Thanks Sagar for writing this PR. I think twice about the options that have been proposed in https://issues.apache.org/jira/browse/KAFKA-13152, and feel that at the moment it's simpler to just do the even distribution of the configured total bytes. My rationale is that right now we have a static tasks -> threads mapping, and hence each partition would only be fetched by a single thread / consumer at a given time. If in the future we break that static mapping into dynamic mapping, then we would not be able to do this even distribution. Instead we would have other threads polling from consumer only, and those threads would be responsible for checking the config and pause non-empty partitions if it goes beyond the threshold. But since at that time we would not change the config but just how it would be implemented behind the scenes we would not need another KIP to change it.
Some more comments: 1. We need to discuss a bit about the default value of this new config. Personally I think we need to be a bit conservative with large values so that it would not have any perf regression compared with old configs especially with large topology and large number of partitions. 2. I looked at the existing metrics, and do not have corresponding sensors. How about also adding a task level metric indicating the current totally aggregated metrics. The reason I do not suggest this metric on the per-thread level is that in the future we may break the static mapping of tasks -> threads. [optional] As an orthogonal thought, I'm thinking maybe we can rename the other "*cache.max.bytes.buffering*" as "statestore.cache.max.bytes" (via deprecation of course), piggy-backed in this KIP? Would like to hear others' thoughts. Guozhang On Sun, Aug 22, 2021 at 9:29 AM Sagar <sagarmeansoc...@gmail.com> wrote: > Hi All, > > I would like to start a discussion on the following KIP: > https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=186878390 > > Thanks! > Sagar. > -- -- Guozhang