheduler.runJob

  def runJob[T, U: ClassTag](
      rdd: RDD[T],
      func: (TaskContext, Iterator[T]) => U,
      partitions: Seq[Int],
      callSite: CallSite,
      allowLocal: Boolean,
      resultHandler: (Int, U) => Unit,
      properties: Properties = null)
  {
    val start = System.nanoTime
    val waiter = submitJob(rdd, func, partitions, callSite, allowLocal, resultHandler, properties)
    waiter.awaitResult() match {
      case JobSucceeded => {
        logInfo("Job %d finished: %s, took %f s".format
          (waiter.jobId, callSite.shortForm, (System.nanoTime - start) / 1e9))
      }
      case JobFailed(exception: Exception) =>
        logInfo("Job %d failed: %s, took %f s".format
          (waiter.jobId, callSite.shortForm, (System.nanoTime - start) / 1e9))
        throw exception
    }
  }
(1)调用submitJob函数,返回JobWaiter对象;
(2)由JobWaiter来等待Job的完成或失败。

1.3. DAGScheduler.submitJob

    // Check to make sure we are not launching a task on a partition that does not exist.
    val maxPartitions = rdd.partitions.length
    partitions.find(p => p >= maxPartitions || p < 0).foreach { p =>
      throw new IllegalArgumentException(
        "Attempting to access a non-existent partition: " + p + ". " +
          "Total number of partitions: " + maxPartitions)
    }
    val jobId = nextJobId.getAndIncrement()
    if (partitions.size == 0) {
      return new JobWaiter[U](this, jobId, 0, resultHandler)
    }
    assert(partitions.size > 0)
    val func2 = func.asInstanceOf[(TaskContext, Iterator[_]) => _]
    val waiter = new JobWaiter(this, jobId, partitions.size, resultHandler)
    eventProcessLoop.post(JobSubmitted(
      jobId, rdd, func2, partitions.toArray, allowLocal, callSite, waiter, properties))
    waiter
(1)创建jobId;
(2)创建JobWaiter对象;
(3)将jobId,JobWriter等对象封入JobSubmitted消息并塞入DAGSchedulerEventProcessLoop的消息队列,DAGSchedulerEventProcessLoop对象包含一个消息队列及读取消息的线程。线程从消息队列中读取消息,根据消息的类型调用DAGScheduler的不同方法(具体见DAGSchedulerEventProcessLoop.onReceive)。 JobSubmitted消息将触发DAGScheduler.handleJobSubmitted方法的调用。

2. RDDs的Stage划分

2.1. DAGScheduler.handleJobSubmitted

    var finalStage: Stage = null
    try {
      // New stage creation may throw an exception if, for example, jobs are run on a
      // HadoopRDD whose underlying HDFS files have been deleted.
      finalStage = newStage(finalRDD, partitions.size, None, jobId, callSite)
    } catch {
      ......
    }
    ......
调用newStage方法创建RDD对应的finalStage,该RDD是调用action操作的RDD。

2.2. Stage划分流程

Spark学习之7:Job触发及Stage划分_第2张图片

2.2.1. Stage

private[spark] class Stage(
    val id: Int,
    val rdd: RDD[_],
    val numTasks: Int,
    val shuffleDep: Option[ShuffleDependency[_, _, _]],  // Output shuffle if stage is a map stage
    val parents: List[Stage],
    val jobId: Int,
    val callSite: CallSite)
Stage有两种类型:
(1)shuffle map stage;
(2)result stage;
类结构说明:
(1)rdd,表示一个Stage中的最后一个RDD;
(2)numTasks,就是分区的数量;
(3)shuffleDep,如注释所说,如果该stage是shuffle map stage,则该字段表示输出shuffle,即有个RDD shuffle依赖该Stage的结果;
(4)parents,该Stage的父Stages。

2.2.2. DAGScheduler.newStage

  private def newStage(
      rdd: RDD[_],
      numTasks: Int,
      shuffleDep: Option[ShuffleDependency[_, _, _]],
      jobId: Int,
      callSite: CallSite)
    : Stage =
  {
    val parentStages = getParentStages(rdd, jobId)
    val id = nextStageId.getAndIncrement()
    val stage = new Stage(id, rdd, numTasks, shuffleDep, parentStages, jobId, callSite)
    stageIdToStage(id) = stage
    updateJobIdStageIdMaps(jobId, stage)
    stage
  }
创建一个新的Stage。
从Stage的定义可以看出,创建Stage需要知道其父Stage信息。所以:
(1)先获取RDD(所在Stage)的父Stage;
(2)创建Stage id;
(3)创建Stage
从流程图知道,newStage可能会是一个递归的过程,在获取父Stage时,也需要获取其祖父Stage。

2.2.3. 举例

Spark学习之7:Job触发及Stage划分_第3张图片

有这样一个RDD图,其中红色箭头表示RDD之间的Shuffle依赖(宽依赖),其他颜色箭头表示窄依赖。
在RDD9上执行了action操作。
我们要创建RDD9的stage。
Stage以ShuffleDependency为界进行划分。

