[
https://issues.apache.org/jira/browse/HBASE-14867?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Enis Soztutar reopened HBASE-14867:
-----------------------------------
Reopening this since I think the patch has a problem.
This logic:
{code}
while (candidateIdx < tableRegions.size()-1) {
if (Math.abs(regionsWithSize.get(candidateIdx).getThird() -
regionsWithSize.get(candidateIdx + 1).getThird()) == 1) {
break;
}
candidateIdx++;
}
{code}
only looks for regions who are neighbors and also their region sizes are
neighbors in the regionsWithSize sorted array. This does not seem like it will
be useful in real world region distributions.
I think we should look for all neighboring region pairs and decide to split or
merge for all possible pairs. What about the other suggestions in my first
comment?
> SimpleRegionNormalizer needs to have better heuristics to trigger merge
> operation
> ---------------------------------------------------------------------------------
>
> Key: HBASE-14867
> URL: https://issues.apache.org/jira/browse/HBASE-14867
> Project: HBase
> Issue Type: Bug
> Components: master
> Affects Versions: 1.2.0
> Reporter: Romil Choksi
> Assignee: Ted Yu
> Fix For: 2.0.0, 1.2.0, 1.3.0
>
> Attachments: 14867-v2.txt, 14867-v3.txt, 14867-v4.txt
>
>
> SimpleRegionNormalizer needs to have better heuristics to trigger merge
> operation. SimpleRegionNormalizer is not able to trigger a merge action if
> the table's smallest region has neighboring regions that are larger than
> table's average region size, whereas there are other smaller regions whose
> combined size is less than the average region size.
> For example,
> - Consider a table with six region, say r1 to r6.
> - Keep r1 as empty and create some data say, 100K rows of data for each of
> the regions r2, r3 and r4. Create smaller amount of data for regions r5 and
> r6, say about 27K rows of data.
> - Run the normalizer. Verify the number the regions for that table and also
> check the master log to see if any merge action was triggered as a result of
> normalization.
> In such scenario, it would be better to have a merge action triggered for
> those two smaller regions r5 and r6 even though either of them is not the
> smallest one
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
