[
https://issues.apache.org/jira/browse/SPARK-60373?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18128946#comment-18128946
]
Jungtaek Lim commented on SPARK-60373:
--------------------------------------
Will submit a PR soon.
> Streaming outer joins underreport removed state rows
> ----------------------------------------------------
>
> Key: SPARK-60373
> URL: https://issues.apache.org/jira/browse/SPARK-60373
> Project: Spark
> Issue Type: Bug
> Components: Structured Streaming
> Affects Versions: 4.2.0, 4.3.0, 4.1.3, 4.0.4
> Reporter: Jungtaek Lim
> Priority: Major
>
> The bug is an ancient one introduced in Spark 3.2.0.
>
> Structured Streaming outer joins underreport
> {{{}StateOperatorProgress.numRowsRemoved{}}}.
>
> State removed while producing unmatched outer-join output is not counted.
> Left/right outer joins omit removals from the preserved side; full outer
> joins can report zero despite removing state from both sides.
>
> An end-to-end full-outer regression buffers five rows per side, including
> three matching pairs. After a checkpoint restart and watermark advancement,
> all ten old state rows are removed, while four unmatched rows produce
> null-extended output. The metric reports {*}0{*}, but should report {*}10{*}.
>
> The fix counts every removed state row before filtering matched rows from
> output. It changes metric accounting only; query results, eviction decisions,
> state layout, and checkpoint compatibility remain unchanged.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
