[
https://issues.apache.org/jira/browse/SPARK-23423?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16368261#comment-16368261
]
Susan X. Huynh commented on SPARK-23423:
----------------------------------------
I'll check if the task updates might be dropped under heavy load.
[~igor.berman] Normally, you should see the TASK_KILLED updates in the logs,
something like:
{noformat}
15:38:47 INFO TaskSchedulerImpl: Executor 2 on 10.0.1.201 killed by driver.
15:38:47 INFO DAGScheduler: Executor lost: 2 (epoch 0)
15:38:47 INFO BlockManagerMasterEndpoint: Trying to remove executor 2 from
BlockManagerMaster.
15:38:47 INFO BlockManagerMasterEndpoint: Removing block manager
BlockManagerId(2, 10.0.1.201, 42805, None)
15:38:47 INFO BlockManagerMaster: Removed 2 successfully in removeExecutor
15:38:47 INFO ExecutorAllocationManager: Existing executor 2 has been removed
(new total is 1)
15:38:48 INFO MesosCoarseGrainedSchedulerBackend: Mesos task 2 is now
TASK_KILLED
15:38:48 INFO BlockManagerMaster: Removal of executor 2 requested
15:38:48 INFO CoarseGrainedSchedulerBackend$DriverEndpoint: Asked to remove
non-existent executor 2
{noformat}
> Application declines any offers when killed+active executors rich
> spark.dynamicAllocation.maxExecutors
> ------------------------------------------------------------------------------------------------------
>
> Key: SPARK-23423
> URL: https://issues.apache.org/jira/browse/SPARK-23423
> Project: Spark
> Issue Type: Bug
> Components: Mesos, Spark Core
> Affects Versions: 2.2.1
> Reporter: Igor Berman
> Priority: Major
> Labels: Mesos, dynamic_allocation
>
> Hi
> Mesos Version:1.1.0
> I've noticed rather strange behavior of MesosCoarseGrainedSchedulerBackend
> when running on Mesos with dynamic allocation on and limiting number of max
> executors by spark.dynamicAllocation.maxExecutors.
> Suppose we have long running driver that has cyclic pattern of resource
> consumption(with some idle times in between), due to dyn.allocation it
> receives offers and then releases them after current chunk of work processed.
> Since at
> [https://github.com/apache/spark/blob/master/resource-managers/mesos/src/main/scala/org/apache/spark/scheduler/cluster/mesos/MesosCoarseGrainedSchedulerBackend.scala#L573]
> the backend compares numExecutors < executorLimit and
> numExecutors is defined as slaves.values.map(_.taskIDs.size).sum and slaves
> holds all slaves ever "met", i.e. both active and killed (see comment
> [https://github.com/apache/spark/blob/master/resource-managers/mesos/src/main/scala/org/apache/spark/scheduler/cluster/mesos/MesosCoarseGrainedSchedulerBackend.scala#L122)]
>
> On the other hand, number of taskIds should be updated due to statusUpdate,
> but suppose this update is lost(actually I don't see logs of 'is now
> TASK_KILLED') so this number of executors might be wrong
>
> I've created test that "reproduces" this behavior, not sure how good it is:
> {code:java}
> //MesosCoarseGrainedSchedulerBackendSuite
> test("max executors registered stops to accept offers when dynamic allocation
> enabled") {
> setBackend(Map(
> "spark.dynamicAllocation.maxExecutors" -> "1",
> "spark.dynamicAllocation.enabled" -> "true",
> "spark.dynamicAllocation.testing" -> "true"))
> backend.doRequestTotalExecutors(1)
> val (mem, cpu) = (backend.executorMemory(sc), 4)
> val offer1 = createOffer("o1", "s1", mem, cpu)
> backend.resourceOffers(driver, List(offer1).asJava)
> verifyTaskLaunched(driver, "o1")
> backend.doKillExecutors(List("0"))
> verify(driver, times(1)).killTask(createTaskId("0"))
> val offer2 = createOffer("o2", "s2", mem, cpu)
> backend.resourceOffers(driver, List(offer2).asJava)
> verify(driver, times(1)).declineOffer(offer2.getId)
> }{code}
>
>
> Workaround: Don't set maxExecutors with dynamicAllocation on
>
> Please advice
> Igor
> marking you friends since you were last to touch this piece of code and
> probably can advice something([~vanzin], [~skonto], [~susanxhuynh])
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
