[jira] [Updated] (CURATOR-7) Session ids not preserved if EnsembleProvider has changed the ensemble

["Mike Drob (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/CURATOR-7?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Mike Drob updated CURATOR-7:
----------------------------

    Attachment: CURATOR-7.patch.txt

Attaching a version of Shevek's patch against the apache code base. It looks like it depends on ZK-3.5, based on his original naming, so it does not compile as applied to master. I have not looked at it in enough detail to figure out if that is something that can be aligned with ZK-3.4 or not.

> Session ids not preserved if EnsembleProvider has changed the ensemble
> ----------------------------------------------------------------------
>
>                 Key: CURATOR-7
>                 URL: https://issues.apache.org/jira/browse/CURATOR-7
>             Project: Apache Curator
>          Issue Type: Improvement
>          Components: Client
>    Affects Versions: 2.0.0-incubating
>            Reporter: Shevek
>             Fix For: awaiting-response
>
>         Attachments: CURATOR-7.patch.txt, EnsembleProvider-zookeeper-3.5.x.patch
>
>
> See https://github.com/Netflix/curator/issues/266
> InterProcessMutex, LeaderLatch, etc use an ephemeral node. If ZooKeeper gives a Disconnected event, the native client reconnects with the same session id, and the ephemeral node is preserved. If the ensemble changes at any point before a Disconnect, Curator's ConnectionState#checkState() calls handleNewConnectionString(), which constructs a new native ZK client, discarding the previous session id, and losing all locks.
> This can be expensive.
> Can ConnectionState be made to preserve the session id, or be more conservative about discarding the entire native client on a Disconnected event?
> Clarifications:
> An ensemble change (e.g. adding a new node to a cluster) does not mean that all session ids are now invalid. The next network glitch should not break locks.
> There is no assumption in this description that the reconfiguration and the glitch/disconnection were related - they need not be.



--
This message was sent by Atlassian JIRA
(v6.2#6252)