[
https://issues.apache.org/jira/browse/HBASE-15504?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15239627#comment-15239627
]
Elliott Clark commented on HBASE-15504:
---------------------------------------
bq.But if config hbase.master.balancer.stochastic.maxMovePercent to 1.0
That's not the default and it's not a good default. From experience the more
regions are moving at a time the longer each move takes. So it's better for
data availability to move just a few regions each time until the cluster gets
balanced. That results in less IO spikes, fewer spikes to meta, and more
reliability. If anything we the balancer should go even farther and only allow
< 1% of regions to move, but that gets really tricky when there aren't many
regions, and I don't really want to add more tuning nobs to the balancer.
> Fix Balancer in 1.3 not moving regions off overloaded regionserver
> ------------------------------------------------------------------
>
> Key: HBASE-15504
> URL: https://issues.apache.org/jira/browse/HBASE-15504
> Project: HBase
> Issue Type: Bug
> Affects Versions: 1.3.0
> Reporter: Elliott Clark
> Assignee: Elliott Clark
> Fix For: 2.0.0, 1.3.0, 1.4.0
>
> Attachments: HBASE-15504-v3.patch
>
>
> We pushed 1.3 to a couple of clusters. In some cases the regions were
> assigned VERY un-evenly and the regions would not move after that.
> We ended up with one rs getting thousands of regions and most servers getting
> 0. Running balancer would do nothing. The balancer would say that it couldn't
> find a solution with less than the current cost.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
