Thanks for posting this here, Luke. Thinking about this more, I'll just go back 
to the old behavior where the synonyms are ignored if you set them as dynamic 
configurations. I will fix the problem where static synonyms are ignored 
sometimes, though, since that seems more "surprising" and bad.

Maybe we should have removed those synonyms in 4.0, but I guess that ship has 
sailed. Perhaps 5.0 :)

best,
Colin


On Tue, Sep 24, 2024, at 02:52, Luke Chen wrote:
> Hi all,
>
> KAFKA-17584 <https://issues.apache.org/jira/browse/KAFKA-17584> is about a
> bug related to the incorrect synonym handling for dynamic log
> configurations. While there is already a PR
> <https://github.com/apache/kafka/pull/17258> to fix the issue, we found
> there are some log configs that should be "cluster-wide" configs, but due
> to the bug in KAFKA-17584
> <https://issues.apache.org/jira/browse/KAFKA-17584>, we incorrectly treat
> them as "read-only" configs, since v3.0.0 (or even earlier). This issue is
> that we only treat 1 config as "cluster-wide" in the same synonym group.
> The list is as follows:
>
> Configs is now "read-only" but should be "cluster-wide" (i.e. dynamically
> changeable):
> - log.retention.hours
>   log.retention.minutes
>   (log.retention.ms is already a "cluster-wide" config)
> - log.roll.hours
>   (log.roll.ms is already a "cluster-wide" config)
> - log.roll.jitter.hours
>   (log.roll.jitter.ms is already a "cluster-wide" config)
> - log.flush.scheduler.interval.ms
>   (log.flush.interval.ms is already a "cluster-wide" config)
>
> Note that, in all the groups, there is 1 config being treated as a
> "cluster-wide" config. So, it should not block any operation.
> After this fix, all the configs above will become "cluster-wide" configs.
> Although these are public interface changes, I think it is more like a bug
> fix. But if people think we should go through the KIP process to have
> proper discussion, I can open one for it. But for v3.9.0, I think we should
> include this fix to avoid any possible unexpected data loss as described in
> KAFKA-17584 <https://issues.apache.org/jira/browse/KAFKA-17584>, even if
> KIP is required and not accepted.
>
> Please let me know what you think.
>
> Thank you.
> Luke

Reply via email to