[SPARK-15865][CORE] Blacklist should not result in job hanging with l…

…ess than 4 executors

## What changes were proposed in this pull request?

Before this change, when you turn on blacklisting with `spark.scheduler.executorTaskBlacklistTime`, but you have fewer than `spark.task.maxFailures` executors, you can end with a job "hung" after some task failures.

Whenever a taskset is unable to schedule anything on resourceOfferSingleTaskSet, we check whether the last pending task can be scheduled on *any* known executor.  If not, the taskset (and any corresponding jobs) are failed.
* Worst case, this is O(maxTaskFailures + numTasks).  But unless many executors are bad, this should be small
* This does not fail as fast as possible -- when a task becomes unschedulable, we keep scheduling other tasks.  This is to avoid an O(numPendingTasks * numExecutors) operation
* Also, it is conceivable this fails too quickly.  You may be 1 millisecond away from unblacklisting a place for a task to run, or acquiring a new executor.

## How was this patch tested?

Added unit test which failed before the change, ran new test 5k times manually, ran all scheduler tests manually, and the full suite via jenkins.

Author: Imran Rashid <[email protected]>

Closes apache#13603 from squito/progress_w_few_execs_and_blacklist.

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

@@ -30,6 +30,10 @@ import org.apache.spark.annotation.DeveloperApi
 @DeveloperApi
 class TaskInfo(
     val taskId: Long,
+    /**
+     * The index of this task within its task set. Not necessarily the same as the ID of the RDD
+     * partition that the task is computing.
+     */
     val index: Int,
     val attemptNumber: Int,
     val launchTime: Long,

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

@@ -279,6 +279,9 @@ private[spark] class TaskSchedulerImpl(
         }
       }
     }
+    if (!launchedTask) {
+      taskSet.abortIfCompletelyBlacklisted(executorIdToHost.keys)
+    }
     return launchedTask
   }
 

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

@@ -83,7 +83,7 @@ private[spark] class TaskSetManager(
   val copiesRunning = new Array[Int](numTasks)
   val successful = new Array[Boolean](numTasks)
   private val numFailures = new Array[Int](numTasks)
-  // key is taskId, value is a Map of executor id to when it failed
+  // key is taskId (aka TaskInfo.index), value is a Map of executor id to when it failed
   private val failedExecutors = new HashMap[Int, HashMap[String, Long]]()
 
   val taskAttempts = Array.fill[List[TaskInfo]](numTasks)(Nil)
@@ -270,7 +270,7 @@ private[spark] class TaskSetManager(
    * Is this re-execution of a failed task on an executor it already failed in before
    * EXECUTOR_TASK_BLACKLIST_TIMEOUT has elapsed ?
    */
-  private def executorIsBlacklisted(execId: String, taskId: Int): Boolean = {
+  private[scheduler] def executorIsBlacklisted(execId: String, taskId: Int): Boolean = {
     if (failedExecutors.contains(taskId)) {
       val failed = failedExecutors.get(taskId).get
 
