You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@spark.apache.org by "Apache Spark (JIRA)" <ji...@apache.org> on 2017/01/17 17:03:26 UTC
[jira] [Assigned] (SPARK-19263) DAGScheduler should handle stage's
pendingPartitions properly in handleTaskCompletion.
[ https://issues.apache.org/jira/browse/SPARK-19263?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Apache Spark reassigned SPARK-19263:
------------------------------------
Assignee: Apache Spark
> DAGScheduler should handle stage's pendingPartitions properly in handleTaskCompletion.
> --------------------------------------------------------------------------------------
>
> Key: SPARK-19263
> URL: https://issues.apache.org/jira/browse/SPARK-19263
> Project: Spark
> Issue Type: Bug
> Components: Scheduler
> Affects Versions: 2.1.0
> Reporter: jin xing
> Assignee: Apache Spark
>
> In current *DAGScheduler handleTaskCompletion* code, when *event.reason* is *Success*, it will first do *stage.pendingPartitions -= task.partitionId*, which maybe a bug when *FetchFailed* happens. Think about below:
> 1. There are 2 executors A and B, executorA got assigned with ShuffleMapTask1 and ShuffleMapTask2;
> 2. ShuffleMapTask1 want's to fetch blocks from local but failed;
> 3. Driver receives the *FetchFailed* caused by ShuffleMapTask1 on executorA and marks executorA as lost and updates *failedEpoch*;
> 4. Driver resubmits stages, containing ShuffleMapTask1x and ShuffleMapTask2x;
> 5. ShuffleMapTask2 is successfully finished on executorA and sends *Success* back to driver;
> 6. Driver receives *Success* and do *stage.pendingPartitions -= task.partitionId*, but then driver finds task's epoch is not big enough *<= failedEpoch(execId)* and just takes it as bogus, does not add the *MapStatus* to stage;
> 7. ShuffleMapTask1x is successfully finished on executorB;
> 8. Driver receives *Success* from ShuffleMapTask1x on executorB and does *stage.pendingPartitions -= task.partitionId*, thus no pending partitions, but then finds not all partitions are available because of step 6;
> 9. Driver resubmits stage; but at this moment ShuffleMapTask2x is still running; in *TaskSchedulerImpl submitTasks*, it finds *conflictingTaskSet*, then throw *IllegalStateException*
> 10. Failed.
> To reproduce the bug:
> 1. We need to do some modification in *ShuffleBlockFetcherIterator*: check whether the task's index in *TaskSetManager* and stage attempt equal to 0 at the same time, if so, throw FetchFailedException;
> 2. Rebuild spark then submit following job:
> {code}
> val rdd = sc.parallelize(List((0, 1), (1, 1), (2, 1), (3, 1), (1, 2), (0, 3), (2, 1), (3, 1)), 2)
> rdd.reduceByKey {
> (v1, v2) => {
> Thread.sleep(10000)
> v1 + v2
> }
> }.map {
> keyAndValue => {
> (keyAndValue._1 % 2, keyAndValue._2)
> }
> }.reduceByKey {
> (v1, v2) => {
> Thread.sleep(10000)
> v1 + v2
> }
> }.collect
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org