Hi, @Kurt @Jark thanks again for your comments.
@Kurt Separating key and non-key for UDTAGG can definitely provide more information for the system, however, it will also add more burden to our users and bring some code reuse problems. BTW, approach 3 can also be used to separate UDTAGG into keyed or non-keyed as we can check whether the key list is empty. So from this point of view, we can use approach 3 to solve your problem. @Jark It's great that the TopN in Blink can decide the key automatically. But, I'd like to point out another case that the keys cannot be decided by the system, i.e., can only be decided by the user. For example, for the TopN, let's say top1 for better understanding. Support the Top1 outputs three columns(rankid, value, seller_name), and the user wants to upsert the result either with key of rankid or with the key of rankid+seller_name. 1. With the key of rankid: In this case, the user just wants to get the top 1 record. 2. With the key of rankid+seller_name: In this case, the user wants to get all seller_names that have ever belong to top1. This can not be solved by the approach 3 if using only one function. However, it is very easy to implement with the withKey approach. Even though, I have thought more clearly about these things and find more interesting things that I want to share with you all. For the TopN example which I listed above, it may also lead to a problem in which batch and streaming are not unified. To make it worse, the upsert sink is not supported in batch and we even don't have any clear implementation plan about how to support upsert on the batch, the unification problem for `withKeys` approach becomes hang in doubt. So, to avoid the unification problem, I think we can also use the approach 3. It is more rigorous although less flexible compared to the `withKeys` approach. Meanwhile, I will think more about the unification problem later. Maybe new ideas about it may come through. :) Best， Jincheng Jark Wu <imj...@gmail.com> 于2019年7月5日周五 上午10:48写道： > Hi Hequn, > > > If the TopN table aggregate function > > outputs three columns(rankid, time, value), either rankid or rankid+time > could be > > used as the key. Which one to be chosen is more likely to be decided by > the user > > according to his business. > In this case, the TopN table aggregate function should return two sets of > unique key, one is "rankid", another is "rankid, time". > This will be more align with current TopN node in blink planner and let > optimizer to decide which key based on the downstream information (column > selection, sink's primary key). > > > Best, > Jark > > On Fri, 5 Jul 2019 at 00:05, Hequn Cheng <chenghe...@gmail.com> wrote: > >> Hi Kurt and Jark, >> >> Thanks a lot for your great inputs! >> >> The keys of the query may not strongly be related to the UDTAGG. >> It may also be related to the corresponding scenarios that a user wants >> to achieve. >> >> For example, take TopN again as an example. If the TopN table aggregate >> function >> outputs three columns(rankid, time, value), either rankid or rankid+time >> could be >> used as the key. Which one to be chosen is more likely to be decided by >> the user >> according to his business. >> >> Best, Hequn >> >> On Thu, Jul 4, 2019 at 8:11 PM Jark Wu <imj...@gmail.com> wrote: >> >>> Hi jingcheng, >>> >>> I agree with Kurt's point. As you said "the user must know the keys of >>> the output of UDTAGG clearly". >>> If I understand correctly, the key information is strongly relative to >>> the UDTAGG implementation. >>> Users may call `flatAggregate` on a UDTAGG instance with different keys >>> which may result in a wrong result. >>> So I think it would be better to couple key information with UDTAGG >>> interface (i.e. "Approach 3" in your design doc). >>> >>> Regards, >>> Jark >>> >>> On Thu, 4 Jul 2019 at 18:06, Kurt Young <ykt...