You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@spark.apache.org by "Hyukjin Kwon (Jira)" <ji...@apache.org> on 2022/12/16 10:18:00 UTC
[jira] [Resolved] (SPARK-41541) Fix wrong child call in SQLShuffleWriteMetricsReporter.decRecordsWritten()
[ https://issues.apache.org/jira/browse/SPARK-41541?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Hyukjin Kwon resolved SPARK-41541.
----------------------------------
Fix Version/s: 3.3.2
3.1.4
3.2.3
3.4.0
Resolution: Fixed
Issue resolved by pull request 39086
[https://github.com/apache/spark/pull/39086]
> Fix wrong child call in SQLShuffleWriteMetricsReporter.decRecordsWritten()
> --------------------------------------------------------------------------
>
> Key: SPARK-41541
> URL: https://issues.apache.org/jira/browse/SPARK-41541
> Project: Spark
> Issue Type: Improvement
> Components: Shuffle
> Affects Versions: 3.0.0
> Reporter: Josh Rosen
> Assignee: Josh Rosen
> Priority: Major
> Fix For: 3.3.2, 3.1.4, 3.2.3, 3.4.0
>
>
> In the {{SQLShuffleWriteMetricsReporter.decRecordsWritten}} method, a call to a child accidentally decrements _bytesWritten_ instead of {_}recordsWritten{_}:
> {code:java}
> override def decRecordsWritten(v: Long): Unit = {
> metricsReporter.decBytesWritten(v)
> _recordsWritten.set(_recordsWritten.value - v)
> } {code}
> One of the situations where {{decRecordsWritten}} is called while reverting shuffle writes from failed/canceled tasks. Due to the mixup in these calls, the _recordsWritten_ metric ends up being _v_ records too high (since it wasn't decremented) and the _bytesWritten_ metric ends up _v_ records too low, causing some failed tasks' write metrics to look like
> {code:java}
> {"Shuffle Bytes Written":-2109,"Shuffle Write Time":2923270,"Shuffle Records Written":2109} {code}
> instead of
> {code:java}
> {"Shuffle Bytes Written":0,"Shuffle Write Time":2923270,"Shuffle Records Written":0} {code}
> I'll submit a fix for this.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org