Re: [RFC PATCH 14/15] gitlab-ci: Allow forks to use different set of jobs

[Philippe Mathieu-Daudé]

On 4/19/21 5:57 PM, Alex Bennée wrote:
> Philippe Mathieu-Daudé <f4bug@amsat.org> writes:
> 
>> Forks run the same jobs than mainstream, which might be overkill.
>> Allow them to easily rebase their custom set, while keeping using
>> the mainstream templates, and ability to pick specific jobs from
>> the mainstream set.
>>
>> To switch to your set, simply add your .gitlab-ci.yml as
>> .gitlab-ci.d/${CI_PROJECT_NAMESPACE}.yml (where CI_PROJECT_NAMESPACE
>> is your gitlab 'namespace', usually username). This file will be
>> used instead of the default mainstream set.
>>
>> Signed-off-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
>> ---
>>  .gitlab-ci.yml | 7 ++++++-
>>  1 file changed, 6 insertions(+), 1 deletion(-)
>>
>> diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
>> index 718c8e004be..35fd35075db 100644
>> --- a/.gitlab-ci.yml
>> +++ b/.gitlab-ci.yml
>> @@ -9,7 +9,12 @@ generate-config:
>>      paths:
>>        - generated-config.yml
>>    script:
>> -    - cp .gitlab-ci.d/qemu-project.yml generated-config.yml
>> +    - if test -e .gitlab-ci.d/${CI_PROJECT_NAMESPACE}.yml ;
>> +      then
>> +        cp .gitlab-ci.d/${CI_PROJECT_NAMESPACE}.yml generated-config.yml ;
>> +      else
>> +        cp .gitlab-ci.d/qemu-project.yml generated-config.yml ;
>> +      fi
> 
> This is going to be a little clunky. I can see a use for the static
> forks that Danial proposes but I guess what is needed is a little
> expressiveness. So how to express things like:
> 
>  - I've only touched stuff in linux-user, so run only linux-user tests
> 
>  - I'm working on KVM, run all KVM enabled builds and tests
> 
>  - I've changed the core TCG code, run everything that exercises that
> 
>  - I'm working on ARM, only build and run jobs that have ARM targets
> 
> This sounds like tags I guess but the documentation indicates they are
> used for runner selection. Could we come up with a subset that could be
> used to select from all our build fragments when constructing the
> generated-config? I could even imagine a script analysing a diffstat and
> guessing the tags based on that.

Ahah this is just what we were discussing with Willian 2h ago after
looking again at stefanha analysis
(https://www.mail-archive.com/qemu-devel@nongnu.org/msg795905.html).

. diff-stat -> files modified
. files modified | get_maintainers -> subsystem maintained sections

I suggested Willian to add support for 'tags' entries to MAINTAINERS,
so we could have:

./get_maintainer --tags file1 file2 ...
-> virtio, migration, kvm

Then we could run all the tests tagged 'virtio, migration, kvm'
(unit tests, iotests, qtests, integration tests).


The transposed use is when a test fails, we can list its tags and
from here get the subsystem maintained sections tracking these tags
and notify them a test using their subsystem failed.

> I think we should define a minimum set of lightweight smoke tests that
> get the most bang for buck for catching sillies. I think checkpatch and
> dco checking should probably be in there - and maybe one of the bog
> standard build everything builds (maybe a random ../configure; make;
> make check on one of the supported LTS targets).
> 
> Then there is the question of defaults. Should we default to a minimised
> set unless asked or should the default be the full fat run everything?
> We could I guess only switch to running everything for the staging
> branch and anything that is associated with a tag or a branch that has
> pull in the name?

Yes, this is a community problem that need to be discussed. Not all the
community members have the same requirements and expectations.

What I'm trying to do here is ease random contributor fork workflow,
not uptimizing mainstream /master gating CI, which is suppose to have
way more resources than a random contributor.

Also I don't believe 1 set of CI jobs will ever make all users happy
together. We have all different needs. I'm looking for a solution
which include every contributors from the community.

I'm brainstorming about a setup where a maintainer might have extra
resources provided by the project (such access to dedicated hardware).
Let's use 'virtio' for example. The maintainer might want to use 2
different pipelines:
- one to run all its 'virtio' tagged tests each time patches are queued
  from the subsystem the contributors (this is the subsystem "gating"
  side).
- one to run extra set more complex, run only before sending a pull
  request.

Regards,

Phil.