[
https://issues.apache.org/jira/browse/FLINK-4942?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15615623#comment-15615623
]
ASF GitHub Bot commented on FLINK-4942:
---------------------------------------
Github user kl0u commented on the issue:
https://github.com/apache/flink/pull/2702
Are we sure that this is going to lead to improvements? I understand that
both cases have pros and cons, as in one case we pay the price at each arriving
timer, while in the other you pay it at once upon checkpointing. Actually,
there was the same discussion when designing how to make the timers rescalable
and we ended up with the existing version.
> Improve processing performance of HeapInternalTimerService with key groups
> --------------------------------------------------------------------------
>
> Key: FLINK-4942
> URL: https://issues.apache.org/jira/browse/FLINK-4942
> Project: Flink
> Issue Type: Improvement
> Reporter: Stefan Richter
> Assignee: Stefan Richter
>
> Currently, key groups awareness in `HeapInternalTimerService` is basically
> implemented as (hash) map of (hash) sets. Purpose of this is grouping key
> groups together in a way that allows easy serialization into key groups.
> However, this data layout comes along with some significant performance
> decrease, in particular when the number of key groups is high.
> I suggest to keep all timers in one set again at runtime, thus being as fast
> as in previous versions without key groups.
> Instead, we can perform a very fast online partitioning into key groups
> before a snapshot.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
