On Wed, Mar 4, 2020 at 7:37 AM Ismaël Mejía <[email protected]> wrote:
> > Bounded SDFs are allowed to have a method signature which has void as the > > return type OR a ProcessContinuation. Unbounded SDFs must use a > > ProcessContinuation as the return type. The "void" return case improves > ease > > of use since it is likely to be the common case for bounded SDFs. > > Luke, thanks for the answer, so I can assume that in the processElement of > an > existing Bounded SDF like HBaseReadSplittableDoFn I can return also > ProcessContinuation? If so maybe it is worth to make it explicit (at least > in > some example) instead of the 'ease of use' approach. Maybe to clarify my > point > (and to connect with the WatermarkEstimator concept) what I am trying to > understand is how to evolve an existing Bounded SDF into an Unbounded one? > Is > this just that signature + assigning a different WatermarkEstimator. Again > I > think an example is worth (a thousand words). > Why do you want to evolve a bounded SDF into an unbounded SDF (is the restriction truly unbounded)? For example, FileIO should be a bounded SDF while Watch should be unbounded. The combination of the two (Watch followed by FileIO) leads to having an unbounded pipeline. How does the watermark integrate in this case? If FileIO doesn't report a watermark, then the records will all be produced at the timestamp of the "file descriptor" that Watch produced and FileIO will effectively track the watermark of Watch. However, let's say Watch was scanning a directory for hourly log files then FileIO could hold back the watermark based upon the data within those files. For example: Watch starts scanning files in the range [1pm, 2pm) and reports the watermark of [1pm] Each FileIO file descriptor starts reading those file descriptors and is reporting a watermark that is monotonically increasing in the range of [1pm, 2pm) Even though all known fileIO instances are reporting watermarks that could be >1pm, the system holds the downstream FileIO watermark back at 1pm since the Watch could still produce a new file descriptor that could produce records at 1pm that aren't late. Let's say that Watch says its finished scanning the [1pm, 2pm) range and now starts scanning the [2pm, 3pm) range and advances the watermark to 2pm and FileIO has now started processing all known inputs that were produced at the 1pm watermark. This allows the Watermark downstream of FileIO to advance to the minimum watermark of all FileIO "file descriptor" reported watermarks upto 2pm since that is the upper bound that the upstream Watch transform still imposes. Note that in this example, it is important that Watch advance the watermark to 2pm as soon as it thinks/knows there will be no more files. If Watch advances too early and some file appears then that "file descriptor" will be considered late and could be dropped or it might be read, it all depends on the windowing strategy the pipeline author has chosen. > > There is SplittableParDoViaKeyedWorkItems[1] that can be used by > classical > > runners but it has limitations. > > Can you please elaborate on the limitations? Will we have a default > translation > override for Unbounded SDF? The existing override for Bounded SDF brought > support for almost every runner (even if in a naive way) and unlock IO > authors > to develop IOs that could be tested easily, it looks like worth to have a > default override for the Unbounded case. > pr/10897[1] added an UnboundedSource -> SDF wrapper but is still lacking deduplication support (I have some preliminary work in a branch[2]) and progress reporting. (Dynamic work rebalancing for unbounded sources is not supported in any Runner) SplittableParDoViaKeyedWorkItems doesn't support bundle finalization which is necessary to support the finalizeCheckpoint callback from CheckpointMark[3] in UnboundedSource and since SplittableParDoViaKeyedWorkItems is built on top of DoFns/PTransform APIs, it would require adding support for the BundleFinalizer[4] to all classic runners. I would greatly appreciate the help in getting BundleFinalizer working on classic runners which I believe would unblock naive support for unbounded SDFs on classic runners. 