Hey there, Ivan! In typical fashion, I'm going to make a somewhat outlandish proposal. I'm hoping that we can side-step some of the complications that have arisen. Please bear with me.
It seems like `distinct()` is not fundamentally unlike other windowed "aggregation" operations. Your concern about unnecessary repartitioning seems to apply just as well to `count()` as to `distinct()`. This has come up before, but I don't remember when: what if we introduce a parameter to `selectKey()` that specifies that the caller asserts that the new key does _not_ change the data partitioning? The docs on that parameter would of course spell out all the "rights and responsibilities" of setting it. In that case, we could indeed get back to `selectKey(A).windowBy(B).distinct(...)`, where we get to compose the key mapper and the windowing function without having to carve out a separate domain just for `distinct()`. All the rest of the KStream operations would also benefit. What do you think? Thanks, John On Sun, May 23, 2021, at 08:09, Ivan Ponomarev wrote: > Hello everyone, > > let me revive the discussion for KIP-655. Now I have some time again and > I'm eager to finalize this. > > Based on what was already discussed, I think that we can split the > discussion into three topics for our convenience. > > The three topics are: > > - idExtractor (how should we extract the deduplication key for the record) > > - timeWindows (what time windows should we use) > > - miscellaneous (naming etc.) > > ---- idExtractor ---- > > Original proposal: use (k, v) -> f(k, v) mapper, defaulting to (k, v) -> > k. The drawback here is that we must warn the user to choose such a > function that sets different IDs for records from different partitions, > otherwise same IDs might be not co-partitioned (and not deduplicated as > a result). Additional concern: what should we do when this function > returns null? > > Matthias proposed key-only deduplication: that is, no idExtractor at > all, and if we want to use `distinct` for a particular identifier, we > must `selectKey()` before. The drawback of this approach is that we will > always have repartitioning after the key selection, while in practice > repartitioning will not always be necessary (for example, when the data > stream is such that different values infer different keys). > > So here we have a 'safety vs. performance' trade-off. But 'safe' variant > is also not very convenient for developers, since we're forcing them to > change the structure of their records. > > A 'golden mean' here might be using composite ID with its first > component equals to k and its second component equals to some f(v) (f > defaults to v -> null, and null value returned by f(v) means > 'deduplicate by the key only'). The nuance here is that we will have > serializers only for types of k and f(v), and we must correctly > serialize a tuple (k, f(v)), but of course this is doable. > > What do you think? > > ---- timeWindows ---- > > Originally I proposed TimeWindows only just because they solved my > particular case :-) but agree with Matthias' and Sophie's objections. > > I like the Sophie's point: we need both epoch-aligned and data-aligned > windows. IMO this is absolutely correct: "data-aligned is useful for > example when you know that a large number of updates to a single key > will occur in short bursts, and epoch-aligned when you specifically want > to get just a single update per discrete time interval." > > I just cannot agree right away with Sophie's > .groupByKey().windowedBy(...).distinct() proposal, as it implies the > key-only deduplication -- see the previous topic. > > Epoch-aligned windows are very simple: they should forward only one > record per enumerated time window. TimeWindows are exactly what we want > here. I mentioned in the KIP both tumbling and hopping windows just > because both are possible for TimeWindows, but indeed I don't see any > real use case for hopping windows, only tumbling windows make sence IMO. > > For data-aligned windows SlidingWindow interface seems to be a nearly > valid choice. Nearly. It should forward a record once when it's first > seen, and then not again for any identical records that fall into the > next N timeUnits. However, we cannot reuse SlidingWindow as is, because > just as Matthias noted, SlidingWindows go backward in time, while we > need a windows that go forward in time, and are not opened while records > fall into an already existing window. We definitely should make our own > implementation, maybe we should call it ExpirationWindow? WDYT? > > > ---- miscellaneous ---- > > Persistent/in-memory stores. Matthias proposed to pass Materialized > parameter next to DistinctParameters (and this is necessary, because we > will need to provide a serializer for extracted id). This is absolutely > valid point, I agree and I will fix it in the KIP. > > Naming. Sophie noted that the Streams DSL operators are typically