hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Andrew Purtell (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HBASE-1583) Start/Stop of large cluster untenable
Date Fri, 26 Jun 2009 00:12:07 GMT

    [ https://issues.apache.org/jira/browse/HBASE-1583?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12724350#action_12724350

Andrew Purtell commented on HBASE-1583:

My understanding for 0.21 is the region assignment process is going to be largely unmediated
by the master, except for the case where the master finds an unassigned region in META and
puts up a node into the "to be assigned" queue out in ZK. My opinion this is the way to go,
but how a regionserver is to judge its load relative to others, or even learn about the load
of others, is an unanswered question. Furthermore, there must be a mechanism in place such
that some regionserver will take on a new region if all others have passed on it.  Is there
an issue up for this type of stuff yet? 

> Start/Stop of large cluster untenable
> -------------------------------------
>                 Key: HBASE-1583
>                 URL: https://issues.apache.org/jira/browse/HBASE-1583
>             Project: Hadoop HBase
>          Issue Type: Bug
>            Reporter: stack
>             Fix For: 0.20.0
> Starting and stopping a loaded large cluster is way too flakey and takes too long.  This
is 0.19.x but same issues apply to TRUNK I'd say.
> At pset with our > 100 nodes carrying 6k regions:
> + shutdown takes way too long.... maybe ten minutes or so.  We compact regions inline
with shutdown.  We should just go down.  It doesn't seem like all regionservers go down everytime
> + startup is a mess with our assigning out regions an rebalancing at same time.  By time
that the compactions on open run, it can be near an hour before whole thing settles down and
becomes useable

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message