[IPython-dev] Docker IPython
Matthias Bussonnier
bussonniermatthias at gmail.com
Tue Aug 5 13:36:02 EDT 2014
Le 5 août 2014 à 18:11, Jon Wilson <jsw at fnal.gov> a écrit :
> If any substantial fraction of your users will want
> scipy/numpy/matplotlib, I would (almost, see below) recommend conda.
Adrew’s IHaskell users will definitively not want the scipy stack, they want Haskell things.
> Conda was, as I understand it, created because pip left too many
> barriers in place against the use of scipy/numpy etc. Specifically,
> experience indicated that many people who might otherwise have casually
> investigated scientific python tools did not do so because pip required
> them to have a proper FORTRAN development environment set up, and they
> did not wish to figure out how to do this.
>
> Conda distributes binaries rather than exclusively source, which is an
> effective way around this sort of problem.
>
> OTOH, a pure-python package that is hosted on PyPI (and therefore
> installable via pip) can (usually) be trivially made into a conda
> package via `conda skeleton pypi <package-name>`. So making pip-style
> packages tends to get you conda packages for almost free.
I suppose this « easy » way to make PyPi package from conda package explain
why continuum package are outdated by more that a year using pip :-)
Even if conda seem great, I still feel sad that there is little effort to help fixing
python packaging (no, replacing is not fixing), but I understand that starting
from scratch might be easier. Installig SciPy was much more a pain even a few
month ago than now.
Keep also in mind than miniconda will be smaller if you decide to use it.
I would also reming you that with 3.0, the IPython notebook can start many kernels in different languages,
so nothing prevent you from installing the notebook using pip, and having a conda kernel.
(it is even possible with 2.x)
So +1 for pip which should be enough, or even maybe Julia Taylor PPA if compatible.
—
M
More information about the IPython-dev
mailing list