[
https://issues.apache.org/jira/browse/KAFKA-3556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16865825#comment-16865825
]
Guozhang Wang commented on KAFKA-3556:
--------------------------------------
Just my two cents:
1. I think latency metrics averaged over all groups may not be very helpful on
the broker side, since one buggy group's signal may be buried with many other
normal groups. On the other hand, average group size / metadata cache and
rebalance / commit rate, as well as member session timeout rate can be
maintained at broker-side.
2. On the consumer side, we can keep the rebalance rate, commit rate, and
heartbeat timeout rate as well, plus the latency metrics, including rebalance
latency, commit latency; and for rebalance latency I'd rather have a
finer-grained split-up as well, e.g.: partition-revocation-callback latency,
join-group-round-trip latency, sync-group-round-trip latency (would include the
performAssignment callback as well), and partition-assignment-callback latency.
This would also help testing the effectiveness of KIP-429 as well.
> Improve group coordinator metrics
> ---------------------------------
>
> Key: KAFKA-3556
> URL: https://issues.apache.org/jira/browse/KAFKA-3556
> Project: Kafka
> Issue Type: Improvement
> Reporter: Jason Gustafson
> Assignee: Liquan Pei
> Priority: Major
>
> We currently don't have many metrics to track the behavior of the group
> coordinator (especially with respect to the new consumer). On a quick pass, I
> only saw a couple gauges in GroupMetadataManager for the number of groups and
> the number of cached offsets. Here are some interesting metrics that may be
> worth tracking:
> 1. Session timeout rate
> 2. Rebalance latency/rate
> 3. Commit latency/rate
> 4. Average group size
> 5. Size of metadata cache
> Some of these may also be interesting to track per group.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
