Piotr, With FLIP-27, Iceberg source already implemented alignment by tracking watermark and holding back split assignment when necessary.
The purpose of this discussion is to see if Iceberg source can leverage some of the watermark alignment work from Flink framework. Thanks, Steven On Thu, May 5, 2022 at 1:10 AM Piotr Nowojski <pnowoj...@apache.org> wrote: > Ok, I see. Thanks to both of you for the explanation. > > Do we need changes to Apache Flink for this feature? Can it be implemented > in the Sources without changes in the framework? I presume source can > access min/max watermark from the split, so as long as it also knows > exactly which splits have finished, it would know which splits to hold back. > > Best, > Piotrek > > śr., 4 maj 2022 o 20:03 Steven Wu <stevenz...@gmail.com> napisał(a): > >> Piotr, thanks a lot for your feedback. >> >> > I can see this being an issue if the existence of too many blocked >> splits is occupying too many resources. >> >> This is not desirable. Eagerly assigning many splits to a reader can >> defeat the benefits of pull based dynamic split assignments. Iceberg >> readers request one split at a time upon start or completion of a split. >> Dynamic split assignment is better for work sharing/stealing as Becket >> mentioned. Limiting number of active splits can be handled by the FLIP-27 >> Iceberg source and is somewhat orthogonal to watermark alignment. >> >> > Can not Iceberg just emit all splits and let FLIP-182/FLIP-217 handle >> the watermark alignment and block the splits that are too much into the >> future? >> >> The enumerator just assigns the next split to the requesting reader >> instead of holding back the split assignment. Let the reader handle the >> pause (if the file split requires alignment wait). This strategy might >> work and leverage more from the framework. >> >> We probably need the following to make this work >> * extract watermark/timestamp only at the completion of a split (not at >> record level). Because records in a file aren't probably not sorted by the >> timestamp field, the pause or watermark advancement is probably better done >> at file level. >> * source readers checkpoint the watermark. otherwise, upon restart >> readers won't be able to determine the local watermark and pause for >> alignment. We don't want to emit records upon restart due to unknown >> watermark info. >> >> All, >> >> Any opinion on different timestamp for source alignment (vs Flink >> application watermark)? For Iceberg source, we might want to enforce >> alignment on kafka timestamp but Flink application watermark may use event >> time field from payload. >> >> Thanks, >> Steven >> >> On Wed, May 4, 2022 at 7:02 AM Becket Qin <becket....@gmail.com> wrote: >> >>> Hey Piotr, >>> >>> I think the mechanism FLIP-182 provided is a reasonable default one, >>> which >>> ensures the watermarks are only drifted by an upper bound. However, >>> admittedly there are also other strategies for different purposes. >>> >>> In the Iceberg case, I am not sure if a static strictly allowed watermark >>> drift is desired. The source might just want to finish reading the >>> assigned >>> splits as fast as possible. And it is OK to have a drift of "one split", >>> instead of a fixed time period. >>> >>> As another example, if there are some fast readers whose splits are >>> always >>> throttled, while the other slow readers are struggling to keep up with >>> the >>> rest of the splits, the split enumerator may decide to reassign the slow >>> splits so all the readers have something to read. This would need the >>> SplitEnumerator to be aware of the watermark progress on each reader. So >>> it >>> seems useful to expose the WatermarkAlignmentEvent information to the >>> SplitEnumerator as well. >>> >>> Thanks, >>> >>> Jiangjie (Becket) Qin >>> >>> >>> >>> On Tue, May 3, 2022 at 7:58 PM Piotr Nowojski <pnowoj...