Hi Wouter, The JIRA is https://issues.apache.org/jira/browse/FLINK-23309. `bundle time` is from the perspective of your e2e latency. Regarding the `bundle size`, generally larger value will provide better throughput, but it should not be set too large, which may cause no output to be seen downstream for a long time and the pressure will be too great during checkpoint.
Best, Xingbo Wouter Zorgdrager <zorgdrag...@gmail.com> 于2021年7月8日周四 下午4:32写道: > Hi Xingbo, all, > > That is good to know, thank you. Is there any Jira issue I can track? I'm > curious to follow this progress! Do you have any recommendations with > regard to these two configuration values, to get somewhat reasonable > performance? > > Thanks a lot! > Wouter > > On Thu, 8 Jul 2021 at 10:26, Xingbo Huang <hxbks...@gmail.com> wrote: > >> Hi Wouter, >> >> In fact, our users have encountered the same problem. Whenever the >> `bundle size` or `bundle time` is reached, the data in the buffer needs to >> be sent from the jvm to the pvm, and then waits for the pym to be processed >> and sent back to the jvm to send all the results to the downstream >> operator, which leads to a large delay, especially when it is a small size >> event as small messages are hard to be processed in pipeline. >> >> I have been solving this problem recently and I plan to make this >> optimization to release-1.14. >> >> Best, >> Xingbo >> >> Wouter Zorgdrager <zorgdrag...@gmail.com> 于2021年7月8日周四 下午3:41写道: >> >>> Hi Dian, all, >>> >>> I will come back to the other points asap. However, I’m still confused >>> about this performance. Is this what I can expect in PyFlink in terms of >>> performance? ~ 1000ms latency for single events? I also had a very simple >>> setup where I send 1000 events to Kafka per second and response >>> times/latencies was around 15 seconds for single events. I understand there >>> is some Python/JVM overhead but since Flink is so performant, I would >>> expect much better numbers. In the current situation, PyFlink would just be >>> unusable if you care about latency. Is this something that you expect to be >>> improved in the future? >>> >>> I will verify how this works out for Beam in a remote environment. >>> >>> Thanks again! >>> Wouter >>> >>> >>> On Thu, 8 Jul 2021 at 08:28, Dian Fu <dian0511...@gmail.com> wrote: >>> >>>> Hi Wouter, >>>> >>>> 1) Regarding the performance difference between Beam and PyFlink, I >>>> guess it’s because you are using an in-memory runner when running it >>>> locally in Beam. In that case, the code path is totally differently >>>> compared to running in a remote cluster. >>>> 2) Regarding to `flink run`, I’m surprising that it’s running locally. >>>> Could you submit a java job with similar commands to see how it runs? >>>> 3) Regarding to `flink run-application`, could you share the exception >>>> stack? >>>> >>>> Regards, >>>> Dian >>>> >>>> 2021年7月6日 下午4:58,Wouter Zorgdrager <zorgdrag...@gmail.com> 写道: >>>> >>>> uses >>>> >>>> >>>>