Hi devs,
Flink benchmarks are periodically executed on
http://codespeed.dak8s.net:8080 to monitor Flink performance. In late
Oct'22, a new slack channel #flink-dev-benchmarks was created for
notifications of performance regressions. It helped us find 2 build
failures[1,2] and 5 performance regressions[3,4,5,6,7] in the past 3
months, which is very meaningful for ensuring the quality of the code.
I am checking the slack notifications once a week now, and if more
people come to monitor together, we can check once a day in the future
to find out regressions in a timely manner.
According to some contributors and my own experience, I have
summarized a document on how to handle performance regressions. The
following is just a draft, which can be continuously iterated and
improved later.
When a benchmark regression is detected, the following steps will help
to deal with regressions:
1. Create a Jira ticket(one per group of related benchmarks). Set
effects and fix versions to the current Flink version,
component=Benchmarks, type=Bug.
2. Post the ticket in the slack channel(replying in a thread).
3. Verify that the regression is real and investigate the cause. Take
FLINK-30623[5] as an example:
3.1 Inspect the timeline following the
link(http://codespeed.dak8s.net:8000/timeline/#/?exe=1&ben=checkpointSingleInput.UNALIGNED&extr=on&quarts=on&equid=off&env=2&revs=200)
from the notification. Suspicious commit ranges can be obtained from
the figure, for this example, the suspicious range is
13ef498172b...fb272D2cdebf.
3.2 Narrow down the commit range via git log. You can directly
locate a specific commit based on experience or compare the benchmark
results of each commit in this range, a commit would be found if this
regression is real. See instructions for using benchmark-request, you
can also try to benchmark locally. http://codespeed.dak8s.net:8080
benchmarking infrastructure is hosted using resources provided by
Ververica(Alibaba) and maintained by PMCs and Ververica, please
contact one of Apache Flink PMCs to get access. For example, two
benchmark requests had been submitted to verify whether FLINK-30533
caused the regression.
> Before FLINK-30533:
> http://codespeed.dak8s.net:8080/job/flink-benchmark-request/177
>
> - checkpointSingleInput.UNALIGNED: 333.635178(+-8.169488)
>
> - checkpointSingleInput.UNALIGNED_1: 213.837107(+-7.282883)
>
> # After FLINK-30533:
> http://codespeed.dak8s.net:8080/job/flink-benchmark-request/178
>
> - checkpointSingleInput.UNALIGNED: 61.536982（+-3.581509）
>
> - checkpointSingleInput.UNALIGNED_1: 38.207438（+-2.937051）
3.3 Changes in flink-benchmarks[8] may also cause a regression,
don't forget to check if flink-benchmarks have changed recently.
3.4 If a regression cannot be reproduced stably which is caused by
the error in results or the issues of physical machines (like
FLINK-18614[9]), this means the regression is not real.
4. Post benchmark results under the Jira ticket, and ping the authors
of the commit(or relevant developers) to investigate the regression if
the regression is real. Otherwise, set the resolution of Jira ticket
as "Not a bug", post the conclusion and close the ticket.
5. If a regression is not fixed within a week of confirming that one
commit is the root cause of the regression, contact the release
manager to revert it (after confirming that reverting the changes
resolves the issue using benchmark-request[10]).
If the above process is considered acceptable, I can draft a version
and put it in the community wiki[10]. @Matthias had proposed to
incorporate performance regression monitoring into the release
management, and make the regression testing be monitored regularly by
release managers or volunteers. I‘m glad to be one of the volunteers.
Hope to hear your advice and opinions!
[1] https://issues.apache.org/jira/browse/FLINK-29883
[2] https://issues.apache.org/jira/browse/FLINK-30015
[3] https://issues.apache.org/jira/browse/FLINK-29886
[4] https://issues.apache.org/jira/browse/FLINK-30181
[5] https://issues.apache.org/jira/browse/FLINK-30623
[6] https://issues.apache.org/jira/browse/FLINK-30624
[7] https://issues.apache.org/jira/browse/FLINK-30625
[8] https://github.com/apache/flink-benchmarks
[9] https://issues.apache.org/jira/browse/FLINK-18614
[10] https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=115511847
Best regards,
Yanfei
Ververica(Alibaba)
Yanfei Lei <fredia...@gmail.com> 于2023年1月12日周四 17:46写道：
>
> Hi all,
>
> Thanks for the reminder.
>
> @Matthias
>
> any updates on the performance tests? ...or more specifically, any updates
> on the script for alerting on performance regressions?
>
>
> I create a PR for FLINK-27571[1] but it's still under review, would you like
> to help take a look?
>
> FLINK-27571 is just for the new benchmarks, for the old existing benchmarks,
> their information is stored
>
> in codespeed's database which can't be updated by URL request, so I also
> logged into the Jenkins master
>
> and modified the codespeed's database, currently "less is better" can be
> displayed normally on the timeline[2].
>
>
> Does it make sense to formalize/document the process?
>
> Certainly, I'm preparing a draft to share my experience of finding commits
> that caused regressions.
>
> Originally, I wanted to wait for FLINK-27571 to be merged before starting a
> discussion, and I will put
>
> a draft of the document later.
>
>
> This slack channel can only provide notice of regression and some experience
> on how to locate regression,
>
> but we also need some people to take action after the regression happens. It
> is mainly a few people who volunteer to do these things,
>
> like FLINK-30015[3] and FLINK-30623[4], many thanks for Martijn's
> contribution.
>
> As for whether to add the responsibilities to the release manager, I think it
> needs to see other people's opinions.
>
> @Martijn
>
> Thanks for creating these tickets. For FLINK-30623 and FLINK-30624[5],
> @Hangxiang and I have located the corresponding commit
>
> and pinged the corresponding submitter. Regression may not be avoided, I
> totally do agree that this work needs to be formalized as soon as possible to
> fix regressions.
