Re: Testing with Bluestore and Filestore

[Sage Weil <sweil@redhat.com>]

On Thu, 2 Mar 2017, Mark Nelson wrote:
> On 03/02/2017 12:40 PM, Jason Dillaman wrote:
> > On Thu, Mar 2, 2017 at 1:36 PM, Gregory Farnum <gfarnum@redhat.com> wrote:
> > > On Thu, Mar 2, 2017 at 8:36 AM, Brett Niver <bniver@redhat.com> wrote:
> > > > Sage,
> > > > 
> > > > Perhaps this has already been discussed, but Jason had a good idea,
> > > > that perhaps instead of testing both BS and FS upstream independently,
> > > >  and doubling our testing, we might have teuthology randomly pick
> > > > which to use.
> > > > 
> > > > That would potentially reduce TCs and also give us the testing that
> > > > would help prepare for working thru the migration/transition once BS
> > > > is ready.
> > > 
> > > This is really what the subset functionality is for, although I think
> > > John had started working on a feature in this area that would let the
> > > subset functionality go across the rados configuration options while
> > > guaranteeing full combinatorial explosion on the FS tests.
> > > -Greg
> > 
> > I was actually thinking that OSDs should be randomly assigned an
> > object store by teuthology (i.e. roughly 50% FileStore vs 50%
> > BlueStore in the same cluster in the same test run). Right now, the
> > tests (at least for RBD) running 100% FileStore or 100% BlueStore with
> > no intermixing (and no verification that PGs comprised of FileStore
> > and BlueStore OSDs behave properly). That would also have the added
> > benefit of eliminating half the test cases in the RBD suite even
> > before you apply a subset.
> 
> Seems reasonable to me.  If the clusters are small we'll get pure bluestore
> and pure filestore every once and a while too which would be good.

Yeah, we can add a 'random' option that will pick between filestore or 
bluestore.  I'm inclined to keep them separate in rados so that we can 
identify failures easily, but for other suites this probably makes sense.

s