Track active shuffles by stage (apache-spark-on-k8s#446)

Author: rynorris

core/src/main/scala/org/apache/spark/ExecutorAllocationManager.scala

@@ -677,6 +677,7 @@ private[spark] class ExecutorAllocationManager(
     override def onJobEnd(jobEnd: SparkListenerJobEnd): Unit = {
       // At the end of a job, trigger the callbacks for idle executors again to clean up executors
       // which we were keeping around only because they held active shuffle blocks.
+      logDebug("Checking for idle executors at end of job")
       allocationManager.checkForIdleExecutors()
     }
 
@@ -727,6 +728,11 @@ private[spark] class ExecutorAllocationManager(
         if (stageIdToNumTasks.isEmpty && stageIdToNumSpeculativeTasks.isEmpty) {
           allocationManager.onSchedulerQueueEmpty()
         }
+
+        // Trigger the callbacks for idle executors again to clean up executors
+        // which we were keeping around only because they held active shuffle blocks.
+        logDebug("Checking for idle executors at end of stage")
+        allocationManager.checkForIdleExecutors()
       }
     }
 

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

@@ -25,7 +25,7 @@ import java.util.function.BiFunction
 
 import scala.annotation.tailrec
 import scala.collection.Map
-import scala.collection.mutable.{ArrayStack, HashMap, HashSet}
+import scala.collection.mutable.{ArrayStack, HashMap, HashSet, Set}
 import scala.concurrent.duration._
 import scala.language.existentials
 import scala.language.postfixOps
@@ -149,6 +149,13 @@ private[spark] class DAGScheduler(
    * the shuffle data will be in the MapOutputTracker).
    */
   private[scheduler] val shuffleIdToMapStage = new HashMap[Int, ShuffleMapStage]
+
+  /**
+   * Mapping from shuffle dependency ID to the IDs of the stages which depend on the shuffle data.
+   * Used to track when shuffle data becomes no longer active.
+   */
+  private[scheduler] val shuffleIdToDependentStages = new HashMap[Int, Set[Int]]
+
   private[scheduler] val jobIdToActiveJob = new HashMap[Int, ActiveJob]
 
   // Stages we need to run whose parents aren't done
@@ -396,6 +403,7 @@ private[spark] class DAGScheduler(
     stageIdToStage(id) = stage
     shuffleIdToMapStage(shuffleDep.shuffleId) = stage
     updateJobIdStageIdMaps(jobId, stage)
+    updateShuffleDependenciesMap(stage)
 
     if (mapOutputTracker.containsShuffle(shuffleDep.shuffleId)) {
       mapOutputTracker.markShuffleActive(shuffleDep.shuffleId)
@@ -455,6 +463,7 @@ private[spark] class DAGScheduler(
     val stage = new ResultStage(id, rdd, func, partitions, parents, jobId, callSite)
     stageIdToStage(id) = stage
     updateJobIdStageIdMaps(jobId, stage)
+    updateShuffleDependenciesMap(stage)
     stage
   }
 
@@ -601,6 +610,17 @@ private[spark] class DAGScheduler(
     updateJobIdStageIdMapsList(List(stage))
   }
 
+  /**
+   * Registers the shuffle dependencies of the given stage.
+   */
+  private def updateShuffleDependenciesMap(stage: Stage): Unit = {
+    getShuffleDependencies(stage.rdd).foreach { shuffleDep =>
+      val shuffleId = shuffleDep.shuffleId
+      logDebug("Tracking that stage " + stage.id + " depends on shuffle " + shuffleId)
+      shuffleIdToDependentStages.getOrElseUpdate(shuffleId, Set.empty) += stage.id
+    }
+  }
+
   /**
    * Removes state for job and any stages that are not needed by any other job.  Does not
    * handle cancelling tasks or notifying the SparkListener about finished jobs/stages/tasks.
@@ -1893,6 +1913,24 @@ private[spark] class DAGScheduler(
       case Some(t) => "%.03f".format((clock.getTimeMillis() - t) / 1000.0)
       case _ => "Unknown"
     }
+
+    getShuffleDependencies(stage.rdd).foreach { shuffleDep =>
+      val shuffleId = shuffleDep.shuffleId
+      if (!shuffleIdToDependentStages.contains(shuffleId)) {
+        logDebug("Stage finished with untracked shuffle dependency " + shuffleId)
+      } else {
+        var dependentStages = shuffleIdToDependentStages(shuffleId)
+        dependentStages -= stage.id;
+        logDebug("Stage " + stage.id + " finished.  " +
+          "Shuffle " + shuffleId + " now has dependencies " + dependentStages)
+        if (dependentStages.isEmpty) {
+          logDebug("Shuffle " + shuffleId + " is no longer needed.  Marking it inactive.")
+          shuffleIdToDependentStages.remove(shuffleId)
+          mapOutputTracker.markShuffleInactive(shuffleId)
+        }
+      }
+    }
+
     if (errorMessage.isEmpty) {
       logInfo("%s (%s) finished in %s s".format(stage, stage.name, serviceTime))
       stage.latestInfo.completionTime = Some(clock.getTimeMillis())

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

@@ -36,7 +36,7 @@ import org.apache.spark.rdd.{DeterministicLevel, RDD}
 import org.apache.spark.scheduler.SchedulingMode.SchedulingMode
 import org.apache.spark.shuffle.{FetchFailedException, MetadataFetchFailedException}
 import org.apache.spark.storage.{BlockId, BlockManagerId, BlockManagerMaster}
-import org.apache.spark.util.{AccumulatorContext, AccumulatorV2, CallSite, LongAccumulator, Utils}
+import org.apache.spark.util._
 
 class DAGSchedulerEventProcessLoopTester(dagScheduler: DAGScheduler)
   extends DAGSchedulerEventProcessLoop(dagScheduler) {
@@ -2195,6 +2195,102 @@ class DAGSchedulerSuite extends SparkFunSuite with LocalSparkContext with TimeLi
     assertDataStructuresEmpty()
   }
 