@apache.org> >>> wrote: >>> >>> > Hi Steven, >>> > >>> > Isn't this redundant to FLIP-182 and FLIP-217? Can not Iceberg just >>> emit >>> > all splits and let FLIP-182/FLIP-217 handle the watermark alignment and >>> > block the splits that are too much into the future? I can see this >>> being an >>> > issue if the existence of too many blocked splits is occupying too many >>> > resources. >>> > >>> > If that's the case, indeed SourceCoordinator/SplitEnumerator would >>> have to >>> > decide on some basis how many and which splits to assign in what >>> order. But >>> > in that case I'm not sure how much you could use from FLIP-182 and >>> > FLIP-217. They seem somehow orthogonal to me, operating on different >>> > levels. FLIP-182 and FLIP-217 are working with whatever splits have >>> already >>> > been generated and assigned. You could leverage FLIP-182 and FLIP-217 >>> and >>> > take care of only the problem to limit the number of parallel active >>> > splits. And here I'm not sure if it would be worth generalising a >>> solution >>> > across different connectors. >>> > >>> > Regarding the global watermark, I made a related comment sometime ago >>> > about it [1]. It sounds to me like you also need to solve this problem, >>> > otherwise Iceberg users will encounter late records in case of some >>> race >>> > conditions between assigning new splits and completions of older. >>> > >>> > Best, >>> > Piotrek >>> > >>> > [1] >>> > >>> https://issues.apache.org/jira/browse/FLINK-21871?focusedCommentId=17495545&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-17495545 >>> > >>> > pon., 2 maj 2022 o 04:26 Steven Wu <stevenz...@gmail.com> napisał(a): >>> > >>> >> add dev@ group to the thread as Thomas suggested >>> >> >>> >> Arvid, >>> >> >>> >> The scenario 3 (Dynamic assignment + temporary no split) in the >>> FLIP-180 >>> >> (idleness) can happen to Iceberg source alignment, as readers can be >>> >> temporarily starved due to the holdback by the enumerator when >>> assigning >>> >> new splits upon request. >>> >> >>> >> Totally agree that we should decouple this discussion with the >>> FLIP-217, >>> >> which addresses the split level watermark alignment problem as a >>> follow-up >>> >> of FLIP-182 >>> >> >>> >> Becket, >>> >> >>> >> Yes, currently Iceberg source implemented the alignment leveraging the >>> >> dynamic split assignment from FLIP-27 design. Basically, the >>> enumerator >>> >> can >>> >> hold back split assignments to readers when necessary. Everything are >>> >> centralized in the enumerator: (1) watermark extraction and >>> aggregation >>> >> (2) >>> >> alignment decision and execution >>> >> >>> >> The motivation of this discussion is to see if Iceberg source can >>> leverage >>> >> some of the watermark alignment solutions (like FLIP-182) from Flink >>> >> framework. E.g., as mentioned in the doc, Iceberg source can >>> potentially >>> >> leverage the FLIP-182 framework to do the watermark extraction and >>> >> aggregation. For the alignment decision and execution, we can keep >>> them in >>> >> the centralized enumerator. >>> >> >>> >> Thanks, >>> >> Steven >>> >> >>> >> On Thu, Apr 28, 2022 at 2:05 AM Becket Qin <becket....@gmail.com> >>> wrote: >>> >> >>> >> > Hi Steven, >>> >> > >>> >> > Thanks for pulling me into this thread. I think the timestamp >>> >> > alignment use case here is a good example of what FLIP-27 was >>> designed >>> >> for. >>> >> > >>> >> > Technically speaking, Iceberg source can already implement the >>> timestamp >>> >> > alignment in the Flink new source even without FLIP-182. However, I >>> >> > understand the rationale here because timestamp alignment is also >>> >> trying to >>> >> > orchestrate the consumption of splits. However, it looks like >>> FLIP-182 >>> >> was >>> >> > not designed in a way that it can be easily extended for other use >>> >> cases. >>> >> > It may probably worth thinking of a more general mechanism to >>> answer the >>> >> > following