[SPARK-1396] Properly cleanup DAGScheduler on job cancellation.

Previously, when jobs were cancelled, not all of the state in the
DAGScheduler was cleaned up, leading to a slow memory leak in the
DAGScheduler.  As we expose easier ways to cancel jobs, it's more
important to fix these issues.

This commit also fixes a second and less serious problem, which is that
previously, when a stage failed, not all of the appropriate stages
were cancelled.  See the "failure of stage used by two jobs" test
for an example of this.  This just meant that extra work was done, and is
not a correctness problem.

This commit adds 3 tests.  “run shuffle with map stage failure” is
a new test to more thoroughly test this functionality, and passes on
both the old and new versions of the code.  “trivial job
cancellation” fails on the old code because all state wasn’t cleaned
up correctly when jobs were cancelled (we didn’t remove the job from
resultStageToJob).  “failure of stage used by two jobs” fails on the
old code because taskScheduler.cancelTasks wasn’t called for one of
the stages (see test comments).

This should be checked in before apache#246, which makes it easier to
cancel stages / jobs.

Author: Kay Ousterhout <[email protected]>

Closes apache#305 from kayousterhout/incremental_abort_fix and squashes the following commits:

f33d844 [Kay Ousterhout] Mark review comments
9217080 [Kay Ousterhout] Properly cleanup DAGScheduler on job cancellation.

Author: kayousterhout

core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala

@@ -982,15 +982,7 @@ class DAGScheduler(
     if (!jobIdToStageIds.contains(jobId)) {
       logDebug("Trying to cancel unregistered job " + jobId)
     } else {
-      val independentStages = removeJobAndIndependentStages(jobId)
-      independentStages.foreach(taskScheduler.cancelTasks)
-      val error = new SparkException("Job %d cancelled".format(jobId))
-      val job = jobIdToActiveJob(jobId)
-      job.listener.jobFailed(error)
-      jobIdToStageIds -= jobId
-      activeJobs -= job
-      jobIdToActiveJob -= jobId
-      listenerBus.post(SparkListenerJobEnd(job.jobId, JobFailed(error, job.finalStage.id)))
+      failJobAndIndependentStages(jobIdToActiveJob(jobId), s"Job $jobId cancelled")
     }
   }
 
@@ -1007,19 +999,39 @@ class DAGScheduler(
     stageToInfos(failedStage).completionTime = Some(System.currentTimeMillis())
     for (resultStage <- dependentStages) {
       val job = resultStageToJob(resultStage)
-      val error = new SparkException("Job aborted: " + reason)
-      job.listener.jobFailed(error)
-      jobIdToStageIdsRemove(job.jobId)
-      jobIdToActiveJob -= resultStage.jobId
-      activeJobs -= job
-      resultStageToJob -= resultStage
-      listenerBus.post(SparkListenerJobEnd(job.jobId, JobFailed(error, failedStage.id)))
+      failJobAndIndependentStages(job, s"Job aborted due to stage failure: $reason")
     }
     if (dependentStages.isEmpty) {
       logInfo("Ignoring failure of " + failedStage + " because all jobs depending on it are done")
     }
   }
 
+  /**
+   * Fails a job and all stages that are only used by that job, and cleans up relevant state.
+   */
+  private def failJobAndIndependentStages(job: ActiveJob, failureReason: String) {
+    val error = new SparkException(failureReason)
+    job.listener.jobFailed(error)
+
+    // Cancel all tasks in independent stages.
+    val independentStages = removeJobAndIndependentStages(job.jobId)
+    independentStages.foreach(taskScheduler.cancelTasks)
+
+    // Clean up remaining state we store for the job.
+    jobIdToActiveJob -= job.jobId
+    activeJobs -= job
+    jobIdToStageIds -= job.jobId
+    val resultStagesForJob = resultStageToJob.keySet.filter(
+      stage => resultStageToJob(stage).jobId == job.jobId)
+    if (resultStagesForJob.size != 1) {
+      logWarning(
+        s"${resultStagesForJob.size} result stages for job ${job.jobId} (expect exactly 1)")
+    }
+    resultStageToJob --= resultStagesForJob
+
+    listenerBus.post(SparkListenerJobEnd(job.jobId, JobFailed(error, job.finalStage.id)))
+  }
+
   /**
    * Return true if one of stage's ancestors is target.
    */

core/src/test/scala/org/apache/spark/scheduler/DAGSchedulerSuite.scala

@@ -18,7 +18,7 @@
 package org.apache.spark.scheduler
 
 import scala.Tuple2
-import scala.collection.mutable.{HashMap, Map}
+import scala.collection.mutable.{HashSet, HashMap, Map}
 
 import org.scalatest.{BeforeAndAfter, FunSuite}
 
@@ -43,6 +43,10 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
   val conf = new SparkConf
   /** Set of TaskSets the DAGScheduler has requested executed. */
   val taskSets = scala.collection.mutable.Buffer[TaskSet]()
+
+  /** Stages for which the DAGScheduler has called TaskScheduler.cancelTasks(). */
+  val cancelledStages = new HashSet[Int]()
+
   val taskScheduler = new TaskScheduler() {
     override def rootPool: Pool = null
     override def schedulingMode: SchedulingMode = SchedulingMode.NONE
@@ -53,7 +57,9 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
       taskSet.tasks.foreach(_.epoch = mapOutputTracker.getEpoch)
       taskSets += taskSet
     }
-    override def cancelTasks(stageId: Int) {}
+    override def cancelTasks(stageId: Int) {
+      cancelledStages += stageId
+    }
     override def setDAGScheduler(dagScheduler: DAGScheduler) = {}
     override def defaultParallelism() = 2
   }
@@ -91,6 +97,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
   before {
     sc = new SparkContext("local", "DAGSchedulerSuite")
     taskSets.clear()
+    cancelledStages.clear()
     cacheLocations.clear()
     results.clear()
     mapOutputTracker = new MapOutputTrackerMaster(conf)
@@ -174,22 +181,28 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
     }
   }
 
