Lianet Magrans created KAFKA-20778:
--------------------------------------
Summary: Improve classic consumer rack-aware double rebalance on
replica bounce
Key: KAFKA-20778
URL: https://issues.apache.org/jira/browse/KAFKA-20778
Project: Kafka
Issue Type: Improvement
Components: clients, consumer
Reporter: Lianet Magrans
Assignee: Lianet Magrans
When a replica is bounced , the classic consumer goes through a double
rebalance because when it notices a rack ID changed from rackX to null, then
from null to rackX, unneeded rebalance noise.
Consider if we could avoid this double rebalance by identifying that the case
is that the replica is down (based on the existing flow that knows that the
replica lost its "host", not only the rack).
This issue only affects the classic consumer, where rebalances are triggered
client-side based on the metadata (a replica temporarily shows as "emtpy node",
no host or rack), but the client blindly takes that as rack=null and compares
that null to the previous rack ID. With the new consumer group protocol there
is no rebalance triggered by the GC in this case of temporary replica bounce,
as the rack info remains unchanged (replica metadata available, unchanged,
while temporarily down)
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