named > as verbs, so she proposes `deduplicate` in favour of `distinct`. I think > that while it's important to stick to the naming conventions, it is also > important to think of the experience of those who come from different > stacks/technologies. People who are familiar with SQL and Java Streams > API must know for sure what does 'distinct' mean, while data > deduplication in general is a more complex task and thus `deduplicate` > might be misleading. But I'm ready to be convinced if the majority > thinks otherwise. > > > Regards, > > Ivan > > > > 14.09.2020 21:31, Sophie Blee-Goldman пишет: > > Hey all, > > > > I'm not convinced either epoch-aligned or data-aligned will fit all > > possible use cases. > > Both seem totally reasonable to me: data-aligned is useful for example when > > you know > > that a large number of updates to a single key will occur in short bursts, > > and epoch- > > aligned when you specifically want to get just a single update per discrete > > time > > interval. > > > > Going a step further, though, what if you want just a single update per > > calendar > > month, or per year with accounting for leap years? Neither of those are > > serviced that > > well by the existing Windows specification to windowed aggregations, a > > well-known > > limitation of the current API. There is actually a KIP > > <https://cwiki.apache.org/confluence/display/KAFKA/KIP-645%3A+Replace+Windows+with+a+proper+interface> > > going > > on in parallel to fix this > > exact issue and make the windowing interface much more flexible. Maybe > > instead > > of re-implementing this windowing interface in a similarly limited fashion > > for the > > Distinct operator, we could leverage it here and get all the benefits > > coming with > > KIP-645. > > > > Specifically, I'm proposing to remove the TimeWindows/etc config from the > > DistinctParameters class, and move the distinct() method from the KStream > > interface > > to the TimeWindowedKStream interface. Since it's semantically similar to a > > kind of > > windowed aggregation, it makes sense to align it with the existing windowing > > framework, ie: > > > > inputStream > > .groupKyKey() > > .windowedBy() > > .distinct() > > > > Then we could use data-aligned windows if SlidingWindows is specified in > > the > > windowedBy(), and epoch-aligned (or some other kind of enumerable window) > > if a Windows is specified in windowedBy() (or an EnumerableWindowDefinition > > once KIP-645 is implemented to replace Windows). > > > > *SlidingWindows*: should forward a record once when it's first seen, and > > then not again > > for any identical records that fall into the next N timeUnits. This > > includes out-of-order > > records, ie if you have a SlidingWindows of size 10s and process records at > > time > > 15s, 20s, 14s then you would just forward the one at 15s. Presumably, if > > you're > > using SlidingWindows, you don't care about what falls into exact time > > boxes, you just > > want to deduplicate. If you do care about exact time boxing then you should > > use... > > > > *EnumerableWindowDefinition* (eg *TimeWindows*): should forward only one > > record > > per enumerated time window. If you get a records at 15s, 20s,14s where the > > windows > > are enumerated at [5,14], [15, 24], etc then you forward the record at 15s > > and also > > the record at 14s > > > > Just an idea: not sure if the impedance mismatch would throw users off > > since the > > semantics of the distinct windows are slightly different than in the > > aggregations. > > But if we don't fit this into the existing windowed framework, then we > > shouldn't use > > any existing Windows-type classes at all, imo. ie we should create a new > > DistinctWindows config class, similar to how stream-stream joins get their > > own > > JoinWindows class > > > > I also think that non-windowed deduplication could be useful, in which case > > we > > would want to also have the distinct() operator on the KStream interface. > > > > > > One quick note regarding the naming: it seems like the Streams DSL operators > > are typically named as verbs rather than adjectives, for example. #suppress > > or > > #aggregate. I get that there's some precedent for 'distinct' specifically, > > but > > maybe something like 'deduplicate' would be more appropriate for the Streams > > API. > > > > WDYT? > > > > > > On Mon, Sep 14, 2020 at 10:04 AM Ivan Ponomarev <iponoma...