[
https://issues.apache.org/jira/browse/KAFKA-19652?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18024066#comment-18024066
]
Jainil Rana edited comment on KAFKA-19652 at 10/1/25 11:47 PM:
---------------------------------------------------------------
Hi [~gargsha] , [~mjsax] , I tried reproducing this, in 3.9.0 the consumer
throughput drops a lot (~250MB/s → ~90MB/s) when {{{}min.insync.replicas=2{}}}.
I noticed {{findPreferredReadReplica}} only picks from ISR. When followers drop
out, consumers end up reading only from the leader, which seems to cause the
slowdown.
Would it be safe to let consumers read from out-of-ISR followers if they have
the data (with HW checks), or is that too risky?
was (Author: JIRAUSER310849):
Hi [~gargsha] , I’ve been reproducing KAFKA-19652 (consumer throughput
regression in 3.9.0 vs 3.5.1, ZK mode) using the perf test tools. For accurate
benchmarking, should I fully load the topic first and then run the consumer, or
is it fine to run producer and consumer in parallel if I set
auto.offset.reset=earliest?
How is it usually done ?
> Consumer throughput drops by 10 times with Kafka v3.9.0 in ZK mode
> ------------------------------------------------------------------
>
> Key: KAFKA-19652
> URL: https://issues.apache.org/jira/browse/KAFKA-19652
> Project: Kafka
> Issue Type: Bug
> Components: clients, consumer
> Affects Versions: 3.9.0
> Reporter: Sharad Garg
> Priority: Blocker
>
> Kafka consumer throughput in best-case drops by ~10 times after upgrading to
> kafka v3.9.0 from v3.5.1. Note that this is in ZK mode and KRAFT migration is
> not done yet.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
