Hi Jark Wu, I agree about the non-windowed aggregates. If there are actual use cases for this operator, we should definitely support it. Since it does not depend on windows or time, we can immediately start to work on it. In principle, it should be rather easy to implement. However, we have to check how well it integrates with the current state of Calcite.
I think forking off a feature branch is a good idea. We have done that before (e.g., for porting the Table API on top of Calcite), but it is not so common in the Flink community. So I would first send a note to the dev list and check that nobody objects. I think we can decouple the development of the Table API and SQL. Although it is desirable to have the same feature set in both APIs, I would not be strict about it. However, the Table API does also depend on Calcite because all Table API queries go through Calcite's logical plan representation and optimizer. By decoupling the SQL and Table API feature development, we do not need to wait for the SQL parser but still might still need certain features in the logical plan or optimizer. I hope we can solve a lot with custom RelNodes and optimizer rules which should eventually be contributed back to Calcite. Best, Fabian 2016-06-17 9:48 GMT+02:00 Jark Wu <wuchong...@alibaba-inc.com>: > Hi Fabian, > > There are a lot of our business are using non-windowed aggregations. And > there is a little difference between non-windowed aggregate and Row window > operator, as the later is bound to a certain window and emit the result of > the N rows preceding for every incoming row. However the former emit the > aggregate result of the whole elements. So I suggest to add them for more > complete semantic. > > Regarding the windowed aggregate task, I’m agree with that and I'm looking > forward as soon as possible to see the corresponding JIRA issues created. > After that, we can start working on an independent branch without waiting > for 1.1 released. But I’m still a little concerned about Calcite’s support, > as we must waiting for Calcite supporting correspond syntax and the > version released. If we can separate the task into Table API and SQL , we > may not be blocked by Calcite too much. > > What do you think? > > - Jark Wu > > > 在 2016年6月16日，下午8:37，Fabian Hueske <fhue...@gmail.com> 写道： > > > > Hi Jark, > > > > thanks for sharing Blink's Streaming Table API. It seems to be close to > the > > DataStream API, while the Table API draft I shared is more similar to > > Calcite's proposal. > > You are right, the current draft does not include running (non-windowed) > > aggregates. We were not sure how useful these are since these aggregates > > are unbound and might become meaningless after being applied on a very > long > > stream. However, we can certainly add them, if users request them. > > An alternative to running aggregates could be what I called "Row window > > operators" in the streaming Table API draft. These operators emit an > > aggregate for each incoming row, however the aggregate is bound to a > > certain window around the row like the 10 rows preceding the row for > which > > the aggregate is computed. Calcite calls these windows "Sliding windows" > > (Attention: This is different from Flink's terminology, in Flink sliding > > windows are something different). Row windows are similar to running > > aggregates in that they emit a row for each incoming row. You can also > > think of them as a (Flink) sliding count window which is evaluate for > each > > incoming record. > > > > Further differences are the support of Scalar UDFs in the Table API and > the > > support for joins which have not been drafted for the Table API yet. > > Scalar UDFs are definitely also on our roadmap and with upcoming support > > for side inputs, the DataStream API will also support more types of > joins. > > > > Regarding the current state of Stream SQL in Calcite I am not up to date. > > > > I would propose to start with the effort of adding support for windowed > > aggregates as follows: > > > > 1) Add support to define a timestamp / watermark extractor to tables. > This > > includes to define a "quasi-monotone" column in a Table's schema. Calcite > > will use this information to reason about the validity of a query (making > > sure that grouping includes at least one monotone attribute). > > 2) Add support for sorting a stream on the timestamp attribute. While > > sorting itself is not very exciting, it is an easy operation and can be > > immediately implemented without worrying about API questions. This will > > also show how well Calcite supports the reasoning about monotone > attributes. > > 3) Add support for tumbling windows. > > > > In each of these steps we might need to get involved with the Calcite > > community, depending on Calcite's current support for "quasi-monotone" > > attributes, etc. > > > > What do you think? > > > > Best, Fabian > > > > > > 2016-06-14 11:03 GMT+02:00 Jark Wu <wuchong...