solver does not handle meta-packages

[soumith]

The following commands install the incorrect habitat-sim package if mamba is used. When conda is used, it handles things correctly.

conda create -n habitat2 python==3.7 habitat-sim headless -c aihabitat -c conda-forge

Correctly installs aihabitat/linux-64::habitat-sim-0.2.1-py3.7_headless_linux_fc7fb11ccec407753a73ab810d1dbb5f57d0f9b9

mamba install habitat-sim headless -c aihabitat

Incorrectly installs py3.7_linux_fc7fb11ccec407753a73ab810d1dbb5f57d0f9b9

Installing the metapackage headless is supposed to change which binary of habitat-sim gets installed (as you see in the conda command).

[wolfv]

Hi @soumith

yeah that is unfortunate. I just checked the repodata.json of the aihabitat channel at https://conda.anaconda.org/aihabitat/linux-64/repodata.json and it seems that it's using features and track_features to decide between the two variants.

Unfortunately track_features have been deprecated since a while and we haven't implemented full support for them in mamba (they are still used, but only to down-weight variants). In conda-forge and defaults channel, only mutex packages are used these days.

pytorch has now also adopted mutex packages: #336

[soumith]

gotcha. I did suspect so, indeed, I remember pytorch used to be on track-features and we moved to mutex

cc: @erikwijmans @mathfac , you should move to mutex based selection for aihabitat

[soumith]

Thanks @wolfv , please feel free to close the issue as wontfix, It is probably appropriate.