>
>
> [1] https://issues.apache.org/jira/browse/FLINK-27571
>
> [2]
> http://codespeed.dak8s.net:8000/timeline/#/?ben=createScheduler.BATCH&extr=on&quarts=on&equid=off&env=2&revs=200&exe=1,3,5,6,8,9
>
> [3] https://issues.apache.org/jira/browse/FLINK-30015
>
> [4] https://issues.apache.org/jira/browse/FLINK-30623
>
> [5] https://issues.apache.org/jira/browse/FLINK-30624
>
>
> Best regards,
>
> Yanfei
>
>
> Martijn Visser <martijnvis...@apache.org> 于2023年1月11日周三 01:11写道：
>>
>> Hi all,
>>
>> Related to Matthias' email, I've checked the notifications in the Slack
>> channel and noticed three major benchmark regressions. In the end, I've
>> decided to create Jira tickets for it [1] [2] [3] but I do agree that this
>> work needs to be formalized as soon as possible to avoid regressions. It
>> would also be great to include a process on how these regressions will be
>> fixed, because I have no idea who to ping/notify that these regressions
>> have occurred.
>>
>> Best regards,
>>
>> Martijn
>>
>> [1] https://issues.apache.org/jira/browse/FLINK-30623
>> [2] https://issues.apache.org/jira/browse/FLINK-30624
>> [3] https://issues.apache.org/jira/browse/FLINK-30625
>>
>> On Tue, Jan 10, 2023 at 1:56 PM Matthias Pohl
>> <matthias.p...@aiven.io.invalid> wrote:
>>
>> > Hi Yanfei,
>> > any updates on the performance tests? ...or more specifically, any updates
>> > on the script for alerting on performance regressions?
>> >
>> > Does it make sense to formalize/document the process? Currently, the
>> > release management doesn't do anything in terms of performance
>> > test monitoring. Therefore, performance regressions are not necessarily
>> > identified actively (in contrast to CI instabilities). Or is this covered
>> > by the PMC? It would be interesting to know whether there's someone to
>> > reach out to who's monitoring the regression tests regularly. Would it make
>> > sense for this person to join the release calls?
>> >
>> > Or shall we work on formalizing/documenting the process and integrating
>> > this responsibility into what the release manager(s) are in charge of? My
>> > concern with that approach is that contributors might be less willing to
>> > volunteer in the release management if we collect everything in one role.
>> > Alternatively, we could split the release manager role up into sub-roles
>> > that contributors can volunteer for in a release (e.g. CI monitoring,
>> > performance test monitoring, Jira maintenance, ... just coming up with
>> > random tasks here).
>> >
>> > Alternatively, we could leave everything as is and just respond if there's
>> > some complaint. I'm curious about your (and other's) opinions.
>> >
>> > Matthias
>> >
>> > On Tue, Nov 29, 2022 at 2:13 PM Yanfei Lei <fredia...@gmail.com> wrote:
>> >
>> > > Hi Martijn,
>> > >
>> > > Thanks for bringing this up.
>> > >
>> > > In the past two months, this channel has helped us find many benchmark
>> > fail
>> > > issues, like FLINK-29883
>> > > <https://issues.apache.org/jira/browse/FLINK-29883>[1],
>> > > FLINK-29886 <https://issues.apache.org/jira/browse/FLINK-29886>[2],
>> > > FLINK-30015 <https://issues.apache.org/jira/browse/FLINK-30015>[3] and
>> > > FLINK-30181 <https://issues.apache.org/jira/browse/FLINK-30181>[4]. I
>> > also
>> > > have tried investigating several of the frequently reported regressions
>> > and
>> > > replied under the notification in slack channel(copy them here):
>> > >
>> > > 1. serializerHeavyString
>> > > <
>> > >
>> > http://codespeed.dak8s.net:8000/timeline/#/?exe=6&ben=serializerHeavyString&extr=on&quarts=on&equid=off&env=2&revs=200
>> > > >:
>> > > It is unstable for a long time, see [5]
>> > > https://issues.apache.org/jira/browse/FLINK-27165 for possible
>> > reasons.
>> > > 2. Regressions are detected by a simple script which may have false
>> > > positives and false negatives, especially for benchmarks with small
>> > > absolute values, small value changes cause large percentage changes.
>> > see
>> > > [6] for details.
>> > >
>> > > Maybe slidingWindow
>> > > <
>> > >
>> > http://codespeed.dak8s.net:8000/timeline/#/?exe=6&ben=slidingWindow&extr=on&quarts=on&equid=off&env=2&revs=200
>> > > >(value~=600),
>> > > stateBackends.ROCKS
>> > > <
>> > >
>> > http://codespeed.dak8s.net:8000/timeline/#/?exe=6&ben=stateBackends.ROCKS&extr=on&quarts=on&equid=off&env=2&revs=200
>> > > >
>> > > (value~=260) and serializerHeavyString
>> > > <
>> > >
>> > http://codespeed.dak8s.net:8000/timeline/#/?exe=6&ben=serializerHeavyString&extr=on&quarts=on&equid=off&env=2&revs=200
>> > > >(value~=170)
>> > > are
>> > > not true regressions.
>> > >
>> > > 1. For deployAllTasks.STREAMING
>> > > <
>> > >
>> > http://codespeed.dak8s.net:8000/timeline/#/?exe=8&ben=deployAllTasks.STREAMING&extr=on&quarts=on&equid=off&env=2&revs=200
>> > > >,
>> > > this benchmark result is how much time it takes to deploy job, the
>> > less
>> > > value the better performance, see [7] for details. FLINK-27571
>> > > <https://issues.apache.org/jira/browse/FLINK-27571>[8] would fix this
>> > > problem.
>> > >
>> > >
>> > > As mentioned before, regressions are detected by a simple script that is
>> > > less stable, FLINK-29825 <
>> > > https://issues.apache.org/jira/browse/FLINK-29825>[9]
>> > > is created to improve the benchmark's stability. I planned to invite more
>> > > volunteers to monitor it after the checking of regression became more
>> > > stable, but I've been stuck with something else lately, sorry for the
>> > late
>> > > response. Any suggestions on handling benchmark regressions/fails are
>> > > welcome.