@gmail.com> wrote: >>> >>>> Hi Jincheng, >>>> >>>> Thanks for the clarification. I think you just pointed out my concern by >>>> yourself: >>>> >>>> > When a user uses a User-defined table aggregate function (UDTAGG), he >>>> must understand the behavior of the UDTAGG, including the return type >>>> and >>>> the characteristics of the returned data. such as the key fields. >>>> >>>> This indicates that the UDTAGG is somehow be classified to different >>>> types, >>>> one will no key, one with key information. So the developer of the >>>> UDTAGG >>>> should choose which type of this function should be. In this case, >>>> my question would be, why don't we have explicit information about keys >>>> such as we split UDTAGG to keyed UDTAGG and non-keyed UDTAGG. So the >>>> user >>>> and the framework will have a better understanding of >>>> this UDTAGG. `withKeys` solution is letting user to choose the key and >>>> it >>>> seems it will only work correctly only if the user choose the *right* >>>> key >>>> this UDTAGG has. >>>> >>>> Let me know if this makes sense to you. >>>> >>>> Best, >>>> Kurt >>>> >>>> >>>> On Thu, Jul 4, 2019 at 4:32 PM jincheng sun <sunjincheng...@gmail.com> >>>> wrote: >>>> >>>> > Hi All, >>>> > >>>> > @Kurt Young <ykt...@gmail.com> one user-defined table aggregate >>>> function >>>> > can be used in both with(out) keys case, and we do not introduce any >>>> other >>>> > aggregations. just like the explanation from @Hequn. >>>> > >>>> > @Hequn Cheng <chenghe...@gmail.com> thanks for your explanation! >>>> > >>>> > One thing should be mentioned here: >>>> > >>>> > When a user uses a User-defined table aggregate function (UDTAGG), he >>>> must >>>> > understand the behavior of the UDTAGG, including the return type and >>>> the >>>> > characteristics of the returned data. such as the key fields. So >>>> although >>>> > using `withKeys` approach it is not rigorous enough(we do not need) >>>> but >>>> > intuitive enough, considering that if `flatAggregate` is followed by >>>> an >>>> > `upsertSink`, then the user must know the keys of the output of UDTAGG >>>> > clearly, otherwise the keys of `upsertSink` cannot be defined. So I >>>> still >>>> > prefer the `withKeys` solution by now. >>>> > >>>> > Looking forward to any feedback from all of you! >>>> > >>>> > Best, >>>> > Jincheng >>>> > >>>> > >>>> > >>>> > Hequn Cheng <chenghe...@gmail.com> 于2019年7月1日周一 下午5:35写道： >>>> > >>>> >> Hi Kurt, >>>> >> >>>> >> Thanks for your questions. Here are my thoughts. >>>> >> >>>> >> > if I want to write such kind function, should I make sure that this >>>> >> function is used with some keys? >>>> >> The key information may not be used. We can also use RetractSink to >>>> emit >>>> >> the table directly. >>>> >> >>>> >> > If I need a use case to calculate topn without key, should I write >>>> >> another function or I can reuse previous one. >>>> >> For the TopN example, you can reuse the previous function if you >>>> don't >>>> >> care >>>> >> about the key information. >>>> >> >>>> >> So, the key information is only an indicator(or a description), not >>>> an >>>> >> operator, as Jincheng mentioned above. >>>> >> We do not need to change the function logic and it will not add any >>>> other >>>> >> aggregations. >>>> >> >>>> >> BTW, we have three approaches in the document. Approach 1 defines >>>> keys on >>>> >> API level as we think it's more common to define keys on Table. >>>> >> While approach 3 defines keys in the TableAggregateFunction which is >>>> more >>>> >> precise but it is not very clear for Table users. So, we should take >>>> all >>>> >> these into consideration, and make the decision in this discussion >>>> thread. >>>> >> >>>> >> You can take a look at the document and welcome any suggestions or >>>> other >>>> >> better solutions. >>>> >> >>>> >> Best, Hequn >>>> >> >>>> >> >>>> >> On Mon, Jul 1, 2019 at 12:13 PM Kurt Young <ykt...