- it either blows state in batch case or the time complexity of the sort would be O(n^2) (and reprocessing several years of dense time-series data makes it a no go)
- it is not reusable for different time-ordering needs, because the logic implemented purely in user-space cannot be transferred to different problem (there are two states needed, one for buffer, the other for user-state) and extending DoFns does not work (cannot create abstract SortedDoFn, because of the state annotation definitions)
Jan On 11/26/19 12:56 PM, David Morávek wrote:
Hi,I think what Jan has in mind would look something like this <https://gist.github.com/dmvk/3ea32eb36c6406fa72d70b9b1df1d878>, if implemented in user code. Am I right?D.On Tue, Nov 26, 2019 at 10:23 AM Jan Lukavský <je...@seznam.cz <mailto:je...@seznam.cz>> wrote:On 11/25/19 11:45 PM, Kenneth Knowles wrote:On Mon, Nov 25, 2019 at 1:56 PM Jan Lukavský <je...@seznam.cz <mailto:je...@seznam.cz>> wrote: Hi Rui, > Hi Kenn, you think stateful DoFn based join can emit joined rows that never to be retracted because in stateful DoFn case joined rows will be controlled by timers and emit will be only once? If so I will agree with it. Generally speaking, if only emit once is the factor of needing retraction or not. that would imply buffering elements up until watermark, then sorting and so reduces to the option a) again, is that true? This also has to deal with allowed lateness, that would mean, that with allowed lateness greater than zero, there can still be multiple firings and so retractions are needed. Specifically, when I say "bi-temporal join" I mean unbounded-to-unbounded join where one of the join conditions is that elements are within event time distance d of one another. An element at time t will be saved until time t + 2d and then garbage collected. Every matching pair can be emitted immediately.OK, this might simplify things a little. Is there a design doc for that? If there are multiple LHS elements within event time distance from RHS element, which one should be joined? I suppose all of them, but that is not "(time-varying-)relational" join semantics. In that semantics only the last element must be joined, because that is how a (classical) relational database would see the relation at time T (the old record would have been overwritten and not be part of the output). Because of the time distance constraint this is different from the join I have in mind, because that simply joins every LHS element(s) to most recent RHS element(s) and vice versa, without any additional time constraints (that is the RHS "update" can happen arbitrarily far in past). JanIn the triggered CoGBK + join-product implementation, you do need retractions as a model concept. But you don't need full support, since they only need to be shipped as deltas and only from the CoGBK to the join-product transform where they are all consumed to create only positive elements. Again a delay is not required; this yields correct results with the "always" trigger. Neither case requires waiting or time sorting a whole buffer. The bi-temporal join requires something more, in a way, since you need to query by time range and GC time prefixes. Kenn Jan On 11/25/19 10:17 PM, Rui Wang wrote:On Mon, Nov 25, 2019 at 11:29 AM Jan Lukavský <je...@seznam.cz <mailto:je...@seznam.cz>> wrote: On 11/25/19 7:47 PM, Kenneth Knowles wrote:On Sun, Nov 24, 2019 at 12:57 AM Jan Lukavský <je...@seznam.cz <mailto:je...@seznam.cz>> wrote: I can put down a design document, but before that I need to clarify some things for me. I'm struggling to put all of this into a bigger picture. Sorry if the arguments are circulating, but I didn't notice any proposal of how to solve these. If anyone can disprove any of this logic it would be very much appreciated as I might be able to get from a dead end: a) in the bi-temporal join you can either buffer until watermark, or emit false data that has to be retracted This is not the case. A stateful DoFn based join can emit immediately joined rows that will never need to be retracted. The need for retractions has to do with CoGBK-based implementation of a join.I fail to see how this could work. If I emit joined rows immediately without waiting for watermark to pass, I can join two elements, that don't belong to each other, because later can arrive element with lower time distance, that should have been joint in the place of the previously emitted one. This is wrong result that has to be retracted. Or what I'm missing? Hi Kenn, you think stateful DoFn based join can emit joined rows that never to be retracted because in stateful DoFn case joined rows will be controlled by timers and emit will be only once? If so I will agree with it. Generally speaking, if only emit once is the factor of needing retraction or not. In the past brainstorming, even having retractions ready, streaming join with windowing are likely be implemented by a style of CoGBK + stateful DoFn.I suggest that you work out the definition of the join you are interested in, with a good amount of mathematical rigor, and then consider the ways you can implement it. That is where a design doc will probably clarify things. Kenn b) until retractions are 100% functional (and that is sort of holy grail for now), then the only solution is using a buffer holding data up to watermark *and then sort by event time* c) even if retractions were 100% functional, there would have to be special implementation for batch case, because otherwise this would simply blow up downstream processing with insanely many false additions