Re: Project direction with testing changes (branches and patches)

[Ludovic Courtès]

Hello!

Christopher Baines <mail@cbaines.net> skribis:

> So, I think I've recently switched to thinking about the problem as one
> of testing changes, rather than just testing patches. Since both patch
> series, and branches are used to propose changes, I think this makes
> sense.
>
> In abstract, when testing a change, I would break down the problem as
> follows:
>
>   - You need to work out what's affected by the change, so that you can
>     assess the impact
>
>   - Once you know what's effected, you can then build those
>     packages/system tests/... and compare the build statuses and outputs
>     against some baseline
>
>   - Then there's the general UI component, ideally a first time
>     contributor would be able to take advantage of automatic feedback
>     about a patch they submit. There's multiple other groups of users
>     though, like patch reviewers, and committers for example.

Makes sense to me.

I agree that the first problem, seeing what’s affected by a change, is
solved, but it’s still hard to get that info.  I think we could have a
special “skin” for the Guix Data Service to make it easier for people to
view specifically this information.  IMO the current UI has the upside
that it’s generic and exposes all the available information, but it has
the downside that it’s generic and exposes all the available
information.  :-)

Or we could extend Julien’s Gitile¹ to include links from commits to
lists of changed packages.

The UI doesn’t have to be a web UI actually; we could use the Data
Service client interface at
<https://lists.gnu.org/archive/html/guix-devel/2021-06/msg00228.html>
and write a new CLI, Emacs mode (similar to ‘M-x build-farm’), or
something.

¹ https://git.lepiller.eu/gitile

> I think the first two sub-problems are effectively solved. The Guix Data
> Service is able to determine the changes between two revisions (assuming
> it's processed them). The Guix Build Coordinator can then be used to
> build the relevant packages/system tests, and report that information
> back to the Guix Data Service.
>
> The UI part is much less certain, I've done some work with Patchwork,
> and I do have some ideas in mind, but there's still more thinking and
> work to do in this area.
>
> Before pressing on though, I think it would be good to know if this is a
> viable direction?

I think we desperately need more automation, even more than when you
started working on this!

I think a first step could be to make info from the Guix Data Service
more readily available, as suggested above.  And from there we could
address #2 and #3.

The Patchwork instance you maintain at
<https://patches.guix-patches.cbaines.net/project/guix-patches/list/>
does a large part of what we want, though the UI is not my favorite I
must say.  ;-)  I wonder if we could again make minimal changes to Mumi
so that it includes links to the relevant bits at the Data Service.
That’d make it more readily available.  WDYT?

> Currently, there's no automated testing of patches, and testing of
> branches is limited to the information that Cuirass provides on failed
> builds. What I'm proposing for the future is: using the Guix Data
> Service together with the Guix Build Coordinator to analyse the effects
> of changes, whether that be from a patch series or a branch. I realise
> that I've already been experimenting with this, what I'm mostly
> referring to here is moving towards this being the documented approach,
> maintained by the project, not just me.

>From an administrative standpoint, I very much agree that this sort of
infrastructure should be financially supported by the project rather
than by an individual, and documented so we can maintain it
collectively.  Of course that involves a bit of overhead, and sometimes
we’re not all responsive when it comes to paperwork or sysadmin (perhaps
these are not our preferred activities?), but still, it makes sense to
build that collectively.

That’s my take.  Thanks for the update!

Ludo’.