提交stage-创新互联

//提交stage，为stage创建一批task，task数量和partition数量相同

 private def submitMissingTasks(stage: Stage, jobId: Int) {

  logDebug("submitMissingTasks(" + stage + ")")

  // Get our pending tasks and remember them in our pendingTasks entry

  stage.pendingTasks.clear()

  // First figure out the indexes of partition ids to compute.

//获取要创建的task的数量

  val partitionsToCompute: Seq[Int] = {

   if (stage.isShuffleMap) {

    (0 until stage.numPartitions).filter(id => stage.outputLocs(id) == Nil)

   } else {

    val job = stage.resultOfJob.get

    (0 until job.numPartitions).filter(id => !job.finished(id))

   }

  }

  val properties = if (jobIdToActiveJob.contains(jobId)) {

   jobIdToActiveJob(stage.jobId).properties

  } else {

   // this stage will be assigned to "default" pool

   null

  }

//将stage加入runningstage队列

  runningStages += stage

  // SparkListenerStageSubmitted should be posted before testing whether tasks are

  // serializable. If tasks are not serializable, a SparkListenerStageCompleted event

  // will be posted, which should always come after a corresponding SparkListenerStageSubmitted

  // event.

  stage.latestInfo = StageInfo.fromStage(stage, Some(partitionsToCompute.size))

  outputCommitCoordinator.stageStart(stage.id)

  listenerBus.post(SparkListenerStageSubmitted(stage.latestInfo, properties))

  // TODO: Maybe we can keep the taskBinary in Stage to avoid serializing it multiple times.

  // Broadcasted binary for the task, used to dispatch tasks to executors. Note that we broadcast

  // the serialized copy of the RDD and for each task we will deserialize it, which means each

  // task gets a different copy of the RDD. This provides stronger isolation between tasks that

  // might modify state of objects referenced in their closures. This is necessary in Hadoop

  // where the JobConf/Configuration object is not thread-safe.

  var taskBinary: Broadcast[Array[Byte]] = null

  try {

   // For ShuffleMapTask, serialize and broadcast (rdd, shuffleDep).

   // For ResultTask, serialize and broadcast (rdd, func).

   val taskBinaryBytes: Array[Byte] =

    if (stage.isShuffleMap) {

     closureSerializer.serialize((stage.rdd, stage.shuffleDep.get) : AnyRef).array()

    } else {

     closureSerializer.serialize((stage.rdd, stage.resultOfJob.get.func) : AnyRef).array()

    }

   taskBinary = sc.broadcast(taskBinaryBytes)

  } catch {

   // In the case of a failure during serialization, abort the stage.

   case e: NotSerializableException =>

    abortStage(stage, "Task not serializable: " + e.toString)

    runningStages -= stage

    return

   case NonFatal(e) =>

    abortStage(stage, s"Task serialization failed: $e\n${e.getStackTraceString}")

    runningStages -= stage

    return

  }

//为stage创建指定数量的task

  val tasks: Seq[Task[_]] = if (stage.isShuffleMap) {

   partitionsToCompute.map { id =>

//给每个partition创建一个task

//给每个task计算最佳位置

    val locs = getPreferredLocs(stage.rdd, id)

    val part = stage.rdd.partitions(id)

//对于finalstage之外的stage的isShuffleMap都是true

//所以会创建ShuffleMapTask

    new ShuffleMapTask(stage.id, taskBinary, part, locs)

   }

  } else {

//如果不是ShuffleMap，就会创建finalstage

//finalstage是穿件resultTask

   val job = stage.resultOfJob.get

   partitionsToCompute.map { id =>

    val p: Int = job.partitions(id)

    val part = stage.rdd.partitions(p)

//获取task计算的最佳位置的方法 getPreferredLocs

    val locs = getPreferredLocs(stage.rdd, p)

    new ResultTask(stage.id, taskBinary, part, locs, id)

   }

  }

  if (tasks.size > 0) {

   logInfo("Submitting " + tasks.size + " missing tasks from " + stage + " (" + stage.rdd + ")")

   stage.pendingTasks ++= tasks

   logDebug("New pending tasks: " + stage.pendingTasks)

   taskScheduler.submitTasks(

    new TaskSet(tasks.toArray, stage.id, stage.newAttemptId(), stage.jobId, properties))

   stage.latestInfo.submissionTime = Some(clock.getTimeMillis())

  } else {

   // Because we posted SparkListenerStageSubmitted earlier, we should post

   // SparkListenerStageCompleted here in case there are no tasks to run.

   outputCommitCoordinator.stageEnd(stage.id)

   listenerBus.post(SparkListenerStageCompleted(stage.latestInfo))

   logDebug("Stage " + stage + " is actually done; %b %d %d".format(

    stage.isAvailable, stage.numAvailableOutputs, stage.numPartitions))

   runningStages -= stage

  }

 }

 def getPreferredLocs(rdd: RDD[_], partition: Int): Seq[TaskLocation] = {

  getPreferredLocsInternal(rdd, partition, new HashSet)

 }

//task对应partition的最佳位置

//就是从stage的最后一个RDD开始，找哪个RDD是被持久化了或者checkpoint

//那么task的最佳位置就是缓存的/checkpoint 的 partition的位置

//因为这样的话，task就在那个节点上执行，不需要计算之前的RDD

 private def getPreferredLocsInternal(

   rdd: RDD[_],

   partition: Int,

   visited: HashSet[(RDD[_],Int)])

  : Seq[TaskLocation] =

 {

  // If the partition has already been visited, no need to re-visit.

  // This avoids exponential path exploration.  SPARK-695

  if (!visited.add((rdd,partition))) {

   // Nil has already been returned for previously visited partitions.

   return Nil

  }

  // If the partition is cached, return the cache locations

//寻找当前RDD是否缓存了

  val cached = getCacheLocs(rdd)(partition)

  if (!cached.isEmpty) {

   return cached

  }

  // If the RDD has some placement preferences (as is the case for input RDDs), get those

//寻找当前RDD是否checkpoint了

  val rddPrefs = rdd.preferredLocations(rdd.partitions(partition)).toList

  if (!rddPrefs.isEmpty) {

   return rddPrefs.map(TaskLocation(_))

  }

  // If the RDD has narrow dependencies, pick the first partition of the first narrow dep

  // that has any placement preferences. Ideally we would choose based on transfer sizes,

  // but this will do for now.

//递归调用，看看父RDD是否缓存或者checkpoint

  rdd.dependencies.foreach {

   case n: NarrowDependency[_] =>

    for (inPart <- n.getParents(partition)) {

     val locs = getPreferredLocsInternal(n.rdd, inPart, visited)

     if (locs != Nil) {

      return locs

     }

    }

   case _ =>

  }

//如果从第一个RDD到最后一个RDD都没有缓存或者checkpoint，那最佳位置就是Nil，也就是没有最佳位置

//那他的位置就要由taskscheduler来分配

  Nil

 }

另外有需要云服务器可以了解下创新互联scvps.cn，海内外云服务器15元起步，三天无理由+7*72小时售后在线，公司持有idc许可证，提供“云服务器、裸金属服务器、高防服务器、香港服务器、美国服务器、虚拟主机、免备案服务器”等云主机租用服务以及企业上云的综合解决方案，具有“安全稳定、简单易用、服务可用性高、性价比高”等特点与优势，专为企业上云打造定制，能够满足用户丰富、多元化的应用场景需求。