Open Stack

On Thu, May 01, 2014 at 06:18:10PM +0900, Ken'ichi Ohmichi wrote:
> # Sorry for sending this again, previous mail was unreadable.
> 
> 2014-04-28 11:54 GMT+09:00 Ken'ichi Ohmichi <ken1ohmichi at gmail.com>:
> >>
> >> This is also why there are a bunch of nova v2 extensions that just add
> >> properties to an existing API. I think in v3 the proposal was to do this with
> >> microversioning of the plugins. (we don't have a way to configure
> >> microversioned v3 api plugins in tempest yet, but we can cross that bridge when
> >> the time comes) Either way it will allow tempest to have in config which
> >> behavior to expect.
> >
> > Good point, my current understanding is:
> > When adding new API parameters to the existing APIs, these parameters should
> > be API extensions according to the above guidelines. So we have three options
> > for handling API extensions in Tempest:
> >
> > 1. Consider them as optional, and cannot block the incompatible
> > changes of them. (Current)
> > 2. Consider them as required based on tempest.conf, and can block the
> > incompatible changes.
> > 3. Consider them as required automatically with microversioning, and
> > can block the incompatible changes.
> 
> I investigated the way of the above option 3, then have one question
> about current Tempest implementation.
> 
> Now verify_tempest_config tool gets API extension list from each
> service including Nova and verifies API extension config of tempest.conf
> based on the list.
> Can we use the list for selecting what extension tests run instead of
> the verification?
> As you said In the previous IRC meeting, current API tests will be
> skipped if the test which is decorated with requires_ext() and the
> extension is not specified in tempest.conf. I feel it would be nice
> that Tempest gets API extension list and selects API tests automatically
> based on the list.

So we used to do this type of autodiscovery in tempest, but we stopped because
it let bugs slip through the gate. This topic has come up several times in the
past, most recently in discussing reorganizing the config file. [1] This is why
we put [2] in the tempest README. I agree autodiscovery would be simpler, but
the problem is because we use tempest as the gate if there was a bug that caused
autodiscovery to be different from what was expected the tests would just
silently skip. This would often go unnoticed because of the sheer volume of
tempest tests.(I think we're currently at ~2300) I also feel that explicitly
defining what is a expected to be enabled is a key requirement for branchless
tempest for the same reason.

The verify_tempest_config tool was an attempt at a compromise between being
explicit and also using auto discovery. By using the APIs to help create a
config file that reflected the current configuration state of the services. It's
still a WIP though, and it's really just meant to be a user tool. I don't ever
see it being included in our gate workflow.

> In addition, The methods which are decorated with requires_ext() are
> test methods now, but I think it would be better to decorate client
> methods(get_hypervisor_list, etc.) because each extension loading
> condition affects available APIs.

So my concern with decorating the client methods directly is that it might raise
the skip too late and we'll end up leaking resources. But, I haven't tried it so
it might work fine without leaking anything. I agree that it would make skipping
based on extensions easier because it's really the client methods that depend on
the extensions. So give it a shot and lets see if it works. The only other
complication is the scenario, and cli tests because they don't use the tempest
clients. But, we can just handle that by decorating the test methods like we do
now.


Thanks,

Matt Treinish

[1] http://lists.openstack.org/pipermail/openstack-dev/2013-October/016859.html
[2] http://git.openstack.org/cgit/openstack/tempest/tree/README.rst#n16