Github user kayousterhout commented on a diff in the pull request:
https://github.com/apache/spark/pull/11996#discussion_r64655524
--- Diff:
core/src/test/scala/org/apache/spark/scheduler/TaskSetManagerSuite.scala ---
@@ -789,6 +791,51 @@ class TaskSetManagerSuite extends SparkFunSuite with
LocalSparkContext with Logg
assert(TaskLocation("executor_host1_3") ===
ExecutorCacheTaskLocation("host1", "3"))
}
+ test("Kill other task attempts when one attempt belonging to the same
task succeeds") {
+ sc = new SparkContext("local", "test")
+ val sched = new FakeTaskScheduler(sc, ("exec1", "host1"), ("exec2",
"host2"))
+ val taskSet = FakeTask.createTaskSet(4)
+ val manager = new TaskSetManager(sched, taskSet, MAX_TASK_FAILURES)
+ val accumUpdatesByTask: Array[Seq[AccumulableInfo]] =
taskSet.tasks.map { task =>
+ task.initialAccumulators.map { a => a.toInfo(Some(0L), None) }
+ }
+ // Offer resources for 4 tasks to start
+ for ((k, v) <- List(
+ "exec1" -> "host1",
+ "exec1" -> "host1",
+ "exec2" -> "host2",
+ "exec2" -> "host2")) {
+ val taskOption = manager.resourceOffer(k, v, NO_PREF)
+ assert(taskOption.isDefined)
+ val task = taskOption.get
+ assert(task.executorId === k)
+ }
+ assert(sched.startedTasks.toSet === Set(0, 1, 2, 3))
+ // Complete the 3 tasks and leave 1 task in running
+ for (id <- Set(0, 1, 2)) {
+ manager.handleSuccessfulTask(id, createTaskResult(id,
accumUpdatesByTask(id)))
+ assert(sched.endedTasks(id) === Success)
+ }
+
+ // Wait for the threshold time to start speculative attempt for the
running task
+ Thread.sleep(100)
--- End diff --
This does seem a little trickier than I'd anticipated. I think the best
thing to do is (1) change the Schedulable class to take a minTimeToSpeculation
required argument. This looks pretty simple to do -- you just need to change
two implementations, and this is a private spark class, so we're not changing a
public or developer API. (2) add a constant MIN_TIME_TO_SPECULATION in the
TaskSchedulerImpl object, and pass that value in when TaskSchedulerImpl calls
checkSpeculatableTasks.
I think overtime, we should have more tests to verify the speculation
behavior, so this relatively small change to make this code path more testable
seems worthwhile.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