@alibaba-inc.com>: > > > >> Hi Fabian, > >> > >> It’s great to hear that we are going to start it! > >> > >> I’m glad to share our current Streaming Table API [1]. I find that that > >> all aggregation functions are scoped to the defined window in Flink > Stream > >> Table API design [2] and Calcite StreamSQL desgin [3]. I’m thinking > that do > >> we need global aggregation? The global aggregation means that > aggregation > >> is applied only on grouped key not including window which is supported > in > >> DataStream `datastream.keyBy(f1).sum(f2)`. > >> > >> Since the window syntax of StreamSQL is not implemented yet, so will we > >> help Calcite community with that first or work code for window+agg Table > >> API first ? > >> > >> > >> [1] > >> > https://docs.google.com/document/d/1KMUzvBAWSyQ39T8MyxUi0zNHyvLUnyGMPA7_RLSDpFw/edit?usp=sharing > >> < > >> > https://docs.google.com/document/d/1KMUzvBAWSyQ39T8MyxUi0zNHyvLUnyGMPA7_RLSDpFw/edit?usp=sharing > >>> > >> [2] > >> > https://docs.google.com/document/d/19kSOAOINKCSWLBCKRq2WvNtmuaA9o3AyCh2ePqr3V5E/edit# > >> < > >> > https://docs.google.com/document/d/19kSOAOINKCSWLBCKRq2WvNtmuaA9o3AyCh2ePqr3V5E/edit# > >>> > >> [3] https://calcite.apache.org/docs/stream.html#tumbling-windows < > >> https://calcite.apache.org/docs/stream.html#tumbling-windows> > >> > >> > >> - Jark Wu > >> > >>> 在 2016年6月14日，上午1:10，Fabian Hueske <fhue...@gmail.com> 写道： > >>> > >>> Hi Jark, > >>> > >>> wow, that's good news! > >>> You are right, the streaming Table API is currently very limited. In > the > >>> last month's we changed the internal architecture and put everything on > >> top > >>> of Apache Calcite. > >>> For the upcoming 1.1 release, we won't add new features to the Table > API > >> / > >>> SQL. However for the 1.2 release, it we plan to focus on the streaming > >>> Table API and Stream SQL to add support for windowed aggregates and > >> joins, > >>> which corresponds to Task 7 and 9 in the design document. You are > >>> completely right, that we should start to add tickets to JIRA for > this. I > >>> will do that tomorrow. > >>> > >>> It is great that you have already working code for windowed aggregates > >> and > >>> joins! Here is a link to our current API draft for windows in the Table > >> API > >>> [1]. Would be great if you could share how your API looks like. As you > >>> said, the internals have changed a lot by now, but we might want to > reuse > >>> your API for Table API windows and maybe the code of the runtime. > >> However, > >>> we need to go through Calcite for optimization and SQL support, so some > >>> parts need to be definitely changed. Stream SQL is also on the roadmap > of > >>> the Calcite community, but it might be that some features that we will > >> need > >>> are not completed yet. So, maybe we help the Calcite community with > that > >> as > >>> well. > >>> > >>> If you want to contribute, you should first read the how to contribute > >>> guide [2] and guide for code contributions [3]. > >>> The general rule is to first open a JIRA and later a pull request. > >> Changes > >>> that are extensive or modify current behavior (except bugs) should be > >>> discussed before starting to work on them. > >>> > >>> Looking forward to work with you on Flink, > >>> Fabian > >>> > >>> [1] > >>> > >> > https://docs.google.com/document/d/19kSOAOINKCSWLBCKRq2WvNtmuaA9o3AyCh2ePqr3V5E/edit#heading=h.3iw7frfjdcb2 > >>> [2] http://flink.apache.org/how-to-contribute.html > >>> [3] http://flink.apache.org/contribute-code.html > >>> > >>> > >>> 2016-06-13 11:31 GMT+02:00 Jark Wu <wuchong...@alibaba-inc.com>: > >>> > >>>> Hi, > >>>> > >>>> We have a great interest in the new Table API & SQL. In Alibaba, we > have > >>>> added a lot of features (groupBy, agg, window, join, UDF …) to > Streaming > >>>> Table API (base on Flink 1.0). Now, many jobs run on Table API in > >>>> production environment. But we want to keep pace with the community, > >> and we > >>>> have noticed that Flink Community reworked the Table API and also > >> supported > >>>> SQL. That is really cool. However, the Streaming Table API is still so > >>>> weak. So we want to contribute to accelerate the Streaming Table API > and > >>>> StreamSQL growth. > >>>> > >>>> It seems that we have complete Task-5 and Task-6 mentioned in the Work > >>>> Plan < > >>>> > >> > https://docs.google.com/document/d/1TLayJNOTBle_-m1rQfgA6Ouj1oYsfqRjPcp1h2TVqdI/edit# > >>> . > >>>> So can we start Task-7 and Task-9 now? Is there any more specific > >> plans? I > >>>> think it’s better to create an umbrella JIRA like FLINK-3221 to make > the > >>>> develop plan clearer. > >>>> > >>>> If I want to contribute code for groupBy and agg function, what > should I > >>>> do? As I didn’t find related JIRAs, can I create JIRA and pull a > request > >>>> directly? > >>>> > >>>> Sorry for so many questions at a time. > >>>> > >>>> > >>>> > >>>> - Jark Wu (wuchong) > >>>> > >>>> > >> > >> > >
