[infinispan-issues] [JBoss JIRA] (ISPN-7749) A node leaving during rebalance causes extra segment transfers

[Tristan Tarrant (JIRA)]

     [ https://issues.jboss.org/browse/ISPN-7749?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Tristan Tarrant updated ISPN-7749:
----------------------------------
    Fix Version/s: 9.2.0.Final
                       (was: 9.1.0.Final)


> A node leaving during rebalance causes extra segment transfers
> --------------------------------------------------------------
>
>                 Key: ISPN-7749
>                 URL: https://issues.jboss.org/browse/ISPN-7749
>             Project: Infinispan
>          Issue Type: Bug
>          Components: Core, State Transfer
>    Affects Versions: 9.0.0.Final
>            Reporter: Dan Berindei
>            Assignee: Dan Berindei
>             Fix For: 9.2.0.Final
>
>
> If a node leaves during a rebalance, the only change is that other nodes will no longer request segments from that node. Otherwise the rebalance proceeds as usual, and at the end a node may delete its copy of a segment even if that leaves less than {{numOwners}} copies in the cluster. 
> When the leaver is the joiner, a 2nd rebalance is very likely to return to the initial CH, but only after transferring some segments twice:
> {noformat}
> Rebalance starts: current_owners(s) = AB, pending_owners(s) = AC
> C leaves: current_owners(s) = AB, pending_owners(s) = A
> Rebalance finishes: current_owners(s) = A, pending_owners(s) = A
> 2nd rebalance starts: current_owners(s) = A, pending_owners(s) = AB
> {noformat}
> Even if the leaver is one of the old owners and the 2 segment transfers are necessary, the cluster stays for too long with less than {{numOwners}} copies of the segment:
> {noformat}
> Rebalance starts: current_owners(s) = AB, pending_owners(s) = AC
> A leaves: current_owners(s) = B, pending_owners(s) = C
> C transfers segment s from B
> Rebalance finishes: current_owners(s) = C, pending_owners(s) = C
> 2nd rebalance starts: current_owners(s) = C, pending_owners(s) = DC
> {noformat}
> We can fix this by making the pending CH a union of the current and pending CH whenever a node leavers, and only removing extra segment copies after the 2nd rebalance.



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