Open Stack

On Oct 11, 2013, at 19:29 , Russell Bryant <rbryant at redhat.com<mailto:rbryant at redhat.com>>
 wrote:

On 10/11/2013 12:04 PM, John Griffith wrote:



On Fri, Oct 11, 2013 at 9:12 AM, Bob Ball <bob.ball at citrix.com<mailto:bob.ball at citrix.com>
<mailto:bob.ball at citrix.com>> wrote:

-----Original Message-----
From: Russell Bryant [mailto:rbryant at redhat.com<http://redhat.com>
   <mailto:rbryant at redhat.com>]
Sent: 11 October 2013 15:18
To: openstack-dev at lists.openstack.org<mailto:openstack-dev at lists.openstack.org>
   <mailto:openstack-dev at lists.openstack.org>
Subject: Re: [openstack-dev] [Hyper-V] Havana status

As a practical example for Nova: in our case that would simply
   include the
following subtrees: "nova/virt/hyperv" and "nova/tests/virt/hyperv".

If maintainers of a particular driver would prefer this sort of
autonomy, I'd rather look at creating new repositories.  I'm
   completely
open to going that route on a per-driver basis.  Thoughts?

   I think that all drivers that are officially supported must be
   treated in the same way.

   If we are going to split out drivers into a separate but still
   official repository then we should do so for all drivers.  This
   would allow Nova core developers to focus on the architectural side
   rather than how each individual driver implements the API that is
   presented.

   Of course, with the current system it is much easier for a Nova core
   to identify and request a refactor or generalisation of code written
   in one or multiple drivers so they work for all of the drivers -
   we've had a few of those with XenAPI where code we have written has
   been pushed up into Nova core rather than the XenAPI tree.

   Perhaps one approach would be to re-use the incubation approach we
   have; if drivers want to have the fast-development cycles uncoupled
   from core reviewers then they can be moved into an incubation
   project.  When there is a suitable level of integration (and
   automated testing to maintain it of course) then they can graduate.
    I imagine at that point there will be more development of new
   features which affect Nova in general (to expose each hypervisor's
   strengths), so there would be fewer cases of them being restricted
   just to the virt/* tree.

   Bob

   _______________________________________________
   OpenStack-dev mailing list
   OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
   <mailto:OpenStack-dev at lists.openstack.org>
   http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev


I've thought about this in the past, but always come back to a couple of
things.

Being a community driven project, if a vendor doesn't want to
participate in the project then why even pretend (ie having their own
project/repo, reviewers etc).  Just post your code up in your own github
and let people that want to use it pull it down.  If it's a vendor
project, then that's fine; have it be a vendor project.

In my opinion pulling out and leaving things up to the vendors as is
being described has significant negative impacts.  Not the least of
which is consistency in behaviors.  On the Cinder side, the core team
spends the bulk of their review time looking at things like consistent
behaviors, missing features or paradigms that are introduced that
"break" other drivers.  For example looking at things like, are all the
base features implemented, do they work the same way, are we all using
the same vocabulary, will it work in an multi-backend environment.  In
addition, it's rare that a vendor implements a new feature in their
driver that doesn't impact/touch the core code somewhere.

Having drivers be a part of the core project is very valuable in my
opinion.  It's also very important in my view that the core team for
Nova actually has some idea and notion of what's being done by the
drivers that it's supporting.  Moving everybody further and further into
additional private silos seems like a very bad direction to me, it makes
things like knowledge transfer, documentation and worst of all bug
triaging extremely difficult.

I could go on and on here, but nobody likes to hear anybody go on a
rant.  I would just like to see if there are other alternatives to
improving the situation than fragmenting the projects.

Really good points here.  I'm glad you jumped in, because the underlying
issue here applies well to other projects (especially Cinder and Neutron).

So, the alternative to the split official repos is to either:

1) Stay in tree, participate, and help share the burden of maintenance
of the project


Which means getting back to the status quo with all the problems we had. I hope we'll be able to find something better than that.

or

2) Truly be a vendor project, and to make that more clear, split out
into your own (not nova) repository.

I explained in my previous relpy some points about why it would be IMO totally counterproductive to have a fork outside of OpenStack.
Our goal is to have more and more independent community members contributing to the driver and forking would definitely go in the opposite direction. What would be the official version of the driver at that point? The old copy in Nova that nobody practically mantains or the shiny new version that a vendor mantains w/o community contribution?

I agreed on your initial proposal: a separate OpenStack project, independent from Nova. What's the difference at this point from the "split out into your own" option that you are suggesting now from a management perspective?

Talking about new community involvements, newcomers are getting very frustrated to have to wait for weeks to get a meaningful review and I cannot blame them if they don't want to get involved anymore after the first patch!
This makes appear public bureocracy here in eastern Europe a lightweight process in comparison! :-)

Let me add another practical reason about why a separate OpenStack project would be a good idea:

Anytime that we commit a driver specific patch, a lot of Tempests tests are executed on Libvirt and XenServer (for Icehouse those will be joined by another pack of CIs, including Hyper-V).
On the jenkins side, we have to wait for regression tests that have nothing to do with the code that we are pushing. During the H3 push, this meant waiting for hours and hoping not to have to issue the 100th "recheck / revery bug xxx".

A separate project would obviously include only the required tests and be definitely more lightweight, offloading quite some work from the SmokeStack / Jenkins job for everybody's happiness.



#2 really isn't so bad if that's what you want, and it honestly sounds
like this may be the case for the Hyper-V team.  You could still be very
close to the OpenStack community by using the same tools.  Use
stackforge for the code (same gerrit, jenkins, etc), and have your own
launchpad project.  If you go that route, you get all of the control you
want, but the project is still very closely tied to the rest of the code
in the OpenStack community.

--
Russell Bryant

_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20131011/9408e868/attachment.html>