[ 
https://issues.apache.org/jira/browse/HBASE-15065?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15086396#comment-15086396
 ] 

Mikhail Antonov commented on HBASE-15065:
-----------------------------------------

On the patch..

 - do we not need regionSizeComparator anymore, and 
Collections.sort(regionsWithSize, regionSizeComparator) before we consider 
candidates for the merge? Now we just merge any two adjacent (since the 
original list of regions is sorted the way so they are all adjacent, right?) 
regions where r1+r2<avg. I assume that's fine.
  -One note here..we compute bunch of plans in one pass, without recomputing 
the avg size. So say we have regions with sizes 10, 10, 10, 60. Avg size would 
be 22.5, so with this logic we merge 2 out of 3 10-size regions and split 60. 
Do we want that, or should we recompute avg on each loop iteration, pretending 
that we already applied the proposed change?
 - I think if we compute plans all in once, we can run them concurrently more 
than 1 at a time. Running serial is probably good enough for now, but may be 
too slow to keep up with ongoing data distribution skew on large write-heavy 
clusters? Ideally we would run configurable N  number of normalization ops at 
the moment, or not more than X percent of total number of regions in this table?
 - Regarding priority, I think with computing and running multiple plans 
concurrently it's not necessary to give splits more priority. No harm to leave 
it sorted either. I'm agnostic.

> SimpleRegionNormalizer should return multiple normalization plans in one run
> ----------------------------------------------------------------------------
>
>                 Key: HBASE-15065
>                 URL: https://issues.apache.org/jira/browse/HBASE-15065
>             Project: HBase
>          Issue Type: Bug
>            Reporter: Ted Yu
>            Assignee: Ted Yu
>         Attachments: 15065-v1.txt
>
>
> This is follow up to HBASE-14867 w.r.t. SimpleRegionNormalizer
> Outline for enhancements:
> * adjustment to the period when SimpleRegionNormalizer runs
> * explore merge opportunities among all neighboring region pairs
> * return multiple normalization plans



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to