IRC channel logs


back to list of logs

<rekado_>zimoun: thanks for sharing the article on the supercooled water controversy!
<zimoun>rekado_: thanks go to Konrad, from their Monday’s talk. :-)
<civodul>zimoun: hey! did you attend the container talk this morning? how was it?
<rekado_>zimoun: heh, I read some of the comments and found one to be really well-written … and then I saw it was Konrad’s comment :)
<zimoun>rekado_: oh, I have not seen the comments… going to read.
<zimoun>civodul: yeah, I was. Late as usual, but there. :-)
<zimoun>it was interesting.
<civodul>just checked the slides:
<civodul>(in Frenglish)
<zimoun>not as Spack ;-) I mean it explains the pros/cons about containers. Conclusion Docker sucks for HPC but Singularity is okish
<civodul>"Conclusion: containers don't guarantee reproducibility... but they have other qualities" lol
<civodul>(from the last slide)
<civodul>i find it problematic that many illustrations come from the marketing depts of these companies
<civodul>suggests the talking points come from there as well
<zimoun>the good point is the presenter explained that containers are the the graal for reproducibility
<zimoun>not the graal
<civodul>got it :-)
<zimoun>well, they explained security issues etc.
<zimoun>well, from my understanding, it was an “honnest“ presentation of containers in the scope of Reproducible Science targetting HPC.
<zimoun>they said that is poor as Konrad explained on Monday, etc.
<zimoun>On the other hand, Guix cannot fight against “packages ready-to-use” of Tensorflow and PyTorch &co. Even with more manpower, some beasts are still ugly beasts…
*zimoun 'll be back!
<civodul>yeah not having tensorflow, pytorch, & co is more and more of a proble
<zimoun>Well, Debian folks have hard time with “machine learning stuff” too. I am not reading carefully their mailing-lists (debian-med and debian-ai, e.g., but their recent messages seem underline that tensorflow and pytorch are ugly beast.
<zimoun>From Debian dev: «Currently, the most practical recommendation to serious deep learning users are "pip or conda", or upstream binary releases.»
<rekado_>for us the biggest obstacle to tensorflow is bazel
<rekado_>there are no fundamental problems with packaging bazel; it’s just a lot of busy work for little gain in itself.
<rekado_>by that I mean that we would override the inputs anyway and not let bazel download things, so using bazel to build tensorflow is merely an inconvenient convenience.
<rekado_>we should have a hackathon to map out the open problems and chip away at it.
<rekado_>keras, pytorch, and tensorflow
<rekado_>another hackathon may be needed for some CRAN stragglers…
<rekado_>r-metap, for example
<civodul>rekado_: we could definitely plan for an on-line ML packaging hackathon, like zimoun organized a while back
<zimoun>yeah it is doable… maybe in June. :-)
<zimoun>rekado_: are you aware about MRAN, a CRAN snappshots archive run by Microsoft
<rekado_>zimoun: yes, I know that it exists, but I hadn’t considered using it to download CRAN source archives.
<zimoun>it could be nice to add support to fallback
<rekado_>yes, let’s see if it’s possible!
*rekado_ builds all CRAN updates now
<zimoun>an example is r-foreign at commit d81fb2a. Hash mismatch, so the time-machine is broken.
<rekado_>do you have an idea how to fix missing source code in the time machine?
<zimoun>no, yet. :-) I am reading the source code
<rekado_>can we perhaps provide a hash-indexed alist of *current* URLs that satisfy the vanished sources?
<rekado_>the *current* Guix would inject that as a fallback to download the requested sources.
<rekado_>this way we could retroactively fix old versions of Guix
<zimoun>an alist fixing in-place upstream replacement. yeah maybe
<rekado_>hmm, trying to run pigx-rnaseq.w and … it’s not working at a really basic level.
<rekado_>looks like the profile isn’t actually built.
<rekado_>or rather: I get one profile that’s correct, and a reference to another one that doesn’t actually exist
<zimoun>there is 2 issues: fixing the current past (as you said retroactively) and fixing the future past (fallback to MRAN in addition to SWH)
<rekado_>this stuff is hard to test automatically
<rekado_>zimoun: yes
<rekado_>they are independent
<rekado_>once my big block of CRAN updates is done, I’ll look into adding MRAN support.
<rekado_>it may not even be difficult
<zimoun>speaking about CRAN, could you review ? which fixes r-v8
<rekado_>zimoun: yes, I’ll get on to it after the CRAN updates :)
<rekado_>and I also want to take a look at your recursive importer patches
<zimoun>the recursive importer patches just changed the returned value by <foo>->guix-name to put all as (values #f ’()) instead of #f when failing. Otherwise, some corner cases are broken. There are trivial patches. :-)
<zimoun>about your profile, it looks like recent lfam issue, see bug#45992
<zimoun>civodul: are you following TPs sessions from the training?
<zimoun>TP other than yours :-)
<civodul>zimoun: nope, at least not today
***rekado_ is now known as rekado
<rekado>trying to untangle this ball of yarn now… –>
<rekado>that’s also the source of the profile bug I encountered
<rekado>#$profile here is an existing directory:
<rekado>but somewhere else a profile is referenced that hasn’t even been built
<rekado>I’m hoping that most of this code is obsolete and could be replaced with computed-file
<civodul>rekado: tip of the day: did you know there's now a lowerable <profile> record type? :-)
<civodul>at first sight, it could be useful here
<civodul>would allow you to de-monadify this code
<civodul>but then, you prolly shouldn't call build-derivations right here; the result can be gc'd right away
<rekado>yeah, this is a horrible mess
<rekado>all I remember is that I was trying to get everything built and return a launcher script that then uses those things.
<rekado>but I barely understand how it was supposed to work
<rekado>AFAIU I’ll still need “run-with-store” and “mlet %store-monad” to actually build the derivation.
<civodul>no, you could use 'build-derivation', which is non-monadic
<civodul>but in general, 'build-derivation' should be done as the last thing of the program
*rekado has left the comfort zone behind
<civodul>in between you should only pass file-like objects around
<civodul>if the "gwl" command returns a script, there should also be an option to register a GC root
<civodul>(as with "guix build -r")
*rekado takes notes
<rekado>so, now that the profile-compiler exists, I don’t need to call profile-derivation but can just embed a profile directly in a gexp?
<rekado>I’m feeling especially dense right now…
<rekado>“computed-file” is the declarative counterpart of gexp->derivation
<rekado>but … how can I pass it to build-derivations, which expects … well, a derivation?
*rekado reads more
*rekado thinks it’s done via lower-object
<civodul>which is monadic
<rekado>ah, it clicked! Thanks!
<zimoun`>oh, I miss how a derivation could be built wihtout monade… Need to read more.
<rekado>zimoun: you’ll be able to see it in process->script later
<rekado>the basics already work, but I’m now working on adding all the other features back in
<rekado>(e.g. containerization)
<rekado>for containers I need to get the closure of the script, so that all modules are mapped into the container
<rekado>it’s a little unfortunate that I have to provide this manually; maybe I can find a shortcut
<zimoun`>thanks for the explanations