[
https://issues.apache.org/jira/browse/ZOOKEEPER-1413?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13662605#comment-13662605
]
Thawan Kooburat commented on ZOOKEEPER-1413:
--------------------------------------------
The main benefit of this patch is to reduce the chance that the leader need to
send a snapshot to the learner. So it tries to fix various corner cases in
synchronization logic as reported in JIRAs like ZOOKEEPER-1465, ZOOKEEPER-876
(I just found this one)
The most common corner case that cause the leader to send a snapshot to
learners is when the learner fall behind by more than 1 epoch. This is possible
if the leader fail to form a quorum and move to a new epoch in a quick
succession (let say you have situation like ZOOKEEPER-1697).
For us, it is prohibitively expensive to send a snapshot because we have a
large number of observers and sending a snapshot will kick-out client sessions
on RO-observers (ZOOKEEPER-1607). I don't have a concrete number at the moment,
but this is one of the critical feature that kept our infrastructure running.
> Use on-disk transaction log for learner sync up
> -----------------------------------------------
>
> Key: ZOOKEEPER-1413
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-1413
> Project: ZooKeeper
> Issue Type: Improvement
> Components: server
> Affects Versions: 3.4.3
> Reporter: Thawan Kooburat
> Assignee: Thawan Kooburat
> Priority: Minor
> Labels: performance
> Fix For: 3.5.0
>
> Attachments: ZOOKEEPER-1413.patch, ZOOKEEPER-1413.patch,
> ZOOKEEPER-1413.patch
>
>
> Motivation:
> The learner syncs up with leader by retrieving committed log from the leader.
> Currently, the leader only keeps 500 entries of recently committed log in
> memory. If the learner falls behind more than 500 updates, the leader will
> send the entire snapshot to the learner.
> With the size of the snapshot for some of our Zookeeper deployments (~10G),
> it is prohibitively expensive to send the entire snapshot over network.
> Additionally, our Zookeeper may serve more than 4K updates per seconds. As a
> result, a network hiccups for less than a second will cause the learner to
> use snapshot transfer.
> Design:
> Instead of looking only at committed log in memory, the leader will also look
> at transaction log on disk. The amount of transaction log kept on disk is
> configurable and the current default is 100k. This will allow Zookeeper to
> tolerate longer temporal network failure before initiating the snapshot
> transfer.
> Implementation:
> We plan to add interface to the persistence layer will can be use to retrieve
> proposals from on-disk transaction log. These proposals can then be used to
> send to the learner using existing protocol.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
