[SPARK-23816][CORE] Killed tasks should ignore FetchFailures.

squito · squito · commit d886ba3840ab · 2018-04-05T16:21:47.000-05:00
SPARK-19276 ensured that FetchFailures do not get swallowed by other layers of exception handling, but it also meant that a killed task could look like a fetch failure. This is particularly a problem with speculative execution, where we expect to kill tasks as they are reading shuffle data. The fix is to ensure that we always check for killed tasks first. Added a new unit test which fails before the fix, ran it 1k times to check for flakiness. Full suite of tests on jenkins.
diff --git a/core/src/main/scala/org/apache/spark/executor/Executor.scala b/core/src/main/scala/org/apache/spark/executor/Executor.scala
@@ -480,6 +480,19 @@ private[spark] class Executor(
         execBackend.statusUpdate(taskId, TaskState.FINISHED, serializedResult)
 
       } catch {
+        case t: TaskKilledException =>
+          logInfo(s"Executor killed $taskName (TID $taskId), reason: ${t.reason}")
+          setTaskFinishedAndClearInterruptStatus()
+          execBackend.statusUpdate(taskId, TaskState.KILLED, ser.serialize(TaskKilled(t.reason)))
+
+        case _: InterruptedException | NonFatal(_) if
+            task != null && task.reasonIfKilled.isDefined =>
+          val killReason = task.reasonIfKilled.getOrElse("unknown reason")
+          logInfo(s"Executor interrupted and killed $taskName (TID $taskId), reason: $killReason")
+          setTaskFinishedAndClearInterruptStatus()
+          execBackend.statusUpdate(
+            taskId, TaskState.KILLED, ser.serialize(TaskKilled(killReason)))
+
         case t: Throwable if hasFetchFailure && !Utils.isFatalError(t) =>
           val reason = task.context.fetchFailed.get.toTaskFailedReason
           if (!t.isInstanceOf[FetchFailedException]) {
@@ -494,19 +507,6 @@ private[spark] class Executor(
           setTaskFinishedAndClearInterruptStatus()
           execBackend.statusUpdate(taskId, TaskState.FAILED, ser.serialize(reason))
 
-        case t: TaskKilledException =>
-          logInfo(s"Executor killed $taskName (TID $taskId), reason: ${t.reason}")
-          setTaskFinishedAndClearInterruptStatus()
-          execBackend.statusUpdate(taskId, TaskState.KILLED, ser.serialize(TaskKilled(t.reason)))
-
-        case _: InterruptedException | NonFatal(_) if
-            task != null && task.reasonIfKilled.isDefined =>
-          val killReason = task.reasonIfKilled.getOrElse("unknown reason")
-          logInfo(s"Executor interrupted and killed $taskName (TID $taskId), reason: $killReason")
-          setTaskFinishedAndClearInterruptStatus()
-          execBackend.statusUpdate(
-            taskId, TaskState.KILLED, ser.serialize(TaskKilled(killReason)))
-
         case CausedBy(cDE: CommitDeniedException) =>
           val reason = cDE.toTaskCommitDeniedReason
           setTaskFinishedAndClearInterruptStatus()
diff --git a/core/src/test/scala/org/apache/spark/executor/ExecutorSuite.scala b/core/src/test/scala/org/apache/spark/executor/ExecutorSuite.scala
@@ -139,7 +139,7 @@ class ExecutorSuite extends SparkFunSuite with LocalSparkContext with MockitoSug
     // the fetch failure.  The executor should still tell the driver that the task failed due to a
     // fetch failure, not a generic exception from user code.
     val inputRDD = new FetchFailureThrowingRDD(sc)
-    val secondRDD = new FetchFailureHidingRDD(sc, inputRDD, throwOOM = false)
+    val secondRDD = new FetchFailureHidingRDD(sc, inputRDD, throwOOM = false, interrupt = false)
     val taskBinary = sc.broadcast(serializer.serialize((secondRDD, resultFunc)).array())
     val serializedTaskMetrics = serializer.serialize(TaskMetrics.registered).array()
     val task = new ResultTask(
@@ -173,8 +173,26 @@ class ExecutorSuite extends SparkFunSuite with LocalSparkContext with MockitoSug
   }
 
