[infinispan-issues] [JBoss JIRA] Commented: (ISPN-1365) Data left in inconsistent state on rehash.

[Dan Berindei (JIRA)]

    [ https://issues.jboss.org/browse/ISPN-1365?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12627763#comment-12627763 ] 

Dan Berindei commented on ISPN-1365:
------------------------------------

It turns out Erik's problem is not caused by the merge, but by the partitioning itself. He is using ISPN locks to ensure that external resources can only have a single a single owner, so any partitioning will allow multiple owners and break his model.

His conclusion was that in production the load balancer will only connect to the nodes that are part of the cluster, so when a node starts up in a cluster by itself it will not receive any client requests until it has finished merging with the initial cluster. This will ensure the integrity of the model.

A subsequent partition in the cluster ca still break his model, but the rehashing process can't help there.

Note that the attached pull request solves a related problem when the node joins the cluster normally, *not* the 2) scenario in my Sep 4 comment. Somehow my environment changed and I'm not seeing any merge views, although my JGroups configuration is the same.


> Data left in inconsistent state on rehash.
> ------------------------------------------
>
>                 Key: ISPN-1365
>                 URL: https://issues.jboss.org/browse/ISPN-1365
>             Project: Infinispan
>          Issue Type: Bug
>          Components: Core API
>    Affects Versions: 5.0.0.FINAL
>            Reporter: Erik Salter
>            Assignee: Dan Berindei
>             Fix For: 5.0.1.FINAL, 5.1.0.ALPHA1, 5.1.0.FINAL
>
>         Attachments: cacheTest_rehash.zip
>
>
> I'm seeing a lot of data inconsistencies on a rehash, especially if there is a lot of lock contention for keys on caches participating in a transaction.
> Attached is a unit test that can reproduce the problem quite readily.  This uses the grouping API, eager locking of a single node, and the distribution framework to effect "local" transactions.

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira