[
https://issues.apache.org/jira/browse/FLINK-6204?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15944998#comment-15944998
]
hongyuhong commented on FLINK-6204:
-----------------------------------
I think the memory of 'sortedTimestamps' is the same as the memory of timer you
register for each timestamp. We suppose that one second has 1 million records,
suppose there is just one key, and the sortedTimestamps at most will have 1000
elements. So as the timer, at most you will register 1000 timer, It seems no
difference. What do you think?
> Improve Event-Time OVER ROWS BETWEEN UNBOUNDED PRECEDING aggregation to SQL
> ---------------------------------------------------------------------------
>
> Key: FLINK-6204
> URL: https://issues.apache.org/jira/browse/FLINK-6204
> Project: Flink
> Issue Type: Bug
> Components: Table API & SQL
> Reporter: sunjincheng
> Assignee: sunjincheng
>
> Currently `event time OVER ROWS BETWEEN UNBOUNDED PRECEDING aggregation to
> SQL` implementation class: ` UnboundedEventTimeOverProcessFunction` use
> data size uncontrollable memory data structures`sortedTimestamps:
> util.LinkedList [Long] cache data timestamps and sort timestamps. IMO,It's
> not a good way, because in the production environment there are millions of
> window data pre millisecond in our application scenario.So, I want to remove
> `util.LinkedList [Long] `. Welcome anyone to give me feedback.
> What do you think? [~fhueske] and [~Yuhong_kyo]
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
