In the original version of the dataflow model, windowing was not annotated on each PCollection, rather it was inferred based on tracing up the graph to the latest WindowInto operation. This tracing logic was put in the SDK for simplicity.
I agree that there is room for a variety of SDK/DSL choices, but would strongly argue that for SDKs that implicitly specify triggering, the rules should be consistent and defined by the model. This is consistent with the principle of least surprise, as well as fact that the "beam:transform:group_by_key:v1" transform (should such an operation be provided), when applied to a PCollection with specific windowing strategy, should produce a PCollection with a well specified windowing strategy (and similarly for other well-known transforms). Likewise, I see sink triggers, once we figure them out, as semantic definitions belonging to the model (with likely some flexibility in implementation), not a choice each SDK should make on its own (though some may be able to declare/support them sooner than others). On Tue, Apr 30, 2019 at 6:24 PM Kenneth Knowles <[email protected]> wrote: > > +dev@ since this has taken a turn in that direction > > SDK/DSL consistency is nice. But each SDK/DSL being the best thing it can be > is more important IMO. I'm including DSLs to be clear that this is a > construction issue having little/nothing to do with SDK in the sense of the > per-run-time coprocessor we call the SDK harness, because that part of the > construction-time decision is not executed by the harness. > > So, for example, I am supportive of all of these: > > - SDK/DSL where every aggregation has an explicit trigger configuration > - SDK/DSL where the default trigger is "always" and explicit triggers are > used for throttling > - SDK/DSL that implements sink triggers and assigns triggering in the > pipeline graph as an implementation detail of that > > Each of these will have technical challenges to overcome (most notably > retractions) and won't look like Beam's original Java SDK and that is fine > with me. Python and Go already look very different, and it sounds like their > behavior has diverged as well, to say nothing of Scio, Euphoria, SQL. FWIW I > think this is somewhat comparable to how SDKs handle coders - they do the > best thing in their context and the proto/model makes many things possible. > > To go in the direction of consistency amongst the core SDKs, we could make > all triggers downstream of an initial GBK use the "repeat(always)" trigger. I > think we've discussed and this is simpler and more reliable than today's > continuation trigger, while keeping its intent. Well, the default, after watermark trigger probably shouldn't become repeat(always). > On Tue, Apr 30, 2019 at 2:41 AM Maximilian Michels <[email protected]> wrote: >> >> While it might be debatable whether "continuation triggers" are part of >> the model, the goal should be to provide a consistent experience across >> SDKs. I don't see a reason why the Java SDK would use continuation >> triggers while the Python SDK doesn't. >> >> This makes me think that trigger behavior across transforms should >> actually be part of the model. Or at least be standardized for SDK >> authors. This would also imply that it is documented for end users. >> >> In the end, users do not care about whether it's part of the model or >> not, but they like having no surprises :) >> >> On 29.04.19 09:20, Robert Bradshaw wrote: >> > I would say that the triggering done in stacked GBKs, with windowings >> > in between, is part of the model (at least in the sense that it's not >> > something that we'd want different SDKs to do separately.) >> > >> > OTOH, I'm not sure the continuation trigger should be part of the >> > model. Much easier to either let WindowInto