1: https://github.com/apache/beam/pull/10897 2: https://github.com/lukecwik/incubator-beam/tree/splittabledofn3 3: https://github.com/apache/beam/blob/ded686a58ad4747e91a26d3e59f61019b641e655/sdks/java/core/src/main/java/org/apache/beam/sdk/io/UnboundedSource.java#L130 4: https://github.com/apache/beam/blob/ded686a58ad4747e91a26d3e59f61019b641e655/sdks/java/core/src/main/java/org/apache/beam/sdk/transforms/DoFn.java#L1174 > On Tue, Mar 3, 2020 at 7:03 PM Luke Cwik <[email protected]> wrote: > > > > > > > > On Tue, Mar 3, 2020 at 9:11 AM Ismaël Mejía <[email protected]> wrote: > >> > >> > the unification of bounded/unbounded within SplittableDoFn has always > been a goal. > >> > >> I am glad to know that my intuition is correct and that this was > envisioned, the > >> idea of checkpoints for bounded inputs sounds super really useful. > Eager to try > >> that on practice. > >> > >> An explicit example (with a WatermarkEstimator for a bounded case would > be > >> really nice to see, for learning purposes), also with the unification > goal what > >> if we align then the Bounded SDFs to have similar signatures no? I mean > the > >> method that returns a continuation even for the Bounded case. > > > > > > Bounded SDFs are allowed to have a method signature which has void as > the return type OR a ProcessContinuation. Unbounded SDFs must use a > ProcessContinuation as the return type. The "void" return case improves > ease of use since it is likely to be the common case for bounded SDFs. > > > >> > >> > Currently the watermark that is reported as part of the PollResult is > passed > >> > to the ProcessContext.updateWatermark [1, 2] function and instead > that call > >> > would be redirected to the ManualWatermarkEstimator.setWatermark > function [3]. > >> > >> Is there a JIRA for the Watch adjustments so we don't forget to > integrate the > >> WatermarkEstimators in? I am really curious on the implementation to > see if I > >> finally understand the internals of Watch too. > > > > > > Migrating from ProcessContext#updateWatermark to WatermarkEstimators > will require updating Watch. Filed BEAM-9430. > > > >> > >> Extra question: Do you think we can have a naive version of Unbounded > SDF like > >> we have the naive one on classical runners (if I understood correctly > the > >> current one is only for portable runners). I worry about the adoption > potential. > > > > > > There is SplittableParDoViaKeyedWorkItems[1] that can be used by > classical runners but it has limitations. > > > > 1: > https://github.com/apache/beam/blob/master/runners/core-java/src/main/java/org/apache/beam/runners/core/SplittableParDoViaKeyedWorkItems.java > > > >> > >> On Tue, Mar 3, 2020 at 1:41 AM Robert Bradshaw <[email protected]> > wrote: > >> > > >> > I don't have a strong preference for using a provider/having a set of > >> > tightly coupled methods in Java, other than that we be consistent (and > >> > we already use the methods style for restrictions). > >> > > >> > On Mon, Mar 2, 2020 at 3:32 PM Luke Cwik <[email protected]> wrote: > >> > > > >> > > Jan, there are some parts of Apache Beam the watermarks package > will likely rely on (@Experimental annotation, javadoc links) but > fundamentally should not rely on core and someone could create a separate > package for this. > >> > > >> > I think it does make sense for a set of common watermark trackers to > >> > be shipped with core (e.g. manual, monotonic, and eventually a > >> > probabilistic one). > >> > > >> > > Ismael, the unification of bounded/unbounded within SplittableDoFn > has always been a goal. There are a set of features that BoundedSources are > unlikely to use but would still be allowed to use them. For example, > bounded sources may want to have support for checkpointing since I could > foresee an BoundedSource that can notice that a certain resource becomes > unavailable and can only process it later. The choice of which watermark > estimator to use is a likely point of difference between bounded and > unbounded SDFs since bounded SDFs would be able to use a very simple > estimator where the watermark is held at -infinity and only advances to > +infinity once there is no more data to process. But even though unbounded > SDFs are likely to be the most common users of varied watermark estimators, > a bounded SDF may still want to advance the watermark as they read records > so that runners that are more "streaming" (for example micro batch) could > process the entire pipeline in parallel vs other runners that execute one > whole segment of the pipeline at a time. > >> > > >> > Put another way, the value of watermark trackers is to allow > >> > processing to continue downstream before the source has completed > >> > reading. This is of course essential for streaming, but If the source > >> > is read to completion before downstream stages start (as is the case > >> > for most batch runners) it is not needed. What this unification does > >> > allow, however, is a source to be written in such a way that can be > >> > efficiently used in both batch and streaming mode. > >> > > >> > > Currently the watermark that is reported as part of the PollResult > is passed to the ProcessContext.updateWatermark [1, 2] function and instead > that call would be redirected to the ManualWatermarkEstimator.setWatermark > function [3]. > >> > > > >> > > 1: > https://github.com/apache/beam/blob/a16725593b84b84b37bc67cd202d1ac8b724c6f4/sdks/java/core/src/main/java/org/apache/beam/sdk/transforms/Watch.java#L757 > >> > > 2: > https://github.com/apache/beam/blob/a16725593b84b84b37bc67cd202d1ac8b724c6f4/sdks/java/core/src/main/java/org/apache/beam/sdk/transforms/DoFn.java#L275 > >> > > 3: > https://github.com/apache/beam/blob/fb42666a4e1aec0413f161c742d8f010ef9fe9f2/sdks/java/core/src/main/java/org/apache/beam/sdk/transforms/splittabledofn/ManualWatermarkEstimator.java#L45 > >> > > > >> > > On Fri, Feb 28, 2020 at 6:09 AM Ismaël Mejía <[email protected]> > wrote: > >> > >> > >> > >> I just realized that the HBaseIO example is not a good one because > we can > >> > >> already have Watch like behavior as we do for Partition discovery > in HCatalogIO. > >> > >> Still I am interested on your views on bounded/unbounded > unification. > >> > >> > >> > >> Interesting question2: How this will annotations connect with the > Watch > >> > >> transform Polling patterns? > >> > >> > https://github.com/apache/beam/blob/650e6cd9c707472c34055382a1356cf22d14ee5e/sdks/java/core/src/main/java/org/apache/beam/sdk/transforms/Watch.java#L178 > >> > >> > >> > >> > >> > >> On Fri, Feb 28, 2020 at 10:47 AM Ismaël Mejía <[email protected]> > wrote: > >> > >>> > >> > >>> Really interesting! Implementing correctly the watermark has been > a common > >> > >>> struggle for IO authors, to the point that some IOs still have > issues around > >> > >>> that. So +1 for this, in particular if we can get to reuse common > patterns. > >> > >>> I was not aware of Boyuan's work around this, really nice. > >> > >>> > >> > >>> One aspect I have always being confused about since I read the > SDF proposal > >> > >>> documents is if we could get to have a single API for both > Bounded and Unbounded > >> > >>> IO by somehow assuming that with a BoundedSDF is an UnboundedSDF > special case. > >> > >>> Could WatermarkEstimator help in this direction? > >> > >>> > >> > >>> One quick case that I can think is to make the current HBaseIO > SDF to work in an > >> > >>> unbounded manner, for example to 'watch and read new tables'. > >> > >>> > >> > >>> > >> > >>> On Thu, Feb 27, 2020 at 11:43 PM Luke Cwik <[email protected]> > wrote: > >> > >>>> > >> > >>>> See this doc[1] and blog[2] for some context about > SplittableDoFns. > >> > >>>> > >> > >>>> To support watermark reporting within the Java SDK for > SplittableDoFns, we need a way to have SDF authors to report watermark > estimates over the element and restriction pair that they are processing. > >> > >>>> > >> > >>>> For UnboundedSources, it was found to be a pain point to ask > each SDF author to write their own watermark estimation which typically > prevented re-use. Therefore we would like to have a "library" of watermark > estimators that help SDF authors perform this estimation similar to how > there is a "library" of restrictions and restriction trackers that SDF > authors can use. For SDF authors where the existing library doesn't work, > they can add additional ones that observe timestamps of elements or choose > to directly report the watermark through a "ManualWatermarkEstimator" > parameter that can be supplied to @ProcessElement methods. > >> > >>>> > >> > >>>> The public facing portion of the DoFn changes adds three new > annotations for new DoFn style methods: > >> > >>>> GetInitialWatermarkEstimatorState: Returns the initial watermark > state, similar to GetInitialRestriction > >> > >>>> GetWatermarkEstimatorStateCoder: Returns a coder compatible with > watermark state type, similar to GetRestrictionCoder for restrictions > returned by GetInitialRestriction. > >> > >>>> NewWatermarkEstimator: Returns a watermark estimator that either > the framework invokes allowing it to observe the timestamps of output > records or a manual watermark estimator that can be explicitly invoked to > update the watermark. > >> > >>>> > >> > >>>> See [3] for an initial PR with the public facing additions to > the core Java API related to SplittableDoFn. > >> > >>>> > >> > >>>> This mirrors a bunch of work that was done by Boyuan within the > Pyhon SDK [4, 5] but in the style of new DoFn parameter/method invocation > we have in the Java SDK. > >> > >>>> > >> > >>>> 1: https://s.apache.org/splittable-do-fn > >> > >>>> 2: https://beam.apache.org/blog/2017/08/16/splittable-do-fn.html > >> > >>>> 3: https://github.com/apache/beam/pull/10992 > >> > >>>> 4: https://github.com/apache/beam/pull/9794 > >> > >>>> 5: https://github.com/apache/beam/pull/10375 >
