[
https://issues.apache.org/jira/browse/CASSANDRA-2454?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Sylvain Lebresne updated CASSANDRA-2454:
----------------------------------------
Attachment: 0001-Submit-counters-update-on-mutation-stage-only-if-not.patch
Looking closer that this there is 2 places from which we execute counter write:
# if the coordinator is a replica, from the thrift thread.
# otherwise in the CounterMutationVerbHandler on a replica, that is on the
MUTATION stage.
In the latter case, we must indeed avoid re-submitting to the MUTATION stage to
avoid deadlock. But in the former we should not skip the stage.
Attaching a v2 patch that distinguishes between the 2 cases and 'do the right
thing'.
Note that another way to fix this would be to make CounterMutationVerbHandler
execute on some other stage that the MUTATION one. Even though that would
simpler in the number of line modified, I don't think an existing stage would
fit the bill and creating a new one for that doesn't feel right.
> Possible deadlock for counter mutations
> ---------------------------------------
>
> Key: CASSANDRA-2454
> URL: https://issues.apache.org/jira/browse/CASSANDRA-2454
> Project: Cassandra
> Issue Type: Bug
> Reporter: Stu Hood
> Fix For: 0.8
>
> Attachments: 0001-Don-t-re-submit-to-the-mutation-stage.txt,
> 0001-Submit-counters-update-on-mutation-stage-only-if-not.patch
>
>
> {{StorageProxy.applyCounterMutation}} is executed on the mutation stage, but
> it also submits tasks to the mutation stage, and then blocks for them. If
> there are more than a few concurrent mutations, this can lead to deadlock.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
