[
https://issues.apache.org/jira/browse/SPARK-13872?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15194014#comment-15194014
]
Ian commented on SPARK-13872:
-----------------------------
A spark plan illustrating the scenario was attached.
1. A cartesian is probably dumb in first place, but it is valid query.
A slow performance is expected but not OOM.
2. The plan is also special, in that, the SortMergeJoin is done with the
Cartesian Product at the same stage. If the SortMergeJoin is done in a separate
stage the OOM can be avoided. From query planning point of view,
is it optimized to run SortMergeJoin with Cartesian Product on the sam
stage?
From result correctness point view, no matter how the execution is planned,
OOM should not happen.
> Memory leak SortMergeOuterJoin
> ------------------------------
>
> Key: SPARK-13872
> URL: https://issues.apache.org/jira/browse/SPARK-13872
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 1.6.1
> Reporter: Ian
> Attachments: Screen Shot 2016-03-11 at 5.42.32 PM.png
>
>
> SortMergeJoin composes its partition/iterator from
> org.apache.spark.sql.execution.Sort, which in turns designates the sorting to
> UnsafeExternalRowSorter.
> UnsafeExternalRowSorter's implementation cleans up the resources when:
> 1. org.apache.spark.sql.catalyst.util.AbstractScalaRowIterator is fully
> iterated.
> 2. task is done execution.
> In case of outer join case of SortMergeJoin, when the left or right iterator
> is not fully iterated, the only only occasion for the recources to be cleaned
> up is at the end of the spark task. This probably ok most of the time,
> however when a SortMergeOuterJoin is nested within a CartesianProduct, the
> "deferred" resources cleanup becomes an memory leak amplified by the loop
> driven by the CartesianRdd's outter loop iteration.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
