[
https://issues.apache.org/jira/browse/HBASE-71?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12572713#action_12572713
]
Bryan Duxbury commented on HBASE-71:
------------------------------------
One problem is that when a region server checks in, if it's overloaded, we
don't know which regions we should tell it to offline. The current flow
requires us to tell it to close specific regions by name. However, unless
they're meta regions, we won't know their names, as a region server only
reports its aggregate load, not the individual regions and loads.
We have a few options here:
* Have regionservers report in with N region names along with their normal
reports. If a region server is overloaded, then the master can pick from those
regions and choose some to offline. At first, the way that the region server
picks the N regions could be essentially arbitrary, but in the future we could
make it do something like send the 10 most requested regions, or some other
load-based factor. Downside of this approach is that it would potentially bloat
the reports, even when rebalancing isn't required.
* Determine what regions to relocate during meta scans. This way we have
access to all the region names we need as well as who has them. Downside is
that it would really increase the amount of stuff MetaScanner does, maybe
complicating it significantly. We'd also have to figure out a way to do this
without having to load all of the region->server assignments into memory at
once, so it would be non-trivial.
* Have a new message type that just tells a region server to close a number of
regions, letting it choose which to close. It would then report what regions it
chose to close. Downside of this is that we'd have to assume a reported close
that wasn't specifically marked for closing means rebalancing. We wouldn't know
where those regions are in terms of the overall flow at that moment; they'd
just be offlining.
Are there any alternatives that I missed?
> [hbase] Master should rebalance region assignments periodically
> ---------------------------------------------------------------
>
> Key: HBASE-71
> URL: https://issues.apache.org/jira/browse/HBASE-71
> Project: Hadoop HBase
> Issue Type: Improvement
> Components: master
> Reporter: Bryan Duxbury
> Assignee: Bryan Duxbury
> Fix For: 0.2.0
>
>
> The master currently only does region assignments at startup or when there
> are splits or dead regionservers. This means that if you join a new
> regionserver to the cluster after startup, it does not get assigned a fair
> share of the already-served regions as you would expect. It only gets a share
> of new regions being served.
> The master should periodically check the balance of regions, based on
> whatever assignment function, instead of in reaction to the above listed
> events.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