questions: >>> >> > >>> >> > 1. What information whose source of truth is the Flink framework >>> should >>> >> be >>> >> > exposed to the SplitEnumerator and SourceReader? And how? >>> >> > 2. What control actions in the Flink framework are worth exposing >>> to the >>> >> > SplitEnumerators and SourceReaders? And how? >>> >> > >>> >> > In the context of timestamp alignment, the first question is more >>> >> > relevant. For example, instead of hardcode the ReportWatermarkEvent >>> >> > handling logic in the SourceCoordinator, should we expose this to >>> the >>> >> > SplitEnumerator? So basically there will be some information, such >>> as >>> >> > subtask local watermark, whose source of truth is Flink runtime, but >>> >> useful >>> >> > to the user provided pluggables. >>> >> > >>> >> > I think there are a few control flow patterns to make a complete >>> design: >>> >> > >>> >> > a. Framework space information (e.g. watermark) --> User space >>> >> Pluggables >>> >> > (e.g. SplitEnumerator) --> User space Actions (e.g. Pause reading a >>> >> split). >>> >> > b. Framework space information (e.g. task failure) --> User space >>> >> > pluggables (e.g. SplitEnumerator) --> Framework space actions (e.g. >>> exit >>> >> > the job) >>> >> > c. User space information (e.g. a custom workload metric) --> User >>> space >>> >> > pluggables (e.g. SplitEnumerator) --> User space actions (e.g. >>> rebalance >>> >> > the workload across the source readers). >>> >> > d. Use space information (e.g. a custom stopping event in the >>> stream) >>> >> --> >>> >> > User space pluggables (e.g. SplitEnumerator) --> Framework space >>> actions >>> >> > (e.g. stop the job). >>> >> > >>> >> > So basically for any user provided pluggables, the input >>> information may >>> >> > either come from another user provided logic or from the framework, >>> and >>> >> > after receiving that information, the pluggable may either want the >>> >> > framework or another pluggable to take an action. So this gives the >>> >> above 4 >>> >> > combinations. >>> >> > >>> >> > In our case, when the pluggables are SplitEnumerator and >>> SourceReader, >>> >> the >>> >> > control flows that only involve user space actions are fully >>> supported. >>> >> But >>> >> > it seems that when it comes to control flows involving framework >>> space >>> >> > information, some of the information has not been exposed to the >>> >> pluggable, >>> >> > and some framework actions might also be missing. >>> >> > >>> >> > Thanks, >>> >> > >>> >> > Jiangjie (Becket) Qin >>> >> > >>> >> > On Thu, Apr 28, 2022 at 3:44 PM Arvid Heise <ar...@apache.org> >>> wrote: >>> >> > >>> >> >> Hi folks, >>> >> >> >>> >> >> quick input from my side. I think this is from the implementation >>> >> >> perspective what Piotr and I had in mind for a global min watermark >>> >> that >>> >> >> helps in idleness cases. See also point 3 in >>> >> >> >>> >> >>> https://cwiki.apache.org/confluence/display/FLINK/FLIP-180%3A+Adjust+StreamStatus+and+Idleness+definition >>> >> >> . >>> >> >> >>> >> >> Basically, we would like to empower source enumerators to >>> determine the >>> >> >> global min watermark for all source readers factoring in even >>> future >>> >> >> splits. Not all sources can supply that information (think of a >>> general >>> >> >> file source) but most should be able to. Basically, Flink should >>> know >>> >> for a >>> >> >> given source at a given point in time what the min watermark >>> across all >>> >> >> source subtasks is. >>> >> >> >>> >> >> Here is some background: >>> >> >> In the context of idleness, we can deterministically advance the >>> >> >> watermark. In the pre-FLIP-27 era, we had heuristic approaches in >>> >> sources >>> >> >> to switch to idleness and thus allow watermarks to increase in >>> cases >>> >> where >>> >> >> fewer splits than source tasks are available. However, for sources >>> with >>> >> >> dynamic split discovery