[
https://issues.apache.org/jira/browse/FLINK-25614?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17477524#comment-17477524
]
Jing Zhang edited comment on FLINK-25614 at 1/18/22, 2:23 AM:
--------------------------------------------------------------
[~wenlong.lwl][~lmagics]Thanks a lot for explain. I understand.
In current implementation, we use getLong because there must exists a shuffle
before localWindowAggregate. So it's safe to use getLong instead of
getTimestamp for better performance.
However after set chain strategy of the local window aggregate to alway, the
assumption does not exists anymore. We should use getTimestamp to replace
getLong.
BTW, this change would affect the job topology, we should to move it to new
version.
WDYT, [~jark][~lmagics][~wenlong.lwl]
was (Author: qingru zhang):
[~wenlong.lwl][~lmagics]Thanks a lot for explain. I understand.
I agree to use getTimestamp instead of getLong after set chain strategy of
local window aggregate.
> Let LocalWindowAggregate be chained with upstream
> -------------------------------------------------
>
> Key: FLINK-25614
> URL: https://issues.apache.org/jira/browse/FLINK-25614
> Project: Flink
> Issue Type: Improvement
> Components: Table SQL / Runtime
> Affects Versions: 1.14.2
> Reporter: Q Kang
> Assignee: Q Kang
> Priority: Minor
> Labels: pull-request-available
>
> When enabling two-phase aggregation (local-global) strategy for Window TVF,
> the physical plan is shown as follows:
> {code:java}
> TableSourceScan -> Calc -> WatermarkAssigner -> Calc
> ||
> || [FORWARD]
> ||
> LocalWindowAggregate
> ||
> || [HASH]
> ||
> GlobalWindowAggregate
> ||
> ||
> ...{code}
> We can let the `LocalWindowAggregate` node be chained with upstream operators
> in order to improve efficiency, just like the non-windowing counterpart
> `LocalGroupAggregate`.
>
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
