[jira] [Commented] (ZOOKEEPER-2748) Admin command to voluntarily drop client connections

["Marco P. (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/ZOOKEEPER-2748?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15964644#comment-15964644 ] 

Marco P. commented on ZOOKEEPER-2748:
-------------------------------------

[~mhan810][~shralex] Thanks for the feedback.
I'm looking at doing this via ZooKeeperAdmin and have a question:

I could handle this command in ZooKeeperServer#processPacket and never pass this request to the processor pipeline.
Or I could handle it somewhere in the processors pipeline.

Which one would you recommend?
And if you recommend the latter, which processor is the most appropriate to take action and terminate connections?


> Admin command to voluntarily drop client connections
> ----------------------------------------------------
>
>                 Key: ZOOKEEPER-2748
>                 URL: https://issues.apache.org/jira/browse/ZOOKEEPER-2748
>             Project: ZooKeeper
>          Issue Type: New Feature
>          Components: server
>            Reporter: Marco P.
>            Assignee: Marco P.
>            Priority: Minor
>
> In certain circumstances, it would be useful to be able to move clients from one server to another.
> One example: a quorum that consists of 3 servers (A,B,C) with 1000 active client session, where 900 clients are connected to server A, and the remaining 100 are split over B and C (see example below for an example of how this can happen).
> A will do a lot more work than B, C. 
> Overall throughput will benefit by having the clients more evenly divided.
> In case of A failure, all its client will create an avalanche by migrating en masse to a different server.
> There are other possible use cases for a mechanism to move clients: 
>  - Migrate away all clients before a server restart
>  - Migrate away part of clients in response to runtime metrics (CPU/Memory usage, ...)
>  - Shuffle clients after adding more server capacity (i.e. adding Observer nodes)
> The simplest form of rebalancing which does not require major changes of protocol or client code consists of requesting a server to voluntarily drop some number of connections.
> Clients should be able to transparently move to a different server.
> Patch introducing 4-letter commands to shed clients:
> https://github.com/apache/zookeeper/pull/215
> -- -- --
> How client imbalance happens in the first place, an example.
> Imagine servers A, B, C and 1000 clients connected.
> Initially clients are spread evenly (i.e. 333 clients per server).
> A: 333 (restarts: 0)
> B: 333 (restarts: 0)
> C: 334 (restarts: 0)
> Now restart servers a few times, always in A, B, C order (e.g. to pick up a software upgrades or configuration changes).
> Restart A:
> A: 0 (restarts: 1)
> B: 499 (restarts: 0)
> C: 500 (restarts: 0)
> Restart B:
> A: 250 (restarts: 1)
> B: 0 (restarts: 1)
> C: 750 (restarts: 0)
> Restart C:
> A: 625 (restarts: 1)
> B: 375 (restarts: 1)
> C: 0 (restarts: 1)
> The imbalance is pretty bad already. C is idle while A has a lot of work.
> A second round of restarts makes the situation even worse:
> Restart A:
> A: 0 (restarts: 2)
> B: 688 (restarts: 1)
> C: 313 (restarts: 1)
> Restart B:
> A: 344 (restarts: 2)
> B: 657 (restarts: 1)
> C: 0 (restarts: 1)
> Restart C:
> A: 673 (restarts: 2)
> B: 328 (restarts: 1)
> C: 0 (restarts: 1)
> Large cluster (5, 7, 9 servers) make the imbalance even more evident.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)