You are viewing a plain text version of this content. The canonical link for it is here.
Posted to jira@kafka.apache.org by "Preston Price (Jira)" <ji...@apache.org> on 2021/11/16 22:38:00 UTC

[jira] [Commented] (KAFKA-10413) rebalancing leads to unevenly balanced connectors

    [ https://issues.apache.org/jira/browse/KAFKA-10413?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17444846#comment-17444846 ] 

Preston Price commented on KAFKA-10413:
---------------------------------------

I am experiencing this issue as well, and it's particularly nasty because we're running in kubernetes so we have to WAY over limit our CPU because we end up with very uneven distribution of tasks resulting in CPU throttling on the workers that get work over-allocated to them. I've tried the mitigation of using connect.protocl=eager, but it seems my connector always gets back into an uneven state through pause/deploy/resume cycles.

> rebalancing leads to unevenly balanced connectors
> -------------------------------------------------
>
>                 Key: KAFKA-10413
>                 URL: https://issues.apache.org/jira/browse/KAFKA-10413
>             Project: Kafka
>          Issue Type: Bug
>          Components: KafkaConnect
>    Affects Versions: 2.5.1
>            Reporter: yazgoo
>            Assignee: rameshkrishnan muthusamy
>            Priority: Major
>             Fix For: 2.4.2, 2.5.2, 2.8.0, 2.7.1, 2.6.2
>
>         Attachments: connect_worker_balanced.png
>
>
> GHi,
> With CP 5.5, running kafka connect s3 sink on EC2 whith autoscaling enabled, if a connect instance disappear, or a new one appear, we're seeing unbalanced consumption, much like mentionned in this post:
> [https://stackoverflow.com/questions/58644622/incremental-cooperative-rebalancing-leads-to-unevenly-balanced-connectors]
> This usually leads to one kafka connect instance taking most of the load and consumption not being able to keep on.
> Currently, we're "fixing" this by deleting the connector and re-creating it, but this is far from ideal.
> Any suggestion on what we could do to mitigate this ?



--
This message was sent by Atlassian Jira
(v8.20.1#820001)