@mail.ru.invalid> > > wrote: > > > >> Hi Matthias, > >> > >> Thanks for your review! It made me think deeper, and indeed I understood > >> that I was missing some important details. > >> > >> To simplify, let me explain my particular use case first so I can refer > >> to it later. > >> > >> We have a system that collects information about ongoing live sporting > >> events from different sources. The information sources have their IDs > >> and these IDs are keys of the stream. Each source emits messages > >> concerning sporting events, and we can have many messages about each > >> sporing event from each source. Event ID is extracted from the message. > >> > >> We need a database of event IDs that were reported at least once by each > >> source (important: events from different sources are considered to be > >> different entities). The requirements are: > >> > >> 1) each new event ID should be written to the database as soon as possible > >> > >> 2) although it's ok and sometimes even desired to repeat the > >> notification about already known event ID, but we wouldn’t like our > >> database to be bothered by the same event ID more often than once in a > >> given period of time (say, 15 minutes). > >> > >> With this example in mind let me answer your questions > >> > >> > (1) Using the `idExtractor` has the issue that data might not be > >> > co-partitioned as you mentioned in the KIP. Thus, I am wondering if it > >> > might be better to do deduplication only on the key? If one sets a new > >> > key upstream (ie, extracts the deduplication id into the key), the > >> > `distinct` operator could automatically repartition the data and thus > >> we > >> > would avoid user errors. > >> > >> Of course with 'key-only' deduplication + autorepartitioning we will > >> never cause problems with co-partitioning. But in practice, we often > >> don't need repartitioning even if 'dedup ID' is different from the key, > >> like in my example above. So here we have a sort of 'performance vs > >> security' tradeoff. > >> > >> The 'golden middle way' here can be the following: we can form a > >> deduplication ID as KEY + separator + idExtractor(VALUE). In case > >> idExtractor is not provided, we deduplicate by key only (as in original > >> proposal). Then idExtractor transforms only the value (and not the key) > >> and its result is appended to the key. Records from different partitions > >> will inherently have different deduplication IDs and all the data will > >> be co-partitioned. As with any stateful operation, we will repartition > >> the topic in case the key was changed upstream, but only in this case, > >> thus avoiding unnecessary repartitioning. My example above fits this > >> perfectly. > >> > >> > (2) What is the motivation for allowing the `idExtractor` to return > >> > `null`? Might be good to have some use-case examples for this feature. > >> > >> Can't think of any use-cases. As it often happens, it's just came with a > >> copy-paste from StackOverflow -- see Michael Noll's answer here: > >> > >> https://stackoverflow.com/questions/55803210/how-to-handle-duplicate-messages-using-kafka-streaming-dsl-functions > >> > >> But, jokes aside, we'll have to decide what to do with nulls. If we > >> accept the above proposal of having deduplication ID as KEY + postfix, > >> then null can be treated as no postfix at all. If we don't accept this > >> approach, then treating nulls as 'no-deduplication' seems to be a > >> reasonable assumption (we can't get or put null as a key to a KV store, > >> so a record with null ID is always going to look 'new' for us). > >> > >> > >> > (2) Is using a `TimeWindow` really what we want? I was wondering if a > >> > `SlidingWindow` might be better? Or maybe we need a new type of window? > >> > >> Agree. It's probably not what we want. Once I thought that reusing > >> TimeWindow is a clever idea, now I don't. > >> > >> Do we need epoch alignment in our use case? No, we don't, and I don't > >> know if anyone going to need this. Epoch alignment is good for > >> aggregation, but deduplication is a different story. > >> > >> Let me describe the semantic the way I see it now and tell me what you > >> think: > >> > >> - the only parameter that defines the deduplication logic is 'expiration > >> period' > >> > >> - when a deduplication ID arrives and we cannot find it in the store, we > >> forward the message downstream and store the ID + its timestamp. > >> > >> - when an out-of-order ID arrives with an older timestamp and we find a > >> 'fresher' record, we do nothing and don't forward the message (??? OR > >> NOT? In what case would we want to forward an out-of-order message?) > >> > >> - when an ID with fresher timestamp arrives we check if it falls into > >> the expiration period and either forward it or not, but in both cases we > >> update the timestamp of the message in the store > >> > >> - the WindowStore retention mechanism should clean up very old records > >> in order not to run out of space. > >> > >> > (3) `isPersistent` -- instead of using this flag, it seems better to > >> > allow users to pass in a `Materialized` parameter next to > >> > `DistinctParameters` to configure the state store? > >> > >> Fully agree! Users might also want to change the retention time. > >> > >> > (4) I am wondering if we should really have 4 overloads for > >> > `DistinctParameters.with()`? It might be better to have one overload > >> > with all require parameters, and