You are viewing a plain text version of this content. The canonical link for it is here.
Posted to jira@kafka.apache.org by "Jeff Kim (Jira)" <ji...@apache.org> on 2020/08/31 20:53:00 UTC
[jira] [Updated] (KAFKA-10452) Only expire preferred read replica
if a leader is alive for the topic
[ https://issues.apache.org/jira/browse/KAFKA-10452?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Jeff Kim updated KAFKA-10452:
-----------------------------
Description:
Fetch from follower functionality periodically expires and refreshes preferred read replica (at `metadata.max.age.ms` interval). This allows a client to discover a better follower to fetch from if one becomes available.
However the expiration is done even if the current partition has no leader. It makes sense to get the new preferred replica information and update existing one, instead of expiring existing one and then fetching new one.
Doing this will allow clients to keep on fetching from a follower instead of failing to find leader when all ISR replicas go offline.
was:
Fetch from follower functionality periodically expires and refreshes preferred read replica (at `metadata.max.age.ms` interval). This allows a client to discover a better follower to fetch from if one becomes available.
However the expiration is done even if the current partition has no leader (can happen in DR scenario with observers). It makes sense to get the new preferred replica information and update existing one, instead of expiring existing one and then fetching new one.
Doing this will allow clients to keep on fetching from a follower/observer instead of failing to find leader when all ISR replicas go offline.
> Only expire preferred read replica if a leader is alive for the topic
> ---------------------------------------------------------------------
>
> Key: KAFKA-10452
> URL: https://issues.apache.org/jira/browse/KAFKA-10452
> Project: Kafka
> Issue Type: Bug
> Components: clients
> Reporter: Jeff Kim
> Assignee: Jeff Kim
> Priority: Major
>
> Fetch from follower functionality periodically expires and refreshes preferred read replica (at `metadata.max.age.ms` interval). This allows a client to discover a better follower to fetch from if one becomes available.
> However the expiration is done even if the current partition has no leader. It makes sense to get the new preferred replica information and update existing one, instead of expiring existing one and then fetching new one.
> Doing this will allow clients to keep on fetching from a follower instead of failing to find leader when all ISR replicas go offline.
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)