[
https://issues.apache.org/jira/browse/HBASE-4194?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ted Yu updated HBASE-4194:
--------------------------
Description:
When running RegionSplitter, our app devs noticed that they were getting a lot
of NSREs. This is caused by 2 factors:
1. the split itself will cause an NSRE
2. any load balancing will cause one.
The former cannot be helped. We can more tightly control load balancing
though. Instead of doing a name-sorted round-robin split across RS in the
tier, we could sort the RS's by region count. That way, we only split an RS
with 10 regions after there are no more RS with 9 regions. This will prevent
the load balancing slop from kicking in and will fix the problem where
restarting RegionSplitter always starts splitting at RS #1.
was:
When running RegionSplitter, our app devs noticed that they were getting a lot
of NSREs. This is caused by 2 factors:
1. the split itself will cause an NSRE
2. any load balancing will cause one.
The former cannot be helped. We can more tightly control load balancing
though. Instead of doing a name-sorted round-robin split across RS in the
tier, we could sorted the RS by region count. That way, we only split an RS
with 10 regions after there are no more RS with 9 regions. This will prevent
the load balancing slop from kicking in and will fix the problem where
restarting RegionSplitter always starts splitting at RS #1.
Summary: RegionSplitter: Split on under-loaded region servers first
(was: RegionSplitter: Split on low-loaded regions first)
> RegionSplitter: Split on under-loaded region servers first
> ----------------------------------------------------------
>
> Key: HBASE-4194
> URL: https://issues.apache.org/jira/browse/HBASE-4194
> Project: HBase
> Issue Type: Improvement
> Reporter: Nicolas Spiegelberg
> Assignee: Nicolas Spiegelberg
> Priority: Trivial
> Fix For: 0.92.0
>
> Attachments: HBASE-4194.patch
>
>
> When running RegionSplitter, our app devs noticed that they were getting a
> lot of NSREs. This is caused by 2 factors:
> 1. the split itself will cause an NSRE
> 2. any load balancing will cause one.
> The former cannot be helped. We can more tightly control load balancing
> though. Instead of doing a name-sorted round-robin split across RS in the
> tier, we could sort the RS's by region count. That way, we only split an RS
> with 10 regions after there are no more RS with 9 regions. This will prevent
> the load balancing slop from kicking in and will fix the problem where
> restarting RegionSplitter always starts splitting at RS #1.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