add optional parameters using the > >> > builder pattern? This seems to follow the DSL Grammer proposal. > >> > >> Oh, I can explain. We can't fully rely on the builder pattern because of > >> Java type inference limitations. We have to provide type parameters to > >> the builder methods or the code won't compile: see e. g. this > >> https://twitter.com/inponomarev/status/1265053286933159938 and following > >> discussion with Tagir Valeev. > >> > >> When we came across the similar difficulties in KIP-418, we finally > >> decided to add all the necessary overloads to parameter class. So I just > >> reproduced that approach here. > >> > >> > (5) Even if it might be an implementation detail (and maybe the KIP > >> > itself does not need to mention it), can you give a high level overview > >> > how you intent to implement it (that would be easier to grog, compared > >> > to reading the PR). > >> > >> Well as with any operation on KStreamImpl level I'm building a store and > >> a processor node. > >> > >> KStreamDistinct class is going to be the ProcessorSupplier, with the > >> logic regarding the forwarding/muting of the records located in > >> KStreamDistinct.KStreamDistinctProcessor#process > >> > >> ---- > >> > >> Matthias, if you are still reading this :-) a gentle reminder: my PR for > >> already accepted KIP-418 is still waiting for your review. I think it's > >> better for me to finalize at least one KIP before proceeding to a new > >> one :-) > >> > >> Regards, > >> > >> Ivan > >> > >> 03.09.2020 4:20, Matthias J. Sax пишет: > >>> Thanks for the KIP Ivan. Having a built-in deduplication operator is for > >>> sure a good addition. > >>> > >>> Couple of questions: > >>> > >>> (1) Using the `idExtractor` has the issue that data might not be > >>> co-partitioned as you mentioned in the KIP. Thus, I am wondering if it > >>> might be better to do deduplication only on the key? If one sets a new > >>> key upstream (ie, extracts the deduplication id into the key), the > >>> `distinct` operator could automatically repartition the data and thus we > >>> would avoid user errors. > >>> > >>> (2) What is the motivation for allowing the `idExtractor` to return > >>> `null`? Might be good to have some use-case examples for this feature. > >>> > >>> (2) Is using a `TimeWindow` really what we want? I was wondering if a > >>> `SlidingWindow` might be better? Or maybe we need a new type of window? > >>> > >>> It would be helpful if you could describe potential use cases in more > >>> detail. -- I am mainly wondering about hopping window? Each record would > >>> always falls into multiple window and thus would be emitted multiple > >>> times, ie, each time the window closes. Is this really a valid use case? > >>> > >>> It seems that for de-duplication, one wants to have some "expiration > >>> time", ie, for each ID, deduplicate all consecutive records with the > >>> same ID and emit the first record after the "expiration time" passed. In > >>> terms of a window, this would mean that the window starts at `r.ts` and > >>> ends at `r.ts + windowSize`, ie, the window is aligned to the data. > >>> TimeWindows are aligned to the epoch though. While `SlidingWindows` also > >>> align to the data, for the aggregation use-case they go backward in > >>> time, while we need a window that goes forward in time. It's an open > >>> question if we can re-purpose `SlidingWindows` -- it might be ok the > >>> make the alignment (into the past vs into the future) an operator > >>> dependent behavior? > >>> > >>> (3) `isPersistent` -- instead of using this flag, it seems better to > >>> allow users to pass in a `Materialized` parameter next to > >>> `DistinctParameters` to configure the state store? > >>> > >>> (4) I am wondering if we should really have 4 overloads for > >>> `DistinctParameters.with()`? It might be better to have one overload > >>> with all require parameters, and add optional parameters using the > >>> builder pattern? This seems to follow the DSL Grammer proposal. > >>> > >>> (5) Even if it might be an implementation detail (and maybe the KIP > >>> itself does not need to mention it), can you give a high level overview > >>> how you intent to implement it (that would be easier to grog, compared > >>> to reading the PR). > >>> > >>> > >>> > >>> -Matthias > >>> > >>> On 8/23/20 4:29 PM, Ivan Ponomarev wrote: > >>>> Sorry, I forgot to add [DISCUSS] tag to the topic > >>>> > >>>> 24.08.2020 2:27, Ivan Ponomarev пишет: > >>>>> Hello, > >>>>> > >>>>> I'd like to start a discussion for KIP-655. > >>>>> > >>>>> KIP-655: > >>>>> > >> https://cwiki.apache.org/confluence/display/KAFKA/KIP-655%3A+Windowed+Distinct+Operation+for+Kafka+Streams+API > >>>>> > >>>>> > >>>>> I also opened a proof-of-concept PR for you to experiment with the API: > >>>>> > >>>>> PR#9210: https://github.com/apache/kafka/pull/9210 > >>>>> > >>>>> Regards, > >>>>> > >>>>> Ivan Ponomarev > >>>> > >>> > >> > >> > > > >
