Open Stack

Excerpts from Matthew Treinish's message of 2016-06-14 15:12:45 -0400:
> On Tue, Jun 14, 2016 at 02:41:10PM -0400, Doug Hellmann wrote:
> > Excerpts from Matthew Treinish's message of 2016-06-14 14:21:27 -0400:
> > > On Tue, Jun 14, 2016 at 10:57:05AM -0700, Chris Hoge wrote:
> > > > Last year, in response to Nova micro-versioning and extension updates[1],
> > > > the QA team added strict API schema checking to Tempest to ensure that
> > > > no additional properties were added to Nova API responses[2][3]. In the
> > > > last year, at least three vendors participating the the OpenStack Powered
> > > > Trademark program have been impacted by this change, two of which
> > > > reported this to the DefCore Working Group mailing list earlier this year[4].
> > > > 
> > > > The DefCore Working Group determines guidelines for the OpenStack Powered
> > > > program, which includes capabilities with associated functional tests
> > > > from Tempest that must be passed, and designated sections with associated
> > > > upstream code [5][6]. In determining these guidelines, the working group
> > > > attempts to balance the future direction of development with lagging
> > > > indicators of deployments and user adoption.
> > > > 
> > > > After a tremendous amount of consideration, I believe that the DefCore
> > > > Working Group needs to implement a temporary waiver for the strict API
> > > > checking requirements that were introduced last year, to give downstream
> > > > deployers more time to catch up with the strict micro-versioning
> > > > requirements determined by the Nova/Compute team and enforced by the
> > > > Tempest/QA team.
> > > 
> > > I'm very much opposed to this being done. If we're actually concerned with
> > > interoperability and verify that things behave in the same manner between multiple
> > > clouds then doing this would be a big step backwards. The fundamental disconnect
> > > here is that the vendors who have implemented out of band extensions or were
> > > taking advantage of previously available places to inject extra attributes
> > > believe that doing so means they're interoperable, which is quite far from
> > > reality. **The API is not a place for vendor differentiation.**
> > 
> > This is a temporary measure to address the fact that a large number
> > of existing tests changed their behavior, rather than having new
> > tests added to enforce this new requirement. The result is deployments
> > that previously passed these tests may no longer pass, and in fact
> > we have several cases where that's true with deployers who are
> > trying to maintain their own standard of backwards-compatibility
> > for their end users.
> 
> That's not what happened though. The API hasn't changed and the tests haven't
> really changed either. We made our enforcement on Nova's APIs a bit stricter to
> ensure nothing unexpected appeared. For the most these tests work on any version
> of OpenStack. (we only test it in the gate on supported stable releases, but I
> don't expect things to have drastically shifted on older releases) It also
> doesn't matter which version of the API you run, v2.0 or v2.1. Literally, the
> only case it ever fails is when you run something extra, not from the community,
> either as an extension (which themselves are going away [1]) or another service
> that wraps nova or imitates nova. I'm personally not comfortable saying those
> extras are ever part of the OpenStack APIs.
>
> > We have basically three options.
> > 
> > 1. Tell deployers who are trying to do the right for their immediate
> >    users that they can't use the trademark.
> > 
> > 2. Flag the related tests or remove them from the DefCore enforcement
> >    suite entirely.
> > 
> > 3. Be flexible about giving consumers of Tempest time to meet the
> >    new requirement by providing a way to disable the checks.
> > 
> > Option 1 goes against our own backwards compatibility policies.
> 
> I don't think backwards compatibility policies really apply to what what define
> as the set of tests that as a community we are saying a vendor has to pass to
> say they're OpenStack. From my perspective as a community we either take a hard
> stance on this and say to be considered an interoperable cloud (and to get the
> trademark) you have to actually have an interoperable product. We slowly ratchet
> up the requirements every 6 months, there isn't any implied backwards
> compatibility in doing that. You passed in the past but not in the newer stricter
> guidelines.
> 
> Also, even if I did think it applied, we're not talking about a change which
> would fall into breaking that. The change was introduced a year and half ago
> during kilo and landed a year ago during liberty:
> 
> https://review.openstack.org/#/c/156130/
> 
> That's way longer than our normal deprecation period of 3 months and a release
> boundary.
> 
> > 
> > Option 2 gives us no winners and actually reduces the interoperability
> > guarantees we already have in place.
> > 
> > Option 3 applies our usual community standard of slowly rolling
> > forward while maintaining compatibility as broadly as possible.
> 
> Except in this case there isn't actually any compatibility being maintained.
> We're saying that we can't make the requirements for interoperability testing
> stricter until all the vendors who were passing in the past are able to pass
> the stricter version.
> 
> >
> > No one is suggesting that a permanent, or even open-ended, exception
> > be granted.
> 
> Sure, I agree an permanent or open-ended exception would be even worse. But, I
> still think as a community we need to draw a hard line in the sand here. Just
> because this measure is temporary doesn't make it any more palatable.
> 
> By doing this, even as a temporary measure, we're saying it's ok to call things
> an OpenStack API when you add random gorp to the responses. Which is something we've
> very clearly said as a community is the exact opposite of the case, which the
> testing reflects. I still contend just because some vendors were running old
> versions of tempest and old versions of openstack where their incompatible API
> changes weren't caught doesn't mean they should be given pass now.