   test("SPARK-19276: OOMs correctly handled with a FetchFailure") {
+    val (failReason, uncaughtExceptionHandler) = testFetchFailureHandling(true)
+    assert(failReason.isInstanceOf[ExceptionFailure])
+    val exceptionCaptor = ArgumentCaptor.forClass(classOf[Throwable])
+    verify(uncaughtExceptionHandler).uncaughtException(any(), exceptionCaptor.capture())
+    assert(exceptionCaptor.getAllValues.size === 1)
+    assert(exceptionCaptor.getAllValues().get(0).isInstanceOf[OutOfMemoryError])
+  }
+
+  test(s"SPARK-23816: interrupts are not masked by a FetchFailure") {
+    // If killing the task causes a fetch failure, we still treat it as a task that was killed,
+    // as the fetch failure could easily be caused by interrupting the thread.
+    val (failReason, _) = testFetchFailureHandling(false)
+    assert(failReason.isInstanceOf[TaskKilled])
+  }
+
+  def testFetchFailureHandling(oom: Boolean): (TaskFailedReason, UncaughtExceptionHandler) = {
     // when there is a fatal error like an OOM, we don't do normal fetch failure handling, since it
     // may be a false positive.  And we should call the uncaught exception handler.
+    // SPARK-23816 also handle interrupts the same way, as killing an obsolete speculative task
+    // does not represent a real fetch failure.
     val conf = new SparkConf().setMaster("local").setAppName("executor suite test")
     sc = new SparkContext(conf)
     val serializer = SparkEnv.get.closureSerializer.newInstance()
@@ -183,7 +201,13 @@ class ExecutorSuite extends SparkFunSuite with LocalSparkContext with MockitoSug
     // Submit a job where a fetch failure is thrown, but then there is an OOM.  We should treat
     // the fetch failure as a false positive, and just do normal OOM handling.
     val inputRDD = new FetchFailureThrowingRDD(sc)
-    val secondRDD = new FetchFailureHidingRDD(sc, inputRDD, throwOOM = true)
+    if (!oom) {
+      // we are trying to setup a case where a task is killed after a fetch failure -- this
+      // is just a helper to coordinate between the task thread and this thread that will
+      // kill the task
+      ExecutorSuiteHelper.latches = new ExecutorSuiteHelper()
+    }
+    val secondRDD = new FetchFailureHidingRDD(sc, inputRDD, throwOOM = oom, interrupt = !oom)
     val taskBinary = sc.broadcast(serializer.serialize((secondRDD, resultFunc)).array())
     val serializedTaskMetrics = serializer.serialize(TaskMetrics.registered).array()
     val task = new ResultTask(
@@ -200,15 +224,8 @@ class ExecutorSuite extends SparkFunSuite with LocalSparkContext with MockitoSug
     val serTask = serializer.serialize(task)
     val taskDescription = createFakeTaskDescription(serTask)
 
-    val (failReason, uncaughtExceptionHandler) =
-      runTaskGetFailReasonAndExceptionHandler(taskDescription)
-    // make sure the task failure just looks like a OOM, not a fetch failure
-    assert(failReason.isInstanceOf[ExceptionFailure])
-    val exceptionCaptor = ArgumentCaptor.forClass(classOf[Throwable])
-    verify(uncaughtExceptionHandler).uncaughtException(any(), exceptionCaptor.capture())
-    assert(exceptionCaptor.getAllValues.size === 1)
-    assert(exceptionCaptor.getAllValues.get(0).isInstanceOf[OutOfMemoryError])
-  }
+    runTaskGetFailReasonAndExceptionHandler(taskDescription, killTask = !oom)
+ }
 
   test("Gracefully handle error in task deserialization") {
     val conf = new SparkConf
@@ -257,19 +274,32 @@ class ExecutorSuite extends SparkFunSuite with LocalSparkContext with MockitoSug
   }
 
