zookeeper-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Timothy Ward (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (ZOOKEEPER-2368) Client watches are not disconnected on close
Date Mon, 09 Jan 2017 14:45:58 GMT

    [ https://issues.apache.org/jira/browse/ZOOKEEPER-2368?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15811936#comment-15811936
] 

Timothy Ward commented on ZOOKEEPER-2368:
-----------------------------------------

I'm not sure that I agree with this. If the ZooKeeper handle is already disconnected (e.g.
due to a network failure) before it is closed then the session may not actually be expired
when the handle is closed. If the session timeout is a long one it is perfectly possible that
a client could be recreated and re-attach to the previous session. The one thing that is absolutely
guaranteed on a call to close is that the ZooKeeper handle is disconnected, hence I proposed
using that KeeperState.

> Client watches are not disconnected on close
> --------------------------------------------
>
>                 Key: ZOOKEEPER-2368
>                 URL: https://issues.apache.org/jira/browse/ZOOKEEPER-2368
>             Project: ZooKeeper
>          Issue Type: Improvement
>    Affects Versions: 3.4.0, 3.5.0
>            Reporter: Timothy Ward
>            Assignee: Timothy Ward
>             Fix For: 3.5.3, 3.6.0
>
>         Attachments: ZOOKEEPER-2368.patch
>
>
> If I have a ZooKeeper client connected to an ensemble then obviously I can register watches.

> If the client is disconnected (for example by a failing ensemble member) then I get a
disconnection event for all of my watches. If, on the other hand, my client is closed then
I *do not* get a disconnection event. This asymmetry makes it really hard to clear up properly
when using the asynchronous API, as there is no way to "fail" data reads/updates when the
client is closed.
> I believe that the correct behaviour should be for all watchers to receive a disconnection
event when the client is closed. The watchers can then respond as appropriate, and can differentiate
between a "server disconnect" and a "client disconnect" by checking the ZooKeeper#getState()
method. 
> This would not be a breaking behaviour change as Watchers are already required to handle
disconnection events.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message