>> > >
>> > > [1] https://issues.apache.org/jira/browse/FLINK-29883
>> > >
>> > > [2] https://issues.apache.org/jira/browse/FLINK-29886
>> > >
>> > > [3] https://issues.apache.org/jira/browse/FLINK-30015
>> > >
>> > > [4] https://issues.apache.org/jira/browse/FLINK-30181
>> > >
>> > > [5] https://issues.apache.org/jira/browse/FLINK-27165
>> > >
>> > > [6]
>> > >
>> > >
>> > https://github.com/apache/flink-benchmarks/blob/master/regression_report.py#L132-L136
>> > >
>> > > [7]
>> > >
>> > >
>> > https://github.com/apache/flink-benchmarks/blob/master/src/main/java/org/apache/flink/scheduler/benchmark/deploying/DeployingTasksInStreamingJobBenchmarkExecutor.java#L58
>> > >
>> > > [8] https://issues.apache.org/jira/browse/FLINK-27571
>> > >
>> > > [9] https://issues.apache.org/jira/browse/FLINK-29825
>> > >
>> > >
>> > > Best,
>> > >
>> > > Yanfei
>> > >
>> > > Martijn Visser <martijnvis...@apache.org> 于2022年11月29日周二 15:54写道：
>> > >
>> > > > Hi,
>> > > >
>> > > > Is there any update to be expected on the benchmark? I see results of
>> > the
>> > > > benchmark being posted to Slack, but it appears that it's not being
>> > > > monitored and no follow-up actions are being taken. I think it's
>> > > currently
>> > > > lacking a process on how to interpret the results and what action
>> > should
>> > > > be taken and by whom.
>> > > >
>> > > > Best regards,
>> > > >
>> > > > Martijn
>> > > >
>> > > > On Thu, Nov 3, 2022 at 12:22 PM Jing Ge <j...@ververica.com> wrote:
>> > > >
>> > > > > Thanks yanfei for driving this!
>> > > > >
>> > > > > Looking forward to further discussion w.r.t. the workflow.
>> > > > >
>> > > > > Best regards,
>> > > > > Jing
>> > > > >
>> > > > > On Mon, Oct 31, 2022 at 6:04 PM Mason Chen <mas.chen6...@gmail.com>
>> > > > wrote:
>> > > > >
>> > > > > > +1, thanks for driving this!
>> > > > > >
>> > > > > > On a side note, can we also ensure that a performance summary
>> > report
>> > > > for
>> > > > > > Flink major version upgrades is in release notes, once this
>> > > > > infrastructure
>> > > > > > becomes mature? From the user perspective, it would be nice to know
>> > > > what
>> > > > > > the expected (or unexpected) regressions in a major version upgrade
>> > > > are.
>> > > > > > I've seen the community do something like this before (e.g. the
>> > major
>> > > > > > rocksdb version bump in 1.14?) and it was quite valuable to know
>> > that
>> > > > > > upfront!
>> > > > > >
>> > > > > > Best,
>> > > > > > Mason
>> > > > > >
>> > > > > > On Fri, Oct 28, 2022 at 1:46 AM weijie guo <
>> > > guoweijieres...@gmail.com>
>> > > > > > wrote:
>> > > > > >
>> > > > > > > Thanks Yanfei for driving this.
>> > > > > > >
>> > > > > > > It allows us to easily find the problem of performance
>> > regression.
>> > > > > > > Especially recently, I have made some improvements to the
>> > > scheduling
>> > > > > > > related parts, your work is very important to ensure that these
>> > > > changes
>> > > > > > do
>> > > > > > > not cause some unexpected problems.
>> > > > > > >
>> > > > > > > Best regards,
>> > > > > > >
>> > > > > > > Weijie
>> > > > > > >
>> > > > > > >
>> > > > > > > Congxian Qiu <qcx978132...@gmail.com> 于2022年10月28日周五 16:03写道：
>> > > > > > >
>> > > > > > > > Thanks for driving this and making the performance monitoring
>> > > > public,
>> > > > > > > this
>> > > > > > > > can make us know and resolve the performance problem quickly.
>> > > > > > > >
>> > > > > > > > Looking forward to the workflow and detailed descriptions fo
>> > > > > > > > flink-dev-benchmarks.
>> > > > > > > >
>> > > > > > > > Best,
>> > > > > > > > Congxian
>> > > > > > > >
>> > > > > > > >
>> > > > > > > > Yun Tang <myas...@live.com> 于2022年10月27日周四 12:41写道：
>> > > > > > > >
>> > > > > > > > > Thanks, Yanfei for driving this to monitor the performance in
>> > > the
>> > > > > > > Apache
>> > > > > > > > > Flink Slack Channel.
>> > > > > > > > >
>> > > > > > > > > Look forward to the workflow and detailed descriptions of
>> > > > > > > > > flink-dev-benchmarks.
>> > > > > > > > >
>> > > > > > > > > Best
>> > > > > > > > > Yun Tang
>> > > > > > > > > ________________________________
>> > > > > > > > > From: Hangxiang Yu <master...@gmail.com>
>> > > > > > > > > Sent: Thursday, October 27, 2022 10:59
>> > > > > > > > > To: dev@flink.apache.org <dev@flink.apache.org>
>> > > > > > > > > Subject: Re: [ANNOUNCE] Performance Daily Monitoring Moved
>> > from
>> > > > > > > Ververica
>> > > > > > > > > to Apache Flink Slack Channel
>> > > > > > > > >
>> > > > > > > > > Hi, Yanfei.
>> > > > > > > > > Thanks for driving this.
>> > > > > > > > > It could help us to detect and resolve the regression problem
>> > > > > quickly
>> > > > > > > and
>> > > > > > > > > officially.