   private def runTaskAndGetFailReason(taskDescription: TaskDescription): TaskFailedReason = {
-    runTaskGetFailReasonAndExceptionHandler(taskDescription)._1
+    runTaskGetFailReasonAndExceptionHandler(taskDescription, false)._1
   }
 
   private def runTaskGetFailReasonAndExceptionHandler(
-      taskDescription: TaskDescription): (TaskFailedReason, UncaughtExceptionHandler) = {
+      taskDescription: TaskDescription,
+      killTask: Boolean): (TaskFailedReason, UncaughtExceptionHandler) = {
     val mockBackend = mock[ExecutorBackend]
     val mockUncaughtExceptionHandler = mock[UncaughtExceptionHandler]
     var executor: Executor = null
+    var killingThread: Thread = null
     try {
       executor = new Executor("id", "localhost", SparkEnv.get, userClassPath = Nil, isLocal = true,
         uncaughtExceptionHandler = mockUncaughtExceptionHandler)
       // the task will be launched in a dedicated worker thread
       executor.launchTask(mockBackend, taskDescription)
+      if (killTask) {
+        killingThread = new Thread("kill-task") {
+          override def run(): Unit = {
+            // wait to kill the task until it has thrown a fetch failure
+            ExecutorSuiteHelper.latches.latch1.await()
+            // now we can kill the task
+            executor.killAllTasks(true, "Killed task, eg. because of speculative execution")
+          }
+        }
+        killingThread.start()
+      }
       eventually(timeout(5.seconds), interval(10.milliseconds)) {
         assert(executor.numRunningTasks === 0)
       }
@@ -282,8 +312,9 @@ class ExecutorSuite extends SparkFunSuite with LocalSparkContext with MockitoSug
     val statusCaptor = ArgumentCaptor.forClass(classOf[ByteBuffer])
     orderedMock.verify(mockBackend)
       .statusUpdate(meq(0L), meq(TaskState.RUNNING), statusCaptor.capture())
+    val finalState = if (killTask) TaskState.KILLED else TaskState.FAILED
     orderedMock.verify(mockBackend)
-      .statusUpdate(meq(0L), meq(TaskState.FAILED), statusCaptor.capture())
+      .statusUpdate(meq(0L), meq(finalState), statusCaptor.capture())
     // first statusUpdate for RUNNING has empty data
     assert(statusCaptor.getAllValues().get(0).remaining() === 0)
     // second update is more interesting
@@ -321,7 +352,8 @@ class SimplePartition extends Partition {
 class FetchFailureHidingRDD(
     sc: SparkContext,
     val input: FetchFailureThrowingRDD,
-    throwOOM: Boolean) extends RDD[Int](input) {
+    throwOOM: Boolean,
+    interrupt: Boolean) extends RDD[Int](input) {
   override def compute(split: Partition, context: TaskContext): Iterator[Int] = {
     val inItr = input.compute(split, context)
     try {
@@ -330,6 +362,15 @@ class FetchFailureHidingRDD(
       case t: Throwable =>
         if (throwOOM) {
           throw new OutOfMemoryError("OOM while handling another exception")
+        } else if (interrupt) {
+          // make sure our test is setup correctly
+          assert(TaskContext.get().asInstanceOf[TaskContextImpl].fetchFailed.isDefined)
+          // signal our test is ready for the task to get killed
+          ExecutorSuiteHelper.latches.latch1.countDown()
+          // then wait for another thread in the test to kill the task -- this latch
+          // is never actually decremented, we just wait to get killed.
+          ExecutorSuiteHelper.latches.latch2.await()
+          throw new IllegalStateException("impossible")
         } else {
           throw new RuntimeException("User Exception that hides the original exception", t)
         }
@@ -352,6 +393,11 @@ private class ExecutorSuiteHelper {
   @volatile var testFailedReason: TaskFailedReason = _
 }
 
+// helper for coordinating killing tasks
+private object ExecutorSuiteHelper {
+  var latches: ExecutorSuiteHelper = null
+}
+
 private class NonDeserializableTask extends FakeTask(0, 0) with Externalizable {
   def writeExternal(out: ObjectOutput): Unit = {}
   def readExternal(in: ObjectInput): Unit = {
