[SciPy-dev] a modest proposal for technology previews

[David Cournapeau]

Gael Varoquaux wrote:
>
> Currently Eggs + pypi don't work well-enough for compiled code. There is
> a famous "sandbox violation" bug that strikes every so and then when
> installing source packages that have to be builts. It may be a due to a
> bad interaction between numpy.distutils and setuptools.

Let's not go into the setuptools debate, please. I think I am in a
relatively good position to know about distutils/setuptools
idiosyncraties. Pypi is independant of setuptools, and can already help
quite a bit for the source release part. I think to say: grab this
tarball from pypi is quite a big step toward easier installation
compared to use svn.

It could be a good policy to tell people developing scikits to upload
sources to pypi whenever they do a release (I am as guilty as other in
that department).

Then, there is the problem to build binary releases. On Linux, there are
automated build farms available; I spent quite a good deal of time on
this at some point. I am not sure you can see the following page, but
this shows the state of linux package on various distributions

https://build.opensuse.org/project/monitor?project=home%3Aashigabou

It is almost 100 % automated. For windows and mac os X, this needs to be
done differently. The problem is then lack of resources (of which time
is not the cheapest). But basically, it is not fundamentally different
from how R does it.

cheers,

David