>> > > > > > > > > I'd like to join as a maintainer.
>> > > > > > > > > Looking forward to the workflow.
>> > > > > > > > >
>> > > > > > > > > On Wed, Oct 26, 2022 at 5:18 PM Yuan Mei <
>> > > yuanmei.w...@gmail.com
>> > > > >
>> > > > > > > wrote:
>> > > > > > > > >
>> > > > > > > > > > Thanks, Yanfei, to drive this and make the performance
>> > > > monitoring
>> > > > > > > > > publicly
>> > > > > > > > > > available.
>> > > > > > > > > >
>> > > > > > > > > > Looking forward to seeing the workflow, and more details as
>> > > > > Martijn
>> > > > > > > > > > mentioned.
>> > > > > > > > > >
>> > > > > > > > > > Best
>> > > > > > > > > > Yuan
>> > > > > > > > > >
>> > > > > > > > > > On Wed, Oct 26, 2022 at 2:59 PM Martijn Visser <
>> > > > > > > > martijnvis...@apache.org
>> > > > > > > > > >
>> > > > > > > > > > wrote:
>> > > > > > > > > >
>> > > > > > > > > > > Hi Yanfei Lei,
>> > > > > > > > > > >
>> > > > > > > > > > > Thanks for setting this up! It would be interesting to
>> > also
>> > > > > know
>> > > > > > > > which
>> > > > > > > > > > > aspects of Flink are monitored for "performance". I'm
>> > > > assuming
>> > > > > > > there
>> > > > > > > > > are
>> > > > > > > > > > > specific pieces of functionality that are performance
>> > > tested,
>> > > > > but
>> > > > > > > it
>> > > > > > > > > > would
>> > > > > > > > > > > be great if this would be written down somewhere (next
>> > to a
>> > > > > > > procedure
>> > > > > > > > > how
>> > > > > > > > > > > to detect a regression and what should be next steps).
>> > > > > > > > > > >
>> > > > > > > > > > > Best regards,
>> > > > > > > > > > >
>> > > > > > > > > > > Martijn
>> > > > > > > > > > >
>> > > > > > > > > > > On Wed, Oct 26, 2022 at 8:21 AM Zakelly Lan <
>> > > > > > zakelly....@gmail.com
>> > > > > > > >
>> > > > > > > > > > wrote:
>> > > > > > > > > > >
>> > > > > > > > > > > > Hi yanfei,
>> > > > > > > > > > > >
>> > > > > > > > > > > > Thanks for driving this! It's a great help.
>> > > > > > > > > > > >
>> > > > > > > > > > > > I would like to join as a maintainer.
>> > > > > > > > > > > >
>> > > > > > > > > > > > Best,
>> > > > > > > > > > > > Zakelly
>> > > > > > > > > > > >
>> > > > > > > > > > > > On Wed, Oct 26, 2022 at 11:32 AM yanfei lei <
>> > > > > > fredia...@gmail.com
>> > > > > > > >
>> > > > > > > > > > wrote:
>> > > > > > > > > > > > >
>> > > > > > > > > > > > > Hi everyone,
>> > > > > > > > > > > > >
>> > > > > > > > > > > > > As discussed earlier, we plan to create a benchmark
>> > > > channel
>> > > > > > in
>> > > > > > > > > Apache
>> > > > > > > > > > > > Flink
>> > > > > > > > > > > > > slack[1], but the plan was shelved for a while[2].
>> > So I
>> > > > > went
>> > > > > > on
>> > > > > > > > > with
>> > > > > > > > > > > this
>> > > > > > > > > > > > > work, and created the #flink-dev-benchmarks channel
>> > for
>> > > > > > > > performance
>> > > > > > > > > > > > > regression notifications.
>> > > > > > > > > > > > >
>> > > > > > > > > > > > > We have a regression report script[3] that runs
>> > daily,
>> > > > and
>> > > > > a
>> > > > > > > > > > > notification
>> > > > > > > > > > > > > would be sent to the slack channel when the last few
>> > > > > > benchmark
>> > > > > > > > > > results
>> > > > > > > > > > > > are
>> > > > > > > > > > > > > significantly worse than the baseline.
>> > > > > > > > > > > > > Note, regressions are detected by a simple script
>> > which
>> > > > may
>> > > > > > > have
>> > > > > > > > > > false
>> > > > > > > > > > > > > positives and false negatives. And all benchmarks are
>> > > > > > executed
>> > > > > > > on
>> > > > > > > > > one
>> > > > > > > > > > > > > physical machine[4] which is provided by
>> > > > > > Ververica(Alibaba)[5],
>> > > > > > > > it
>> > > > > > > > > > > might
>> > > > > > > > > > > > > happen that hardware issues affect performance, like
>> > > > > > > > "[FLINK-18614
>> > > > > > > > > > > > > <https://issues.apache.org/jira/browse/FLINK-18614>]
>> > > > > > > Performance
>> > > > > > > > > > > > regression
>> > > > > > > > > > > > > 2020.07.13"[6].
>> > > > > > > > > > > > >
>> > > > > > > > > > > > > After the migration, we need a procedure to watch
>> > over
>> > > > the
>> > > > > > > entire
>> > > > > > > > > > > > > performance of Flink code together. For example, if a
>> > > > > > > regression
>> > > > > > > > > > > > > occurs, investigating the cause and resolving the
>> > > problem
>> > > > > are
>> > > > > > > > > needed.
>> > > > > > > > > > > In
>> > > > > > > > > > > > > the past, this procedure is maintained internally
>> > > within
>> > > > > > > > Ververica,
>> > > > > > > > > > but
>> > > > > > > > > > > > we
>> > > > > > > > > > > > > think making the procedure public would benefit all.
