Thanks for moving the thread out. Please allow me to repeat some of
my points here:
On 2019-07-11 10:53, Steffen Möller wrote:
> On an project-internal mailing list the thread "Conda vs Debian"
> evolved. Sam kindly reminded us to discuss this publicly. So, here you
> go, issues raised were
> * interaction with industry-standard non-free software
> + Intel compilers and libraries
> + NVidia drivers and libraries
To some extent this overlaps with some of the "difficulties in
deep learning framework packaging". Many serious users need
the performance improvement brought by the non-free blobs,
but we Debian always have trouble dealing with these blobs.
We don't have control over them, we became passive once
something goes wrong with them.
So my personal opinion is not to touch these legally
problematic stuff anymore. We do keep providing a solid
base system to the world, on top of which business groups
can do whatever they want with non-free blobs. If some
software upstream doesn't have the intention to be well
integrated into distributions, why do we bother packaging
them? Why not just leave the work to the business groups
such as conda.
Apart from the non-free & performance stuff, as a conda
user I still have some reasons to use conda instead of
python packages provided by Debian:
1. conda doesn't require root permission. This helps
non-privileged users alot.
2. conda's python environment management
3. conda is distribution-agnostic. I can retain the same
python environment with conda across different machines
with different systems.
The question is, what can we do to improve Debian, if
we see it appropriate?