[infinispan-issues] [JBoss JIRA] (ISPN-6187) Nodes should be recognized if they join a cluster again if PartitionHandling is active

[Dan Berindei (JIRA)]

    [ https://issues.jboss.org/browse/ISPN-6187?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13397458#comment-13397458 ] 

Dan Berindei commented on ISPN-6187:
------------------------------------

# If the partition is already AVAILABLE, we will do a rebalance and transfer state to the joiner regardless of the persistent UUID.
# Agree, we could make a DEGRADED partition AVAILABLE if the new joiner makes it a majority partition and the partition already had a copy of each segment (also possible with {{numOwners <= numNodes/2}}, as long as {{numOwners > 2}}). After we mark the partition as AVAILABLE, we're in case #1 and we should do a rebalance with the current members.
# I think the whole point of enabling partition handling is to not allow the application to proceed with incomplete data, so I don't think we should ever rebalance if we don't have a copy of each segment.


> Nodes should be recognized if they join a cluster again if PartitionHandling is active
> --------------------------------------------------------------------------------------
>
>                 Key: ISPN-6187
>                 URL: https://issues.jboss.org/browse/ISPN-6187
>             Project: Infinispan
>          Issue Type: Enhancement
>          Components: Core
>            Reporter: Wolf-Dieter Fink
>              Labels: partition_handling
>
> If a node leave a cluster by crash/power or network problems and is restarted the ID has changed and it is a NEW node.
> In case of PartitionHandling enabled the cluster can be in a state where it is impossible to recover automatically.
> As ISPN8 will have a new feature "persistent" address, which is used for CH only at this moment, this address can be used for PH as well.
> As a result the node join as known and can be
> # filled with data via state-transfer if the remaining cluster is the major partition and AVAILABLE.
> # filled with data if at least one owner is inside the remaining cluster
>   equal splitt with numOwner>numNode/2
> # full cluster rebalance with WARN/ERROR if there is a possible loss of data



--
This message was sent by Atlassian JIRA
(v7.2.3#72005)