>> > I
>> > > > > > > volunteer
>> > > > > > > > to
>> > > > > > > > > > > serve
>> > > > > > > > > > > > > as one of the initial maintainers, and would be glad
>> > if
>> > > > > more
>> > > > > > > > > > > contributors
>> > > > > > > > > > > > > can join me. I'd also prepare some guidelines to help
>> > > > > others
>> > > > > > > get
>> > > > > > > > > > > familiar
>> > > > > > > > > > > > > with the workflow. I will start a new thread to
>> > discuss
>> > > > the
>> > > > > > > > > workflow
>> > > > > > > > > > > > soon.
>> > > > > > > > > > > > >
>> > > > > > > > > > > > >
>> > > > > > > > > > > > > [1]
>> > > > > > > > >
>> > > https://www.mail-archive.com/dev@flink.apache.org/msg58666.html
>> > > > > > > > > > > > > [2]
>> > https://issues.apache.org/jira/browse/FLINK-28468
>> > > > > > > > > > > > > [3]
>> > > > > > > > > > > > >
>> > > > > > > > > > > >
>> > > > > > > > > > >
>> > > > > > > > > >
>> > > > > > > > >
>> > > > > > > >
>> > > > > > >
>> > > > > >
>> > > > >
>> > > >
>> > >
>> > https://github.com/apache/flink-benchmarks/blob/master/regression_report.py
>> > > > > > > > > > > > > [4] http://codespeed.dak8s.net:8080
>> > > > > > > > > > > > > [5]
>> > > > > > > > >
>> > > https://lists.apache.org/thread/jzljp4233799vwwqnr0vc9wgqs0xj1ro
>> > > > > > > > > > > > >
>> > > > > > > > > > > > > [6]
>> > https://issues.apache.org/jira/browse/FLINK-18614
>> > > > > > > > > > > >
>> > > > > > > > > > >
>> > > > > > > > > >
>> > > > > > > > >
>> > > > > > > > >
>> > > > > > > > > --
>> > > > > > > > > Best,
>> > > > > > > > > Hangxiang.
>> > > > > > > > >
>> > > > > > > >
>> > > > > > >
>> > > > > >
>> > > > >
>> > > >
>> > >
>> >
>
>
Yanfei Lei <fredia...@gmail.com> 于2023年1月12日周四 17:46写道：
Yanfei Lei <fredia...@gmail.com> 于2023年1月12日周四 17:46写道：
>
> Hi all,
>
> Thanks for the reminder.
>
> @Matthias
>
> any updates on the performance tests? ...or more specifically, any updates
> on the script for alerting on performance regressions?
>
>
> I create a PR for FLINK-27571[1] but it's still under review, would you like
> to help take a look?
>
> FLINK-27571 is just for the new benchmarks, for the old existing benchmarks,
> their information is stored
>
> in codespeed's database which can't be updated by URL request, so I also
> logged into the Jenkins master
>
> and modified the codespeed's database, currently "less is better" can be
> displayed normally on the timeline[2].
>
>
> Does it make sense to formalize/document the process?
>
> Certainly, I'm preparing a draft to share my experience of finding commits
> that caused regressions.
>
> Originally, I wanted to wait for FLINK-27571 to be merged before starting a
> discussion, and I will put
>
> a draft of the document later.
>
>
> This slack channel can only provide notice of regression and some experience
> on how to locate regression,
>
> but we also need some people to take action after the regression happens. It
> is mainly a few people who volunteer to do these things,
>
> like FLINK-30015[3] and FLINK-30623[4], many thanks for Martijn's
> contribution.
>
> As for whether to add the responsibilities to the release manager, I think it
> needs to see other people's opinions.
>
> @Martijn
>
> Thanks for creating these tickets. For FLINK-30623 and FLINK-30624[5],
> @Hangxiang and I have located the corresponding commit
>
> and pinged the corresponding submitter. Regression may not be avoided, I
> totally do agree that this work needs to be formalized as soon as possible to
> fix regressions.
>
>
> [1] https://issues.apache.org/jira/browse/FLINK-27571
>
> [2]
> http://codespeed.dak8s.net:8000/timeline/#/?ben=createScheduler.BATCH&extr=on&quarts=on&equid=off&env=2&revs=200&exe=1,3,5,6,8,9
>
> [3] https://issues.apache.org/jira/browse/FLINK-30015
>
> [4] https://issues.apache.org/jira/browse/FLINK-30623
>
> [5] https://issues.apache.org/jira/browse/FLINK-30624
>
>
> Best regards,
>
> Yanfei
>
>
> Martijn Visser <martijnvis...@apache.org> 于2023年1月11日周三 01:11写道：
>>
>> Hi all,
>>
>> Related to Matthias' email, I've checked the notifications in the Slack
>> channel and noticed three major benchmark regressions. In the end, I've
>> decided to create Jira tickets for it [1] [2] [3] but I do agree that this
>> work needs to be formalized as soon as possible to avoid regressions. It
>> would also be great to include a process on how these regressions will be
>> fixed, because I have no idea who to ping/notify that these regressions
>> have occurred.
>>
>> Best regards,
>>
>> Martijn
>>
>> [1] https://issues.apache.org/jira/browse/FLINK-30623
>> [2] https://issues.apache.org/jira/browse/FLINK-30624
>> [3] https://issues.apache.org/jira/browse/FLINK-30625
>>
>> On Tue, Jan 10, 2023 at 1:56 PM Matthias Pohl
>> <matthias.p...@aiven.io.invalid> wrote:
>>
>> > Hi Yanfei,
>> > any updates on the performance tests? ...or more specifically, any updates
>> > on the script for alerting on performance regressions?
>> >
>> > Does it make sense to formalize/document the process? Currently, the
>> > release management doesn't do anything in terms of performance
>> > test monitoring. Therefore, performance regressions are not necessarily
>> > identified actively (in contrast to CI instabilities). Or is this covered
>> > by the PMC? It would be interesting to know whether there's someone to
>> > reach out to who's monitoring the regression tests regularly. Would it make
>> > sense for this person to join the release calls?
