Hi all,

Bringing this thread back up after several weeks of quiet.

The FLIP proposes breaking down the task-level numRecordsOut metric per
downstream target vertex so consumers (in particular the Kubernetes
autoscaler) can see per-edge throughput instead of guessing from the
aggregate. Purely additive, the aggregate metric stays unchanged. Full
motivation, design, and alternatives are in the FLIP doc.

Design doc:
https://docs.google.com/document/d/1TlyTc6fvYGG1xlO-IFlGBC97CxilUBFdZBpLOBPdKD8/edit?usp=sharing
Draft implementation: https://github.com/apache/flink/pull/28014

@Marton Balassi, @Gabor Somogyi, and @Peter Huang, based on your vote on
[1] & [2], I would really value your read here.
@Maximilian Michels and @Rui Fan, this metric is autoscaler-adjacent, your
read would be particularly useful.

Thanks,
Dennis

[1] https://lists.apache.org/thread/zffx5r6fjkjldjjxr8oqtkwsobtwc6c3
[2] https://lists.apache.org/thread/pkn8rmdfzfq5vl8smj8kw03jns9gq8ty


On Thu, Apr 23, 2026 at 11:01 PM Dennis-Mircea Ciupitu <
[email protected]> wrote:

> Hi all,
>
> I’d like to start a discussion on FLIP-XXX: Expose per downstream target
> numRecordsOut metric [1].
>
> I’ve also opened a draft PR with a reference implementation to make the
> proposal concrete and easier to review [2].
>
> Feedback and suggestions are very welcome.
>
> Best regards,
> Dennis
>
> [1]
> https://docs.google.com/document/d/1TlyTc6fvYGG1xlO-IFlGBC97CxilUBFdZBpLOBPdKD8/edit?usp=sharing
> [2] https://github.com/apache/flink/pull/28014
>

Reply via email to