[SPARK-27082] Dynamic Allocation: we should consider the scenario that speculative task being killed and never resubmit

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Incomplete
• Affects Version/s: 2.4.0
• Component/s: Spark Core
• Labels: bulk-closed patch
• Flags: Patch
• Assignee: Unassigned
• Reporter: Zhen Fan
• Created: 07/Mar/19 10:52
• Updated: 25/May/21 01:54
• Resolved: 25/May/21 01:44

Description

Issue background:

When we enable dynamic allocation, we expect that the executors can be removed appropriately, especially in some stages with data skew. With speculation enabled, the copying task  can be killed by the original task and vice versa. In TaskSetManager, we set successful(index)=true, and never resubmit the killed tasks. However, in ExecutorAllocationManager which is very related to the dynamic allocation function, doesn't handle this scenario.

See SPARK-8366. However, (SPARK-8366) ignores one scenario that copying task is being killed. When this happens, the TaskSetManager will mark the task index of the stage as success and never resubmit the killed task, so here we shouldn't treat it as pending task.

This can do harm to the computing of  maxNumExecutorsNeeded, as a result, we always retain unnecessary  executors and waste the computing resources of clusters.

Solution:

When the task index is marked as speculative and the mirror task is successful, we won't treat it as pending task. 

Code has been tested.

private val stageIdToSpeculativeTaskIndices = new mutable.HashMap[Int, mutable.HashMap[Int, Boolean]]
... 
val speculativeTaskIndices = stageIdToSpeculativeTaskIndices.get(stageId)
if (taskEnd.reason == Success) {
  if (speculativeTaskIndices.isDefined && speculativeTaskIndices.get.contains(taskIndex)) {
    speculativeTaskIndices.get(taskIndex) = true
  }
} else {
  var resubmitTask = true
  if (taskEnd.taskInfo.killed) {
    resubmitTask = !(speculativeTaskIndices.isDefined &&
        speculativeTaskIndices.get.getOrElse(taskIndex, false))
  }

  if (resubmitTask) {
    if (totalPendingTasks() == 0) {
      allocationManager.onSchedulerBacklogged()
    }
    if (taskEnd.taskInfo.speculative) {
      stageIdToSpeculativeTaskIndices.get(stageId).foreach {_.remove(taskIndex)}
    } else {
      stageIdToTaskIndices.get(stageId).foreach {_.remove(taskIndex)}
    }
  }
}

 Please take a look, Thanks.

Issue Links

• is related to: SPARK-30511 Spark marks intentionally killed speculative tasks as pending leads to holding idle executors