>> >
>> > Or shall we work on formalizing/documenting the process and integrating
>> > this responsibility into what the release manager(s) are in charge of? My
>> > concern with that approach is that contributors might be less willing to
>> > volunteer in the release management if we collect everything in one role.
>> > Alternatively, we could split the release manager role up into sub-roles
>> > that contributors can volunteer for in a release (e.g. CI monitoring,
>> > performance test monitoring, Jira maintenance, ... just coming up with
>> > random tasks here).
>> >
>> > Alternatively, we could leave everything as is and just respond if there's
>> > some complaint. I'm curious about your (and other's) opinions.
>> >
>> > Matthias
>> >
>> > On Tue, Nov 29, 2022 at 2:13 PM Yanfei Lei <fredia...@gmail.com> wrote:
>> >
>> > > Hi Martijn,
>> > >
>> > > Thanks for bringing this up.
>> > >
>> > > In the past two months, this channel has helped us find many benchmark
>> > fail
>> > > issues, like FLINK-29883
>> > > <https://issues.apache.org/jira/browse/FLINK-29883>[1],
>> > > FLINK-29886 <https://issues.apache.org/jira/browse/FLINK-29886>[2],
>> > > FLINK-30015 <https://issues.apache.org/jira/browse/FLINK-30015>[3] and
>> > > FLINK-30181 <https://issues.apache.org/jira/browse/FLINK-30181>[4]. I
>> > also
>> > > have tried investigating several of the frequently reported regressions
>> > and
>> > > replied under the notification in slack channel(copy them here):
>> > >
>> > > 1. serializerHeavyString
>> > > <
>> > >
>> > http://codespeed.dak8s.net:8000/timeline/#/?exe=6&ben=serializerHeavyString&extr=on&quarts=on&equid=off&env=2&revs=200
>> > > >:
>> > > It is unstable for a long time, see [5]
>> > > https://issues.apache.org/jira/browse/FLINK-27165 for possible
>> > reasons.
>> > > 2. Regressions are detected by a simple script which may have false
>> > > positives and false negatives, especially for benchmarks with small
>> > > absolute values, small value changes cause large percentage changes.
>> > see
>> > > [6] for details.
>> > >
>> > > Maybe slidingWindow
>> > > <
>> > >
>> > http://codespeed.dak8s.net:8000/timeline/#/?exe=6&ben=slidingWindow&extr=on&quarts=on&equid=off&env=2&revs=200
>> > > >(value~=600),
>> > > stateBackends.ROCKS
>> > > <
>> > >
>> > http://codespeed.dak8s.net:8000/timeline/#/?exe=6&ben=stateBackends.ROCKS&extr=on&quarts=on&equid=off&env=2&revs=200
>> > > >
>> > > (value~=260) and serializerHeavyString
>> > > <
>> > >
>> > http://codespeed.dak8s.net:8000/timeline/#/?exe=6&ben=serializerHeavyString&extr=on&quarts=on&equid=off&env=2&revs=200
>> > > >(value~=170)
>> > > are
>> > > not true regressions.
>> > >
>> > > 1. For deployAllTasks.STREAMING
>> > > <
>> > >
>> > http://codespeed.dak8s.net:8000/timeline/#/?exe=8&ben=deployAllTasks.STREAMING&extr=on&quarts=on&equid=off&env=2&revs=200
>> > > >,
>> > > this benchmark result is how much time it takes to deploy job, the
>> > less
>> > > value the better performance, see [7] for details. FLINK-27571
>> > > <https://issues.apache.org/jira/browse/FLINK-27571>[8] would fix this
>> > > problem.
>> > >
>> > >
>> > > As mentioned before, regressions are detected by a simple script that is
>> > > less stable, FLINK-29825 <
>> > > https://issues.apache.org/jira/browse/FLINK-29825>[9]
>> > > is created to improve the benchmark's stability. I planned to invite more
>> > > volunteers to monitor it after the checking of regression became more
>> > > stable, but I've been stuck with something else lately, sorry for the
>> > late
>> > > response. Any suggestions on handling benchmark regressions/fails are
>> > > welcome.
>> > >
>> > > [1] https://issues.apache.org/jira/browse/FLINK-29883
>> > >
>> > > [2] https://issues.apache.org/jira/browse/FLINK-29886
>> > >
>> > > [3] https://issues.apache.org/jira/browse/FLINK-30015
>> > >
>> > > [4] https://issues.apache.org/jira/browse/FLINK-30181
>> > >
>> > > [5] https://issues.apache.org/jira/browse/FLINK-27165
>> > >
>> > > [6]
>> > >
>> > >
>> > https://github.com/apache/flink-benchmarks/blob/master/regression_report.py#L132-L136
>> > >
>> > > [7]
>> > >
>> > >
>> > https://github.com/apache/flink-benchmarks/blob/master/src/main/java/org/apache/flink/scheduler/benchmark/deploying/DeployingTasksInStreamingJobBenchmarkExecutor.java#L58
>> > >
>> > > [8] https://issues.apache.org/jira/browse/FLINK-27571
>> > >
>> > > [9] https://issues.apache.org/jira/browse/FLINK-29825
>> > >
>> > >
>> > > Best,
>> > >
>> > > Yanfei
>> > >
>> > > Martijn Visser <martijnvis...@apache.org> 于2022年11月29日周二 15:54写道：
>> > >
>> > > > Hi,
>> > > >
>> > > > Is there any update to be expected on the benchmark? I see results of
>> > the
>> > > > benchmark being posted to Slack, but it appears that it's not being
>> > > > monitored and no follow-up actions are being taken. I think it's
>> > > currently
>> > > > lacking a process on how to interpret the results and what action
>> > should
>> > > > be taken and by whom.
>> > > >
>> > > > Best regards,
>> > > >
>> > > > Martijn
>> > > >
>> > > > On Thu, Nov 3, 2022 at 12:22 PM Jing Ge <j...@ververica.com> wrote:
>> > > >
>> > > > > Thanks yanfei for driving this!
