[
https://issues.apache.org/jira/browse/HDFS-14964?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jinglun updated HDFS-14964:
---------------------------
Attachment: HDFS-14964-show-race-case.patch
Status: Patch Available (was: Open)
Upload HDFS-14964-show-race-case.patch showing the race case.
> Race between RouterUpdateService and RouterAdmin -setQuota
> ----------------------------------------------------------
>
> Key: HDFS-14964
> URL: https://issues.apache.org/jira/browse/HDFS-14964
> Project: Hadoop HDFS
> Issue Type: Bug
> Reporter: Jinglun
> Assignee: Jinglun
> Priority: Minor
> Attachments: HDFS-14964-show-race-case.patch
>
>
> There is a race condition between RouterUpdateService and RouterAdmin
> -setQuota.
> Supposing the mount table is:
> M1: /dir-1 ns0:/dir-1 \{nsQuota, ssQuota}
> # RouterUpdateService is scheduled. It finds the nsUsage of M1 is changed
> and wants to update it at RouterQuotaUpdateService#147.
> # Before RouterUpdateService updates, a RouterAdmin use -setQuota and update
> the Mount Entry Quota.
> # RouterAdmin -setQuota executes successfully and the quota is updated.
> # RouterUpdateService is re-scheduled. It doesn't know the mount entry on zk
> has already been updated. So it continue to update the quota usage and hence
> overwrite RouterAdmin's change.
>
> I think a possible solution would be not letting RouterUpdateService to
> update mount entry on zk(It will still update usage in cache). Let
> RouterAdmin be the only entrance to update mount entry on zk. Even one Router
> restarts the RouterUpdateService would compute and update the usage in cache
> so not getting usage from zk won't be a big problem.
>
> I wrote a unit test to show the race case.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
