[
https://issues.apache.org/jira/browse/ZOOKEEPER-4261?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17314640#comment-17314640
]
Zhewei Hu commented on ZOOKEEPER-4261:
--------------------------------------
Hi Jian, I think one ZK observer should be able to handle 5K to 10K
connections. Therefore, I think you may not need that many observers.
I think the reason we need the server id is that ZK needs to fetch the ports
from the config file for communicating to other ZK servers.
> Observer do not need a serverid
> -------------------------------
>
> Key: ZOOKEEPER-4261
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-4261
> Project: ZooKeeper
> Issue Type: Improvement
> Components: server
> Affects Versions: 3.6.2
> Reporter: Jian Wang
> Priority: Major
>
> I am working on a project which uses Zookeeper as the authority KV store, the
> write TPS would be quite low, but the read TPS would be some like 10K from
> 10K clients. So my idea is quite simple, I will use a cache layer(like http)
> in front of ZK, then my problem become how to keep my cache updated.
> After some investigation, I come up with 2 solutions:
> # Run the Observer inside my cache, so that the cache would be updated
> incrementally, and could be persisted with the stable ZK data dir.
> # [Use
> https://curator.apache.org/curator-recipes/curator-cache.html.|https://curator.apache.org/curator-recipes/curator-cache.html]
> With approach 2, the disconnection of the session would lead to read all the
> keys inside the KV store, and I have to implement a persistent store myself
> if I want to store the data locally.
> For the approach 1, I come up with the question:
> Each server inside ensemble should have a serverid, which is read from
> data/myid, it is limited to range [0, 255]. I my case I would like to have
> thousands of Observer process running, which would easily break the
> limitation.
> After look into the code, I don't think each observer has unique serverid is
> not necessary (I tested the Observer with a quite large number, it works
> though). I think we should remove the serverid of Observer so that we could
> build application on Observer easily and benefit from the streaming of
> transaction log.
>
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
