Hi all,

On 2026-11-11 at 10:52+01:00, Andreas Enge wrote:
> On Tue, Nov 10, 2026 at 08:22:34PM +0000, Christopher Baines wrote:
> > I don't know where this coupling of branches and teams has come from,
> > the package graph is far more important in my mind, and at the moment
> > that isn't structured around "teams" (although we can move in that
> > direction).
>
> The reason I am advocating for team branches is that responsibility and
> agency come together. I am still traumatised by the previous core-updates
> branch! There random people pushed random commits (slightly exaggerating
> of course). Once something broke, nobody was really responsible.
> With an xxx-team branch, it is clear that the xxx team is expected to
> work together to get the branch through.

Since I subscribed to guix-patches, I've noticed the pace we're
having with the topic branches largely has the regular monitoring
by you (Andreas and Chris) to thank.  In other words, anything to go
in topic branches means more work for you two, and this is not healthy.

On 2026-11-11 at 10:52+01:00, Andreas Enge wrote:
> But we can always try something else. The world-rebuild branches are/were
> such a trial. The way they were carried out, they depended a lot on an
> individual, so I am not sure they could completely replace team branches.

Note that for the last two iterations when I herded world-rebuild
branch, I also receive a lot of help from you two and the people
who submitted to original patches.  In addition, unlike the way
it was done last year, the branch mostly took in trivial updates
to orphaned packages.  They were nowhere near the volume
of many team branches.

Maybe it's just because I don't have much experience doing
topic branches, but I don't like their workflow at all,
i.e. the rushed fixes to not delaying the branches behind,
and forced pushes and rebasing that makes cooporating so difficult.

I'd guess this goes unnoticed because (1) most contributors
don't work on topic branches, (2) most of those who do
have been doing it for years and (3) most team branches
at a time are herded by 1-2 members whose daily schedule
don't overlap.

Conceptually, I find the notion of filing a MR to reserve a spot
in the queue before stablizing the changes to be hinky,
but it does not work the other way around as relevant packages
will change when a branch gets to the front of the queue.

Meanwhile, dependent channels are right there for the grab.
I've proposed it before, and perhaps the process was too complex:
https://lists.gnu.org/archive/html/guix-devel/2026-06/msg00184.html

I'll try to simplify it this time.  Currently, around 2k packages
have >300 dependents so their patches cannot be applied directly
on master.  Imma call them core packages in this message,
and the other 31k the crust.  If core packages (+ guix itself)
are split to a separate channel, let's try to construct a process
to solve the following problems.

1. Excessive rebuilds of crust
2. Barrier of entry to, and collaboration in patching core
3. Outdated core
4. Pinned variants of core packages
5. Security releases for Guix packaged by other distros
6. Broken subset of crust after core updates

Much of these have to do with our rolling release model,
so let's put a few restriction on it.  First, crust must have
two branches, stable and devel.  Most patches landing on master
at the time of writing would get applied on devel.  Stable must pin
a core release tag.

Core must have regular releases.  If we want to keep the current pace,
there should be a major release every two months.  Let's also reserve
two staging weeks for QA and building crust.  For 7-8 weeks
after each release, changes are reviewed by the relevant teams [2, 3]
and only direct dependants are built in the farm [1].

After this development period, devel crust switches to depend
the core branch and new build failures get fixed where appropriate [2].
When QA meets our standard [6], a major core release is tagged
and a branch is created for its security updates.  All grafts must land
on these release branches [4], and immediately ungrafted on core's
development branch unless the latter contains a newer version.
A patch release is tagged after grafting (one or more packages) [5].

For guix pull, we define a meta channel as an alias to a crust branch.
During core staging, meta points to stable crust and only security updates
in crust packages or core patch release land on it.  The latter updates
must also be cherry-picked to devel crust.  When a major core release
is tagged, devel crust pins to it and is rebased on stable.
When there are sufficient substitutes, meta switches to devel crust.
Next time when staging begins, stable is reset to devel, its descendant,
and the cycle continues.

I will try to send out a GCD draft after the upcoming 1.6,
and any early feedback is also appreciated.

Best wishes,
Phong

Reply via email to