nvidia-experimental package with expedited SRU process?

[Bryce Harrington]

On Mon, Sep 17, 2012 at 04:44:46PM +0100, Colin Watson wrote:
> On Mon, Sep 10, 2012 at 03:14:49PM -0700, Kees Cook wrote:
> > +1, I think this sounds good. I think the benefits outweigh the risk.
> > 
> > The risk I see is in wondering what percentage of the install base will
> > end up on it as a result of a game install. If it's large, we run a
> > larger risk of breaking someone in the face of a update regression.
> 
> I share this concern.  The problem with introducing this kind of
> "bleeding-edge, but important things may need it" package is that over
> time enough important things need it that you find that it's become
> critical-must-not-break without you noticing, and then you end up back
> at the start of the cycle.
> 
> Bryce, you mention "used when needed by specific games" above.  I'm not
> really familiar with the details here and I'd like to ensure I know
> exactly what you mean.  Is this something that's per-X-session, so we're
> talking about using it when any of a set of specific games are
> installed, or per-application, so you could run multiple games in the
> same session with different drivers?

I mean the former.  Our expectation is that this would come into play
only for certain newly released commercial (paid-for) games.

> Even in the latter case, I can imagine an uncomfortable situation where
> we do an -experimental update for Important Game Vendor A and find that
> it breaks Important Game Vendor B's best-selling title from last month.
> Would we need to arrange testing with all sufficiently important vendors
> before releasing updates?  (But then we have the problem where we can't
> release a fix for Important Game Vendor A because Important Game Vendor
> B hasn't responded, and the incentives here may well be perverse ...)

Some extra background, apologies if you already know this.  NVIDIA
typically puts out a series of beta drivers over the course of several
weeks, numbered like 123.11, 123.22, 123.33.  Then at some point NVIDIA
declares it stable and puts out the official release, say, 123.44.  Then
sometimes they will put out some bugfix-only stable releases, 123.55,
123.66.  At some point (a few weeks or months) after the official
release they'll start a new beta series.

The intent of the nvidia-experimental package is to track the beta
drivers (123.11, 123.22, 123.33).  So there's two ways the situation you
describe could come about:  1) Game A requires 123.11 and breaks on
123.22, while Game B requires 123.22 and newer; or 2) Game A requires
123.11 or newer but breaks on 130.* but Game B requires 130.08 or newer.

I think #2 is a more credible risk than #1 but let's examine both.

In the first case, both companies (and NVIDIA) are going to be motivated
to get games A and B supported in the official driver release, and
likely will be testing internally and giving NVIDIA feedback directly.
Thus I think it would be entirely appropriate for us to request they
provide us with testing feedback.  We could also require that they
provide us with NDA access to the game so in a worst case situation like
you describe, we can do the testing ourselves.  This won't scale well if
we have more than a few game vendors wanting to do this, but I guess we
can cross that bridge when we get to it.

In the second case, Vendor A may well have moved on to other projects so
can't be depended on for giving testing feedback.  If we have access to
the game we can test it ourselves, as above.  However, it concerns me
that if we did find problems, it effectively blocks us being able to do
future updates of nvidia-experimental.  That suggests that perhaps what
we should be thinking about is not one but a series of experimental
packages.  E.g. nvidia-experimental-123, nvidia-experimental-130, ...
Of course, if the user wants to have both games installed and play
simultaneously then they're going to have a problem, but at least this
way they're opting-in to the breakage rather than it coming as an
unexpected update.

So to summarize, the changes to the proposals would be:

 1.  Require game vendors that wish to participate in this provide us
     with testing feedback on subsequent beta versions
 2.  Require game vendors provide us with copies of the game.
 3.  Make separate nvidia-experimental-NNN packages for each major
     series we wish to track.

Does that address the concerns you raised?

Bryce