[Mesa-dev] How to merge Mesa changes which require corresponding piglit changes

[Michel Dänzer]

On 2019-12-02 3:15 p.m., Tapani Pälli wrote:
> On 11/15/19 8:41 PM, Mark Janes wrote:
>> Michel Dänzer <michel at daenzer.net> writes:
>>
>>> On 2019-11-15 4:02 p.m., Mark Janes wrote:
>>>> Michel Dänzer <michel at daenzer.net> writes:
>>>>
>>>>> Now that the GitLab CI pipeline tests a snapshot of piglit with
>>>>> llvmpipe
>>>>> (https://gitlab.freedesktop.org/mesa/mesa/merge_requests/2468), the
>>>>> question has come up how to deal with inter-dependent Mesa/piglit
>>>>> changes (where merging only one or the other causes some piglit
>>>>> regressions).
>>>>>
>>>>>
>>>>> First of all, let it be clear that just merging the Mesa changes as-is
>>>>> and breaking the GitLab CI pipeline is not acceptable.
>>>>>
>>>>>
>>>>>  From the Mesa POV, the easiest solution is:
>>>>>
>>>>> 1. Merge the piglit changes
>>>>> 2. In the Mesa MR (merge request), add a commit which updates
>>>>> piglit[0]
>>>>> 3. If the CI pipeline is green, the MR can be merged
>>>>>
>>>>>
>>>>> In case one wants to avoid alarms from external CI systems, another
>>>>> possibility is:
>>>>
>>>> For the Intel CI, no alarm is generated if the piglit test is pushed
>>>> first.  Normal development process includes writing a piglit test to
>>>> illustrate the bug that is being fixed.
>>>
>>> Cool, but what if the piglit changes affect the results of existing
>>> tests? That was the situation yesterday which prompted this thread.
>>
>> We attribute the status change to piglit in the CI config, within a few
>> hours.  The test shows up as a failure in CI until it is triaged.
> 
> I think we have a problem with current gitlab CI process.
> 
> Right now if someone needs to update piglit commit used by CI, he also
> ends up fixing and editing the .gitlab-ci/piglit/quick_gl.txt (and
> glslparser+quick_shader.txt) as CI reports numerous failures because of
> completely unrelated stuff as meanwhile people added other tests,
> removed tests and modified them.

This is at least somewhat intentional, as the results of any newly added
tests should be carefully checked for plausibility.


> I think we should turn such warnings on only when we have more
> sophisticated algorithm to detect actual regression (not just 'state
> change', like additional test or removed test).

It's unclear what exactly you're proposing. In order to catch
regressions (e.g. pass -> warn, pass -> fail, pass -> skip, pass ->
crash), we need a list of all tests on at least one side of each
transition. We're currently keeping the list of all
warning/failing/skipped/crashing tests, but not passing tests (to keep
the lists as small as possible).

One possibility might be to remove the summary at the end of the lists.
That would allow new passing tests to be silently added, but it would
mean we could no longer catch pass -> notrun regressions.


-- 
Earthling Michel Dänzer               |               https://redhat.com
Libre software enthusiast             |             Mesa and X developer