[SPARK-5259] Do not submit stage until its dependencies map outputs are registered

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.1.1, 1.2.0
• Fix Version/s: 1.6.0
• Component/s: Spark Core
• Target Version/s: 1.6.0
• Assignee: SuYan
• Reporter: SuYan
• Created: 15/Jan/15 04:06
• Updated: 09/Dec/16 12:06
• Resolved: 21/Sep/15 19:26

Description

We should track pending tasks by partition ID instead of Task objects.

Before this, failure & retry could result in a case where a stage got submitted before the map output from its dependencies get registered. This was due to an error in the condition for registering map outputs.

More complete explanation of the original problem:

1. while shuffle stage was retry, there may have 2 taskSet running.

we call the 2 taskSet:taskSet0.0, taskSet0.1, and we know, taskSet0.1 will re-run taskSet0.0's un-complete task

if taskSet0.0 was run all the task that the taskSet0.1 not complete yet but covered the partitions.

then stage is Available is true.

  def isAvailable: Boolean = {
    if (!isShuffleMap) {
      true
    } else {
      numAvailableOutputs == numPartitions
    }
  } 

but stage.pending task is not empty, to protect register mapStatus in mapOutputTracker.

because if task is complete success, pendingTasks is minus Task in reference-level because the task is not override hashcode() and equals()
pendingTask -= task

but numAvailableOutputs is according to partitionID.

here is the testcase to prove:

  test("Make sure mapStage.pendingtasks is set() " +
    "while MapStage.isAvailable is true while stage was retry ") {
    val firstRDD = new MyRDD(sc, 6, Nil)
    val firstShuffleDep = new ShuffleDependency(firstRDD, null)
    val firstShuyffleId = firstShuffleDep.shuffleId
    val shuffleMapRdd = new MyRDD(sc, 6, List(firstShuffleDep))
    val shuffleDep = new ShuffleDependency(shuffleMapRdd, null)
    val shuffleId = shuffleDep.shuffleId
    val reduceRdd = new MyRDD(sc, 2, List(shuffleDep))
    submit(reduceRdd, Array(0, 1))
    complete(taskSets(0), Seq(
      (Success, makeMapStatus("hostB", 1)),
      (Success, makeMapStatus("hostB", 2)),
      (Success, makeMapStatus("hostC", 3)),
      (Success, makeMapStatus("hostB", 4)),
      (Success, makeMapStatus("hostB", 5)),
      (Success, makeMapStatus("hostC", 6))
    ))
    complete(taskSets(1), Seq(
      (Success, makeMapStatus("hostA", 1)),
      (Success, makeMapStatus("hostB", 2)),
      (Success, makeMapStatus("hostA", 1)),
      (Success, makeMapStatus("hostB", 2)),
      (Success, makeMapStatus("hostA", 1))
    ))
    runEvent(ExecutorLost("exec-hostA"))
    runEvent(CompletionEvent(taskSets(1).tasks(0), Resubmitted, null, null, null, null))
    runEvent(CompletionEvent(taskSets(1).tasks(2), Resubmitted, null, null, null, null))
    runEvent(CompletionEvent(taskSets(1).tasks(0),
      FetchFailed(null, firstShuyffleId, -1, 0, "Fetch Mata data failed"),
      null, null, null, null))
    scheduler.resubmitFailedStages()
    runEvent(CompletionEvent(taskSets(1).tasks(0), Success,
      makeMapStatus("hostC", 1), null, null, null))
    runEvent(CompletionEvent(taskSets(1).tasks(2), Success,
      makeMapStatus("hostC", 1), null, null, null))
    runEvent(CompletionEvent(taskSets(1).tasks(4), Success,
      makeMapStatus("hostC", 1), null, null, null))
    runEvent(CompletionEvent(taskSets(1).tasks(5), Success,
      makeMapStatus("hostB", 2), null, null, null))
    val stage = scheduler.stageIdToStage(taskSets(1).stageId)
    assert(stage.attemptId == 2)
    assert(stage.isAvailable)
    assert(stage.pendingTasks.size == 0)
  }

Issue Links

• blocks: SPARK-10370 After a stages map outputs are registered, all running attempts should be marked as zombies
• links to: [Github] Pull Request #4055 (suyanNone)
• links to: [Github] Pull Request #7699 (squito)