@@ -575,6 +575,56 @@ private[spark] class TaskSetManager(
     index
   }
 
+  /**
+   * Check whether the given task set has been blacklisted to the point that it can't run anywhere.
+   *
+   * It is possible that this taskset has become impossible to schedule *anywhere* due to the
+   * blacklist.  The most common scenario would be if there are fewer executors than
+   * spark.task.maxFailures. We need to detect this so we can fail the task set, otherwise the job
+   * will hang.
+   *
+   * There's a tradeoff here: we could make sure all tasks in the task set are schedulable, but that
+   * would add extra time to each iteration of the scheduling loop. Here, we take the approach of
+   * making sure at least one of the unscheduled tasks is schedulable. This means we may not detect
+   * the hang as quickly as we could have, but we'll always detect the hang eventually, and the
+   * method is faster in the typical case. In the worst case, this method can take
+   * O(maxTaskFailures + numTasks) time, but it will be faster when there haven't been any task
+   * failures (this is because the method picks on unscheduled task, and then iterates through each
+   * executor until it finds one that the task hasn't failed on already).
+   */
+  private[scheduler] def abortIfCompletelyBlacklisted(executors: Iterable[String]): Unit = {
+
+    val pendingTask: Option[Int] = {
+      // usually this will just take the last pending task, but because of the lazy removal
+      // from each list, we may need to go deeper in the list.  We poll from the end because
+      // failed tasks are put back at the end of allPendingTasks, so we're more likely to find
+      // an unschedulable task this way.
+      val indexOffset = allPendingTasks.lastIndexWhere { indexInTaskSet =>
+        copiesRunning(indexInTaskSet) == 0 && !successful(indexInTaskSet)
+      }
+      if (indexOffset == -1) {
+        None
+      } else {
+        Some(allPendingTasks(indexOffset))
+      }
+    }
+
+    // If no executors have registered yet, don't abort the stage, just wait.  We probably
+    // got here because a task set was added before the executors registered.
+    if (executors.nonEmpty) {
+      // take any task that needs to be scheduled, and see if we can find some executor it *could*
+      // run on
+      pendingTask.foreach { taskId =>
+        if (executors.forall(executorIsBlacklisted(_, taskId))) {
+          val execs = executors.toIndexedSeq.sorted.mkString("(", ",", ")")
+          val partition = tasks(taskId).partitionId
+          abort(s"Aborting ${taskSet} because task $taskId (partition $partition)" +
+            s" has already failed on executors $execs, and no other executors are available.")
+        }
+      }
+    }
+  }
+
   /**
    * Marks the task as getting result and notifies the DAG Scheduler
    */

core/src/test/scala/org/apache/spark/executor/ExecutorSuite.scala

@@ -53,7 +53,7 @@ class ExecutorSuite extends SparkFunSuite {
     when(mockEnv.closureSerializer).thenReturn(serializer)
     val serializedTask =
       Task.serializeWithDependencies(
-        new FakeTask(0),
+        new FakeTask(0, 0),
         HashMap[String, Long](),
         HashMap[String, Long](),
         serializer.newInstance())

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

@@ -57,7 +57,8 @@ class BlacklistIntegrationSuite extends SchedulerIntegrationSuite[MultiExecutorM
   testScheduler(
     "With blacklist on, job will still fail if there are too many bad executors on bad host",
     extraConfs = Seq(
-      // just set this to something much longer than the test duration
+      // set this to something much longer than the test duration so that executors don't get
+      // removed from the blacklist during the test
       ("spark.scheduler.executorTaskBlacklistTime", "10000000")
     )
   ) {
@@ -74,7 +75,8 @@ class BlacklistIntegrationSuite extends SchedulerIntegrationSuite[MultiExecutorM
   testScheduler(
     "Bad node with multiple executors, job will still succeed with the right confs",
     extraConfs = Seq(
-      // just set this to something much longer than the test duration
+      // set this to something much longer than the test duration so that executors don't get
+      // removed from the blacklist during the test
       ("spark.scheduler.executorTaskBlacklistTime", "10000000"),
       // this has to be higher than the number of executors on the bad host
       ("spark.task.maxFailures", "5"),
@@ -91,6 +93,33 @@ class BlacklistIntegrationSuite extends SchedulerIntegrationSuite[MultiExecutorM
     assertDataStructuresEmpty(noFailure = true)
   }
 
+  // Make sure that if we've failed on all executors, but haven't hit task.maxFailures yet, the job
+  // doesn't hang
+  testScheduler(
+    "SPARK-15865 Progress with fewer executors than maxTaskFailures",
+    extraConfs = Seq(
+      // set this to something much longer than the test duration so that executors don't get
+      // removed from the blacklist during the test
+      "spark.scheduler.executorTaskBlacklistTime" -> "10000000",
+      "spark.testing.nHosts" -> "2",
+      "spark.testing.nExecutorsPerHost" -> "1",
+      "spark.testing.nCoresPerExecutor" -> "1"
+    )
+  ) {
+    def runBackend(): Unit = {
+      val (taskDescription, _) = backend.beginTask()
+      backend.taskFailed(taskDescription, new RuntimeException("test task failure"))
+    }
+    withBackend(runBackend _) {
+      val jobFuture = submit(new MockRDD(sc, 10, Nil), (0 until 10).toArray)
+      Await.ready(jobFuture, duration)
+      val pattern = ("Aborting TaskSet 0.0 because task .* " +
+        "already failed on executors \\(.*\\), and no other executors are available").r
+      assert(pattern.findFirstIn(failure.getMessage).isDefined,
+        s"Couldn't find $pattern in ${failure.getMessage()}")
+    }
+    assertDataStructuresEmpty(noFailure = false)
+  }
 }
 