and subsequent retractions Property b) means that if we want this feature now, we must sort by event time and there is no way around. Property c) shows that even in the future, we must make (in certain cases) distinction between batch and streaming code paths, which seems weird to me, but it might be an option. But still, there is no way to express this join in batch case, because it would require either buffering (up to) whole input on local worker (doesn't look like viable option) or provide a way in user code to signal the need for ordering of data inside GBK (and we are there again :)). Yes, we might shift this need from stateful dofn to GBK like input.apply(GroupByKey.sorted()) I cannot find a good reasoning why this would be better than giving this semantics to (stateful) ParDo. Maybe someone can help me out here? Jan On 11/24/19 5:05 AM, Kenneth Knowles wrote:I don't actually see how event time sorting simplifies this case much. You still need to buffer elements until they can no longer be matched in the join, and you still need to query that buffer for elements that might match. The general "bi-temporal join" (without sorting) requires one new state type and then it has identical API, does not require any novel data structures or reasoning, yields better latency (no sort buffer delay), and discards less data (no sort buffer cutoff; watermark is better). Perhaps a design document about this specific case would clarify. Kenn On Fri, Nov 22, 2019 at 10:08 PM Jan Lukavský <je...@seznam.cz <mailto:je...@seznam.cz>> wrote: I didn't want to go too much into detail, but to describe the idea roughly (ignoring the problem of different window fns on both sides to keep it as simple as possible): rhs ----- \ flatten (on global window) ---- stateful par do (sorted by event time) ---- output lhs ----- / If we can guarantee event time order arrival of events into the stateful pardo, then the whole complexity reduces to keep current value of left and right element and just flush them out each time there is an update. That is the "knob" is actually when watermark moves, because it is what tells the join operation that there will be no more (not late) input. This is very, very simplified, but depicts the solution. The "classical" windowed join reduces to this if all data in each window is projected onto window end boundary. Then there will be a cartesian product, because all the elements have the same timestamp. I can put this into a design doc with all the details, I was trying to find out if there is or was any effort around this. I was in touch with Reza in the PR #9032, I think that it currently suffers from problems with running this on batch. I think I can even (partly) resolve the retraction issue (for joins), as described on the thread [1]. Shortly, there can be two copies of the stateful dofn, one running at watermark and the other at (watermark - allowed lateness). One would produce ON_TIME (maybe wrong) results, the other would produce LATE but correct ones. Being able to compare them, the outcome would be that it would be possible to retract the wrong results. Yes, this is also about providing more evidence of why I think event-time sorting should be (somehow) part of the model. :-) Jan [1] https://lists.apache.org/thread.html/a736ebd6b0913d2e06dfa54797716d022adf2c45140530d5c3bd538d@%3Cdev.beam.apache.org%3E On 11/23/19 5:54 AM, Kenneth Knowles wrote:+Mikhail Gryzykhin <mailto:mig...@google.com> +Rui Wang <mailto:ruw...@google.com> +Reza Rokni <mailto:r...@google.com> who have all done some investigations here. On Fri, Nov 22, 2019 at 11:48 AM Jan Lukavský <je...@seznam.cz <mailto:je...@seznam.cz>> wrote: On 11/22/19 7:54 PM, Reuven Lax wrote:On Fri, Nov 22, 2019 at 10:19 AM Jan Lukavský <je...@seznam.cz <mailto:je...@seznam.cz>> wrote: Hi Reuven, I didn't investigate that particular one, but looking into that now, it looks that is (same as the "classic" join library) builds around CoGBK. Is that correct? If yes, then it essentially means that it: - works only for cases where both sides have the same windowfn (that is limitation of Flatten that precedes CoGBK) Correct. Did you want to join different windows? If so what are the semantics? If the lhs has FixedWindows and the rhs has SessionWindows, what do you want the join semantics to be? The only thing I could imagine would be for the user to provide some function telling the join how to map the windows together, but that could be pretty complicated.I don't want to go too far into details, but generally both lhs and rhs can be put onto time line and then full join can be defined as each pair of (lhs, first preceding rhs) and (rhs, first preceding lhs). Then the end of window is semantically just clearing the joined value (setting it to null, thus at the end of window there will be pair (lhs, null) or (null, rhs) in case of full outer join). This way any combination of windows is possible, because all window does is that it "scopes" validity of respective values (lhs, rhs). I think it is very valid to hope to do a join in the sense of a relational join where it is row-to-row. In this case, Beam's concept of windowing may or may not make sense. It is just a tool for the job. It is just a grouping key that provides a time when state can be deleted. So I would say your use case is more global window to global window join. That is what I think of as a true stream-to-stream join anyhow. You probably don't want to wait forever for output. So you'll need to use some knob other than Beam windows or triggers. Reza has prototyped a join like you describe here: https://github.com/apache/beam/pull/9032 If your join condition explicitly includes the event time distance between elements, then it could "just work". If that isn't really part of your join condition, then you will have to see this restriction as a "knob" that you tweak on your results.