https://pantsbuild.org/ logo
Powered by
#general
  • r

    rough-minister-58256

    07/21/2015, 7:02 PM
    maybe we just need to go a step further in bootstrapping the environment to make these builds more fluid
  • r

    rough-minister-58256

    07/21/2015, 7:03 PM
    for bad actors that leverage non-setup.py plumbing, maybe could also do some degree of network isolation/mangling to build them up in a vacuum w/o them hitting pypi or other internet-based services directly
  • e

    enough-analyst-54434

    07/21/2015, 7:03 PM
    Well, presumably you used modern versions there - I know the user-reported issue was in scipy==0.15.1 which definitely does not build under pex 1.0.1
  • e

    enough-analyst-54434

    07/21/2015, 7:04 PM
    Its got a busted setup.py, so it can't build under pip either
  • r

    rough-minister-58256

    07/21/2015, 7:04 PM
    scipy-0.15.1-cp27-none-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl
  • e

    early-mechanic-59271

    07/21/2015, 7:04 PM
    The issue I ran into with scikit-learn was that it requires numpy already be installed.
  • e

    early-mechanic-59271

    07/21/2015, 7:04 PM
    Similar to maven provided scope
  • r

    rough-minister-58256

    07/21/2015, 7:04 PM
    thats exactly the version in my pex!
  • e

    enough-analyst-54434

    07/21/2015, 7:04 PM
    exactly!
  • e

    enough-analyst-54434

    07/21/2015, 7:04 PM
    kwlzn, scipy had this numpy issue to
  • r

    rough-minister-58256

    07/21/2015, 7:04 PM
    $ pex scipy==0.15.1 Python 2.7.8 (default, Jul 17 2014, 174308) [GCC 4.2.1 (Apple Inc. build 5666) (dot 3)] on darwin Type "help", "copyright", "credits" or "license" for more information. (InteractiveConsole)
    >>
  • e

    enough-analyst-54434

    07/21/2015, 7:04 PM
    it assumes you have numpy installed
  • e

    early-mechanic-59271

    07/21/2015, 7:05 PM
    That’s not something I’ve seen before with python, but makes sense given numpy has all the atlas/blas/etc deps.
  • e

    enough-analyst-54434

    07/21/2015, 7:05 PM
    The funny thing about scipy is it seems to try hard in its setup.py to deal with numpy not being on the system
  • e

    enough-analyst-54434

    07/21/2015, 7:05 PM
    but messes up some details
  • e

    enough-analyst-54434

    07/21/2015, 7:06 PM
    @rough-minister-58256: if you have to check out https://github.com/pantsbuild/pants/issues/1751 maybe you can find something simple I missed
  • e

    enough-analyst-54434

    07/21/2015, 7:07 PM
    I had to have an involved vagrant provisioning step to use pre-build scipy only in an venv that had numpy already installed to make things work
  • e

    enough-analyst-54434

    07/21/2015, 7:16 PM
    @stu - killed the osx CI and yours is flowing now - they had announced OSX elevated wait times earlier today, still an issue in progress
  • r

    rough-minister-58256

    07/21/2015, 7:23 PM
    ah hmm, on a special python distro
  • w

    witty-crayon-22786

    07/21/2015, 7:23 PM
    @enough-analyst-54434: thanks
  • w

    witty-crayon-22786

    07/21/2015, 7:23 PM
    didn't realize they ran in lockstep
  • w

    witty-crayon-22786

    07/21/2015, 7:24 PM
    killed a few other attempts.
  • e

    enough-analyst-54434

    07/21/2015, 7:24 PM
    @rough-minister-58256: the issues he had were numerous, I reproed issues on non-special distros too
  • e

    enough-analyst-54434

    07/21/2015, 7:24 PM
    plain old 2.7.10 for example
  • e

    enough-analyst-54434

    07/21/2015, 7:25 PM
    @witty-crayon-22786: yeah - they had an announce a few weeks back that made it clear an org shares 1 queue, until that announce I did not realize this either
  • r

    rough-minister-58256

    07/21/2015, 7:34 PM
    ah, was this 
    pyquante
    ultimately the main problem child? in this case at Twitter (and in most native compile cases - for perf), I would normally advise pre-building building a bdist and letting pants resolve that vs building the sdist every time. seems to be the best work around for these sorts of libs.
  • r

    rough-minister-58256

    07/21/2015, 7:35 PM
    you chew through the bad setup.py once, manually then let pants just grab the product
  • e

    enough-analyst-54434

    07/21/2015, 7:36 PM
    Nope, matplotlib and scipy both expect numpy to be installed and do not express a proper requires on it - pyquante aside
  • e

    enough-analyst-54434

    07/21/2015, 7:37 PM
    And yes, exactly, pre-building is the only viable option I see for now
  • w

    witty-crayon-22786

    07/21/2015, 7:37 PM
    hm
1...666768...2420Latest