Open Stack

On 04/04/2014 07:37 AM, Sean Dague wrote:
> An interesting conversation has cropped up over the last few days in -qa
> and -infra which I want to bring to the wider OpenStack community. When
> discussing the use of Tempest as part of the Defcore validation we came
> to an interesting question:
>
> Why does Tempest have stable/* branches? Does it need them?
>
> Historically the Tempest project has created a stable/foo tag the week
> of release to lock the version of Tempest that will be tested against
> stable branches. The reason we did that is until this cycle we had
> really limited nobs in tempest to control which features were tested.
> stable/havana means - test everything we know how to test in havana. So
> when, for instance, a new API extension landed upstream in icehouse,
> we'd just add the tests to Tempest. It wouldn't impact stable/havana,
> because we wouldn't backport changes.
>
> But is this really required?
>
> For instance, we don't branch openstack clients. They are supposed to
> work against multiple server versions. Tempest, at some level, is
> another client. So there is some sense there.
>
> Tempest now also have flags on features, and tests are skippable if
> services, or even extensions aren't enabled (all explicitly setable in
> the tempest.conf). This is a much better control mechanism than the
> course grained selection of stable/foo.
>
>
> If we decided not to set a stable/icehouse branch in 2 weeks, the gate
> would change as follows:
>
> Project masters: no change
> Project stable/icehouse: would be gated against Tempest master
> Tempest master: would double the gate jobs, gate on project master and
> project stable/icehouse on every commit.
>
> (That last one needs infra changes to work right, those are all in
> flight right now to assess doability.)
>
> Some interesting effects this would have:
>
>   * Tempest test enhancements would immediately apply on stable/icehouse *
>
> ... giving us more confidence. A large amount of tests added to master
> in every release are enhanced checking for existing function.
>
>   * Tempest test changes would need server changes in master and
> stable/icehouse *
>
> In trying tempest master against stable/havana we found a number of
> behavior changes in projects that there had been a 2 step change in the
> Tempest tests to support. But this actually means that stable/havana and
> stable/icehouse for the same API version are different. Going forward
> this would require master + stable changes on the projects + Tempest
> changes. Which would provide much more friction in changing these sorts
> of things by accident.
>
>   * Much more stable testing *
>
> If every Tempest change is gating on stable/icehouse, the week long
> stable/havana can't pass tests won't happen. There will be much more
> urgency to keep stable branches functioning.
>
>
> If we got rid of branches in Tempest the path would be:
>   * infrastructure to support this in infra - in process, probably
> landing today
>   * don't set stable/icehouse - decision needed by Apr 17th
>   * changes to d-g/devstack to be extra explicit about what features
> stable/icehouse should support in tempest.conf
>   * see if we can make master work with stable/havana to remove the
> stable/havana Tempest branch (if this is doable in a month, great, if
> not just wait for havana to age out).
>
>
> I think we would still want to declare Tempest versions from time to
> time. I'd honestly suggest a quarterly timebox. The events that are
> actually important to Tempest are less the release itself, but the eol
> of branches, as that would mean features which removed completely from
> any supported tree could be removed.
>
>
> My current leaning is that this approach would be a good thing, and
> provide a better experience for both the community and the defcore
> process. However it's a big enough change that we're still collecting
> data, and it would be interesting to hear other thoughts from the
> community at large on this approach.
>
> 	-Sean

With regard to havana, the problems with DefCore using stable/havana are 
the same as many of us have felt with testing real deployments of havana.
Master (now icehouse) has many more tests, is more robust to individual 
test failures, and is more configurable. But the work to backport 
improvements is difficult or impossible due to many refactorings on 
master, api changes, and the tempest backport policy that we don't want 
to spend our review time looking backwards. The reality is that almost 
nothing has been backported to stable/havana tempest, and we don't want 
to start doing that now. As defcore/refstack becomes a reality, more 
bugs and desired features in tempest will be found and it would be good 
if issues could be addressed on master.

The approach advocated here would prevent this from happening again with 
icehouse and going forward. That still leaves havana as an important 
case for many folks. I did an initial run of master tempest against 
havana using nova-network but no heat/ceilo/swift). 148 out of 2009 
tests failed. The failures seemed to be in these categories:

1. An api change occurred such as change in response code, added fields 
in return dicts, and others I have not yet categorized.
2. A new feature was added in icehouse and the tempest test was not 
behind a config option to see if it was enabled.
3. A bug was fixed in icehouse which is still there in havana and the 
tempest test was changed or unskipped.
4. A new test was added that never ran against havana, but could have, 
and it fails.

Even if we adopt this approach going forward, case (3,4) will continue 
to exist in future iterations. That implies the need to have a config 
option and associated test tag to say which release is targeted so that 
the test will not run against an older release. This will create some 
ugliness around such cases but seems to me less ugly than what we have 
now, which is one giant switch (new branch) that controls everything.

The easiest way to get master running against havana would be to add 
such a config and then simply skip all of the failing tests when running 
against havana. There could also be conditionals in tempest saying what 
behaviour is expected but I'm not sure we want to go there.

This approach will add clarity to the notion that OpenStack releases are 
decoupled from API versions. As Sean said, doing a "tempest-two-step" to 
implement an api change would now need to also be done on a stable branch.

  -David
>
>
> _______________________________________________
> OpenStack-dev mailing list
> 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/20140404/3dca0407/attachment.html>