that actually yields incorrect results. >>> Think >>> >> of a >>> >> >> Kinesis consumer where a shard is split. Then a previously idle >>> source >>> >> >> subtask may receive a new split with time t0 as the lowest >>> timestamp. >>> >> Since >>> >> >> the source subtask did not participate in the global watermark >>> >> generation >>> >> >> (because it was idle), the previously emitted watermark may be >>> past t0 >>> >> and >>> >> >> thus results in late records potentially being discarded. A rerun >>> of >>> >> the >>> >> >> same pipeline on historic data would not render the source subtask >>> >> idle and >>> >> >> not result in late records. The solution was to not render source >>> >> subtasks >>> >> >> automatically idle by the framework if there are no spits. That >>> leads >>> >> to >>> >> >> confusion for Kafka users with static topic subscription where >>> #splits >>> >> < >>> >> >> #parallelism stalls pipelines because the watermark is not >>> advancing. >>> >> Here, >>> >> >> your sketched solution can be transferred to KafkaSource to let >>> Flink >>> >> know >>> >> >> that min global watermark on a static assignment is determined by >>> the >>> >> >> slowest partition. Hence, all idle readers emit that min global >>> >> watermark >>> >> >> and the user sees progress. >>> >> >> This whole idea is related to FLIP-182 watermark alignment but I'd >>> go >>> >> >> with another FLIP as the goal is quite different even though the >>> >> >> implementation overlaps. >>> >> >> >>> >> >> Now Iceberg seems to use the same information to actually pause the >>> >> >> consumption of files and create some kind of orderness guarantees >>> as >>> >> far as >>> >> >> I understood. This probably can be applied to any source with >>> dynamic >>> >> split >>> >> >> discovery. However, I wouldn't mix up the concepts and hence I >>> >> appreciate >>> >> >> you not chiming into the FLIP-182 and ff. threads. The goal of >>> >> FLIP-182 is >>> >> >> to pause readers while consuming a split, while your approach >>> pauses >>> >> >> readers before processing another split. So it feels more closely >>> >> related >>> >> >> to the global min watermark - so it could either be part of that >>> FLIP >>> >> or a >>> >> >> FLIP of its own. Afaik API changes should actually happen only on >>> the >>> >> >> enumerator side both for your ideas and for global min watermark. >>> >> >> >>> >> >> Best, >>> >> >> >>> >> >> Arvid >>> >> >> >>> >> >> On Wed, Apr 27, 2022 at 7:31 PM Thomas Weise <t...@apache.org> >>> wrote: >>> >> >> >>> >> >>> Hi Steven, >>> >> >>> >>> >> >>> Would it be better to bring this as a separate thread related to >>> >> Iceberg >>> >> >>> source to the dev@ list? I think this could benefit from broader >>> >> input? >>> >> >>> >>> >> >>> Thanks >>> >> >>> >>> >> >>> On Wed, Apr 27, 2022 at 9:36 AM Steven Wu <stevenz...@gmail.com> >>> >> wrote: >>> >> >>> >>> >> >>>> + Becket and Sebastian >>> >> >>>> >>> >> >>>> It is also related to the split level watermark alignment >>> discussion >>> >> >>>> thread. Because it is already very long, I don't want to further >>> >> complicate >>> >> >>>> the ongoing discussion there. But I can move the discussion to >>> that >>> >> >>>> existing thread if that is preferred. >>> >> >>>> >>> >> >>>> >>> >> >>>> On Tue, Apr 26, 2022 at 10:03 PM Steven Wu <stevenz...@gmail.com >>> > >>> >> >>>> wrote: >>> >> >>>> >>> >> >>>>> Hi all, >>> >> >>>>> >>> >> >>>>> We are thinking about how to align with the Flink community and >>> >> >>>>> leverage the FLIP-182 watermark alignment in the Iceberg >>> source. I >>> >> put some >>> >> >>>>> context in this google doc. Would love to get hear your >>> thoughts on >>> >> this. >>> >> >>>>> >>> >> >>>>> >>> >> >>>>> >>> >> >>> https://docs.google.com/document/d/1zfwF8e5LszazcOzmUAOeOtpM9v8dKEPlY_BRFSmI3us/edit# >>> >> >>>>> >>> >> >>>>> Thanks, >>> >> >>>>> Steven >>> >> >>>>> >>> >> >>>> >>> >> >>> > >>> >>