[
https://issues.apache.org/jira/browse/HBASE-9095?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13728266#comment-13728266
]
Devaraj Das commented on HBASE-9095:
------------------------------------
Thanks folks for the reviews. [~stack] the failure is a real one. I can
reproduce it.. The problem is that some tests rely on the executor service as a
way to register listeners, and the AM executes the listener as part of running
the handler code within the executor. Removing the executor invocation makes
those tests hang (for example, TestZKBasedOpenCloseRegion#testCloseRegion). I
am trying to see how to make such tests work.
> AssignmentManager's handleRegion should respect the single threaded nature of
> the processing
> --------------------------------------------------------------------------------------------
>
> Key: HBASE-9095
> URL: https://issues.apache.org/jira/browse/HBASE-9095
> Project: HBase
> Issue Type: Bug
> Components: Region Assignment
> Reporter: Devaraj Das
> Assignee: Devaraj Das
> Fix For: 0.95.2
>
> Attachments: 9095-1.txt, 9095-1.txt, 9095-1.txt
>
>
> While debugging a case where a region was getting opened on a RegionServer
> and then closed soon after (and then never re-opened anywhere thereafter), it
> seemed like the processing in handleRegion to do with deletion of ZK nodes
> should be non-asynchronous. This achieves two things:
> 1. The synchronous deletion prevents more than one processing on the same
> event data twice. Assuming that we do get more than one notification (on
> let's say, region OPENED event), the subsequent processing(s) in handleRegion
> for the same znode would end up with a zookeeper node not found exception.
> The return value of the data read would be null and that's already handled.
> If it is asynchronous, it leads to issues like - master opens a region on a
> certain RegionServer and soon after it sends that RegionServer a close for
> the same region, and then the znode is deleted.
> 2. The deletion is currently handled in an executor service. This is
> problematic since by design the events for a given region should be processed
> in order. By delegating a part of the processing to executor service we are
> somewhat violating this contract since there is no guarantee of the ordering
> in the executor service executions...
> Thanks to [~jeffreyz] and [~enis] for the discussions on this issue.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
