GLEP 40: Standardizing "arch" keywording across all archs.
|Author||Grant Goodyear <email@example.com>|
|Posting history||2005-09-06, 2005-09-15, 2006-09-03|
Approved by the Gentoo Council on 2005-09-15. As of 2006-09-03 we have a robust x86 arch team, so this GLEP is final.
This GLEP originated from a rather contentious discussion on gentoo-dev about combining the x86 and amd64 keywords. This GLEP attempts to get at the heart of that discontent. The proposed stable-keyword guidelines have been lifted verbatim from The Doc.
It is time for x86 to no longer be an exception to the standard keywording guidelines. Thus, an x86 arch team should be responsible for moving packages from ~x86 to x86.
The original, informal x86 keywording policy, where almost any x86 dev (which were the vast majority of devs) who used a package could mark it stable, arose from a time when there were relatively few Gentoo devs. Adding packages to the tree was the principal concern, as opposed to maintaining existing packages. QA considerations have since modified that policy slightly, and now it is the package maintainers who should mark an x86 package stable. Of course, that policy presumes that package maintainers are generally x86 devs, which is slowly becoming less and less true.
This policy for x86 is quite different from how every other arch marks packages stable. For the non-x86 archs, each arch has a specific "arch team" which is responsible for moving packages from ~arch to arch, although vapier notes that "arch teams generally defer to maintainers (and rightly so) as to when newer versions should go stable." This approach has worked quite well for the non-x86 archs, and this GLEP asserts that the same approach would benefit x86 as well.
For a package to move to stable, the following guidelines must be met:
- The package has spent a reasonable amount of time in ~arch first. Thirty days is the usual figure, although this is clearly only a guideline. For critical packages, a much longer duration is expected. For small packages which have only minor changes between versions, a shorter period is sometimes appropriate.
- The package must not have any non-arch dependencies.
- The package must not have any severe outstanding bugs or issues.
- The package must be widely tested.
- If the package is a library, it should be known not to break any package which depends upon it.
- The relevant arch team must agree to it.
A robust x86 arch team needs to be created. The firstname.lastname@example.org alias already exists, and it merely needs to be used. This team, with the aid of potential non-dev arch testers, has the responsibility of stabling all x86 packages. Current x86 devs who wish to mark their own packages stable must therefore either be members of or make individual arrangements with the x86 arch team.
There will be a considerable one-time cost involved in establishing a robust x86 arch team--a good number of bodies (the amd64 arch team has 19 active devs and 12 active non-dev arch testers) need to be recruited to be part of the new arch team, and convincing devs that it is in their best interests to work in a new fashion is likely to be even harder. Certainly the benefit of consistency between the various archs is obvious, but is it worth the cost involved? Here are the arguments for enduring the pain involved:
- Over time, x86 is likely to become a minority arch as 64-bit systems become the norm. The implicit assumptions that underly the current system (that most devs, users, and package maintainers use x86) will become increasingly less valid.
- Markedly improved QA for x86. Assuming that the author's own is behavior is representative, most x86 devs run ~x86 systems. Thus, the assumption that devs are good x86 testers is not really valid. One obvious consequence is that packages tend to languish in ~x86 for a very long time, since x86 devs running ~x86 have little cause to notice that a package has not been marked stable. The much larger effect, though, is that it is rare for x86 packages to be stabled in the context of a full x86 tree, so the big picture of a stable system, not just a stable package, is lost. This approach of stabling in the context of a full stable arch tree, it has been argued, is the fundamental reason why the non-x86 archs have notably better QA than does the x86 arch.
Creation of a robust x86 team is already underway. The more vital step is the official change in policy, along with a sustained effort to get existing x86 devs to go along with it.
Stuart has suggested the creation of a new arch keyword: "[-]maint", which would exist in tandem with the normal arch keywords, thereby making the package maintainer's intention explicit. Ciaranm has responded that by definition a package in ~arch is a candidate for arch, so a package's mere presence in the tree (without being in package.mask) should indicate the package maintainer's intention. There was a fair bit of discussion about whether the idea should be a "maint" keyword, or named something else, or an entirely different variable, etcetera, but the basic gist didn't change much.
Jstubbs notes that it could be a very good idea if all non-arch devs worked in overlays, but that new portage (gensync) support would be needed to make it truly viable. Stuart pointed out that php5 support was handled just that way. One author's view is that this approach would make the "package in ~arch means that it's a de-facto candidate for arch" interpretation even more valid.
Ciaranm and weeve have noted that it is occasionally necessary for arch teams to override a package maintainer when it comes to stabling a package. Stuart has asserted that in those cases the arch team should be willing to take on the support burden for that package.
Not really an issue here.