[
https://issues.apache.org/jira/browse/SPARK-22471?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16244836#comment-16244836
]
Arseniy Tashoyan commented on SPARK-22471:
------------------------------------------
I am coming up with a simple fix, backportable to 2.2.
> SQLListener consumes much memory causing OutOfMemoryError
> ---------------------------------------------------------
>
> Key: SPARK-22471
> URL: https://issues.apache.org/jira/browse/SPARK-22471
> Project: Spark
> Issue Type: Bug
> Components: SQL, Web UI
> Affects Versions: 2.2.0
> Environment: Spark 2.2.0, Linux
> Reporter: Arseniy Tashoyan
> Labels: memory-leak, sql
> Attachments: SQLListener_retained_size.png,
> SQLListener_stageIdToStageMetrics_retained_size.png
>
> Original Estimate: 72h
> Remaining Estimate: 72h
>
> _SQLListener_ may grow very large when Spark runs complex multi-stage
> requests. The listener tracks metrics for all stages in
> __stageIdToStageMetrics_ hash map. _SQLListener_ has some means to cleanup
> this hash map regularly, but this is not enough. Precisely, the method
> _trimExecutionsIfNecessary_ ensures that __stageIdToStageMetrics_ does not
> have metrics for very old data; this method runs on each execution completion.
> However, if an execution has many stages, _SQLListener_ keeps adding new
> entries to __stageIdToStageMetrics_ without calling
> _trimExecutionsIfNecessary_. The hash map may grow to enormous size.
> Strictly speaking, it is not a memory leak, because finally
> _trimExecutionsIfNecessary_ cleans the hash map. However, the driver program
> has high odds to crash with OutOfMemoryError (and it does).
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
