[
https://issues.apache.org/jira/browse/FLINK-18996?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17703468#comment-17703468
]
Chalres Tan commented on FLINK-18996:
-------------------------------------
+1 [~zicat]. I was pointed to [this design doc|http://goo.gl/VW5Gpd] and
https://issues.apache.org/jira/browse/FLINK-6233.
In the design doc they mention "Considering that, performing cache cleaning too
frequently may affect efficiency. We add a default delay to postpone this
process, i.e., {_}minCleanUpInterval = (LSize + RSize) / 2{_}."
It seems like the minCleanUpInterval is there to prevent the frequency of
cleanup to possibly save compute. I agree with you that delaying cleanup will
cause issues for downstream operators and the default should be that
minCleanUpInterval = 0. If we cannot remove or change the minCleanUpInterval
default value to 0, perhaps we can expose an option to the user to override
this value.
> Avoid disorder for time interval join
> -------------------------------------
>
> Key: FLINK-18996
> URL: https://issues.apache.org/jira/browse/FLINK-18996
> Project: Flink
> Issue Type: Improvement
> Components: Table SQL / Runtime
> Reporter: Benchao Li
> Priority: Major
> Labels: auto-deprioritized-critical, auto-deprioritized-major
> Fix For: 1.17.0
>
>
> Currently, the time interval join will produce data with rowtime later than
> watermark. If we use the rowtime again in downstream, e.t. window
> aggregation, we'll lose some data.
>
> reported from user-zh:
> [http://apache-flink.147419.n8.nabble.com/Re-flink-interval-join-tc4458.html#none]
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
