[
https://issues.apache.org/jira/browse/FLINK-26498?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17505973#comment-17505973
]
Jark Wu commented on FLINK-26498:
---------------------------------
Hi [~paul8263], as I mentioned before, we can introduce a configuration to
control this behavior. But I think always emitting result when cleanup does
change the semantics and is not a good idea.
> The window result may not have been emitted when use window emit feature and
> set allow-latency
> ------------------------------------------------------------------------------------------------
>
> Key: FLINK-26498
> URL: https://issues.apache.org/jira/browse/FLINK-26498
> Project: Flink
> Issue Type: Improvement
> Components: Table SQL / Planner
> Reporter: hehuiyuan
> Priority: Major
> Labels: pull-request-available
> Attachments: image-2022-03-05-23-53-37-086.png,
> image-2022-03-05-23-53-44-196.png, image-2022-03-06-00-03-11-670.png
>
>
> the sql of job :
> {code:java}
> CREATE TABLE tableSource(
> name string,
> age int not null,
> sex string,
> dt TIMESTAMP(3),
> WATERMARK FOR dt AS dt
> ) WITH (
> );
> CREATE TABLE tableSink(
> windowstart timestamp(3),
> windowend timestamp(3),
> name string,
> age int,
> cou bigint
> )
> WITH (
> );
> INSERT INTO tablesink
> SELECT
> TUMBLE_START(dt, INTERVAL '1' HOUR),
> TUMBLE_END(dt, INTERVAL '1' HOUR),
> name,
> age,
> count(sex)
> FROM tableSource
> GROUP BY TUMBLE(dt, INTERVAL '1' HOUR), name,age {code}
>
> and table config:
> {code:java}
> table.exec.emit.allow-lateness = 1 hour
> table.exec.emit.late-fire.delay = 1 min
> table.exec.emit.early-fire.delay = 1min{code}
>
> The data:
> {code:java}
> >hehuiyuan1,22,woman,2022-03-05 00:30:22.000
> >hehuiyuan1,22,woman,2022-03-05 00:40:22.000
> //pause ,wait for the window trigger for earlyTrigger 1 min
> >hehuiyuan1,22,woman,2022-03-05 00:50:22.000
> >hehuiyuan1,22,woman,2022-03-05 00:56:22.000
> //pause ,wait for the window trigger for earlyTrigger 1 min
> >hehuiyuan1,22,woman,2022-03-05 01:00:00.000
> //pause ,wait for the window trigger for earlyTrigger 1 min
> >hehuiyuan1,22,woman,2022-03-05 00:59:20.000 --latency data
> //pause ,wait for the window trigger for earlyTrigger 1 min
> >hehuiyuan1,22,woman,2022-03-05 00:59:20.100 --latency data
> >hehuiyuan1,22,woman,2022-03-05 02:00:00.000 -- window state clean for
> >[0:00:00 1:00:00]
> >hehuiyuan1,22,woman,2022-03-05 02:10:00.000 {code}
>
> The result:
> {code:java}
> > +I(+I[2022-03-05T00:00, 2022-03-05T01:00, hehuiyuan1, 22, 2])
> > -U(-U[2022-03-05T00:00, 2022-03-05T01:00, hehuiyuan1, 22, 2])
> > +U(+U[2022-03-05T00:00, 2022-03-05T01:00, hehuiyuan1, 22, 4])
> > +I(+I[2022-03-05T01:00, 2022-03-05T02:00, hehuiyuan1, 22, 1])
> > -U(-U[2022-03-05T00:00, 2022-03-05T01:00, hehuiyuan1, 22, 4])
> > +U(+U[2022-03-05T00:00, 2022-03-05T01:00, hehuiyuan1, 22, 5])
>
>
> > +I(+I[2022-03-05T02:00, 2022-03-05T03:00, hehuiyuan1, 22, 2]) {code}
>
> The window result is lost when `hehuiyuan1,22,woman,2022-03-05 00:59:20.100`
> arrived， the lateTrigger is not trigger and the window[0:00:00 ,1:00:00] is
> cleaned when the data `hehuiyuan1,22,woman,2022-03-05 02:00:00.000` arrived
> that updated watermark.
>
> The window[0:00:00 ,1:00:00] has 6 pieces of data, but we got 5.
> The trigger is AfterEndOfWindowEarlyAndLate .
> So WindowOpearator may need to emit reuslt when the window cleanupTimer call
> onEventTime.
>
> I think the correct result is as follows:
> {code:java}
> > +I(+I[2022-03-05T00:00, 2022-03-05T01:00, hehuiyuan1, 22, 2])
> > -U(-U[2022-03-05T00:00, 2022-03-05T01:00, hehuiyuan1, 22, 2])
> > +U(+U[2022-03-05T00:00, 2022-03-05T01:00, hehuiyuan1, 22, 4])
> > +I(+I[2022-03-05T01:00, 2022-03-05T02:00, hehuiyuan1, 22, 1])
> > -U(-U[2022-03-05T00:00, 2022-03-05T01:00, hehuiyuan1, 22, 4])
> > +U(+U[2022-03-05T00:00, 2022-03-05T01:00, hehuiyuan1, 22, 5])
> > -U(-U[2022-03-05T00:00, 2022-03-05T01:00, hehuiyuan1, 22, 5])
> > +U(+U[2022-03-05T00:00, 2022-03-05T01:00, hehuiyuan1, 22, 6])
> > +I(+I[2022-03-05T02:00, 2022-03-05T03:00, hehuiyuan1, 22, 2]) {code}
>
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
