[
https://issues.apache.org/jira/browse/HIVE-15562?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Siddharth Seth updated HIVE-15562:
----------------------------------
Attachment: HIVE-15562.01.patch
Patch makes a few changes. Some unrelated to the core issue.
For the core issue, the registration for state change notifications has moved
into the main lock.
Additional changes.
- Avoid excessive debug logging in the ZK registry.
- Handle a potential leak in the AM
- Switch around the maybeRegisterForStateChangeNotifications to work correctly
with the return value (was inverted)
[~sershe] - could you please take a look?
> LLAP TaskExecutorService race can lead to some fragments being permanently
> lost
> -------------------------------------------------------------------------------
>
> Key: HIVE-15562
> URL: https://issues.apache.org/jira/browse/HIVE-15562
> Project: Hive
> Issue Type: Bug
> Reporter: Siddharth Seth
> Assignee: Siddharth Seth
> Priority: Critical
> Attachments: HIVE-15562.01.patch
>
>
> A fragment can be evicted before it's submission completed. A race can cause
> a situation As a result, we end up trying to unregister the fragment for
> stateChangeNotifications before it has been registered - leading to an
> exception, which skips the actual fragment kill.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
