[
https://issues.apache.org/jira/browse/ZOOKEEPER-1032?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13011935#comment-13011935
]
Flavio Junqueira commented on ZOOKEEPER-1032:
---------------------------------------------
Hi Jiangwen, Some thoughts and questions on your proposal:
# I'm not entirely clear on what you're suggesting on (1). In the way I
understand, you're saying that once the follower starts following a leader, if
it has all updates according to the leader, then it doesn't have to take a
snapshot at that point. If this is correct, then the follower would simply keep
appending to the last log it wrote to and take a snapshot eventually. Is this
right?
# I'm not sure what you mean with "DataTree should not be cleared, only
restored from logs/snap" in (2). In what sense should it not be cleared?
# The idea in (3) is interesting, but I have one concern. Given that ZooKeeper
keeps its state in memory, adding a cache of logs might cause some memory
pressure on the state of ZooKeeper. How much would we have to cache to see a
significant improvement? I understand that this can be configurable, though.
> speed up recovery from leader failure
> -------------------------------------
>
> Key: ZOOKEEPER-1032
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-1032
> Project: ZooKeeper
> Issue Type: Improvement
> Components: server
> Reporter: jiangwen wei
> Fix For: 3.4.0
>
>
> when the number of nodes is large, it may take a long time to recover from
> leader failure
> there are some points to improve:
> 1. Follower should take snapshot asynchronously when follower up to date
> 2. Currently Leader/Follower will clear the DataTree on leader failures, and
> then restore it from a snapshot and transaction logs. DataTree should not be
> cleared, only restore it from transaction logs.
> 3. FileTxnLog should store recently transaction logs in memory, so when
> DataTree is not behind the transaction logs a lot, the transaction logs in
> memory can be used to restore DataTree.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
