[
https://issues.apache.org/jira/browse/HBASE-934?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12711408#action_12711408
]
stack commented on HBASE-934:
-----------------------------
I think we need to remove the switch where we decide there is 1 or N servers.
Thats the way to guarantee we don't run into above issue.
> Assigning all regions to one server only
> ----------------------------------------
>
> Key: HBASE-934
> URL: https://issues.apache.org/jira/browse/HBASE-934
> Project: Hadoop HBase
> Issue Type: Bug
> Reporter: stack
>
> From Rong-en who has two regionservers:
> {code}
> When the second HRS starts up, but when all regions
> are still not assigned to the first server. The 2nd HRS keeps
> silence (though it's visible at master's ui). Once all regions are
> assigned to first server, the balancer kicks in. Finally, the
> regions are evenly split to two servers.
> {code}
> And then on IRC:
> {code}
> [21:38] <rafan> st^ack: got my mail?
> [21:48] <st^ack> You have two HRSs?
> [21:49] <st^ack> There is an odd decision made in the master that goes
> something like "is there only one regionserver? or are there more". If it
> decides the answer is 1, it just dumps all to it.
> [21:49] <st^ack> My guess is that in your case, it decided that all
> regions were for server 1.
> [21:50] <st^ack> Mind making a JIRA? Let the master parcel out the
> regions... the rate can vary with the number of OPEN messages it gets back.
> [21:50] <st^ack> Allow that other regions could come in meantime?
> {code}
> If many regions, master should allow that while its assigning, other
> regionservers could come online rather than make the binary decision "There
> is only one regionserver in this cluster"
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
