[Distutils] numpy's distutils fork circa 2001

[Dag Sverre Seljebotn]

On 11/29/2012 10:51 PM, Alex Clark wrote:
> Is anyone familiar with this situation enough to comment on it, in a few
> sentences? Based on:
>
> -
> https://github.com/numpy/numpy/commit/f1a2d6376c430f65550efa235209b86c1a0967e3
>
>
> It looks like:
>
> - Numpy forked distutils (2001) prior to setuptools existence/popularity
> (2004)
> - As a result, SciPy users must now pay a terrible price (i.e. SciPy
> does not `install_requires` numpy as you might expect;  i suspect
> because it relies on numpy's distutils… which it can't rely on until
> it's installed… which it can't `install_requires` until it's installed…)

I believe numpy.distutils is not a fork as such but extends and 
monkey-patches distutils...

The scientific Python community do have problems distributing software 
(and hate distutils with a passion), and building SciPy the right way is 
something for experts, not newbies.

But while I know the problem you mentioned has come up on the SciPy 
list, it really is a trifle in this context. Having to remember to 
install NumPy before SciPy is not what people have issues with.

The pain points SciPy deals with are just from a different world. I 
think the most common problems are of the sort a) linking with the right 
LAPACK in the right way, b) using Fortran compilers with distutils and 
in particular combining various Fortran compilers with various C runtime 
libraries on Windows.

Going beyond SciPy and just to provide another example, a non-trivial 
number of scientific projects use Fortran and Cython and NumPy together; 
that's basically unachievable with either distutils or numpy.distutils 
or Cython.Distutils (so one uses an actual build tool such as CMake or 
scons of waf instead; distutils doesn't really qualify (or was meant to 
be) a build tool after all).

Dag Sverre