[infinispan-issues] [JBoss JIRA] (ISPN-6921) Use intersection of expected members and stable topology during merge

Dan Berindei (JIRA) issues at jboss.org
Thu Apr 26 04:23:00 EDT 2018


    [ https://issues.jboss.org/browse/ISPN-6921?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13567368#comment-13567368 ] 

Dan Berindei commented on ISPN-6921:
------------------------------------

[~rvansa] unfortunately being an owner in the stable topology doesn't guarantee that node still has the data. Because the stable topology update is sent asynchronously after the rebalance ends, B could finish the rebalance for (B, C, D) and send the NO_REBALANCE topology update to C and D, but fail to send the stable topology update before it gets isolated. The rebalance process doesn't guarantee that when a node is removed only that node's segments move, so you could have segment 0 owned by CD in the stable topology and by DB in the rebalanced topology, meaning C has deleted its copy of the segment by the time we're trying to use the stable topology for the (A, C, D) partition.

We could probably move the stable topology from being a completely separate thing to being just a flag or topology phase. I'm not yet sure if that would be enough or if we also have to change the way we update the topology after a node leaves to keep it in the CH until the rebalance ends.

> Use intersection of expected members and stable topology during merge
> ---------------------------------------------------------------------
>
>                 Key: ISPN-6921
>                 URL: https://issues.jboss.org/browse/ISPN-6921
>             Project: Infinispan
>          Issue Type: Bug
>          Components: Core
>    Affects Versions: 9.0.0.Alpha3
>            Reporter: Radim Vansa
>            Assignee: Dan Berindei
>
> During a split (A, B, C, D) -> (A), (B, C, D) -> (A), (B), (C, D) -> (A, C, D), (B)
> There was  topology update with BCD in CH, but stable topology is still ABCD.
> When a merge happens, the merged topology has only CD as newMembers when computing the new availability mode. Therefore, the ACD does not become available though it has enough data from the stable topology.



--
This message was sent by Atlassian JIRA
(v7.5.0#75005)


More information about the infinispan-issues mailing list