Hi David, Matthias, Following up on my October 29 mail: I filed KAFKA-21433 with the reproduction of the controller/broker group-bounds mismatch. The ticket asks whether matching bounds are expected on controllers and brokers, or whether validation should change. If this is already tracked elsewhere, I'm happy to close it as a duplicate.
When you have time, I'd appreciate your thoughts on the remaining scope questions: - David: what scope would make sense for DescribeConfigs to expose effective group values, and would that require a KIP? - Matthias: would that address the use case behind KAFKA-20848, or would a client-side diagnostic need remain? Best, Sunmin 2026년 10월 29일 (목) 오후 4:32, 이선민 <[email protected]>님이 작성: > > Hi David, > > Thanks for the detailed feedback. > > I agree with your points. Most of the list in this KIP is configuration > that the client does not use, so the heartbeat and share fetch responses > would be distributing a group resource rather than carrying something the > member acts on. That is also where most of the complexity comes from (the > token, the per-key cache and its lifetime rules), and it bypasses > DESCRIBE_CONFIGS for values the member does not need. I don't think that > is worth it for logging. > > On concrete use cases: I'm not aware of a report from a production system. > While checking your point that the effective value only differs when an > administrator narrows the broker bounds, I found another way it can > happen. A group config override is validated by the controller against > the controller's own group bounds, but it is clamped by the broker against > the broker's bounds. With a separate controller that has no group.consumer.* > settings (default 45000-60000) and a broker with > group.consumer.min.session.timeout.ms=20000, > group.consumer.session.timeout.ms=25000 and > group.consumer.max.session.timeout.ms=30000, on trunk at 601492157a: > > - --alter consumer.session.timeout.ms=45000 succeeded, and the broker > immediately logged that the value is capped to 30000; > - DescribeConfigs still returned 45000, and session expiry about 28 > seconds after SIGKILL was consistent with the capped timeout; > - 25000, which the broker would accept, was rejected with "must be in the > range 45000 to 60000". > > So when the two sets of bounds differ, an override can be out of effect > from the moment it is set, without anyone changing the bounds, and when > they don't overlap, every accepted override is clamped. I'm not sure > whether operators are expected to set the group bounds on the controllers > as well. Is this known? If not, I can file a ticket with the > reproduction. This reinforces the need to expose effective values through > DescribeConfigs, but does not by itself justify distributing them through > the heartbeat or fetch APIs. > > I'd be glad to work on DescribeConfigs returning the effective values with > your help. Some questions to scope it: > > - Should the value be the effective one, with the stored override still > visible so that a clamped override can be told apart, for example as > the DYNAMIC_GROUP_CONFIG synonym? > - The admin client sends a group config describe to any broker, and the > effective value depends on that broker's bounds. Is reporting the > responding broker's view enough? > - Should streams.assignor.name, when the named assignor is not available, > be covered the same way? > - Would you treat this as a change that needs a KIP? > > Matthias, KAFKA-20848 proposes surfacing all effective group configs for > client-side logging. Given David's concerns, would improving > DescribeConfigs address the use case you had in mind, or is there a > specific client-side diagnostic need that would remain, such as a warning > when an override is not in effect, without the values? > > I'll update the proposal based on the agreed scope. > > Best, > Sunmin
