[jira] [Commented] (GEODE-5925) Server shutdown delays election of new primary bucket owners

["ASF subversion and git services (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/GEODE-5925?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16664140#comment-16664140 ] 

ASF subversion and git services commented on GEODE-5925:
--------------------------------------------------------

Commit df185b1a82fa45f19cf892a343203f6b9bbc6566 in geode's branch refs/heads/feature/GEODE-5925 from [~bschuchardt]
[ https://gitbox.apache.org/repos/asf?p=geode.git;h=df185b1 ]

GEODE-5925 Server shutdown delays election of new primary bucket owners

This PR addresses the problem of AcceptorImpl sending out expensive
profile update messages when it's being stopped during Cache.close().
There is no reason to send these messages if the affected regions are
going to be destroyed since that also sends profile updates
to other members of the cluster.


> Server shutdown delays election of new primary bucket owners
> ------------------------------------------------------------
>
>                 Key: GEODE-5925
>                 URL: https://issues.apache.org/jira/browse/GEODE-5925
>             Project: Geode
>          Issue Type: Improvement
>          Components: regions
>            Reporter: Bruce Schuchardt
>            Assignee: Bruce Schuchardt
>            Priority: Major
>
> The distributed lock service is configured to avoid releasing locks while the cache is closing.  If the cache that is closing has any primary bucket locks this can delay cache operations on those buckets until the cache is completely closed and the DistributedSystem is disconnected.
> I've seen this take over 30 seconds, causing client connections to be timed out on the server-side and clients failing over from one server to another only to be blocked by the same issue in those servers.
> Another thing I observed at the same time is that AcceptorImpl is sending profile updates for all partitioned regions.  Those profile updates take as long as 2 seconds apiece to process.  This also delays election of new primary bucket owners and it's unnecessary since 
> DestroyPartitionRegion messages are later sent that remove the profiles in other servers.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)