[
https://issues.apache.org/jira/browse/HBASE-16765?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15546402#comment-15546402
]
Lars Hofhansl commented on HBASE-16765:
---------------------------------------
I think ideally we want the following axioms:
# quick splitting and spreading of regions as the table is small
# ideally not more than one region of a table per server (MAX_FILESIZE
permitting of course)
#2 is where IncreasingToUpperBoundRegionSplitPolicy falls short.
I'd propose a step function instead: split at 2xflushsize when only one region
of the table is seen, stop splitting (i.e. constant size split policy) when
more than 1 region is seen.
This should be as close to ideal as is possible with local knowledge only
usually not leading to more than 2 regions per server (unless we need to split
more due to MAX_FILESIZE)
[~stack]
> Improve IncreasingToUpperBoundRegionSplitPolicy
> -----------------------------------------------
>
> Key: HBASE-16765
> URL: https://issues.apache.org/jira/browse/HBASE-16765
> Project: HBase
> Issue Type: Bug
> Reporter: Lars Hofhansl
>
> We just did some experiments on some larger clusters and found that while
> using IncreasingToUpperBoundRegionSplitPolicy generally works well and is
> very convenient, it does tend to produce too many regions.
> Since the logic is - by design - local, checking the number of regions of the
> table in question on the local server only, we end with more regions then
> necessary.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