+  test("stage level active shuffle tracking") {
+    // We will have 3 stages depending on each other.
+    // The second stage is composed of 2 RDDs to check we're tracking shuffle up the chain.
+    val shuffleMapRdd1 = new MyRDD(sc, 2, Nil)
+    val shuffleDep1 = new ShuffleDependency(shuffleMapRdd1, new HashPartitioner(1))
+    val shuffleId1 = shuffleDep1.shuffleId
+    val shuffleMapRdd2 = new MyRDD(sc, 2, List(shuffleDep1), tracker = mapOutputTracker)
+    val shuffleDep2 = new ShuffleDependency(shuffleMapRdd2, new HashPartitioner(1))
+    val shuffleId2 = shuffleDep2.shuffleId
+    val intermediateRdd = new MyRDD(sc, 1, List(shuffleDep2), tracker = mapOutputTracker)
+    val intermediateDep = new OneToOneDependency(intermediateRdd)
+    val reduceRdd = new MyRDD(sc, 1, List(intermediateDep), tracker = mapOutputTracker)
+
+    // Submit the job.
+    // Both shuffles should become active.
+    submit(reduceRdd, Array(0))
+    assert(mapOutputTracker.shuffleStatuses(shuffleId1).isActive === true)
+    assert(mapOutputTracker.shuffleStatuses(shuffleId2).isActive === true)
+
+    // Complete the first stage.
+    // Both shuffles remain active.
+    completeShuffleMapStageSuccessfully(0, 0, 2)
+    assert(mapOutputTracker.shuffleStatuses(shuffleId1).isActive === true)
+    assert(mapOutputTracker.shuffleStatuses(shuffleId2).isActive === true)
+
+    // Complete the second stage.
+    // Shuffle 1 is no longer needed and should become inactive.
+    completeShuffleMapStageSuccessfully(1, 0, 1)
+    assert(mapOutputTracker.shuffleStatuses(shuffleId1).isActive === false)
+    assert(mapOutputTracker.shuffleStatuses(shuffleId2).isActive === true)
+
+    // Complete the results stage.
+    // Both shuffles are no longer needed and should become inactive.
+    completeNextResultStageWithSuccess(2, 0)
+    assert(mapOutputTracker.shuffleStatuses(shuffleId1).isActive === false)
+    assert(mapOutputTracker.shuffleStatuses(shuffleId2).isActive === false)
+
+    // Double check results.
+    assert(results === Map(0 -> 42))
+    results.clear()
+    assertDataStructuresEmpty()
+  }
+
+  test("stage level active shuffle tracking with multiple dependents") {
+    // We will have a diamond shape dependency.
+    val shuffleMapRdd = new MyRDD(sc, 2, Nil)
+    val shuffleDep = new ShuffleDependency(shuffleMapRdd, new HashPartitioner(1))
+    val shuffleId = shuffleDep.shuffleId
+    val intermediateRdd1 = new MyRDD(sc, 1, List(shuffleDep), tracker = mapOutputTracker)
+    val intermediateRdd2 = new MyRDD(sc, 1, List(shuffleDep), tracker = mapOutputTracker)
+    val intermediateDep1 = new ShuffleDependency(intermediateRdd1, new HashPartitioner(1))
+    val intermediateDep2 = new ShuffleDependency(intermediateRdd2, new HashPartitioner(1))
+    val reduceRdd =
+      new MyRDD(sc, 1, List(intermediateDep1, intermediateDep2), tracker = mapOutputTracker)
+
+    // Submit the job.
+    // Shuffle becomes active.
+    submit(reduceRdd, Array(0))
+    assert(mapOutputTracker.shuffleStatuses(shuffleId).isActive === true)
+
+    // Complete the shuffle stage.
+    // Shuffle remains active.
+    completeShuffleMapStageSuccessfully(0, 0, 2)
+    assert(mapOutputTracker.shuffleStatuses(shuffleId).isActive === true)
+
+    // Complete first intermediate stage.
+    // Shuffle is still active.
+    val stageAttempt = taskSets(1)
+    checkStageId(1, 0, stageAttempt)
+    complete(stageAttempt, stageAttempt.tasks.zipWithIndex.map {
+      case (task, idx) =>
+        (Success, makeMapStatus("host" + ('A' + idx).toChar, 1))
+    }.toSeq)
+    assert(mapOutputTracker.shuffleStatuses(shuffleId).isActive === true)
+
+    // Complete second intermediate stage.
+    // Shuffle is no longer active.
+    val stageAttempt2 = taskSets(2)
+    checkStageId(2, 0, stageAttempt2)
+    complete(stageAttempt2, stageAttempt2.tasks.zipWithIndex.map {
+      case (task, idx) =>
+        (Success, makeMapStatus("host" + ('A' + idx).toChar, 1))
+    }.toSeq)
+    assert(mapOutputTracker.shuffleStatuses(shuffleId).isActive === false)
+
+    // Complete the results stage.
+    // Shuffle is still inactive.
+    completeNextResultStageWithSuccess(3, 0)
+    assert(mapOutputTracker.shuffleStatuses(shuffleId).isActive === false)
+
+    // Double check results.
+    assert(results === Map(0 -> 42))
+    results.clear()
+    assertDataStructuresEmpty()
+  }
+
   test("map stage submission with fetch failure") {
     val shuffleMapRdd = new MyRDD(sc, 2, Nil)
     val shuffleDep = new ShuffleDependency(shuffleMapRdd, new HashPartitioner(2))