with no trigger specified >> > either keep the existing one or reset it to the default. A runner can >> > mutate this to a continuation trigger under the hood, which should be >> > strictly looser (triggers are a promise about the earliest possible >> > firing, they don't force firings to happen). >> > >> > On Mon, Apr 29, 2019 at 4:34 AM Kenneth Knowles <[email protected]> wrote: >> >> >> >> It is accurate to say that the "continuation trigger" is not documented >> >> in the general programming guide. It shows up in the javadoc only, as far >> >> as I can tell [1]. Technically, this is accurate. It is not part of the >> >> core of Beam - each language SDK is required to explicitly specify a >> >> trigger for every GroupByKey when they submit a pipeline to a runner. >> >> But, of course, this is pretty much an implementation detail. >> >> >> >> Kenn >> >> >> >> [1] >> >> https://www.google.com/search?q="continuation+trigger"+site%3Abeam.apache.org >> >> >> >> On Sun, Apr 28, 2019 at 7:08 PM Reza Rokni <[email protected]> wrote: >> >>> >> >>> +1 I recall a fun afternoon a few years ago figuring this out ... >> >>> >> >>> On Mon, 11 Mar 2019 at 18:36, Maximilian Michels <[email protected]> wrote: >> >>>> >> >>>> Hi, >> >>>> >> >>>> I have seen several users including myself get confused by the "default" >> >>>> triggering behavior. I think it would be worthwhile to update the docs. >> >>>> >> >>>> In fact, Window.into(windowFn) does not override the existing >> >>>> windowing/triggering. It merges the previous input WindowStrategy with >> >>>> the new one. >> >>>> >> >>>> So your w1trigger will still be set when you do not set w2trigger. The >> >>>> default `AfterWatermark.pastEndOfWindow()` trigger will only be used >> >>>> when windowing for the first time, or when you set it explicitly. >> >>>> >> >>>> Thanks, >> >>>> Max >> >>>> >> >>>> On 06.03.19 00:28, Daniel Debrunner wrote: >> >>>>> Thanks Kenn,. >> >>>>> >> >>>>> Is it fair to say that this continuation trigger functionality is not >> >>>>> documented? >> >>>>> >> >>>>> In the Javadoc it has a similar sentence to the programming guide: >> >>>>> >> >>>>>> triggering(Trigger) allows specifying a trigger to control when (in >> >>>>>> processing time) results for the given window can be produced. If >> >>>>>> unspecified, the default behavior is to trigger first when the >> >>>>>> watermark passes the end of the window, and then trigger again every >> >>>>>> time there is late arriving data. >> >>>>> >> >>>>> Thanks, >> >>>>> Dan. >> >>>>> >> >>>>> On Tue, Mar 5, 2019 at 1:46 PM Kenneth Knowles <[email protected]> wrote: >> >>>>>> >> >>>>>> The Window.into transform does not reset the trigger to the default. >> >>>>>> So where you have w2trigger, if you leave it off, then the triggering >> >>>>>> is left as the "continuation trigger" from w1trigger. Basically it >> >>>>>> tries to let any output caused by w1trigger to flow all the way >> >>>>>> through the pipeline without delay. >> >>>>>> >> >>>>>> Kenn >> >>>>>> >> >>>>>> On Tue, Mar 5, 2019 at 1:27 PM Daniel Debrunner <[email protected]> >> >>>>>> wrote: >> >>>>>>> >> >>>>>>> I discover how to fix my issue but not sure I understand why it does. >> >>>>>>> >> >>>>>>> I created a complete sample here: >> >>>>>>> https://gist.github.com/ddebrunner/5d4ef21c255c1d40a4517a0060ff8b99#file-cascadewindows-java-L104 >> >>>>>>> Link points to the area of interest. >> >>>>>>> >> >>>>>>> With the second window I was originally not specifying a trigger so >> >>>>>>> using the default trigger which lead to multiple triggers of the >> >>>>>>> combine on the second window. >> >>>>>>> >> >>>>>>> However changing the trigger to be AfterWatermark.pastEndOfWindow() >> >>>>>>> produced the output I expected, a single combine across all