[
https://issues.apache.org/jira/browse/KAFKA-16985?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Lianet Magrans updated KAFKA-16985:
-----------------------------------
Description: While running some stress tests we found out consumers were
not leaving the group if interrupted while closing (leading to members getting
eventually fenced after the session expired). On close, the consumer generates
an Unsubscribe event to be handled in the background, but we noticed that the
network thread failed with the interruption, seemingly not sending the unsent
requests. We should review this to ensure that a member does a clean leave,
notifying the coordinator with a leave HB, even if in a fire-and-forget mode in
the case of interruption (and validate the legacy consumer behaviour in this
scenario). (Still under investigation, I'll update more info as I discover it)
(was: While running some stress tests we found out consumers were not leaving
the group if interrupted while closing (leading to members getting eventually
fenced after the session expired). On close, the consumer generates an
Unsubscribe event to be handled in the background, but we noticed that the
network thread failed with the interruption, seemingly not sending the unsent
requests. We should review this to ensure that a member does a clean leave,
notifying the coordinator with a leave HB, even if in a fire-and-forget mode in
the case of interruption. (Still under investigation, I'll update more info as
I discover it))
> Ensure consumer sends leave request on close even if interrupted
> ----------------------------------------------------------------
>
> Key: KAFKA-16985
> URL: https://issues.apache.org/jira/browse/KAFKA-16985
> Project: Kafka
> Issue Type: Bug
> Reporter: Lianet Magrans
> Assignee: Lianet Magrans
> Priority: Major
>
> While running some stress tests we found out consumers were not leaving the
> group if interrupted while closing (leading to members getting eventually
> fenced after the session expired). On close, the consumer generates an
> Unsubscribe event to be handled in the background, but we noticed that the
> network thread failed with the interruption, seemingly not sending the unsent
> requests. We should review this to ensure that a member does a clean leave,
> notifying the coordinator with a leave HB, even if in a fire-and-forget mode
> in the case of interruption (and validate the legacy consumer behaviour in
> this scenario). (Still under investigation, I'll update more info as I
> discover it)
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
