Add python-constraint to avoid conda-forge dependencies.

[HackerFoo]

Any packages with a condarc specifying outside channels will require those channels to be available to install them.

Depending in conda-forge is undesirable because it can conflict with many of the default packages. pip cannot be used directly in a conda recipe, so PyPI packages must be converted to Conda packages.

This has caused problems with vtr-optimized (PR), which this PR aims to resolve, but there are other packages using condarc as well.

[PiotrZierhoffer]

Ha, we just discussed adding this, like minutes ago.

LGTM for me, but I'll ask @ajelinski to take a look as well

[litghost]

LGTM

[PiotrZierhoffer]

this should, probably, go to litex-conda-misc, but since it's a direct dependency of vtr-optimized, I don't think it's a big problem to have it here

[PiotrZierhoffer]

is this file autogenerated with something like conda skeleton pypi {packagename}? I actually never used it before, but if this is the result, then it's pretty good

[HackerFoo]

Yeah, but I had to edit a bit, and change from .gz to .bz2.

[acomodi]

There should be another way to fix the python-constraint problem. Basically the condarc file in vtr-optimized should include also the defaults package.

All the dependencies for vtr-optimized will be taken from defaults apart from python-constraint which will be taken from conda-forge. Given that python-constraint does not have any other dependency itself, it should not generate any conflicts.

So basically the condarc file would look like the following:

channels:
  - defaults
  - conda-forge

in the current implementation of the vtr-optimized package, all the various dependencies are taken from conda-forge, hence the conflicts

[HackerFoo]

There should be another way to fix the python-constraint problem. Basically the condarc file in vtr-optimized should include also the defaults package.

All the dependencies for vtr-optimized will be taken from defaults apart from python-constraint which will be taken from conda-forge. Given that python-constraint does not have any other dependency itself, it should not generate any conflicts.

So basically the condarc file would look like the following:

channels:
  - defaults
  - conda-forge

in the current implementation of the vtr-optimized package, all the various dependencies are taken from conda-forge, hence the conflicts

@acomodi I tried this, based on this tip, except reversed to avoid conda-forge, but unfortunately, it didn't work. There were still dependencies from conda-forge in info/about.json.

[HackerFoo]

@ajelinski Any comments?

[ajelinski]

Are you sure you tried with channels in this exact order? The order matters and in the link provided conda-forge is the most important channel.

I've tested how the recipe is rendered after adding the defaults channel on top in condarc like @acomodi mentioned. It definitely solves the problem which is caused by requiring libraries that are not available on the defaults (ie. main) channel:

  run:
  - libstdcxx-ng >=9.3.0
  - libgcc-ng >=9.3.0
  - tbb >=2020.2

This is because conda-build gets all packages it needs from conda-forge and there is gcc/gxx 9.3.0. After making defaults the most important channel (the first one in condarc) it becomes:

  run:
  - libstdcxx-ng >=7.3.0
  - tbb >=2020.3
  - libgcc-ng >=7.3.0

because conda-build will use all packages it can from the defaults channel. Only those not available will be taken from the conda-forge channel.

Those requirements can be satisfied without conda-forge so it won't be needed to install vtr-optimized after such change.

[ajelinski]

Btw. info/about.json doesn't matter during dependency resolving. The depends key from info/index.json inside the package is what matters the most and there you can see this exact problem with requiring libgcc/libstdcxx >=9.3.0.

[ajelinski]

I have added a branch with the aforementioned condarc change and all run requirements of the built package can be satisfied without conda-forge: https://travis-ci.com/github/litex-hub/litex-conda-eda/jobs/461371641#L2925

The package built in that CI can be installed with:

conda install -c litex-hub/label/travis-fix-vtr-optimized-209189450 vtr-optimized

[HackerFoo]

@ajelinski Thanks. I'll try your package and close this if it works, unless there is still value in it.

[HackerFoo]

@ajelinski It seems to work, so made a PR with your branch: #62

[HackerFoo]

Superseded by #62