 class MultiExecutorMockBackend(

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

@@ -21,7 +21,8 @@ import org.apache.spark.TaskContext
 
 class FakeTask(
     stageId: Int,
-    prefLocs: Seq[TaskLocation] = Nil) extends Task[Int](stageId, 0, 0) {
+    partitionId: Int,
+    prefLocs: Seq[TaskLocation] = Nil) extends Task[Int](stageId, 0, partitionId) {
   override def runTask(context: TaskContext): Int = 0
   override def preferredLocations: Seq[TaskLocation] = prefLocs
 }
@@ -40,7 +41,7 @@ object FakeTask {
       throw new IllegalArgumentException("Wrong number of task locations")
     }
     val tasks = Array.tabulate[Task[_]](numTasks) { i =>
-      new FakeTask(i, if (prefLocs.size != 0) prefLocs(i) else Nil)
+      new FakeTask(0, i, if (prefLocs.size != 0) prefLocs(i) else Nil)
     }
     new TaskSet(tasks, 0, stageAttemptId, 0, null)
   }

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

@@ -30,7 +30,7 @@ class PoolSuite extends SparkFunSuite with LocalSparkContext {
   def createTaskSetManager(stageId: Int, numTasks: Int, taskScheduler: TaskSchedulerImpl)
     : TaskSetManager = {
     val tasks = Array.tabulate[Task[_]](numTasks) { i =>
-      new FakeTask(i, Nil)
+      new FakeTask(stageId, i, Nil)
     }
     new TaskSetManager(taskScheduler, new TaskSet(tasks, stageId, 0, 0, null), 0)
   }

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

@@ -279,12 +279,6 @@ private[spark] abstract class MockBackend(
     ThreadUtils.newDaemonSingleThreadScheduledExecutor("driver-revive-thread")
   private val reviveIntervalMs = conf.getTimeAsMs("spark.scheduler.revive.interval", "10ms")
 
-  reviveThread.scheduleAtFixedRate(new Runnable {
-    override def run(): Unit = Utils.tryLogNonFatalError {
-      reviveOffers()
-    }
-  }, 0, reviveIntervalMs, TimeUnit.MILLISECONDS)
-
   /**
    * Test backends should call this to get a task that has been assigned to them by the scheduler.
    * Each task should be responded to with either [[taskSuccess]] or [[taskFailed]].
@@ -348,7 +342,13 @@ private[spark] abstract class MockBackend(
     assignedTasksWaitingToRun.nonEmpty
   }
 
-  override def start(): Unit = {}
+  override def start(): Unit = {
+    reviveThread.scheduleAtFixedRate(new Runnable {
+      override def run(): Unit = Utils.tryLogNonFatalError {
+        reviveOffers()
+      }
+    }, 0, reviveIntervalMs, TimeUnit.MILLISECONDS)
+  }
 
   override def stop(): Unit = {
     reviveThread.shutdown()

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

@@ -281,6 +281,98 @@ class TaskSchedulerImplSuite extends SparkFunSuite with LocalSparkContext with B
     assert(!failedTaskSet)
   }
 
