[
https://issues.apache.org/jira/browse/ZOOKEEPER-2926?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Michael Han resolved ZOOKEEPER-2926.
------------------------------------
Resolution: Fixed
Fix Version/s: 3.6.0
Issue resolved by pull request 447
[https://github.com/apache/zookeeper/pull/447]
> Data inconsistency issue due to the flaw in the session management
> ------------------------------------------------------------------
>
> Key: ZOOKEEPER-2926
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-2926
> Project: ZooKeeper
> Issue Type: Bug
> Components: server
> Affects Versions: 3.5.3, 3.6.0
> Reporter: Fangmin Lv
> Assignee: Fangmin Lv
> Priority: Critical
> Labels: pull-request-available
> Fix For: 3.6.0
>
> Time Spent: 3h
> Remaining Estimate: 0h
>
> The local session upgrading feature will upgrade the session locally before
> receving a quorum commit of creating global session. It's possible that the
> server shutdown before the creating session request being sent to leader, if
> we retained the ZKDatabase or there is Snapshot happened just before
> shutdown, then only this server will have the global session.
> If that server didn't become leader, then it will have more global sessions
> than others, and those global sessions won't expire as the leader doesn't
> know it's existence. If the server became leader, it will accept the client
> renew session request and the client is allowed to create ephemeral nodes,
> which means other servers only have ephemeral nodes but not that global
> session. If there is follower going to have SNAP sync with it, then it will
> also have the global session. If the server without that global session
> becomes new leader, it will check and delete those dangling ephemeral node
> before serving traffic. These could introduce the issues that the ephemeral
> node being exist on some servers but not others.
> There is dangling global session issue even without local session feature,
> because on leader it will update the ZKDatabase when processing
> ConnectionRequest and in the PrepRequestProcessor before it's quorum
> committed, which also has this risk.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