@gmail.com> wrote: >>>> >> >>>> >> > Hi Jincheng, >>>> >> > >>>> >> > Thanks for the clarification. Take 'TopN' for example, if I want to >>>> >> write >>>> >> > such kind function, >>>> >> > should I make sure that this function is used with some keys? If I >>>> need >>>> >> a >>>> >> > use case to calculate >>>> >> > topn without key, should I write another function or I can reuse >>>> >> previous >>>> >> > one. >>>> >> > >>>> >> > I'm not sure about the idea of this does not involve semantic >>>> changes. >>>> >> To >>>> >> > me, it sounds like >>>> >> > we are doing another nested aggregation inside the table >>>> >> > which TableAggregateFunction emits. >>>> >> > >>>> >> > Maybe I'm not familiar with this function enough, hope you can >>>> help me >>>> >> to >>>> >> > understand. >>>> >> > >>>> >> > Best, >>>> >> > Kurt >>>> >> > >>>> >> > >>>> >> > On Mon, Jul 1, 2019 at 11:59 AM jincheng sun < >>>> sunjincheng...@gmail.com> >>>> >> > wrote: >>>> >> > >>>> >> > > Hi Kurt, >>>> >> > > >>>> >> > > Thanks for your questions, I am glad to share my thoughts here: >>>> >> > > >>>> >> > > My question is, will that effect the logic >>>> ofTableAggregateFunction >>>> >> user >>>> >> > > > wrote? Should the user know that there will a key and make some >>>> >> changes >>>> >> > > to >>>> >> > > > this function? >>>> >> > > >>>> >> > > >>>> >> > > No, the keys information depends on the implementation of the >>>> >> > > TableAggregateFunction. >>>> >> > > For example, for a `topN` user defined TableAggregateFunction, >>>> we can >>>> >> > only >>>> >> > > use the `keys` if the `topN` contains `rankid` in the output. >>>> You can >>>> >> > > treat the >>>> >> > > `keys` like an indicator. >>>> >> > > >>>> >> > > If not, how will framework deal with the output of user's >>>> >> > > > TableAggregateFunction. if user output multiple rows with the >>>> same >>>> >> > key, >>>> >> > > > should be latter one replace previous ones? >>>> >> > > >>>> >> > > >>>> >> > > If a TableAggregateFunction outputs multiple rows with the same >>>> key, >>>> >> the >>>> >> > > latter one should replace the previous one, either with upsert >>>> mode or >>>> >> > > retract mode. i.e., Whether the user defines the Key or not, the >>>> Flink >>>> >> > > framework should ensure the correctness of the semantics. >>>> >> > > >>>> >> > > At present, the problem we are discussing does not involve >>>> semantic >>>> >> > > changes. The definition of keys is to support non-window >>>> >> flatAggregate on >>>> >> > > upsert mode. (The upsert mode is already supported in the flink >>>> >> > framework. >>>> >> > > The current discussion only needs to inform the framework that >>>> the >>>> >> keys >>>> >> > > information, which is the `withKeys` API we discussing.) >>>> >> > > >>>> >> > > Welcome any other feedbacks :) >>>> >> > > >>>> >> > > Best, >>>> >> > > Jincheng >>>> >> > > >>>> >> > > Kurt Young <ykt...@gmail.com> 于2019年7月1日周一 上午9:23写道： >>>> >> > > >>>> >> > > > Hi, >>>> >> > > > >>>> >> > > > I have a question about the key information of >>>> >> TableAggregateFunction. >>>> >> > > > IIUC, you need to define >>>> >> > > > something like primary key or unique key in the result table of >>>> >> > > > TableAggregateFunction, and also >>>> >> > > > need a way to let user configure this through the API. My >>>> question >>>> >> is, >>>> >> > > will >>>> >> > > > that effect the logic of >>>> >> > > > TableAggregateFunction user wrote? Should the user know that >>>> there >>>> >> > will a >>>> >> > > > key and make some changes >>>> >> > > > to this function? >>>> >> > > > >>>> >> > > > If so, what's the semantic the user should learned. If not, >>>> how will >>>> >> > > > framework deal with the output of user's >>>> >> > > > TableAggregateFunction. For example, if user output multiple >>>> rows >>>> >> with >>>> >> > > the >>>> >> > > > same key, should be latter one >>>> >> > > > replace previous ones? >>>> >> > > > >>>> >> > > > Best, >>>> >> > > > Kurt >>>> >> > > > >>>> >> > > > >>>> >> > > > On Mon, Jul 1, 2019 at 7:19 AM jincheng sun < >>>> >> sunjincheng...