Re: Feedback from JRES in Dijon

[zimoun]

Hi Julien,

Thank you for this report.

On Thu, 5 Dec 2019 at 15:25, Julien Lepiller <address@hidden> wrote:

> Another wanted to have some kind of doi for guix describe + manifest (a 
> better UI and easier thing to cite in a paper I suppose).

I agree that something is lacking. I was not thinking about DOI but I
was thinking a better UI similar to "git tag" to improve the current
situation. It is not exactly the question that you had but it seems
related, AFAIU. For example, it is hard to know which commit I need to
"guix pull" to have Python 3.6, if we have. I tried to start a
discussion in this topic there [1]. Another concrete example had been
asked here [2].

[1] https://lists.gnu.org/archive/html/guix-devel/2019-11/msg00513.html
[2] https://lists.gnu.org/archive/html/help-guix/2019-06/msg00098.html


> The person I talked to at lunch was a bit skeptical about Guix. They have it 
> and Nix, with module on their cluster. They told me they were more concerned 
> about replicability than reproducibility. Also, their users are lost between 
> the options, they have trouble getting help on guix from their admins and 
> they end up using conda, yet another tool ^^".

Bit-to-bit reproducibility is the first step to go to Replicability, IMHO.
For sure, Conda is a no-go for this goal, again IMHO. :-)
Modulefiles is loosing their time by re-doing what packagers are
doing; it is the dependency hell because they spend their time on
compiling libraries -- the code of interest depends on one lib which
depends on other lib which depends on... So it requires a lot of
manpower. Guix solves that elegantly IMO; considering channels.

And the big issue of replicability is floating point. The same binary
with the same inputs does not compute the same output. For example,
see page 13 of [3].

[3] https://jcad2019.sciencesconf.org/data/2019_JCAD_Repro_Hill_vf.pdf

Well I am not sure that a general solution can solve replicability...
and it is software by software.


Cheers,
simon