[bug#50359] [PATCH 3/3] import: Add 'generic-git' updater.

[iskarian]

Hi,

September 10, 2021 9:21 AM, "Xinglu Chen" <public@yoctocell.xyz> wrote:

> * guix/git.scm (ls-remote-refs): New procedure.
> * tests/git.scm ("remote-refs" "remote-refs: only tags"): New tests.
> * guix/import/git.scm: New file.
> * doc/guix.texi (Invoking guix refresh): Document it.
> * tests/import-git.scm: New test file.
> * Makefile.am (MODULES, SCM_TESTS): Register the new files.
> 
> Co-authored-by: Sarah Morgensen <iskarian@mgsn.dev>

Overall this is looking good.  Thank you for adding tests (for remote-refs as 
well!), much appreciated.  It looks like you've done some good polishing.  I 
see a few nits, which I'll point out in a separate email when I'm not on 
mobile.  I'll also give it a good test.

But... I've been thinking about the overall approach for a couple days, because 
I'm not very happy with the complexity of my heuristic.

There can be a lot of weird tags in a repository--look at the one for 
xf86-video-intel for example.  My heuristic attempts to capture the assumption 
that repostories tend to move from using "_" or "-" to "." but it does fail to 
account for moving to or from dates (because dates don't compare with normal 
versions).

I also realized that we are not using a very useful piece of information--the 
previous version/tag combo.  I expect that in the vast majority of cases, the 
version delimiter for the newest version will be the same as the version 
delimiter for the last known version.  (Perhaps the prefix as well?)  Can we 
use this information to make our guesses better?  What do you think?

Despite saying all that, it's probably better to not try to get it perfect on 
the first go--we can always adjust the internals later.  We just want to avoid 
showing bogus updates.

(Later, I think I'll put together a dataset of tags and current versions to see 
if I can test how well a particular algorithm works.)

--
Sarah (mobile)