Mailing-List: contact issues-help@hbase.apache.org; run by ezmlm
Precedence: bulk
Date: Wed, 6 Nov 2013 19:20:18 +0000 (UTC)
From: "Steve Loughran (JIRA)" <jira@apache.org>
To: issues@hbase.apache.org
Message-ID: <JIRA.12677500.1383622025209.29632.1383765618544@arcas>
In-Reply-To: <JIRA.12677500.1383622025209@arcas>
References: <JIRA.12677500.1383622025209@arcas>
Subject: [jira] [Commented] (HBASE-9892) Add info port to ServerName to
 support multi instances in a node
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit


    [ https://issues.apache.org/jira/browse/HBASE-9892?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13815174#comment-13815174 ] 

Steve Loughran commented on HBASE-9892:
---------------------------------------

As Enis says, currently we know the problem is there but don't try to fix it. 

The issue we have there is not just if YARN assigns >1 region server to the same node (it doesn't currently support anti-affinity in allocation requests), but that someone else may be running their own application, HBase or otherwise, on the same machine. If you hard code a port it can fail -any port. The sole advantage we have is that will trigger a new container request/review

Because this also affects the masters, we have to leave that UI at port 0 too -which is the worst issue. I would really like to get hold of that via ZK, from where we can bootstrap the rest of the cluster information 

> Add info port to ServerName to support multi instances in a node
> ----------------------------------------------------------------
>
>                 Key: HBASE-9892
>                 URL: https://issues.apache.org/jira/browse/HBASE-9892
>             Project: HBase
>          Issue Type: Improvement
>            Reporter: Liu Shaohui
>            Assignee: Liu Shaohui
>            Priority: Minor
>         Attachments: HBASE-9892-0.94-v1.diff, HBASE-9892-0.94-v2.diff, HBASE-9892-0.94-v3.diff
>
>
> The full GC time of  regionserver with big heap(> 30G ) usually  can not be controlled in 30s. At the same time, the servers with 64G memory are normal. So we try to deploy multi rs instances(2-3 ) in a single node and the heap of each rs is about 20G ~ 24G.
> Most of the things works fine, except the hbase web ui. The master get the RS info port from conf, which is suitable for this situation of multi rs  instances in a node. So we add info port to ServerName.
> a. at the startup, rs report it's info port to Hmaster.
> b, For root region, rs write the servername with info port ro the zookeeper root-region-server node.
> c, For meta regions, rs write the servername with info port to root region 
> d. For user regions,  rs write the servername with info port to meta regions 
> So hmaster and client can get info port from the servername.
> To test this feature, I change the rs num from 1 to 3 in standalone mode, so we can test it in standalone mode,
> I think Hoya(hbase on yarn) will encounter the same problem.  Anyone knows how Hoya handle this problem?
> PS: There are  different formats for servername in zk node and meta table, i think we need to unify it and refactor the code.


--
This message was sent by Atlassian JIRA
(v6.1#6144)