Hi Dawid, Piotr, Agree with you that add finish() method to `TableFunction` only. Other `UserDefinedFunction`s (`ScalarFunction`, `AggregateFunction`, `AggregateTableFunction`) are not necessarily to have the finish method(they can not emit records in legacy close() method).
A `TableFunction` is used to correlate with the left table/stream, the following example shows a case that user only select columns from the correlated 'FeatureTF' (no left table column was selected): ``` SELECT feature1, feature2, feature3 FROM MyTable t1 JOIN LATERAL TABLE(FeatureTF(t1.f0, t1.f1)) AS F(feature1, feature2, feature3) ON TRUE ``` the 'FeatureTF' can do some flushing work in legacy close() method and this doesn't break any sql semantics, so I don't see any reason that we can enforce users not do flushing work in new finish() method. I've updated the flip doc to limit the change only for `TableFunction`[1]. For the more powerful `ProcessFunction`, I'd like to share some thoughts: There indeed exists requirements for advanced usage in Table/SQL, even a further UD-Operator, e.g., UD-Join for user controlled join logic which can not simply expressed by SQL. This is an interesting topic, expect more discussions on this. [1] https://cwiki.apache.org/confluence/display/FLINK/FLIP-260%3A+Expose+Finish+Method+For+TableFunction Best, Lincoln Lee Piotr Nowojski <pnowoj...@apache.org> 于2022年9月15日周四 22:39写道： > Hi Dawid, Lincoln, > > I would tend to agree with Dawid. It seems to me like `TableFunction` is > the one that needs to be taken care of. Other types of > `UserDefinedFunction` wouldn't be able to emit anything from the `finish()` > even if we added it. And if we added `finish(Collector<T> out)` to them, it > would create the same problems (how to pass the output type) that prevented > us from adding `finish()` to all functions in the DataStream API. > > However I'm not sure what should be the long term solution for the Table > API. For the DataStream API we wanted to provide a new, better and more > powerful `ProcessFunction` for all of the unusual use cases, that currently > require the use of `StreamOperator` API instead of `DataStream` functions. > I don't know what would be an alternative in the Table API. > > Dawid, who do you think we should ping from the Table API/SQL teams to chip > in? > > Best, > Piotrek > > czw., 15 wrz 2022 o 12:38 Dawid Wysakowicz <dwysakow...@apache.org> > napisał(a): > > > Hey Lincoln, > > > > Thanks for opening the discussion. > > > > To be honest I am not convinced if emitting from close there is a > > contract that was envisioned and thus should be maintained. As far as I > > can see it does affect only the TableFunction, because it has the > > collect method. None of the other UDFs (ScalarFunction, > > AggregateFunction) have means to emit records from close(). > > > > To be honest I am not sure what would be the consequences of interplay > > with other operators which expect TableFunction to emit only when eval > > is called. Not sure if there are such. > > > > If it is a thing that we are certain we want to support, I'd be much > > more comfortable adding finish() to the TableFunction instead. Would be > > happy to hear opinions from the Table API folks. > > > > Best, > > > > Dawid > > > > On 14/09/2022 15:55, Lincoln Lee wrote: > > > Thanks @Piort for your valuable inputs! > > > > > > I did a quick read of the previous discussion you mentioned, seems my > > flip > > > title doesn't give a clear scope here and make some confusions, if my > > > understanding is correct, the UDFs in your context is the user > > > implemented `org.apache.flink.api.common.functions.Function`s, while > the > > > `UserDefinedFunction` I mentioned in the flip is limited to the > > flink-table > > > module which located in `org.apache.flink.table.functions`. > > > > > > Here's an use case we've met recently (which is indeed the motivation > to > > > propose this): > > > one of our user implemented a > > > `org.apache.flink.table.functions.TableFunction`, the simplified > > > pseudo-code is as below: > > > > > > ``` > > > class XFunction extend TableFunction<Out> { > > > > > > open(FunctionContext context){ > > > initMemQueue(); > > > initPythonProc() > > > } > > > > > > eval(In in){ > > > queue.offer(data) > > > Out out = queue.poll() > > > if (out != null) { > > > collect(out) > > > } > > > } > > > > > > close(){ > > > waitForPythonFinish() > > > List<Out> outputs = drainQueue() > > > outputs.foreach(out -> collect(out)) > > > } > > > } > > > ``` > > > It works well in lower flink versions until they attempt to do a > upgrade > > > recently, the 'flush' logic in the legacy close method of > `TableFunction` > > > cannot work properly any more. > > > > > > Before proposing the flip, I also considered the `flush()` extension on > > the > > > `org.apache.flink.api.common.functions.Function`, because some sql > > > operators are also related, but currently not included in the scope of > > this > > > flip, maybe we can discuss it in another thread. > > > > > > Wish this helps explaining the reason and welcome your comments here! > > > > > > Best, > > > Lincoln Lee > > > > > > > > > Piotr Nowojski <pnowoj...@apache.org> 于2022年9月14日周三 16:56写道： > > > > > >> Hi Lincoln, > > >> > > >> Thanks for the proposal. Have you seen the old discussion about adding > > this > > >> `finish()` method? [1] We didn't add it to UDFs, as we didn't see a > > >> motivation (maybe we have missed something), and at the same time it > > wasn't > > >> that easy. Plain `finish()` wouldn't be enough. Users would need a way > > to > > >> output records from the `finish()` call, so it would have to be typed > > with > > >> the user record (`finish(Collector<T> output)`). On the other hand, we > > >> couldn't find an example where a user would actually need the > `finish()` > > >> call in an UDF, as it seemed to us it makes only sense for > > >> operators/functions that are buffering records. Note back then, during > > the > > >> discussion, we were referring to this method as `flush()` or > `drain()`. > > >> > > >> Can you shed some more light and provide more details on the exact > > >> motivating example behind this proposal? > > >> > > >> Best, > > >> Piotrek > > >> > > >> [1] https://lists.apache.org/thread/gmr9r3n3ktojt4bhoxz4t8qho6h7d1rp > > >> > > >> śr., 14 wrz 2022 o 08:22 Lincoln Lee <lincoln.8...@gmail.com> > > napisał(a): > > >> > > >>> Hello everyone, > > >>> > > >>> I’d like to open a discussion on FLIP-260[1]: expose finish method > > for > > >>> UserDefinedFunction, this makes a chance for users who rely on finish > > >> logic > > >>> in the legacy close() method (< 1.14) to migrate to the new finish() > > >>> method. > > >>> > > >>> The task lifecycle was changed in FLINK-22972[2]: a new finish() > > phase > > >>> was introduced (extracted the ‘finish’ part out of the ‘close’) and > > >> removed > > >>> the dispose() method. This change was also done in table module > (e.g., > > >>> `AbstractMapBundleOperator` for mini-batch operation ) but not > covered > > >> the > > >>> UserDefinedFunction which only exposes open() and close() api for > > custom > > >>> usage, those customers who rely on the legacy close() api may > encounter > > >>> wrong result or suffer runtime errors after upgrading to the new > > version. > > >>> Strictly speaking, it is a bug caused by the breaking change, but due > > to > > >>> the public api change, we propose this flip. > > >>> > > >>> Looking forward to your comments or feedback. > > >>> > > >>> [1] > > >>> > > >>> > > >> > > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-260%3A+Expose+Finish+Method+For+UserDefinedFunction > > >>> [2] https://issues.apache.org/jira/browse/FLINK-22972 > > >>> > > >>> Best, > > >>> Lincoln Lee > > >>> > > >
