[
https://issues.apache.org/jira/browse/ZOOKEEPER-3140?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17327110#comment-17327110
]
Brian Nixon commented on ZOOKEEPER-3140:
----------------------------------------
[~liwang], while I was with Facebook I did not get any numbers cleared for
public release and I've since left the company so I no longer have access to
performance data to share.
For metrics, I'd recommend looking at write times with large number of
observers (exactly as you did) and sync times after leader election or failure.
I'd say you're on the right track here with your new ticket.
> Allow Followers to host Observers
> ---------------------------------
>
> Key: ZOOKEEPER-3140
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-3140
> Project: ZooKeeper
> Issue Type: New Feature
> Components: server
> Affects Versions: 3.6.0
> Reporter: Brian Nixon
> Assignee: Brian Nixon
> Priority: Minor
> Labels: pull-request-available
> Fix For: 3.6.0
>
> Time Spent: 8h 20m
> Remaining Estimate: 0h
>
> Observers function simple as non-voting members of the ensemble, sharing the
> Learner interface with Followers and holding only a slightly difference
> internal pipeline. Both maintain connections along the quorum port with the
> Leader by which they learn of all new proposals on the ensemble.
>
> There are benefits to allowing Observers to connect to the Followers to plug
> into the commit stream in addition to connecting to the Leader. It shifts the
> burden of supporting Observers off the Leader and allow it to focus on
> coordinating the commit of writes. This means better performance when the
> Leader is under high load, particularly high network load such as can happen
> after a leader election when many Learners need to sync. It also reduces the
> total network connections maintained on the Leader when there are a high
> number of observers. One the other end, Observer availability is improved
> since it will take shorter time for a high number of Observers to finish
> syncing and start serving client traffic.
>
> The current implementation only supports scaling the number of Observers
> into the hundreds before performance begins to degrade. By opening up
> Followers to also host Observers, over a thousand observers can be hosted on
> a typical ensemble without major negative impact under both normal operation
> and during post-leader election sync.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