>> > > > >
>> > > > > Looking forward to further discussion w.r.t. the workflow.
>> > > > >
>> > > > > Best regards,
>> > > > > Jing
>> > > > >
>> > > > > On Mon, Oct 31, 2022 at 6:04 PM Mason Chen <mas.chen6...@gmail.com>
>> > > > wrote:
>> > > > >
>> > > > > > +1, thanks for driving this!
>> > > > > >
>> > > > > > On a side note, can we also ensure that a performance summary
>> > report
>> > > > for
>> > > > > > Flink major version upgrades is in release notes, once this
>> > > > > infrastructure
>> > > > > > becomes mature? From the user perspective, it would be nice to know
>> > > > what
>> > > > > > the expected (or unexpected) regressions in a major version upgrade
>> > > > are.
>> > > > > > I've seen the community do something like this before (e.g. the
>> > major
>> > > > > > rocksdb version bump in 1.14?) and it was quite valuable to know
>> > that
>> > > > > > upfront!
>> > > > > >
>> > > > > > Best,
>> > > > > > Mason
>> > > > > >
>> > > > > > On Fri, Oct 28, 2022 at 1:46 AM weijie guo <
>> > > guoweijieres...@gmail.com>
>> > > > > > wrote:
>> > > > > >
>> > > > > > > Thanks Yanfei for driving this.
>> > > > > > >
>> > > > > > > It allows us to easily find the problem of performance
>> > regression.
>> > > > > > > Especially recently, I have made some improvements to the
>> > > scheduling
>> > > > > > > related parts, your work is very important to ensure that these
>> > > > changes
>> > > > > > do
>> > > > > > > not cause some unexpected problems.
>> > > > > > >
>> > > > > > > Best regards,
>> > > > > > >
>> > > > > > > Weijie
>> > > > > > >
>> > > > > > >
>> > > > > > > Congxian Qiu <qcx978132...@gmail.com> 于2022年10月28日周五 16:03写道：
>> > > > > > >
>> > > > > > > > Thanks for driving this and making the performance monitoring
>> > > > public,
>> > > > > > > this
>> > > > > > > > can make us know and resolve the performance problem quickly.
>> > > > > > > >
>> > > > > > > > Looking forward to the workflow and detailed descriptions fo
>> > > > > > > > flink-dev-benchmarks.
>> > > > > > > >
>> > > > > > > > Best,
>> > > > > > > > Congxian
>> > > > > > > >
>> > > > > > > >
>> > > > > > > > Yun Tang <myas...@live.com> 于2022年10月27日周四 12:41写道：
>> > > > > > > >
>> > > > > > > > > Thanks, Yanfei for driving this to monitor the performance in
>> > > the
>> > > > > > > Apache
>> > > > > > > > > Flink Slack Channel.
>> > > > > > > > >
>> > > > > > > > > Look forward to the workflow and detailed descriptions of
>> > > > > > > > > flink-dev-benchmarks.
>> > > > > > > > >
>> > > > > > > > > Best
>> > > > > > > > > Yun Tang
>> > > > > > > > > ________________________________
>> > > > > > > > > From: Hangxiang Yu <master...@gmail.com>
>> > > > > > > > > Sent: Thursday, October 27, 2022 10:59
>> > > > > > > > > To: dev@flink.apache.org <dev@flink.apache.org>
>> > > > > > > > > Subject: Re: [ANNOUNCE] Performance Daily Monitoring Moved
>> > from
>> > > > > > > Ververica
>> > > > > > > > > to Apache Flink Slack Channel
>> > > > > > > > >
>> > > > > > > > > Hi, Yanfei.
>> > > > > > > > > Thanks for driving this.
>> > > > > > > > > It could help us to detect and resolve the regression problem
>> > > > > quickly
>> > > > > > > and
>> > > > > > > > > officially.
>> > > > > > > > > I'd like to join as a maintainer.
>> > > > > > > > > Looking forward to the workflow.
>> > > > > > > > >
>> > > > > > > > > On Wed, Oct 26, 2022 at 5:18 PM Yuan Mei <
>> > > yuanmei.w...@gmail.com
>> > > > >
>> > > > > > > wrote:
>> > > > > > > > >
>> > > > > > > > > > Thanks, Yanfei, to drive this and make the performance
>> > > > monitoring
>> > > > > > > > > publicly
>> > > > > > > > > > available.
>> > > > > > > > > >
>> > > > > > > > > > Looking forward to seeing the workflow, and more details as
>> > > > > Martijn
>> > > > > > > > > > mentioned.
>> > > > > > > > > >
>> > > > > > > > > > Best
>> > > > > > > > > > Yuan
>> > > > > > > > > >
>> > > > > > > > > > On Wed, Oct 26, 2022 at 2:59 PM Martijn Visser <
>> > > > > > > > martijnvis...@apache.org
>> > > > > > > > > >
>> > > > > > > > > > wrote:
>> > > > > > > > > >
>> > > > > > > > > > > Hi Yanfei Lei,
>> > > > > > > > > > >
>> > > > > > > > > > > Thanks for setting this up! It would be interesting to
>> > also
>> > > > > know
>> > > > > > > > which
>> > > > > > > > > > > aspects of Flink are monitored for "performance". I'm
>> > > > assuming
>> > > > > > > there
>> > > > > > > > > are
>> > > > > > > > > > > specific pieces of functionality that are performance
>> > > tested,
>> > > > > but
>> > > > > > > it
>> > > > > > > > > > would
>> > > > > > > > > > > be great if this would be written down somewhere (next
>> > to a
>> > > > > > > procedure
>> > > > > > > > > how
>> > > > > > > > > > > to detect a regression and what should be next steps).