- when using global window, there has to be trigger and (afaik) there is no trigger that would guarantee firing after each data element (for early panes) (because triggers are there to express cost-latency tradeoff, not semantics) Can you explain the use case where this matters? If you do trigger elementCountAtLeast(1) on the join, then the consumer will simply see a continuous stream of outputs. I'm not sure I understand why the consumer cares that some of those outputs were in a pane that really held 3 outputs instead of 1.What I'm trying to solve is basically this: - lhs is event stream - rhs is stream of a "state updates" purpose of the join is "take each event, pair it with currently valid state and produce output and possibly modified state". I cannot process two events at a time, because first event can modify the state and the subsequent event should see this. It is not a "simple" stateful pardo either, because the state can be modified externally (not going into too much detail here, but e.g. by writing into kafka topic). Reuven's explanation is missing some detail. If the CoGBK is in discarding mode, then it will miss join results. If the CoGBK is in accumulating mode, it will duplicate join results. This is a known problem and the general solution is retractions. Basically, CoGBK-based joins just don't work with triggers until we have retractions.Moreover, I'd like to define the join semantics so that when there are available elements from both sides, the fired pane should be ON_TIME, not EARLY. That essentially means that the fully general case would not be built around (Co)GBK, but stateful ParDo. There are specific options where this fully general case "degrades" into forms that can be efficiently expressed using (Co)GBK, that is true. BTW building this around stateful DoFn might be a better fit. The main reason I didn't is because we would need a good distributed MapState (something discussed fairly recently on the list), and that is not yet built. Once we had that, I might be inclined to rewrite this join on stateful DoFn.Yes, the sorted state helps for streaming case. But I'd be careful about that for batch case, where this might lead to high pressure on the state (and InMemoryStateInternals might OOME for instance).However can you explain what you are expecting from the pane? An EARLY pane simply means that we are producing output before the end of the window. If you are in the global window triggering every element, then every output is EARLY. It might seem weird if you are interpreting EARLY as "outputting data that isn't ready," however that's not what EARLY is defined to be. Any change to the pane semantics would be a major breaking change to very fundamental semantics. I wonder if you are really objecting to the name EARLY and ON_TIME? Maybe we would've been better off tagging it BEFORE_WINDOW_END instead of EARLY, to make it clear what is meant?Essentially I don't object anything here. I'm missing solution to the "event vs. state" join described above. I was thinking about how to make these types of problems more user friendly and it essentially leads to creating a somewhat more generic semantics of join, where end-of-window is converted into "'value-delete events" and then just joining by the "previous" or "valid" value (yes, this relates to validity windows mentioned on Beam Summit Europe). It actually turns out that with some work we could define quite "naturally" a join on two streams with global window and no trigger. It would even function with lowest latency possible (but yes, with the highest expenses, it is actually the introduction of (same!) windows that enable certain optimizations). It the correctly defines semantics for different windows, although the result would be (probably unexpectedly) windowed using global window. But that doesn't seem to be any breaking change, because it is currently not possible (any such pipeline will not be validated). Maybe for reference, the unwindowed join would be what is described here [1] [1] https://cwiki.apache.org/confluence/display/KAFKA/Kafka+Streams+Join+Semantics#KafkaStreamsJoinSemantics-KStream-KTableJoinJan On 11/22/19 6:47 PM, Reuven Lax wrote:Have you seen the Join library that is part of schemas? I'm curious whether this fits your needs, or there's something lacking there. On Fri, Nov 22, 2019 at 12:31 AM Jan Lukavský <je...@seznam.cz <mailto:je...@seznam.cz>> wrote: Hi, based on roadmap [1], we would like to define and implement a full set of (unified) stream-stream joins. That would include: - joins (left, right, full outer) on global window with "immediate trigger" - joins with different windowing functions on left and right side The approach would be to define these operations in a natural way, so that the definition is aligned with how current joins work (same windows, cartesian product of values with same keys, output timestamp projected to the end of window, etc.). Because this should be a generic approach, this effort should probably be part of join library, that can the be reused by other components, too (e.g. SQL). The question is - is (or was) there any effort that we can build upon? Or should this be designed from scratch? Jan [1] https://beam.apache.org/roadmap/euphoria/
