Hello Dong, thanks a lot for proposing the KIP!
I'm in favor of partition splitting, or more generally consistent hashing
mechanism over partition rehashing for the reasons that Matthias as
summarized well. Another point is that for the downstream stateful
consumers that need to reshard their stat
Hi Litao,
When acks=all is used, we will only require all ISR replicas to ack before
returning, your understanding on this part is right.
What I meant is, that with acks=all, for the above example with {A, B, C},
if the ISR list still contains all three replicas, we will still require
ALL 3 of th
Hello Richard,
Thanks for the proposed KIP. I have a couple of general comments:
1. I'm not sure if piggy-backing the timeout exception on the
existing requestTimeoutMs configured in "request.timeout.ms" is a good idea
since a) it is a general config that applies for all types of requests, and
2)
About partition splitting and load balancing:
I think there are two cases:
1) one partition is a hot-spot: for this case, splitting a partitions is
actually more desirable than adding a new partitions as it ensure the
the load of the hot partitions is reduced. Adding partitions does not
help here
Hi all,
I would like to discuss a potential change which would be made to
KafkaConsumer:
https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=75974886
Thanks,
Richard Yu