Hi, I think your analysis is very thorough and accurate. However, I don't think that https://issues.apache.org/jira/browse/FLINK-7282 <https://issues.apache.org/jira/browse/FLINK-7282> will solve this problem. We're dealing with "time back-pressure" here and not traditional processing back-pressure. (Not sure if there's a term for this so I coined a new one... 😉). The problem is that some Kafka consumers progress faster in event-time than others which leads to windows being spawned at the window operator that only trigger (and get deleted) once all sources catch up to that point in event time.
Simply slowing down all reading mitigates this somewhat but a proper solution would require sources to wait for sources that are "slower in time". Does that make sense to you? Best, Aljoscha > On 3. Jan 2018, at 15:45, Vishal Santoshi <vishal.santo...@gmail.com> wrote: > > To add and an interim solution to the issue. > > I extended the based on the advise "custom source/adapt an existing source" > and put in a RateLimiter ( guava ) that effectively put a cap on each kafka > consumer ( x times the expected incident rqs ). That solved the issue as in > it stabilized the flow into down stream window operation ( I use > ProcessFunction for sessionizing and why is another discussion ) . > > This leads me to these conclusions and either of them could be correct > > > * The kakfa partitions are on different brokers ( the leaders ) and based on > the how efficient the broker is ( whether it has data in OS Cache or whether > there is a skew in leader distribution and thus more stress on a n of m > brokers) the consumption speed can vary. > * The skew caused by number of consumers to number of flink nodes ( if no. of > partitions % no of flink nodes == 0 there is no skew ) the consumption rate > can vary. > * Some TM nodes may be sluggish. > > Either ways any of the above reasons can cause data to be consumed at > different speeds which could lead to an imbalance and b'coz of the paucity of > fine grained back pressure handling leads to more windows that remain open, > windows that reflect the more aggressive consumption ( and thus more variance > from the current WM) than prudent, causing the pathological case described > above. By regulating the consumption rate ( I put the delimiter in the > extractTimestamp method ) , it effectively caused the more aggressive > consumptions to a fixed upper bound, making the rate of consumption across > consumers effectively similar. > > Either ways it seems imperative that > https://issues.apache.org/jira/browse/FLINK-7282 > <https://issues.apache.org/jira/browse/FLINK-7282> should be finalized at the > earliest. The consequences on a shared flink cluster are too huge IMHO. > > Please tell me if my conclusions are problematic or do not make sense. > > > Regards > > Vishal > > > > > > > On Tue, Jan 2, 2018 at 3:04 PM, Vishal Santoshi <vishal.santo...@gmail.com > <mailto:vishal.santo...@gmail.com>> wrote: > Also note that if I were to start 2 pipelines > > 1. Working off the head of the topic and thus not prone to the pathological > case described above > 2. Doing a replay and thus prone to the pathological case described above > > Than the 2nd pipe will stall the 1st pipeline. This seems to to point to > All channels multiplexed into the same TCP connection stall together, as soon > as one channel has backpressure. > > of the jira issue. This has to be a priority IMHO, in a shared VM where jobs > should have at least some isolation. > > On Tue, Jan 2, 2018 at 2:19 PM, Vishal Santoshi <vishal.santo...@gmail.com > <mailto:vishal.santo...@gmail.com>> wrote: > Thank you. > > On Tue, Jan 2, 2018 at 1:31 PM, Nico Kruber <n...@data-artisans.com > <mailto:n...@data-artisans.com>> wrote: > Hi Vishal, > let me already point you towards the JIRA issue for the credit-based > flow control: https://issues.apache.org/jira/browse/FLINK-7282 > <https://issues.apache.org/jira/browse/FLINK-7282> > > I'll have a look at the rest of this email thread tomorrow... > > > Regards, > Nico > > On 02/01/18 17:52, Vishal Santoshi wrote: > > Could you please point me to any documentation on the "credit-based > > flow control" approach.... > > > > On Tue, Jan 2, 2018 at 10:35 AM, Timo Walther <twal...@apache.org > > <mailto:twal...@apache.org> > > <mailto:twal...@apache.org <mailto:twal...@apache.org>>> wrote: > > > > Hi Vishal, > > > > your assumptions sound reasonable to me. The community is currently > > working on a more fine-grained back pressuring with credit-based > > flow control. It is on the roamap for 1.5 [1]/[2]. I will loop in > > Nico that might tell you more about the details. Until then I guess > > you have to implement a custom source/adapt an existing source to > > let the data flow in more realistic. > > > > Regards, > > Timo > > > > [1] > > > > http://flink.apache.org/news/2017/11/22/release-1.4-and-1.5-timeline.html > > <http://flink.apache.org/news/2017/11/22/release-1.4-and-1.5-timeline.html> > > > > <http://flink.apache.org/news/2017/11/22/release-1.4-and-1.5-timeline.html > > <http://flink.apache.org/news/2017/11/22/release-1.4-and-1.5-timeline.html>> > > [2] https://www.youtube.com/watch?v=scStdhz9FHc > > <https://www.youtube.com/watch?v=scStdhz9FHc> > > <https://www.youtube.com/watch?v=scStdhz9FHc > > <https://www.youtube.com/watch?v=scStdhz9FHc>> > > > > > > Am 1/2/18 um 4:02 PM schrieb Vishal Santoshi: > > > > I did a simulation on session windows ( in 2 modes ) and let it > > rip for about 12 hours > > > > 1. Replay where a kafka topic with retention of 7 days was the > > source ( earliest ) > > 2. Start the pipe with kafka source ( latest ) > > > > I saw results that differed dramatically. > > > > On replay the pipeline stalled after good ramp up while in the > > second case the pipeline hummed on without issues. For the same > > time period the data consumed is significantly more in the > > second case with the WM progression stalled in the first case > > with no hint of resolution ( the incoming data on source topic > > far outstrips the WM progression ) I think I know the reasons > > and this is my hypothesis. > > > > In replay mode the number of windows open do not have an upper > > bound. While buffer exhaustion ( and data in flight with > > watermark ) is the reason for throttle, it does not really > > limit the open windows and in fact creates windows that reflect > > futuristic data ( future is relative to the current WM ) . So if > > partition x has data for watermark time t(x) and partition y for > > watermark time t(y) and t(x) << t(y) where the overall watermark > > is t(x) nothing significantly throttles consumption from the y > > partition ( in fact for x too ) , the bounded buffer based > > approach does not give minute control AFAIK as one would hope > > and that implies there are far more open windows than the system > > can handle and that leads to the pathological case where the > > buffers fill up ( I believe that happens way late ) and > > throttling occurs but the WM does not proceed and windows that > > could ease the glut the throttling cannot proceed..... In the > > replay mode the amount of data implies that the Fetchers keep > > pulling data at the maximum consumption allowed by the open > > ended buffer approach. > > > > My question thus is, is there any way to have a finer control of > > back pressure, where in the consumption from a source is > > throttled preemptively ( by for example decreasing the buffers > > associated for a pipe or the size allocated ) or sleeps in the > > Fetcher code that can help aligning the performance to have real > > time consumption characteristics > > > > Regards, > > > > Vishal. > > > > > > > > > > > > > > > >
