On Fri, Apr 14, 2017 at 1:40 PM, Jeandet Alexis
Hi,
On many places users are advised to use "pip install" to install python
packages. As Fedora "evangelist" I also have to help my friends/colleagues
to setup libs/softwares on Fedora. One the most common trap is the "pip
install", many user does "sudo pip install --upgrade whatever" and if this
is a package already installed by the system it may mess up everything since
on Fedora the install path is the same.
I also had to help a colleague who installed anaconda on Fedora 2x and added
it to PYTHONPATH which broke yum, it took me few minutes to figure out what
he did.
He did this because he followed a formation where the guy said "you should
install anaconda even on linux", in my opinion this is a really bad advice.
Been there, done that, *recently* RHEL uses a workable structure for
alternative python releases, packaging them over in /opt/rh/python33
or whatever for out-of-band python tools. I worked with this
extensively for "airflow", for which I wound up publishing over 100
RPM's in a public repo. Then I tried to RPM bundle the "awscli" tool,
with which I was nowhere near so successful. The basic problem was the
documentation building tool, "python-Sphinx", and different
dependencies requiring different, system incompatible, and in many
cases obsolete versions of the tool.
Python and their popular module hosting site, pypi.org, have run
themselves square into the same problem CPAN had for perl modules, and
which ant, maven and gradle all have for Java tools. To wit, they each
think they know better than the operating system how to package, and
they each generate their own dependency chains that are not, and
cannot hope to be, maintained for both current and future versions of
dependencies by the author of any one module.
So my point is first, I think on ubuntu they install with pip on a different
folder than packaged python packages. Should Fedora do something like
this?(IE configure pip to install on some folder like
/usr/local/libXX/pythonX.X)
Should we do more communication on virtualenv usage? Specially for jupyter
stuff which isn't packaged on Fedora(as far as I know), at least my
colleagues engineers and scientists use it a lot.
I have no perfect solution for this issue, anyway I would be happy to get
feedbacks on this topic.
Best regards,
Alexis.
One solution is, for non-RPM packages, use pyvenv. This allows the
creation of local versions of pip installed modules that do not
intermingle with the operating system copies.
Yes, that's also my conclusion for now, even if this is not that user friendly for newcomers.
