On Wed, 22 Dec 2010 11:01:26 -0600 Aaron Griffin <aaronmgriffin@gmail.com> wrote:
On Sat, Dec 18, 2010 at 11:26 PM, Dan McGee <dpmcgee@gmail.com> wrote:
On Sat, Dec 18, 2010 at 11:20 PM, Allan McRae <allan@archlinux.org> wrote:
On 19/12/10 04:59, Dieter Plaetinck wrote:
Last time the topic "moving nilfs/btrfs utilities in core" generated some discussion, but no outcome. (and not even all my questions were answered)
(http://mailman.archlinux.org/pipermail/arch-dev-public/2010-December/018656....)
So I'll try again, and I'll try to be more clear.
question 1: If you install arch linux on a foo-filesystem, should you install foo-utils as well? so that at least your system doesn't break when it needs to run fsck.foo during boot. IMHO the answer is yes.
Yes.
question 2: what exacly is the definition of the core repo? Pierre tells me "The idea of core was to provide a minimal set of packages that are needed by nearly all users to set up a base system." If is true, what separates base from core? Because this definition looks an awful lot like the definition of base to me.
AFAIK core is about bundling all packages which are critical (for the purpose of getting the system to run) for some users, depending on their particular setup. This is the reasoning why we include the core repo on our core install images, because we say "with this repo, users can install what they need in order to make their system boot and run fine". This is also in line with what I see on https://wiki.archlinux.org/index.php/Official_Repositories ("will provide you with a fully functional base system")
This is the problem: - we want to allow users to use new filesystems during installation. Currently this includes nilfs2 and btrfs. (marked as experimental in aif). - On the other hand we apparently don't allow infrequently used packages in core. Even if I would wait and only support these filesystems when they are declared stable, their package usage would still be low. It would be pretty backwards if I would need to wait for users to configure their systems using the filesystems manually, so the package usage goes up, before I can include them on the installation media. Currently, dosfstools, nilfs-utils and btrfs-progs-unstable are not in core and can't be installed by using a networkless install with the core-images.
So, How do we solve this? I see two outcomes: - we add the packages mentioned above to core, and loosen up the signoff requirements for packages with low usage. (because "commonly used" and "needed for a base system" are very different things) - we do not add these packages to core, and add something about "must have high usage" on the wikipage i mentioned above. this also means I will need to include a small repo with "needed for base systems, but with low usage" on the core install images, to allow users to install these packages. And I'll need to adapt aif because it now needs to use additional repositories to install required-to-boot packages.
I vote for outcome 1.
Basically I don't care what the outcome is here. But note that we can probably safely assume that reiserfs is more commonly used than both the two filesystems that you mention and I have yet to receive a signoff for reiserfsprogs (posted 2010-12-04...). So moving the packages for nilfs/btrfs to [core] will only serve to delay updates... Then again, I see no reason why these packages should not be in [core].
My solution would be more like your #2 suggestion. Instead of a "core" CD, make an "install" CD which has a set of packages you decide on it. This would be largely based on [core], but you can include packages from [extra] then too. You can also drop some [core] packages from the CD. For example, there is no point putting gcc-{fortran,objc,ada} on the install CD at all. They are only in [core] because split packages can not go across repos. So get rid of them and gain 40MB more space for your dual arch iso.
I'm of the opposite opinion on this one; I think #1 makes more sense. We're looking at it backwards- "this package can't go in core because it is hard to get signoffs" vs. "perhaps we can adjust our signoff policy for lower-usage core packages". Filesystem utilities, at least to me, seem like the definition of what core should contain. Reiser/JFS/XFS/BTRFS/NILFS, while not the top filesystems in use in Linux, deserve to be supported "out of the box" by us.
For a minor bit of history, "core" is only named the way it is because it linguistically makes sense with a repository named "extra". core was never meant to be anything more or less than "current" used to be - which was a repo full of software that the developers hand-picked to be important.
That doesn't answer the question, though. Now-a-days, core is treated as if it were the mission-critical parts of a running system. That's a good definition, but it also doesn't help answer the filesystem util question. Are they critical to a running system? Not exactly... once I have an installed filesystem, I don't need them anymore (it's a good *suggestion*, but it's not *critical*).
So what do we gain with having the filesystem utils in core? Signoffs, and in theory a more stable installer. But it already appears that signoffs for the more esoteric filesystem utils are hard to get. In my eyes, this seems to make the benefit of gaining signoffs moot.
It seems this question is really about the following: should things needed during install time be part of core? Or rather, is installation an important enough part of Arch Linux to require the (theoretical) added stability of the signoff process?
The signoff procedure is a means to a goal (of providing some stability/trustworthyness indication for packages with high use); putting misc low-usage packages in core just to get the advantage of a free strict signoff procedure will not work, is backwards, and not what i'm suggesting. Like I said, I wouldn't want to push these packages into core without changing the signoff policy. that would make the job of the package maintainer needlessly tough, maybe even plain undoable. in other words, for outcome 1 I would make the signoff policy less strict for packages with low usage. but this brings up the question: if we use such a definition for core (where we say "strict signoff policy for all packages except some"), which point is there in defining a core repository anyway? why do we even have separate repositories? strictly speaking, we could do just fine with 2 repositories. (one for all current packages, one for all testing packages). there are some criteria valid enough to warrant *categorizing* packages (this doesn't necessarily mean putting in different repositories). In our current setup, for example: gcc is in core, this means it's frequently used so there are certain quality indications. (maintained by arch dev, signoff procedure, etc) claws-mail is in extra, so no strict signoff procedure, but still sufficient usage to be an officially supported and maintained package. A community package has even less use, but since it's maintained by a trusted user there is at least some level of trustworthyness, and I can still use the bugtracker. An AUR "package": very low usage, no quality guarantees, no official support. How about multilib? this indicates merely a technical property of the packages, but says nothing about quality indications. (signoff procedure? maintained by devs or by tu's?) => the way we are using repo's is merely to indicate some property of a package (who maintains it, which level of support, which quality indication, etc), but they are by itself no reason to physically separate packages from each other. (although that's how we've been doing it historically). Even worse, they are often cross-concerns making it a bit awkward to classify packages this way (see the multilib example above) just thinking out loud; why don't put packages more together in the same repo, but introduce some new variables per package to indicate these concerns which separate them? for example: maintainer-type : dev/tu/user support: high (signoffs)/medium (no signoffs)/ none (aur) this would also trivialize some things which are currently non-trivial (like moving a package between repo's because a dev maintains it instead of a TU or vice-versa) this would also mean it becomes easier to add new "categorizers" (both as in variablename as in possible values), with very little impact, basicaly more flexibility. the only disadvantage I can see with such an approach is that if users want, say only packages which are currently in core and extra, they would also get the metadata (from pacman -Sy) from packages they are not interested in. but this seems like a small tradeoff to me. Which raises the question: what *are* valid concerns to *physically* separate packages by? Afaik all mirrors have no problem replicating and serving all repo's, so that's not it. unofficial repositories and mirrors could still function pretty much like before, so that's not it either. As far as I can see, there are two reasons to physically separate packages into different repo's: * space constraints on release media * testing packages vs stable, because the packages have the same name. I realise I just suggested some pretty radical changes, but in my mind they make a lot of sense. I hope I didn't miss some obvious downsides (please point them out), I will (and I hope you do too) spend some nights pondering about this. Dieter