[jira] [Commented] (HBASE-21006) Balancer - data locality drops hugely after rolling restarts on cluster, not factoring in data locality enough?

["Hari Sekhon (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HBASE-21006?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16568468#comment-16568468 ] 

Hari Sekhon commented on HBASE-21006:
-------------------------------------

I've linked HBASE-18164 as the Hubspot guys did some excellent work in tuning and balancing - I've done an excessive amount of tuning to this HBase OpenTSDB cluster over the last few weeks, a lot of it based on their research mixed in with a lot of my own.

I don't know if the faster convergence in that ticket would solve this too but suspect it might if the cost calculation chooses to shuffle regions to improve locality - I don't have access to anywhere near the hardware and load comparable to this environment to test it.

> Balancer - data locality drops hugely after rolling restarts on cluster, not factoring in data locality enough?
> ---------------------------------------------------------------------------------------------------------------
>
>                 Key: HBASE-21006
>                 URL: https://issues.apache.org/jira/browse/HBASE-21006
>             Project: HBase
>          Issue Type: Improvement
>          Components: Balancer
>    Affects Versions: 1.1.2
>         Environment: HDP 2.6
>            Reporter: Hari Sekhon
>            Priority: Major
>
> After doing rolling restarts of my HBase cluster the data locality drops by 30-40% every time which implies the stochastic balancer is not optimizing for data locality enough, at least not under the circumstance of rolling restarts.
> The stochastic balancer is supposed to take data locality in to account but if this is the case, surely it should move regions back to their original RegionServer and data locality should return back to around where it was, not drop by 30-40% percent every time I need to do some tuning and a rolling restarts.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)