@gmail.com> >>>> >> > > > wrote: >>>> >> > > > >>>> >> > > > > Hi hequn, Thanks for the reply! I think `withKeys` solution >>>> is our >>>> >> > > better >>>> >> > > > > choice! >>>> >> > > > > >>>> >> > > > > >>>> >> > > > > Hequn Cheng <chenghe...@gmail.com> 于2019年6月26日周三 下午5:11写道： >>>> >> > > > > >>>> >> > > > > > Hi Jincheng, >>>> >> > > > > > >>>> >> > > > > > Thanks for raising the discussion! >>>> >> > > > > > The key information is very important for query >>>> optimizations. >>>> >> It >>>> >> > > would >>>> >> > > > > be >>>> >> > > > > > nice if we can use upsert mode to achieve better >>>> performance. >>>> >> > > > > > >>>> >> > > > > > +1 for the `withKeys` proposal. :) >>>> >> > > > > > >>>> >> > > > > > Best, Hequn >>>> >> > > > > > >>>> >> > > > > > >>>> >> > > > > > On Wed, Jun 26, 2019 at 4:37 PM jincheng sun < >>>> >> > > sunjincheng...@gmail.com >>>> >> > > > > >>>> >> > > > > > wrote: >>>> >> > > > > > >>>> >> > > > > > > Hi all, >>>> >> > > > > > > >>>> >> > > > > > > With the continuous efforts from the community, we >>>> already >>>> >> > > supported >>>> >> > > > > > > `flatAggregate`[1] on TableAPI in retract mode. I think >>>> It's >>>> >> > better >>>> >> > > > to >>>> >> > > > > > add >>>> >> > > > > > > upsert mode for `flatAggregate`. >>>> >> > > > > > > >>>> >> > > > > > > The result table of streaming non-window `flatAggregate` >>>> is a >>>> >> > table >>>> >> > > > > > > contains updates. We can, of course, use a >>>> >> > > RetractStreamTableSink[2] >>>> >> > > > to >>>> >> > > > > > > emit the table, but we can get better performance in >>>> upsert >>>> >> mode. >>>> >> > > > > > However, >>>> >> > > > > > > due to the lack of keys, we can’t use an >>>> >> UpsertStreamTableSink to >>>> >> > > > emit >>>> >> > > > > > the >>>> >> > > > > > > table. We don’t have this problem for a normal aggregate >>>> as it >>>> >> > > emits >>>> >> > > > a >>>> >> > > > > > > single row for each group, so the unique keys are >>>> exactly the >>>> >> > same >>>> >> > > > with >>>> >> > > > > > the >>>> >> > > > > > > group keys. While for a `flatAggregate`, its pretty >>>> difference >>>> >> > that >>>> >> > > > due >>>> >> > > > > > to >>>> >> > > > > > > emits multi rows(a “sub-table”) for a single group. To >>>> solve >>>> >> this >>>> >> > > > > > problem, >>>> >> > > > > > > we need to find a way to define keys on flatAggregate, so >>>> >> that we >>>> >> > > can >>>> >> > > > > > also >>>> >> > > > > > > use upsert sink to emit the result table after >>>> flatAggregate. >>>> >> > > > > > > >>>> >> > > > > > > So, Aljoscha, Hequn and I prepared a design document for >>>> how >>>> >> to >>>> >> > > > define >>>> >> > > > > > the >>>> >> > > > > > > update keys for `flatAggregate` in upsert mode. The >>>> detail >>>> >> can >>>> >> > be >>>> >> > > > > found >>>> >> > > > > > > here: >>>> >> > > > > > > >>>> >> > > > > > > >>>> >> > > > > > > >>>> >> > > > > > >>>> >> > > > > >>>> >> > > > >>>> >> > > >>>> >> > >>>> >> >>>> https://docs.google.com/document/d/183qHo8PDG-xserDi_AbGP6YX9aPY0rVr80p3O3Gyz6U/edit?usp=sharing >>>> >> > > > > > > >>>> >> > > > > > > I appreciate it if you can have look at the document and >>>> any >>>> >> > > comments >>>> >> > > > > are >>>> >> > > > > > > welcome! >>>> >> > > > > > > >>>> >> > > > > > > >>>> >> > > > > > > Best, >>>> >> > > > > > > >>>> >> > > > > > > Jincheng >>>> >> > > > > > > >>>> >> > > > > > > >>>> >> > > > > > > [1] >>>> >> > > > > > > >>>> >> > > > > > >>>> >> > > > > >>>> >> > > > >>>> >> > > >>>> >> > >>>> >> >>>> https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=97552739 >>>> >> > > > > > > >>>> >> > > > > > > [2] >>>> >> > > > > > > >>>> >> > > > > > > >>>> >> > > > > > >>>> >> > > > > >>>> >> > > > >>>> >> > > >>>> >> > >>>> >> >>>> https://ci.apache.org/projects/flink/flink-docs-release-1.8/dev/table/sourceSinks.html#defining-a-streamtablesource >>>> >> > > > > > > >>>> >> > > > > > >>>> >> > > > > >>>> >> > > > >>>> >> > > >>>> >> > >>>> >> >>>> > >>>> >>>
