On 10/18/2016 12:05 PM, Adam Harwell wrote:
> Inline comments.
> On Wed, Oct 19, 2016 at 1:38 AM Thomas Goirand <z...@debian.org
> <mailto:z...@debian.org>> wrote:
> On 10/18/2016 02:37 AM, Ian Cordasco wrote:
> > On Oct 17, 2016 7:27 PM, "Thomas Goirand" <z...@debian.org
> > <mailto:z...@debian.org <mailto:z...@debian.org>>> wrote:
> >> On 10/17/2016 08:43 PM, Adam Harwell wrote:
> >> > Jim, that is exactly my thought -- the main focus of g-r as far
> as I was
> >> > aware is to maintain interoperability between project
> dependencies for
> >> > openstack deploys, and since our amphora image is totally
> separate, it
> >> > should not be restricted to g-r requirements.
> >> The fact that we have a unified version number of a given lib in
> all of
> >> OpenStack is also because that's a requirement of downstream distros.
> >> Imagine that someone would like to build the Octavia image using
> >> exclusively packages from <your-favorite-distro-here>...
> >> > I brought this up, but
> >> > others thought it would be prudent to go the g-r route anyway.
> >> It is, and IMO you should go this route.
> > I'm not convinced by your arguments here, Thomas. If the distributor
> > were packaging Octavia for X but the image is using some other
> > system, say Y, why are X's packages relevant?
> What if operating systems would be the same?
> We still want to install from pypi, because we still want deployers to
> build images for their cloud using our DIB elements. There is absolutely
> no situation in which I can imagine we'd want to install a binary
> packaged version of this. There's a VERY high chance we will soon be
> using a distro that isn't even a supported OpenStack deploy target...
> As a Debian package maintainer, I really prefer if the underlying images
> can also be Debian (and preferably Debian stable everywhere).
> Sure, I love Debian too, but we're investigating things like Alpine and
> Cirros as our base image, and there's pretty much zero chance anyone
> will package ANY of our deps for those distros. Cirros doesn't even have
> a package manager AFAIK.
> > I would think that if this
> > is something inside an image going to be launched by Octavia that
> > co-installibilty wouldn't really be an issue.
> The issue isn't co-instability, but the fact that downstream
> distribution vendors will only package *ONE* version of a given python
> module. If we have Octavia with version X, and another component of
> OpenStack with version Y, then we're stuck with Octavia not being
> packageable in downstream distros.
> Octavia will not use gunicorn for its main OpenStack API layer. It will
> continue to be packagable regardless of whether gunicorn is available.
> Gunicorn is used for our *amphora image*, which is not part of the main
> deployment layer. It is part of our *dataplane*. It is unrelated to any
> part of Octavia that is deployed as part of the main service layer of
> Openstack. In fact, in production, deployers may completely ignore
> gunicorn altogether and use a different solution, that is up to the way
> they build their amphora image (which, again, is not part of the main
> deployment). We just use gunicorn in the image we use for our gate tests.
> > I don't lean either way right now, so I'd really like to
> understand your
> > point of view, especially since right now it isn't making much
> sense to me.
> Do you understand now? :)
> I see what you are saying, but I assert it does not apply to our case at
> all. Do you see how our case is different?
I totally understand, and I can see why it would seem very different.
Consider a few things though:
- OpenStack tries its best to not pick favorites for OS, and I think the
same applies to guest VMs, even if they just seem like appliances. While
we as upstream may be looking at using something like alpine as the base
OS for the service VM appliance, that does not necessarily imply that
all deployers _must_ use Alpine in their service VM, for exactly the
reason you mention (you intend for them to run diskimage-builder themselves)
- If a deployer happens to have a strong preference for a given OS (I
know I've been on customer calls where an OpenStack product having a tie
to a particular OS that is not the one that is in the vetted choice
otherwise at that customer was an issue) - then the use of dib by the
deployer allows them to choose to base their service VM on the OS of
their choice. That's pretty awesome.
- If that deployer similarly has an aversion to deploying any software
that didn't come from distro packages, one could imagine that they would
want their diskimage-builder run to install the python code not from pip
but instead from apt/dnf/zypper. There is obviously nothing stoppping that.
- Finally, if Debian, Ubuntu, Red Hat, SuSE or Gentoo chose to want to
make the parts available in their distro such that a diskimage-builder
run using their OS as a base OS and using packages to install the python
code would work, then it would be the same OS repo as it would for other
things. For the distros that only allow one version of a particular
piece of software at a time, that would mean they would need packages of
the software that you expect to be installed inside the service VM and
So while it's different as a developer in the gate, the principles
behind why we share a set of global-requirements still hold true.
That said - as I mentioned on IRC, I have no personal issue adding
gunicorn to global-requirements. Seems like a fine choice to me. I
mostly barfed the above as a long-winded attempt at explaining how even
though it is different, it's still the same. In fact, I think the fact
that it's the same goes to show you've been doing a good job. :)
OpenStack Development Mailing List (not for usage questions)