hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ted Yu (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-14309) Allow load balancer to operate when there is region in transition by adding force flag
Date Fri, 28 Aug 2015 16:50:47 GMT

    [ https://issues.apache.org/jira/browse/HBASE-14309?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14720221#comment-14720221

Ted Yu commented on HBASE-14309:

I can adopt the suggested warning in balancer.rb in the next patch.

w.r.t. the guard against system table region in transition, the reasoning is that if, e.g.
hbase:meta, is in transition, the results of reassignments can't be written. I can narrow
the condition to checking hbase:meta only if you think that is proper.

> Allow load balancer to operate when there is region in transition by adding force flag
> --------------------------------------------------------------------------------------
>                 Key: HBASE-14309
>                 URL: https://issues.apache.org/jira/browse/HBASE-14309
>             Project: HBase
>          Issue Type: Improvement
>            Reporter: Ted Yu
>            Assignee: Ted Yu
>             Fix For: 2.0.0, 1.3.0
>         Attachments: 14309-branch-1.1.txt, 14309-v1.txt, 14309-v2.txt, 14309-v3.txt,
14309-v4.txt, 14309-v5-branch-1.txt, 14309-v5.txt, 14309-v5.txt, 14309-v6.txt
> This issue adds boolean parameter, force, to 'balancer' command so that admin can force
region balancing even when there is region in transition - assuming RIT being transient.
> This enhancement was requested by some customer.
> The assumption of this change is that the operator has run hbck and has a reasonable
idea why regions are stuck in transition before using the force flag.
> There was a recent event at the customer where a cluster ended up with a small number
of regionservers hosting most of the regions on the cluster (one regionserver had 50% of the
roughly 20,000 regions). The balancer couldn't be run due to the small number of regions that
were stuck in transition. The admin ended up killing the regionservers so that reassignment
would yield a more equitable distribution of the regions.
> On a different cluster, there was a single store file that had corrupt HDFS blocks (the
SSDs on the cluster were known to lose data). However, since this single region (out of 10s
of 1000s of regions on this cluster) was stuck in transition, the balancer couldn't run.
> While the state keeping in HBase isn't so good yet that the admin can kick off the balancer
automatically in such scenarios knowing when it is safe to do so and when it is not, having
this option available for the operator to use as he / she sees fit seems prudent.

This message was sent by Atlassian JIRA

View raw message