[
https://issues.apache.org/jira/browse/SPARK-44389?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17742468#comment-17742468
]
Volodymyr Kot commented on SPARK-44389:
---------------------------------------
Please let me know if there any other information that would be useful here!
> ExecutorDeadException when using decommissioning without external shuffle
> service
> ---------------------------------------------------------------------------------
>
> Key: SPARK-44389
> URL: https://issues.apache.org/jira/browse/SPARK-44389
> Project: Spark
> Issue Type: Question
> Components: Spark Core
> Affects Versions: 3.4.0
> Reporter: Volodymyr Kot
> Priority: Major
>
> Hey, we are trying to use executor decommissioning without external shuffle
> service. We are trying to understand:
> # How often should we expect to see ExecutorDeadException? How is
> information about changes to location of blocks is propagated?
> # Whether the task should be re-submited if we hit that during
> decommissioning?
>
> Current behavior that we observe:
> # Executor 1 is decommissioned
> # Driver successfully removes executor 1's block manager
> [here|https://github.com/apache/spark/blob/87a5442f7ed96b11051d8a9333476d080054e5a0/core/src/main/scala/org/apache/spark/storage/BlockManagerMaster.scala#L44]
> # A task is started on executor 2
> # We hit `ExecutorDeadException` on executor 2 when trying to fetch blocks
> from executor 1
> [here|https://github.com/apache/spark/blob/87a5442f7ed96b11051d8a9333476d080054e5a0/core/src/main/scala/org/apache/spark/network/netty/NettyBlockTransferService.scala#L139-L140]
> # Task on executor 2 fails
> # Stage fails
> # Stage is re-submitted and succeeds
> As far as we understand, this happens because executor 2 has stale [map
> status
> cache|https://github.com/apache/spark/blob/87a5442f7ed96b11051d8a9333476d080054e5a0/core/src/main/scala/org/apache/spark/MapOutputTracker.scala#L1235-L1236]
> Is that expected behavior? Shouldn't the task be retried in that case instead
> of whole stage failing and being retried? This makes Spark job execution
> longer, especially if there are a lot of decommission events.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
