[
https://issues.apache.org/jira/browse/HBASE-15549?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15216108#comment-15216108
]
stack commented on HBASE-15549:
-------------------------------
bq. With a large enough cluster I get double assignments, failed splits, and
every nasty horrible thing if master can't talk to meta.
Not being able to talk to hbase:meta is an issue apart from meta-on-master.
An RPC to do metaupdates will be more prone to issues than going local, yes.
Master 2.0 with its local procedure store mitigates somewhat by only publishing
final state in hbase:meta, not intermediary.
bq. Why would you put that somewhere else?
Scale. Size of hbase:meta, scaling i/o, memory.
Lets have this discussion as other than a comment on JIRA. Can try and restart
the scaling/master discussion out on dev list.
> Undo HMaster carrying regions in master branch as default
> ---------------------------------------------------------
>
> Key: HBASE-15549
> URL: https://issues.apache.org/jira/browse/HBASE-15549
> Project: HBase
> Issue Type: Bug
> Components: Balancer, master
> Reporter: stack
> Attachments: 15549.patch
>
>
> I thought we had an issue to do this but I can't find it so here is a new one.
> Currently, in master branch, HMaster is a RegionServer and carries meta
> regions such as hbase:meta. Disable this facility by default until better
> thought through (I think we should undo master ever carrying regions; FBers
> are thinking we should go forward with notion that master carries meta/system
> regions. TODO). I want to disable it because this facility is not finished
> and meantime I want to test new stuff coming in on master branch w/o this new
> feature getting in the way.
> Parking here a patch that has how to disable master carrying regions and a
> probably redundant test that ensures all works when master is not carrying
> meta regions (added because it WASN'T working for me -- but was pilot error).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