Nobody is saying random gorp is OK, and I'm not sure "line in the
sand" rhetoric is really constructive. The issue is not with the
nature of the API policies, it's with the implementation of those
policies and how they were rolled out.

DefCore defines its rules using named tests in Tempest.  If these
new enforcement policies had been applied by adding new tests to
Tempest, then DefCore could have added them using its processes
over a period of time and we wouldn't have had any issues. That's
not what happened. Instead, the behavior of a bunch of *existing*
tests changed. As a result, deployments that have not changed fail
tests that they used to pass, without any action being taken on the
deployer's part. We've moved the goal posts on our users in a way
that was not easily discoverable, because it couldn't be tracked
through the (admittedly limited) process we have in place for doing
that tracking.

So, we want a way to get the test results back to their existing
status, which will then let us roll adoption forward smoothly instead
of lurching from "pass" to "fail" to "pass".

We should, separately, address the process issues and the limitations
this situation has exposed.  That may mean changing the way DefCore
defines its policies, or tracks things, or uses Tempest.  For
example, in the future, we may want tie versions of Tempest to
versions of the trademark more closely, so that it's possible for
someone running the Mitaka version of OpenStack to continue to use
the Mitaka version of Tempest and not have to upgrade Tempest in
order to retain their trademark (maybe that's how it already works?).
We may also need to consider that test implementation details may
change, and have a review process within DefCore to help expose
those changes to make them clearer to deployers.

Fixing the process issue may also mean changing the way we implement
things in Tempest. In this case, adding a flag helps move ahead
more smoothly. Perhaps we adopt that as a general policy in the
future when we make underlying behavioral changes like this to
existing tests.  Perhaps instead we have a policy that we do not
change the behavior of existing tests in such significant ways, at
least if they're tagged as being used by DefCore. I don't know --
those are things we need to discuss.

