[aur-dev] Using git as a backend for the AUR

Kwpolska kwpolska at gmail.com
Sun Mar 17 08:44:12 EDT 2013

On 16/03/13 23:10, William Giokas wrote:
>  The repositories would be limited to 5M

Measuring by current size of contents or total size?  The latter would
be very unpleasant for packages modified often or packages with binary
files (eg. dropbox/dropbox.png; binary files in git aren’t using
deltas AFAIK).

On Sun, Mar 17, 2013 at 12:43 PM, Jelle van der Waa <jelle at vdwaa.nl> wrote:
> The AUR currently has ~ 41076 packages, which would bring us to an equal
> amount of git repo's. This doesn't seem very efficient to me.

Well, why do you think it is inefficient?  Maybe because of the amount
of repos?  GitHub hosts (as of now) 5,700,361 repos[0], big and small.
 That number doesn’t include gists, and there are a lot of them.  So,
we would be around six million, at the very least.

Because of the repo size?  Let me disprove you.

Currently, the AUR stores stuff in 41078 tarballs (tar+gz), plus 41078
PKGBUILD files stored without any compression on the server (eg.
https://aur.archlinux.org/packages/pk/pkgbuilder/PKGBUILD ).  The
PKGBUILD and the tarball weigh a total of 1241 bytes[1].

Now, we make a local git repo (git init --bare), and get the
aforementioned PKGBUILD into it from a downstream clone (git
{add,commit,push}).  After removing the sample hooks (unnecessary,
although the AUR may make good use of such hooks[2]), we end up with
843 bytes[2].  After an update (pkgver+md5sums) we get 1518 bytes[3].
This is a small price for a complete PKGBUILD modification history.

If somebody wants numbers for Mercurial, it gets 1262 and 1725
bytes[4], respectively.  I couldn’t set SVN up easily, so I don’t have
any numbers.

[0] https://github.com/search; displayed if you are lucky enough
[1] makepkg --source; wc -c PKGBUILD pkgbuilder*.src.tar.gz
[2] For example, Heroku uses post-update hooks to trigger a build of
the website.  GitHub has an awesome Service Hook system, where you can
have a bazillion services do stuff after you push (eg. unit tests via
Travis CI)
[3] wc -c config description HEAD objects/*/* refs/heads/master
[4] wc -c 00changelog.i branch dirstate hgrc last-message.txt requires
undo.bookmarks undo.branch undo.desc undo.dirstate
store/00changelog.i store/00manifest.i store/fncache store/phaseroots
store/undo store/undo.phaseroots store/data/* cache/branchheads-served

> With svn you can put it all in one repo and checkout a single package
> without checking out the whole repo, which is how
> [core],[extra],[community] work.

svn is ugly and human-unfriendly.  Nobody likes it.  And I couldn’t
get it to work properly, as mentioned above.

> Git seems to have sparse checkout for this in 1.7, but I don't know how it works.


Kwpolska <http://kwpolska.tk> | GPG KEY: 5EAAEA16
stop html mail                | always bottom-post
http://asciiribbon.org        | http://caliburn.nl/topposting.html

More information about the aur-dev mailing list