lianetm commented on code in PR #22705:
URL: https://github.com/apache/kafka/pull/22705#discussion_r3529692330
##########
clients/src/main/java/org/apache/kafka/clients/consumer/internals/ConsumerCoordinator.java:
##########
@@ -1656,36 +1677,44 @@ private ConsumerCoordinatorMetrics(MetricsLedger
metrics, String metricGrpPrefix
}
private static class PartitionRackInfo {
- private final Set<String> racks;
+ private final Set<String> knownRacks;
+ private final boolean hasUnknownReplica;
Review Comment:
is a boolean enough or do we need to know the nodes that are down? E.g, what
about there is a rack change (need to rebalance), at the same time that a
replica is down?
Seems safer to be explicit about what we will tolerate and not rebalance. We
have "known replicas" (with their racks, like the line above but wonder if we
need to track a Map<replicaId, rack> instead), and "unknown replicas"
(replicaIds that are down, have no host/rack). Then we can only safely tolerate
a rack missing in one of them but present in the other for the same replica ID
(that would be exactly the condition we want and nothing more: a replica X that
either has the same racks it had before, or it's down). Makes sense?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
