Re: Accuracy of importers?

["Ludovic Courtès" <ludo@gnu.org>]

Hi,

zimoun <zimon.toutoune@gmail.com> skribis:

> On Fri, 29 Oct 2021 at 23:57, Ludovic Courtès <ludo@gnu.org> wrote:

[...]

>> --8<---------------cut here---------------start------------->8---
>> $ SAMPLE_SIZE=200 ./pre-inst-env guile ~/src/guix-debugging/importer-accuracy.scm
>> […]
>> Accuracy for 'pypi' (200 packages):
>>   accurate: 58 (29%)
>>   different inputs: 142 (71%)
>>   different source: 0 (0%)
>>   inconclusive: 0 (0%)
>> Accuracy for 'cran' (200 packages):
>>   accurate: 176 (88%)
>>   different inputs: 23 (12%)
>>   different source: 1 (0%)
>>   inconclusive: 0 (0%)
>> --8<---------------cut here---------------end--------------->8---
>
> [...]
>
>> The script doesn’t do anything useful for crates because they have their
>> own way of representing inputs.  It doesn’t account for changes in
>> ‘arguments’ like zimoun suggested, meaning it’s overestimating
>> accuracy.
>
> It is already quite interesting results.  Because it shows upstream
> stability, IIUC.  Well, it means that running “guix import pypi” one
> months ago and running the sames now, 71% packages have different
> inputs.  Right?  It is because some metadata from PyPI changed, right?

No no; I’m assuming PyPI, CRAN, etc. provide the same info as they did
back when the package was imported (which is probably the case).

> Not because “guix import pypi” was doing wrong and now it does better,
> right?

I’m also assuming that the importer didn’t change significantly in the
meantime, which is probably a good approximation.

What I think those figures show is the amount of manual tweaks necessary
to get a proper package “à la Guix”, with tests running etc.  For PyPI
we often need to add things under ‘native-inputs’, hence the 71%
“different inputs” line.  For CRAN that’s sometimes necessary, but much
less frequently.  There are also cases with non-R/non-Python
dependencies.

> IMHO, it shows how PyPI allows bad practises about packaging, isn’t it?
>
> My understanding of this experiment is about upstream “quality”, not
> about importer “accuracy”.  Do I incorrectly understand?

Yes, in a way, assuming our importers are not lossy, this tells us
whether the upstream repo contains enough information and/or whether
that information is accurate.

Thanks,
Ludo’.