You are viewing a plain text version of this content. The canonical link for it is here.

Posted to jira@kafka.apache.org by "Antony Stubbs (JIRA)" <ji...@apache.org> on 2018/07/03 09:24:00 UTC

[jira] [Commented] (KAFKA-7129) Dynamic default value for number of thread configuration

    [ https://issues.apache.org/jira/browse/KAFKA-7129?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16531083#comment-16531083 ] 

Antony Stubbs commented on KAFKA-7129:
--------------------------------------

One thing to be careful is of course detecting the core count properly. E.g. when running the jvm on k8s, it uses too many threads for GC - it must be set explicitly.

> Dynamic default value for number of thread configuration
> --------------------------------------------------------
>
>                 Key: KAFKA-7129
>                 URL: https://issues.apache.org/jira/browse/KAFKA-7129
>             Project: Kafka
>          Issue Type: Improvement
>          Components: core
>            Reporter: Damien Gasparina
>            Priority: Minor
>
> There are properties in the broker to change the number of thread of a component (e.g. _num.replica.fetchers_ or _num.network.threads_). After discussing with [~astubbs], it seems that the default values are optimized for an 8 CPU machine and might not be optimized for larger machine (e.g. 48 cores). 
> For those larger machine, an admin need to tune them to be able to use all resources of the host.
> Having dynamic default value (e.g. _num.replica.fetchers_ = _ceil(number of core / 8)_, etc...) instead of static (e.g. _num.replica.fetchers =1_) could be a more efficient strategy to have default values optimized for different kind of deployment.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)