Re: Reproducible profiles

[Ludovic Courtès]

David Thompson <address@hidden> skribis:

> Lately I've been wanting to version control the list of packages that I
> install in my user profile so that I can sync it amongst many machines.
> So, I took a stab at adding a new '--apply' option to 'guix package'
> that reads in a package list from a Scheme file and creates a new
> generation of the profile with only those packages are installed.
> Here's an example configuration:
>
>     (use-modules (gnu))
>     (use-package-modules base less guile emacs admin ruby mail pumpio man)
>     
>     (list ruby
>           coreutils
>           less
>           man-db
>           notmuch
>           guile-2.0
>           emacs
>           dmd
>           offlineimap
>           pumpa)

Yes, that sounds very useful.

As usual though, there’s the issue of multiple-output packages.  The
above snippet is nice, but doesn’t allow you to specify a particular
output of a package.

What about instead requiring people to return a manifest:

  (use-modules (guix profiles))
  (use-package-modules base emacs guile)

  (manifest (cons (package->manifest-entry gcc-toolchain "debug")
                  (map package->entry
                       (list gcc-toolchain emacs guile-2.0))))

That means we’ll have to document (guix profiles).

It’s more verbose than what you suggest, though.  If you insist ;-), we
could allow a list of packages instead of a manifest, though I’d prefer
not to do that.

WDYT?

> Below is a naive patch that does the job, but is unideal because it
> doesn't do some nice things like display the diff between generations
> before building.

For that you would need a procedure to infer the manifest transaction:

  (manifests->transaction m1 m2)
  ;; returns a <manifest-transaction>

and then that could be passed to ‘show-manifest-transaction’.

However, I’m not sure it’s very useful.  Perhaps it would be enough to
write “installing new manifest from foo.scm with 42 entries.”
WDYT?

> I'm looking for some guidance to make this option mesh better with the
> rest of the 'guix package' utility.  Any help is appreciated.

Overall it looks OK to me!

[...]

> +         (option '("apply") #t #f
> +                 (lambda (opt name arg result arg-handler)
> +                   (values (alist-cons 'apply (load arg) result)
> +                           arg-handler)))

It would be better to delay loading until after arguments have been
parsed, as in ‘guix system’.  The procedure to load the file should be
similar to ‘read-operating-system’.

We’ll need documentation and tests, too.  :-)

Thank you!

Ludo’.