kafka-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ewen Cheslack-Postava (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (KAFKA-4631) Refresh consumer metadata more frequently for unknown subscribed topics
Date Mon, 16 Jan 2017 04:02:26 GMT

    [ https://issues.apache.org/jira/browse/KAFKA-4631?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15823429#comment-15823429

Ewen Cheslack-Postava commented on KAFKA-4631:

Would this be a constant timeout or yet another config? And especially if the former, tied
to a fixed # of retries? Or maybe retries up until the first "normal" metadata refresh?

I kind of like this idea, but also feel like it's just sort of papering over bigger underlying
issues (i.e. the fact that you can't know when topic creation metadata updates are actually
complete). Those could be limitations we simply want to accept (validating the entire cluster
has updated metadata might be expensive) or something we want to address in a better way (e.g.
ensure that if a metadata request shows a partition isn't known, immediately do a retry but
make sure it goes to the controller).

> Refresh consumer metadata more frequently for unknown subscribed topics
> -----------------------------------------------------------------------
>                 Key: KAFKA-4631
>                 URL: https://issues.apache.org/jira/browse/KAFKA-4631
>             Project: Kafka
>          Issue Type: Bug
>            Reporter: Jason Gustafson
>            Assignee: Matthias J. Sax
>             Fix For:
> By default, the consumer refreshes metadata every 5 minutes. In testing, it can often
happen that a topic is created at about the same time that the consumer is started. In the
worst case, creation finishes after the consumer fetches metadata, and the test must wait
5 minutes for the consumer to refresh metadata in order to discover the topic. To address
this problem, users can decrease the metadata refresh interval, but this means more frequent
refreshes even after all topics are known. An improvement would be to internally let the consumer
fetch metadata more frequently when the consumer encounters unknown topics. Perhaps every
5-10 seconds would be reasonable, for example.

This message was sent by Atlassian JIRA

View raw message