[
https://issues.apache.org/jira/browse/FLINK-27133?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17520385#comment-17520385
]
Matthias Pohl commented on FLINK-27133:
---------------------------------------
Looking at the history of this test, [~pnowojski] and I observed a few more
things:
* Test was stable until mid of January 2022 (FLINK-27165 was created to cover
the investigation)
* There's a big increase in performance beginning of March 2022: {{git log
4f233f86..d212b83f}}
** It's interesting to note that it contains adding the {{flush}} to the
{{FileSystemJobResultStore#createDirtyJobResultEntry}} in {{281b8e74}} which
was then reverted when the drop happened
I have to look into how the JMH runner works. The
{{FileSystemJobResultStore#createDirtyJobResultEntry}} happens when the job
reaches a globally-terminal state. Generally, it doesn't make sense from a
logical point of few because the flush should be also be triggered by the
{{OutputStream#close}} call that happens afterwards.
Other than that, the test results are quite flaky and, therefore, hard to
interpret. My initial test runs didn't produce any conclusions. We came up with
the conclusion that the test performance is kind of back to what it was before
March 2022 and don't consider this a regression. But I'm tempted to add the
flush anyway again and observe the test a bit longer...
> Performance regression in serializerHeavyString
> -----------------------------------------------
>
> Key: FLINK-27133
> URL: https://issues.apache.org/jira/browse/FLINK-27133
> Project: Flink
> Issue Type: Bug
> Components: API / Type Serialization System, Benchmarks
> Affects Versions: 1.15.0, 1.16.0
> Reporter: Piotr Nowojski
> Assignee: Matthias Pohl
> Priority: Blocker
> Fix For: 1.15.0, 1.16.0
>
>
> http://codespeed.dak8s.net:8000/timeline/#/?exe=1&ben=serializerHeavyString&extr=on&quarts=on&equid=off&env=2&revs=200
> Suspected range: 5f21d15a09..caa296b813
> {noformat}
> caa296b813b [8 days ago] [FLINK-26957][runtime] Adds invariant to
> LocalDataOutputStream to verify that no operation is allowed on a closed
> stream [Matthias Pohl]
> c50b0706237 [8 days ago] [FLINK-26957][runtime] Removes flush in
> FileSystemJobResultStore [Matthias Pohl]
> 098dc7eca31 [7 days ago] [hotfix] Fixes error in JavaDoc [Matthias Pohl]
> 116f297478f [7 days ago] [hotfix] Fixes typo in
> NonClosingOutputStreamDecorator [Matthias Pohl]
> 74ecbd13cf8 [7 days ago] [hotfix][runtime] Adds write method for consistency
> reasons [Matthias Pohl]
> e031f82528b [5 days ago] [hotfix] Modify spelling error in IOUtils.java
> [lidefu]
> a5a31de5b30 [4 days ago] [FLINK-26368] [kafka] Add setProperty method to
> KafkaSinkBuilder [Gyula Fora]
> b9c874edae8 [4 days ago] [hotfix][metrics][docs] Update documentation
> [Chesnay Schepler]
> 3416b6440ac [6 days ago] [FLINK-27027][ci] Add default log file suffix
> [Chesnay Schepler]
> 0d105ac910d [6 days ago] [FLINK-27027][ci] Prevent creation of empty log
> files [Chesnay Schepler]
> 7ad47230f22 [4 days ago] [FLINK-27026][build] Upgrade checkstyle plugin
> [Chesnay Schepler]
> db8522cfddb [4 days ago] [FLINK-27034][gcp][tests] Use testcontainers
> [Chesnay Schepler]
> 0097b5a6fae [3 weeks ago] [FLINK-26712][table-planner] Metadata keys should
> not conflict with physical columns [Timo Walther]
> 26aa543b3bb [3 months ago] [FLINK-25907][runtime][security] Add pluggable
> delegation token manager [gabor.g.somogyi]
> cda343349f5 [7 days ago] [FLINK-26987][runtime] Fixes getAllAndLock livelock
> [Matthias Pohl]
> {noformat}
> I've run a benchmark request before FLINK-26957 and it suggests that it is
> indeed the cause for this regression:
> http://codespeed.dak8s.net:8080/job/flink-benchmark-request/77/artifact/jmh-result.csv/*view*/
> CC [~mapohl]
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