2.2.3.1 DAGScheduler.getParentStages

获取RDD的parent stage。
  private def getParentStages(rdd: RDD[_], jobId: Int): List[Stage] = {
    val parents = new HashSet[Stage]
    val visited = new HashSet[RDD[_]]
    // We are manually maintaining a stack here to prevent StackOverflowError
    // caused by recursively visiting
    val waitingForVisit = new Stack[RDD[_]]
    def visit(r: RDD[_]) {
      if (!visited(r)) {
        visited += r
        // Kind of ugly: need to register RDDs with the cache here since
        // we can't do it in its constructor because # of partitions is unknown
        for (dep <- r.dependencies) {
          dep match {
            case shufDep: ShuffleDependency[_, _, _] =>
              parents += getShuffleMapStage(shufDep, jobId)
            case _ =>
              waitingForVisit.push(dep.rdd)
          }
        }
      }
    }
    waitingForVisit.push(rdd)
    while (!waitingForVisit.isEmpty) {
      visit(waitingForVisit.pop())
    }
    parents.toList
  }
(1)将RDD9放入waitingForVisit栈中,并开始遍历该栈;
(2)从栈中取出一个RDD(即RDD9),它有两个依赖,都是窄依赖,所以将RDD4、RDD7压栈;
(3)取出RDD7,它有两个宽依赖,所以要获取宽依赖对应的shuffle map stage;
(4)取出RDD4,它只有一个窄依赖,所以将RDD3压栈;
(5)取出RDD3,它有一个宽依赖,有一个窄依赖,将RDD5压栈,计算RDD3对应宽依赖的shuffle map stage;
一个RDD的parent Stage要么为None,要么是一个shuffle map stage。

注:
在RDD间产生ShuffleDependency依赖的transform操作时,在创建ShuffledRDD过程中将deps初始化为Nil,并没有实际创建ShuffleDependency对象,但窄依赖是在transform操作时就创建好的。
RDD间的ShuffleDependency对象是通过调用RDD.dependencies创建的(如在该方法中调用r.dependencies)。
由于RDD的遍历是从大编号到小编号,因此先遍历的RDD(编号大)对应ShuffleDependency拥有较小的ShuffleId。
另,祖先Stage拥有较小的StageId。

2.2.3.2. DAGScheduler.getShuffleMapStage

  private def getShuffleMapStage(shuffleDep: ShuffleDependency[_, _, _], jobId: Int): Stage = {
    shuffleToMapStage.get(shuffleDep.shuffleId) match {
      case Some(stage) => stage
      case None =>
        // We are going to register ancestor shuffle dependencies
        registerShuffleDependencies(shuffleDep, jobId)
        // Then register current shuffleDep
        val stage =
          newOrUsedStage(
            shuffleDep.rdd, shuffleDep.rdd.partitions.size, shuffleDep, jobId,
            shuffleDep.rdd.creationSite)
        shuffleToMapStage(shuffleDep.shuffleId) = stage
 
        stage
    }
  }
为了划分好Stage的复用,减少Stage划分支出,会将每个shuffle map stage保存起来。
(1)检查shuffle dependency对应的stage是否已经存在;
(2)若存在,直接返回对应的stage;
(3)若不存在,则先注册该 shuffle dependency所有祖先 shuffle dependency对应的stage,然后再创建当前 shuffle dependency对应的stage。

2.2.3.3. DAGScheduler.registerShuffleDependencies

  private def registerShuffleDependencies(shuffleDep: ShuffleDependency[_, _, _], jobId: Int) = {
    val parentsWithNoMapStage = getAncestorShuffleDependencies(shuffleDep.rdd)
    while (!parentsWithNoMapStage.isEmpty) {
      val currentShufDep = parentsWithNoMapStage.pop()
      val stage =
        newOrUsedStage(
          currentShufDep.rdd, currentShufDep.rdd.partitions.size, currentShufDep, jobId,
          currentShufDep.rdd.creationSite)
      shuffleToMapStage(currentShufDep.shuffleId) = stage
    }
  }
(1)获取ShuffleDependency依赖RDD的所有祖先RDD中包含的 ShuffleDependency;
(2)getAncestorShuffleDependencies返回的是一个栈,由于RDD的遍历是从大编号往小编号方向,所以小编号的ShuffleDependency放在栈底,大编号的ShuffleDependency放在栈顶(如果在划分Stage之前就调用过RDD.dependencies,ShuffleDependency的编号就可能是乱序的)。
(3)从栈顶取出的ShuffleDependency的ShuffleId大。然后调用newOrUsedStage方法创建对应的Stage。从流程图中看出,在newOrUsedStage方法中将调用newStage方法。如果RDD位于RDD图的叶子端,对应的parent stage不存在,所以可以直接创建对应的Stage。
(4)保存新创建的Stage。
因此,祖先Stage拥有更小的StageId。
当祖先Stage都创建完成后,并且每个shuffle map stage都保存在以ShuffleId为key的HasMap中,我们可以从该结构获取父Stage。

你可能感兴趣的:(Spark)