+  test("abort stage if executor loss results in unschedulability from previously failed tasks") {
+    // Make sure we can detect when a taskset becomes unschedulable from a blacklisting.  This
+    // test explores a particular corner case -- you may have one task fail, but still be
+    // schedulable on another executor.  However, that executor may fail later on, leaving the
+    // first task with no place to run.
+    val taskScheduler = setupScheduler(
+      // set this to something much longer than the test duration so that executors don't get
+      // removed from the blacklist during the test
+      "spark.scheduler.executorTaskBlacklistTime" -> "10000000"
+    )
+
+    val taskSet = FakeTask.createTaskSet(2)
+    taskScheduler.submitTasks(taskSet)
+    val tsm = taskScheduler.taskSetManagerForAttempt(taskSet.stageId, taskSet.stageAttemptId).get
+
+    val firstTaskAttempts = taskScheduler.resourceOffers(Seq(
+      new WorkerOffer("executor0", "host0", 1),
+      new WorkerOffer("executor1", "host1", 1)
+    )).flatten
+    assert(Set("executor0", "executor1") === firstTaskAttempts.map(_.executorId).toSet)
+
+    // fail one of the tasks, but leave the other running
+    val failedTask = firstTaskAttempts.find(_.executorId == "executor0").get
+    taskScheduler.handleFailedTask(tsm, failedTask.taskId, TaskState.FAILED, TaskResultLost)
+    // at this point, our failed task could run on the other executor, so don't give up the task
+    // set yet.
+    assert(!failedTaskSet)
+
+    // Now we fail our second executor.  The other task can still run on executor1, so make an offer
+    // on that executor, and make sure that the other task (not the failed one) is assigned there
+    taskScheduler.executorLost("executor1", SlaveLost("oops"))
+    val nextTaskAttempts =
+      taskScheduler.resourceOffers(Seq(new WorkerOffer("executor0", "host0", 1))).flatten
+    // Note: Its OK if some future change makes this already realize the taskset has become
+    // unschedulable at this point (though in the current implementation, we're sure it will not)
+    assert(nextTaskAttempts.size === 1)
+    assert(nextTaskAttempts.head.executorId === "executor0")
+    assert(nextTaskAttempts.head.attemptNumber === 1)
+    assert(nextTaskAttempts.head.index != failedTask.index)
+
+    // now we should definitely realize that our task set is unschedulable, because the only
+    // task left can't be scheduled on any executors due to the blacklist
+    taskScheduler.resourceOffers(Seq(new WorkerOffer("executor0", "host0", 1)))
+    sc.listenerBus.waitUntilEmpty(100000)
+    assert(tsm.isZombie)
+    assert(failedTaskSet)
+    val idx = failedTask.index
+    assert(failedTaskSetReason == s"Aborting TaskSet 0.0 because task $idx (partition $idx) has " +
+      s"already failed on executors (executor0), and no other executors are available.")
+  }
+
+  test("don't abort if there is an executor available, though it hasn't had scheduled tasks yet") {
+    // interaction of SPARK-15865 & SPARK-16106
+    // if we have a small number of tasks, we might be able to schedule them all on the first
+    // executor.  But if those tasks fail, we should still realize there is another executor
+    // available and not bail on the job
+
+    val taskScheduler = setupScheduler(
+      // set this to something much longer than the test duration so that executors don't get
+      // removed from the blacklist during the test
+      "spark.scheduler.executorTaskBlacklistTime" -> "10000000"
+    )
+
+    val taskSet = FakeTask.createTaskSet(2, (0 until 2).map { _ => Seq(TaskLocation("host0")) }: _*)
+    taskScheduler.submitTasks(taskSet)
+    val tsm = taskScheduler.taskSetManagerForAttempt(taskSet.stageId, taskSet.stageAttemptId).get
+
+    val offers = Seq(
+      // each offer has more than enough free cores for the entire task set, so when combined
+      // with the locality preferences, we schedule all tasks on one executor
+      new WorkerOffer("executor0", "host0", 4),
+      new WorkerOffer("executor1", "host1", 4)
+    )
+    val firstTaskAttempts = taskScheduler.resourceOffers(offers).flatten
+    assert(firstTaskAttempts.size == 2)
+    firstTaskAttempts.foreach { taskAttempt => assert("executor0" === taskAttempt.executorId) }
+
+    // fail all the tasks on the bad executor
+    firstTaskAttempts.foreach { taskAttempt =>
+      taskScheduler.handleFailedTask(tsm, taskAttempt.taskId, TaskState.FAILED, TaskResultLost)
+    }
+
+    // Here is the main check of this test -- we have the same offers again, and we schedule it
+    // successfully.  Because the scheduler first tries to schedule with locality in mind, at first
+    // it won't schedule anything on executor1.  But despite that, we don't abort the job.  Then the
+    // scheduler tries for ANY locality, and successfully schedules tasks on executor1.
+    val secondTaskAttempts = taskScheduler.resourceOffers(offers).flatten
+    assert(secondTaskAttempts.size == 2)
+    secondTaskAttempts.foreach { taskAttempt => assert("executor1" === taskAttempt.executorId) }
+    assert(!failedTaskSet)
+  }
+
   test("SPARK-16106 locality levels updated if executor added to existing host") {
     val taskScheduler = setupScheduler()