hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Rushabh Shah <rushabh.s...@salesforce.com.INVALID>
Subject Re: Coprocessor restart after region split rollback
Date Fri, 01 Nov 2019 20:40:05 GMT
Looks like we need to open a ticket in hbase to fix this at hbase layer.

Rushabh Shah

   - Software Engineering SMTS | Salesforce
   -
      - Mobile: 213 422 9052



On Fri, Nov 1, 2019 at 1:32 PM Stack <stack@duboce.net> wrote:

> On Fri, Oct 25, 2019 at 12:40 AM Ondrej Kvasnicka <okvasnicka@fastmail.fm>
> wrote:
>
> > Hi all,
> >
> > we are seeing the following behavior in our HBase (version 1.1) and are
> > not sure whether it is an intended behavior or possibly a bug:
> >
> > We have a custom coprocessor loaded (RegionObserver implementation) and
> > see that whenever a region split takes place, prior to the split the
> > coprocessor is stopped automatically (this is expected).
> >
> > However, occasionally the region split will fail and will be rolled
> > back. After a successful rollback, the coprocessor will _not_ be
> > restarted for the region, even though it _was_ running prior to the split
> > attempt.
> >
> > Can someone confirm that this is an intended behavior? We would expect
> > the coprocessor to be automatically re-started after successful region
> > split rollback, but it does not seem to be the case.
> >
> >
> Sounds like a bug to me or more like an oversight in the split rollback
> code; we probably forgot to re-enable any CPs.
>
>
> > Are coprocessors supposed to handle the restart themselves after a
> > region split rollback, e.g. in RegionObserver#postRollbackSplit, such as
> > in the following example?:
> >
> >
> >
> https://github.com/splicemachine/spliceengine/blob/master/hbase_pipeline/src/main/java/com/splicemachine/derby/hbase/SpliceIndexObserver.java#L206
> >
> >
> Looks like they ran into your problem (smile) and this was their
> workaround.
>
> Any chance of upgrading? 1.1. is pretty old by now.
>
> Thanks,
> S
>
>
> > Thanks,
> > Ondrej
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message