IMHO, scheduling a `1ms` punctuation would cause quite some overhead. Alsw, we only need this for "time based" eviction, not for buffer size (num records, num bytes), right?
We should also know, *when* we need to evict earliest -- thus, it should be sufficient to schedule accordingly? I think, we can also exploit cancellation to scheduled punctuation to be more flexible. Also note, that during runtime, we don't check for punctuation execution after each record, but do this only after N records are processed (with N being adjusted dynamically during runtime). We also need to consider, that we fire a lot of punctuations if we "jump ahead" in time what seems to be inefficient. [ Full content available at: https://github.com/apache/kafka/pull/5693 ] This message was relayed via gitbox.apache.org for [email protected]
