Docker and Python Quirks on PowerPC

TL;DR

• pip install on non-x86_64 often requires compilation which takes minutes

• upstream bugs might necessitate you to manually install dependencies first

Testing Setup

We are using an IBM Power9 machine called bu-21-9. Specs:

• Red Hat Enterprise Linux 7.6

• Docker version 1.13.1

System software is out-of-date, which is not a good starting point.

Python

tl;dr it’s slow to install numpy.

Typically (on Intel x86_64, aka amd64) it’s a breeze to do a

pip install numpy

Installation takes a few seconds because PyPi houses precompiled "wheels" (binary packages) for amd64. These are not available for ppc64le (PowerPC 64 Little-Endian) and you have no choice but to compile from source. On bu-21-9 the command

time docker run --rm python:buster pip install numpy

takes 2.6 minutes.

Unlisted Build Dependencies

scikit-learn

tl;dr upstream bug, do pip install numpy && pip install scikit-image

docker run --rm python:buster pip install scikit-image

The command above actually fails on both amd64 and ppc64le.

Collecting scikit-image
  Downloading scikit-image-0.17.2.tar.gz (29.8 MB)
    ERROR: Command errored out with exit status 1:
     command: /usr/local/bin/python -c 'import sys, setuptools, tokenize; sys.argv[0] = '"'"'/tmp/pip-install-_c7n1tuv/scikit-image/setup.py'"'"'; __file__='"'"'/tmp/pip-install-_c7n1tuv/scikit-image/setup.py'"'"';f=getattr(tokenize, '"'"'open'"'"', open)(__file__);code=f.read().replace('"'"'\r\n'"'"', '"'"'\n'"'"');f.close();exec(compile(code, __file__, '"'"'exec'"'"'))' egg_info --egg-base /tmp/pip-pip-egg-info-eud0npsh
         cwd: /tmp/pip-install-_c7n1tuv/scikit-image/
    Complete output (7 lines):
    Traceback (most recent call last):
      File "<string>", line 1, in <module>
      File "/tmp/pip-install-_c7n1tuv/scikit-image/setup.py", line 234, in <module>
        'build_ext': openmp_build_ext(),
      File "/tmp/pip-install-_c7n1tuv/scikit-image/setup.py", line 58, in openmp_build_ext
        from numpy.distutils.command.build_ext import build_ext
    ModuleNotFoundError: No module named 'numpy'
    ----------------------------------------
ERROR: Command errored out with exit status 1: python setup.py egg_info Check the logs for full command output.

scikit-image does not provide the precompiled binary "wheel" for any platform, so every installation must compile it from source. The problem lies with a circular dependency in its setup.py which imports numpy before listing numpy as a requirement.

There is nothing we can do about upstream bugs but to manually install numpy in our Dockerfile

pip install numpy && pip install scikit-image

Numpy

With some versions of pip (known issue with pypip 9.0.1, known working with pip 20.2.4) you get this error:

Traceback (most recent call last):
    File "/tmp/pip-build-v7lch0r1/numpy/tools/cythonize.py", line 59, in process_pyx
        from Cython.Compiler.Version import version as cython_version
ModuleNotFoundError: No module named 'Cython'

Similar to the problem with scikit-image, here Cython is an unspecified build dependency. You don’t get this error on amd64 since a wheel is available for amd64. Again we don’t have a choice but to manually install build dependencies.

pip3 install Cython && pip3 install numpy

Or updating pip might work

pip3 install --upgrade pip && pip install numpy