Hello,
This is Arun from LINE Corporation.
We have a cluster with a large number of brokers (200+), node failures are
bound to happen relatively often. Upon recovery of the machine, or upon
reassignment of the replicas on failed node, we often have a large amount of
lagging replica catch up. Multiple replicas (re-)assigned to a target broker
could start fetching from the same source broker id holding the leader replica.
This occasionally leads to Produce Response Time degradation as illustrated in
https://issues.apache.org/jira/browse/KAFKA-10690 .
Wanted to check if this is faced by anyone else, and if a solution merits a KIP.
With Regards
マテュアルン Mathew Arun
LINE Corporation