-  /** Sends the rdd to the scheduler for scheduling. */
+  /** Sends the rdd to the scheduler for scheduling and returns the job id. */
   private def submit(
       rdd: RDD[_],
       partitions: Array[Int],
       func: (TaskContext, Iterator[_]) => _ = jobComputeFunc,
       allowLocal: Boolean = false,
-      listener: JobListener = listener) {
+      listener: JobListener = listener): Int = {
     val jobId = scheduler.nextJobId.getAndIncrement()
     runEvent(JobSubmitted(jobId, rdd, func, partitions, allowLocal, null, listener))
+    return jobId
   }
 
   /** Sends TaskSetFailed to the scheduler. */
   private def failed(taskSet: TaskSet, message: String) {
     runEvent(TaskSetFailed(taskSet, message))
   }
 
+  /** Sends JobCancelled to the DAG scheduler. */
+  private def cancel(jobId: Int) {
+    runEvent(JobCancelled(jobId))
+  }
+
   test("zero split job") {
     val rdd = makeRdd(0, Nil)
     var numResults = 0
@@ -248,7 +261,15 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
   test("trivial job failure") {
     submit(makeRdd(1, Nil), Array(0))
     failed(taskSets(0), "some failure")
-    assert(failure.getMessage === "Job aborted: some failure")
+    assert(failure.getMessage === "Job aborted due to stage failure: some failure")
+    assertDataStructuresEmpty
+  }
+
+  test("trivial job cancellation") {
+    val rdd = makeRdd(1, Nil)
+    val jobId = submit(rdd, Array(0))
+    cancel(jobId)
+    assert(failure.getMessage === s"Job $jobId cancelled")
     assertDataStructuresEmpty
   }
 
@@ -323,6 +344,67 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
     assertDataStructuresEmpty
   }
 
+  test("run shuffle with map stage failure") {
+    val shuffleMapRdd = makeRdd(2, Nil)
+    val shuffleDep = new ShuffleDependency(shuffleMapRdd, null)
+    val reduceRdd = makeRdd(2, List(shuffleDep))
+    submit(reduceRdd, Array(0, 1))
+
+    // Fail the map stage.  This should cause the entire job to fail.
+    val stageFailureMessage = "Exception failure in map stage"
+    failed(taskSets(0), stageFailureMessage)
+    assert(failure.getMessage === s"Job aborted due to stage failure: $stageFailureMessage")
+    assertDataStructuresEmpty
+  }
+
+  /**
+   * Makes sure that failures of stage used by multiple jobs are correctly handled.
+   *
+   * This test creates the following dependency graph:
+   *
+   * shuffleMapRdd1     shuffleMapRDD2
+   *        |     \        |
+   *        |      \       |
+   *        |       \      |
+   *        |        \     |
+   *   reduceRdd1    reduceRdd2
+   *
+   * We start both shuffleMapRdds and then fail shuffleMapRdd1.  As a result, the job listeners for
+   * reduceRdd1 and reduceRdd2 should both be informed that the job failed.  shuffleMapRDD2 should
+   * also be cancelled, because it is only used by reduceRdd2 and reduceRdd2 cannot complete
+   * without shuffleMapRdd1.
+   */
+  test("failure of stage used by two jobs") {
+    val shuffleMapRdd1 = makeRdd(2, Nil)
+    val shuffleDep1 = new ShuffleDependency(shuffleMapRdd1, null)
+    val shuffleMapRdd2 = makeRdd(2, Nil)
+    val shuffleDep2 = new ShuffleDependency(shuffleMapRdd2, null)
+
+    val reduceRdd1 = makeRdd(2, List(shuffleDep1))
+    val reduceRdd2 = makeRdd(2, List(shuffleDep1, shuffleDep2))
+
+    // We need to make our own listeners for this test, since by default submit uses the same
+    // listener for all jobs, and here we want to capture the failure for each job separately.
+    class FailureRecordingJobListener() extends JobListener {
+      var failureMessage: String = _
+      override def taskSucceeded(index: Int, result: Any) {}
+      override def jobFailed(exception: Exception) = { failureMessage = exception.getMessage }
+    }
+    val listener1 = new FailureRecordingJobListener()
+    val listener2 = new FailureRecordingJobListener()
+
+    submit(reduceRdd1, Array(0, 1), listener=listener1)
+    submit(reduceRdd2, Array(0, 1), listener=listener2)
+
+    val stageFailureMessage = "Exception failure in map stage"
+    failed(taskSets(0), stageFailureMessage)
+
+    assert(cancelledStages.contains(1))
+    assert(listener1.failureMessage === s"Job aborted due to stage failure: $stageFailureMessage")
+    assert(listener2.failureMessage === s"Job aborted due to stage failure: $stageFailureMessage")
+    assertDataStructuresEmpty
+  }
+
   test("run trivial shuffle with out-of-band failure and retry") {
     val shuffleMapRdd = makeRdd(2, Nil)
     val shuffleDep = new ShuffleDependency(shuffleMapRdd, null)