the >> >>>>>>> elements in the window. >> >>>>>>> The gist has comments showing the output and the two code variations. >> >>>>>>> >> >>>>>>> I don't understand why, since according to 8.1.1 [1] I thought >> >>>>>>> AfterWatermark.pastEndOfWindow() was the default. Maybe its due to >> >>>>>>> late data in some way but I'm not sure I understand how the data >> >>>>>>> could >> >>>>>>> be late in this case. >> >>>>>>> >> >>>>>>> This is with Beam 2.7 direct runner btw. >> >>>>>>> >> >>>>>>> Thanks again for your help, >> >>>>>>> Dan. >> >>>>>>> [1] >> >>>>>>> https://beam.apache.org/documentation/programming-guide/#event-time-triggers >> >>>>>>> >> >>>>>>> On Tue, Mar 5, 2019 at 11:48 AM Daniel Debrunner >> >>>>>>> <[email protected]> wrote: >> >>>>>>>> >> >>>>>>>> Thanks Robert, your description is what I'm expecting, I'm working >> >>>>>>>> on >> >>>>>>>> a simple example to see if what I'm seeing is different and then >> >>>>>>>> hopefully use that to clarify my misunderstanding. >> >>>>>>>> >> >>>>>>>> Thanks, >> >>>>>>>> Dan. >> >>>>>>>> >> >>>>>>>> On Tue, Mar 5, 2019 at 11:31 AM Robert Bradshaw >> >>>>>>>> <[email protected]> wrote: >> >>>>>>>>> >> >>>>>>>>> Windows are assigned to elements via the Window.into transform. >> >>>>>>>>> They >> >>>>>>>>> influence grouping operations such as GroupByKey, Combine.perKey, >> >>>>>>>>> and >> >>>>>>>>> Combine.globally. Looking at your example, you start with >> >>>>>>>>> >> >>>>>>>>> PCollection<KV<A,B>> >> >>>>>>>>> >> >>>>>>>>> Presumably via a Read or a Create. These KVs are in a global >> >>>>>>>>> window, >> >>>>>>>>> so the elements are really triples (ignoring PaneInfo) of the form >> >>>>>>>>> >> >>>>>>>>> (KV<A, B>, GlobalWindow, timestamp) >> >>>>>>>>> >> >>>>>>>>> From what I gather, the next step you do is a >> >>>>>>>>> Window.into(FixedWindows.of(...)), yielding a PCollection<KV<A,B>> >> >>>>>>>>> whose elements are, implicitly >> >>>>>>>>> >> >>>>>>>>> (KV<A, B>, IntervalWindow, timestamp) >> >>>>>>>>> >> >>>>>>>>> Now you apply a GroupByKey to get elements of the form >> >>>>>>>>> >> >>>>>>>>> (KV<A, Iterable<B>>, IntervalWindow, timestamp) >> >>>>>>>>> >> >>>>>>>>> where there is one Iterable for each distinct key and window. You >> >>>>>>>>> apply a ParDo to get PCollection<X> which is of the form >> >>>>>>>>> >> >>>>>>>>> (X, IntervalWindow, timestamp) >> >>>>>>>>> >> >>>>>>>>> It looks like your next step is another >> >>>>>>>>> Window.into(FixedWindows.of(...)), yielding >> >>>>>>>>> >> >>>>>>>>> (X, IntervalWindow, timestamp) >> >>>>>>>>> >> >>>>>>>>> where the IntervalWindow here may be different if the parameters to >> >>>>>>>>> FixedWindows were different (e.g. the first was by minute, the >> >>>>>>>>> second >> >>>>>>>>> by hours). If it's the same, this is a no-op. Now you apply >> >>>>>>>>> Combine.globally(CombineFn<X, R>) to get a PCollection<R> whose >> >>>>>>>>> elements are of the form >> >>>>>>>>> >> >>>>>>>>> (R, IntervalWindow, timestamp) >> >>>>>>>>> >> >>>>>>>>> where there is now one R per window (the elements in the same >> >>>>>>>>> window >> >>>>>>>>> being combined, the elements across windows not). >> >>>>>>>>> >> >>>>>>>>> FWIW, internally, Combine.globally is implemented as >> >>>>>>>>> PariWithNullKey + >> >>>>>>>>> CombinePerKey + StripNullKey. >> >>>>>>>>> >> >>>>>>>>> Does this help? >> >>>>>>>>> >> >>>>>>>>> >> >>>>>>>>> On Tue, Mar 5, 2019 at 8:09 PM Daniel Debrunner >> >>>>>>>>> <[email protected]> wrote: >> >>>>>>>>>> >> >>>>>>>>>> Thanks for the reply. >> >>>>>>>>>> >> >>>>>>>>>> As for every element is always associated with a window, when a >> >>>>>>>>>> element is produced due to a window trigger (e.g. the GroupByKey) >> >>>>>>>>>> what >> >>>>>>>>>> window is it associated with? The window it was produced from? >> >>>>>>>>>> Maybe >> >>>>>>>>>> the question is when is a window assigned to an element? >> >>>>>>>>>> >> >>>>>>>>>> I'll see if I can come up with an example, >> >>>>>>>>>> >> >>>>>>>>>> Thanks, >> >>>>>>>>>> Dan. >> >>>>>>>>>> >> >>>>>>>>>> On Tue, Mar 5, 2019 at 10:47 AM Kenneth Knowles <[email protected]> >> >>>>>>>>>> wrote: >> >>>>>>>>>>> >> >>>>>>>>>>> Two pieces to this: >> >>>>>>>>>>> >> >>>>>>>>>>> 1. Every element in a PCollection is always associated with a >> >>>>>>>>>>> window, and GroupByKey (hence CombinePerKey) operates >> >>>>>>>>>>> per-key-and-window (w/ window merging). >> >>>>>>>>>>> 2. If an element is not explicitly a KV, then there is no key >> >>>>>>>>>>> associated with it. >> >>>>>>>>>>> >> >>>>>>>>>>> I'm afraid I don't have any guesses at the problem based on what >> >>>>>>>>>>> you've shared. Can you say more? >> >>>>>>>>>>> >> >>>>>>>>>>> Kenn >> >>>>>>>>>>> >> >>>>>>>>>>> On Tue, Mar 5, 2019 at 10:29 AM Daniel Debrunner >> >>>>>>>>>>> <[email protected]> wrote: >> >>>>>>>>>>>> >> >>>>>>>>>>>> The windowing section of the Beam programming model guide shows >> >>>>>>>>>>>> a >> >>>>>>>>>>>> window defined and used in the GropyByKey transform after a >> >>>>>>>>>>>> ParDo. >> >>>>>>>>>>>> (section 7.1.1). >> >>>>>>>>>>>> >> >>>>>>>>>>>> However I couldn't find any documentation on how long the window >> >>>>>>>>>>>> remains in scope for subsequent transforms. >> >>>>>>>>>>>> >> >>>>>>>>>>>> I have an application with this pipeline: >> >>>>>>>>>>>> >> >>>>>>>>>>>> PCollection<KV<A,B>> -> FixedWindow<KV<A,B>> -> GroupByKey -> >> >>>>>>>>>>>> PCollection<X> -> FixedWindow<X> -> Combine<X,R>.globally -> >> >>>>>>>>>>>> PCollection<R> >> >>>>>>>>>>>> >> >>>>>>>>>>>> The idea is that the first window is aggregating by key but in >> >>>>>>>>>>>> the >> >>>>>>>>>>>> second window I need to combine elements across all keys. >> >>>>>>>>>>>> >> >>>>>>>>>>>> With my initial app I was seeing some runtime errors in/after >> >>>>>>>>>>>> the >> >>>>>>>>>>>> combine where a KV<null,R> was being seen, even though at that >> >>>>>>>>>>>> point >> >>>>>>>>>>>> there should be no key for the PCollection<R>. >> >>>>>>>>>>>> >> >>>>>>>>>>>> In a simpler test I can apply FixedWindow<X> -> >> >>>>>>>>>>>> Combine<X,R>.globally >> >>>>>>>>>>>> -> PCollection<R> to a PCollection without an upstream window >> >>>>>>>>>>>> and the >> >>>>>>>>>>>> combine correctly happens once. >> >>>>>>>>>>>> But then adding the keyed upstream window, the combine occurs >> >>>>>>>>>>>> once per >> >>>>>>>>>>>> key without any final combine across the keys. >> >>>>>>>>>>>> >> >>>>>>>>>>>> So it seems somehow the memory of the key exists even with the >> >>>>>>>>>>>> new >> >>>>>>>>>>>> window transform, >> >>>>>>>>>>>> >> >>>>>>>>>>>> I'm probably misunderstanding some detail of windowing, but I >> >>>>>>>>>>>> couldn't >> >>>>>>>>>>>> find any deeper documentation than the simple examples in the >> >>>>>>>>>>>> programming model guide. >> >>>>>>>>>>>> >> >>>>>>>>>>>> Can anyone point me in the correct direction? >> >>>>>>>>>>>> >> >>>>>>>>>>>> Thanks, >> >>>>>>>>>>>> Dan. >> >>> >> >>> >> >>> >> >>> -- >> >>> >> >>> This email may be confidential and privileged. If you received this >> >>> communication by mistake, please don't forward it to anyone else, please >> >>> erase all copies and attachments, and please let me know that it has >> >>> gone to the wrong person. >> >>> >> >>> The above terms reflect a potential business arrangement, are provided >> >>> solely as a basis for further discussion, and are not intended to be and >> >>> do not constitute a legally binding obligation. No legally binding >> >>> obligations will be created, implied, or inferred until an agreement in >> >>> final form is executed in writing by all parties involved.
