-Re: review request: HBASE-7403 Online Merge
Kevin O'dell 2013-03-16, 13:21
Why would we use that merge tool when hbck will repair that? Should we
throw a warning and tell the user to run repair first?
On Mar 16, 2013 9:17 AM, "Ted" <[EMAIL PROTECTED]> wrote:
> Chunhui replied to this question on review board.
> Basically the force option is to repair overlapping regions or table with
> hole in its regions.
> Personally I think online merge should detect merging regions with hole in
> between them and not require force flag in that case because logically
> they're adjacent.
> On Mar 16, 2013, at 5:52 AM, Jean-Marc Spaggiari <[EMAIL PROTECTED]>
> > Hi Ted,
> > I jut gave it a look.
> > I have updated it on the RB.
> > Overall, this is very good and I'm eager to see that integrated! I'm
> > waiting for this feature since the beginning ;)
> > Regarding non adjacent regions merge? Will the system still be
> > consistent after that? Or will hbck report some regions overlaps?
> > JM
> > 2013/3/16 Ted Yu <[EMAIL PROTECTED]>:
> >> Hi,
> >> On behalf of Chunhui, I am requesting review for HBASE-7403 Online
> >> This JIRA was created 3 months ago.
> >> Chunhui has responded to review comments very promptly, including a
> >> rewrite around the time split transaction was rewritten.
> >> This feature has widely been requested. I feel the patch is mostly
> ready to
> >> go in.
> >> Here is brief recap of the steps.
> >> Process of merging two regions:
> >> a.client sends RPC (dispatch merging regions) to master
> >> b.master moves the regions together (on the regionserver where the more
> >> heavily loaded region resided)
> >> c.master sends RPC (merge regions) to this regionserver
> >> d.Regionserver executes the region merge transaction in the thread pool
> >> I think step b is a nice simplification for the problem. In previous
> >> versions of the patch, the two merging regions stay on respective
> >> which required more complex coordination through zookeeper.
> >> High level comment as well as detailed review are both welcome.
> >> Thanks