IRC channel logs


back to list of logs

<zimoun>thanks rekado_ for the R packages.
<civodul>rekado_: hi! something's wrong with
<civodul>though is the new one
<zimoun>rekado_: r-shiny is broken; as explained in bug#50483; the patch is in <>. Do you have a slot to give a look?
<civodul>a CONDA person on birdsite was angry and mentioned
<civodul>apparently it should have "recipes" to build packages from source but i can't find them
<civodul>ah so it's called "feedstock"
<civodul>i found this one:
<zimoun>civodul: ahah! conda-forge is what we call a channel
<zimoun>bioconda <> is another
<zimoun>I challenge the angry person to pick a random package from conda-forge and then rebuild all from source. For instance, a random package at version from last year.
<zimoun>civodul: long time ago in a galaxy far, far away, I spent countless hours to package a complex C++ scientific software for Conda <>. Just build from source on my laptop (Debian) and on the cluster (Ubuntu LTS 14 at the time), and get something “similar”, it was sweat, pain and workarounds.
<zimoun>civodul: I pick a random package from conda-forge and I get this: <>
<zimoun>{{ PYTHON }} -m pip install . --no-deps -vv
<zimoun>The one of binutils looks like: <>
<zimoun>and I guess the file yum_requirements.txt is the list of dependencies
<civodul>zimoun: yeah, it does look very hacky and brittle
<civodul>though i can concede that it's better than binaries uploaded by the authors, as in the Torch case
<civodul>so i'm thinking of adding a note mentioning conda-forge to placate the person
<dstolfa>the only reason these things ever work is because of the solid base that redhat provides. it's amazing how fragile everything that people bolt on top of it is...
<zimoun>Is PyTorch on conda-forge?
<zimoun>civodul: here it is, I guess <>
<zimoun>I bet it is not straightforward at all to rebuild pytorch from source using this.
<civodul>yeah, i doubt it's reproducible
<zimoun>Well, saying that «opaque binaries [...] lacking reproducible build methodology and tooling» is full true! It is by design of Conda
<zimoun>However, indeed «no signs […] of moving in that direction» is half-true because conda-forge tries to dead-end-y improve… arf!
<zimoun>I do not speak about reproducibility. Simply rebuild at other point in time on another machine. Maybe if first you get a Docker image based on Debian, then inside you run the recipe from conda-forge ;-)
<civodul>heh, yeah
<civodul>zimoun: how 'bout ?
<civodul>is that fair representation?
<zimoun>yeah… even if it cannot be «a possible answer» because Conda is doomed when speaking about rebuild. Anyway, let be sweet to Conda folks. :-)
<civodul>heh, cool, thanks :-)
<civodul>i see Julia packages waiting for you, zimoun: :-)
<zimoun>yeah, thanks :-) I am fighting this «Julia is broken» <>
<zimoun>civodul: <> je me demande comment Guix pourrait-il concourir ? :-)
<civodul>je sais pas !
<civodul>my boss mentioned it to me
<civodul>i wouldn't be surprised if CONDA won that prize ;-)
<civodul>you wanna fill in the application?
<zimoun>Maybe I will do it if Bordeaux folks do not. For sure, it would attract visibility and wider audience, at least. :-)
<zimoun>«La soumission des candidatures sera ouverte début octobre et détaillera le processus de candidature.» Let see when it is open ;-)
<civodul>sounds good!
<civodul>i guess we'd be competing with superstars like scikit-learn and Coq though...
<zimoun>yeah but as Jean-Claude Dusse says: who know? ;-) It might work if the stars are well aligned.
<drakonis>ah todd gamblin
<drakonis>wasnt he on #guix a while back?
<rekado_> was just an internal demo. It’s now
<rekado_>drakonis: todd was here during Fosdem
<drakonis>oh i see
<rekado_>weird comment from todd
<rekado_>zimoun: thanks for the note about r-shiny, I’ll take a look at it tomorrow