Doug

> 
> -Matt Treinish
> 
> [1] http://lists.openstack.org/pipermail/openstack-dev/2016-June/097285.html
> > 
> > Doug
> > 
> > > 
> > > As a user of several clouds myself I can say that having random gorp in a
> > > response makes it much more difficult to use my code against multiple clouds. I
> > > have to determine which properties being returned are specific to that vendor's
> > > cloud and if I actually need to depend on them for anything it makes whatever
> > > code I'm writing incompatible for using against any other cloud. (unless I
> > > special case that block for each cloud) Sean Dague wrote a good post where a lot
> > > of this was covered a year ago when microversions was starting to pick up steam:
> > > 
> > > https://dague.net/2015/06/05/the-nova-api-in-kilo-and-beyond-2
> > > 
> > > I'd recommend giving it a read, he explains the user first perspective more
> > > clearly there.
> > > 
> > > I believe Tempest in this case is doing the right thing from an interoperability
> > > perspective and ensuring that the API is actually the API. Not an API with extra
> > > bits a vendor decided to add. I don't think a cloud or product that does this
> > > to the api should be considered an interoperable OpenStack cloud and failing the
> > > tests is the correct behavior.
> > > 
> > > -Matt Treinish
> > > 
> > > > 
> > > > My reasoning behind this is that while the change that enabled strict
> > > > checking was discussed publicly in the developer community and took
> > > > some time to be implemented, it still landed quickly and broke several
> > > > existing deployments overnight. As Tempest has moved forward with
> > > > bug and UX fixes (some in part to support the interoperability testing
> > > > efforts of the DefCore Working Group), using an older versions of Tempest
> > > > where this strict checking is not enforced is no longer a viable solution
> > > > for downstream deployers. The TC has passed a resolution to advise
> > > > DefCore to use Tempest as the single source of capability testing[7],
> > > > but this naturally introduces tension between the competing goals of
> > > > maintaining upstream functional testing and also tracking lagging
> > > > indicators.
> > > > 
> > > > My proposal for addressing this problem approaches it at two levels:
> > > > 
> > > > * For the short term, I will submit a blueprint and patch to tempest that
> > > >   allows configuration of a grey-list of Nova APIs where strict response
> > > >   checking on additional properties will be disabled. So, for example,
> > > >   if the 'create  servers' API call returned extra properties on that call,
> > > >   the strict checking on this line[8] would be disabled at runtime.
> > > >   Use of this code path will emit a deprecation warning, and the
> > > >   code will be scheduled for removal in 2017 directly after the release
> > > >   of the 2017.01 guideline. Vendors would be required so submit the
> > > >   grey-list of APIs with additional response data that would be
> > > >   published to their marketplace entry.
> > > > 
> > > > * Longer term, vendors will be expected to work with upstream to update
> > > >   the API for returning additional data that is compatible with
> > > >   API micro-versioning as defined by the Nova team, and the waiver would
> > > >   no longer be allowed after the release of the 2017.01 guideline.
> > > > 
> > > > For the next half-year, I feel that this approach strengthens interoperability
> > > > by accurately capturing the current state of OpenStack deployments and
> > > > client tools. Before this change, additional properties on responses
> > > > weren't explicitly disallowed, and vendors and deployers took advantage
> > > > of this in production. While this is behavior that the Nova and QA teams
> > > > want to stop, it will take a bit more time to reach downstream. Also, as
> > > > of right now, as far as I know the only client that does strict response
> > > > checking for Nova responses is the Tempest client. Currently, additional
> > > > properties in responses are ignored and do not break existing client
> > > > functionality. There is currently little to no harm done to downstream
> > > > users by temporarily allowing additional data to be returned in responses.
> > > > 
> > > > Thanks,
> > > > 
> > > > Chris Hoge
> > > > Interop Engineer
> > > > OpenStack Foundation
> > > > 
> > > > [1] https://specs.openstack.org/openstack/nova-specs/specs/kilo/implemented/api-microversions.html
> > > > [2] http://lists.openstack.org/pipermail/openstack-dev/2015-February/057613.html
> > > > [3] https://review.openstack.org/#/c/156130
> > > > [4] http://lists.openstack.org/pipermail/defcore-committee/2016-January/000986.html
> > > > [5] http://git.openstack.org/cgit/openstack/defcore/tree/2015.07.json
> > > > [6] http://git.openstack.org/cgit/openstack/defcore/tree/2016.01.json
> > > > [7] http://git.openstack.org/cgit/openstack/governance/tree/resolutions/20160504-defcore-test-location.rst
> > > > [8] http://git.openstack.org/cgit/openstack/tempest-lib/tree/tempest_lib/api_schema/response/compute/v2_1/servers.py#n39