>> > > > > > > > > > >
>> > > > > > > > > > > Best regards,
>> > > > > > > > > > >
>> > > > > > > > > > > Martijn
>> > > > > > > > > > >
>> > > > > > > > > > > On Wed, Oct 26, 2022 at 8:21 AM Zakelly Lan <
>> > > > > > zakelly....@gmail.com
>> > > > > > > >
>> > > > > > > > > > wrote:
>> > > > > > > > > > >
>> > > > > > > > > > > > Hi yanfei,
>> > > > > > > > > > > >
>> > > > > > > > > > > > Thanks for driving this! It's a great help.
>> > > > > > > > > > > >
>> > > > > > > > > > > > I would like to join as a maintainer.
>> > > > > > > > > > > >
>> > > > > > > > > > > > Best,
>> > > > > > > > > > > > Zakelly
>> > > > > > > > > > > >
>> > > > > > > > > > > > On Wed, Oct 26, 2022 at 11:32 AM yanfei lei <
>> > > > > > fredia...@gmail.com
>> > > > > > > >
>> > > > > > > > > > wrote:
>> > > > > > > > > > > > >
>> > > > > > > > > > > > > Hi everyone,
>> > > > > > > > > > > > >
>> > > > > > > > > > > > > As discussed earlier, we plan to create a benchmark
>> > > > channel
>> > > > > > in
>> > > > > > > > > Apache
>> > > > > > > > > > > > Flink
>> > > > > > > > > > > > > slack[1], but the plan was shelved for a while[2].
>> > So I
>> > > > > went
>> > > > > > on
>> > > > > > > > > with
>> > > > > > > > > > > this
>> > > > > > > > > > > > > work, and created the #flink-dev-benchmarks channel
>> > for
>> > > > > > > > performance
>> > > > > > > > > > > > > regression notifications.
>> > > > > > > > > > > > >
>> > > > > > > > > > > > > We have a regression report script[3] that runs
>> > daily,
>> > > > and
>> > > > > a
>> > > > > > > > > > > notification
>> > > > > > > > > > > > > would be sent to the slack channel when the last few
>> > > > > > benchmark
>> > > > > > > > > > results
>> > > > > > > > > > > > are
>> > > > > > > > > > > > > significantly worse than the baseline.
>> > > > > > > > > > > > > Note, regressions are detected by a simple script
>> > which
>> > > > may
>> > > > > > > have
>> > > > > > > > > > false
>> > > > > > > > > > > > > positives and false negatives. And all benchmarks are
>> > > > > > executed
>> > > > > > > on
>> > > > > > > > > one
>> > > > > > > > > > > > > physical machine[4] which is provided by
>> > > > > > Ververica(Alibaba)[5],
>> > > > > > > > it
>> > > > > > > > > > > might
>> > > > > > > > > > > > > happen that hardware issues affect performance, like
>> > > > > > > > "[FLINK-18614
>> > > > > > > > > > > > > <https://issues.apache.org/jira/browse/FLINK-18614>]
>> > > > > > > Performance
>> > > > > > > > > > > > regression
>> > > > > > > > > > > > > 2020.07.13"[6].
>> > > > > > > > > > > > >
>> > > > > > > > > > > > > After the migration, we need a procedure to watch
>> > over
>> > > > the
>> > > > > > > entire
>> > > > > > > > > > > > > performance of Flink code together. For example, if a
>> > > > > > > regression
>> > > > > > > > > > > > > occurs, investigating the cause and resolving the
>> > > problem
>> > > > > are
>> > > > > > > > > needed.
>> > > > > > > > > > > In
>> > > > > > > > > > > > > the past, this procedure is maintained internally
>> > > within
>> > > > > > > > Ververica,
>> > > > > > > > > > but
>> > > > > > > > > > > > we
>> > > > > > > > > > > > > think making the procedure public would benefit all.
>> > I
>> > > > > > > volunteer
>> > > > > > > > to
>> > > > > > > > > > > serve
>> > > > > > > > > > > > > as one of the initial maintainers, and would be glad
>> > if
>> > > > > more
>> > > > > > > > > > > contributors
>> > > > > > > > > > > > > can join me. I'd also prepare some guidelines to help
>> > > > > others
>> > > > > > > get
>> > > > > > > > > > > familiar
>> > > > > > > > > > > > > with the workflow. I will start a new thread to
>> > discuss
>> > > > the
>> > > > > > > > > workflow
>> > > > > > > > > > > > soon.
>> > > > > > > > > > > > >
>> > > > > > > > > > > > >
>> > > > > > > > > > > > > [1]
>> > > > > > > > >
>> > > https://www.mail-archive.com/dev@flink.apache.org/msg58666.html
>> > > > > > > > > > > > > [2]
>> > https://issues.apache.org/jira/browse/FLINK-28468
>> > > > > > > > > > > > > [3]
>> > > > > > > > > > > > >
>> > > > > > > > > > > >
>> > > > > > > > > > >
>> > > > > > > > > >
>> > > > > > > > >
>> > > > > > > >
>> > > > > > >
>> > > > > >
>> > > > >
>> > > >
>> > >
>> > https://github.com/apache/flink-benchmarks/blob/master/regression_report.py
>> > > > > > > > > > > > > [4] http://codespeed.dak8s.net:8080
>> > > > > > > > > > > > > [5]
>> > > > > > > > >
>> > > https://lists.apache.org/thread/jzljp4233799vwwqnr0vc9wgqs0xj1ro
>> > > > > > > > > > > > >
>> > > > > > > > > > > > > [6]
>> > https://issues.apache.org/jira/browse/FLINK-18614
>> > > > > > > > > > > >
>> > > > > > > > > > >
>> > > > > > > > > >
>> > > > > > > > >
>> > > > > > > > >
>> > > > > > > > > --
>> > > > > > > > > Best,
>> > > > > > > > > Hangxiang.
>> > > > > > > > >
>> > > > > > > >
>> > > > > > >
>> > > > > >
>> > > > >
>> > > >
>> > >
>> >
>
>
--
Best,
Yanfei
