[infinispan-issues] [JBoss JIRA] (ISPN-6107) State transfer should not fetch segments that were added during a rebalance

[Dan Berindei (JIRA)]

     [ https://issues.jboss.org/browse/ISPN-6107?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Dan Berindei updated ISPN-6107:
-------------------------------
              Status: Pull Request Sent  (was: Open)
    Git Pull Request: https://github.com/infinispan/infinispan/pull/3962


> State transfer should not fetch segments that were added during a rebalance
> ---------------------------------------------------------------------------
>
>                 Key: ISPN-6107
>                 URL: https://issues.jboss.org/browse/ISPN-6107
>             Project: Infinispan
>          Issue Type: Bug
>          Components: Core, Test Suite - Core
>    Affects Versions: 8.1.0.Final
>            Reporter: Dan Berindei
>            Assignee: Dan Berindei
>             Fix For: 8.2.0.Beta1
>
>
> When the last owner of a segment leaves the cache, the coordinator will update the consistent hash and replace that owner with {{numOwners}} owners (so that a segment always has at least 1 owner). If there is a rebalance in progress, it could be that both the current and the pending CH lost all the owners of a segment, and the coordinator will assign new owners in both CHs (not necessarily the same).
> Sometimes, this causes tests that create clusters with many nodes to spend a lot of time shutting down the cluster. Here's an example:
> # Cluster ABCDE, coordinator A, topology id = 0, currentCH = \{0: CD, 1: BC\}, pendingCH = null
> # D leaves
> # A broadcasts a REBALANCE_START command with topology id 1, members = ABCE, currentCH = \{0: C, 1: BC\}, pendingCH = \{0: BC, 1: BC\}
> # A and E confirm that they finished the rebalance
> # C leaves before sending the data for segment 0 to B
> # A broadcasts a CH_UPDATE command with topology id 2, members = ABE, currentCH = \{0: AE, 1: B\}, pendingCH = \{0: B, 1: B\}
> # A now owns segment 0 in the writeCH (which is the union of currentCH and pendingCH).
> # A tries to request segment 0 from the other owner in the currentCH, E
> # B confirms that it finished the rebalance
> # A broadcasts a new topology: topology id 3, currentCH = \{0: B, 1: B\}, pendingCH = null
> # E installs topology 3, and throws an IllegalArgumentException when handling A's request for segments
> # A is not able to install topology 3, because it requests the transactions data while holding the lock on the LocalCacheStatus
> # A receives the IllegalArgumentException from E and retries. But because it still has the old topology, it retries on E ad infinitum - using a lot of CPU in the process.
> A requesting segment 0 from E is not a problem in itself - normally E would just send back an empty set of transactions and entries. The problem is that the cluster is able to install a new topology, because A already confirmed receiving all the data, but A is stuck with the old topology.



--
This message was sent by Atlassian JIRA
(v6.4.11#64026)