Closed h-vetinari closed 1 week ago
I don't know of anyone really using the libarrow / pyarrow cuda modules. The one project that I know that was previously using it was cuDF, but it has since moved to not needing the cuda modules.
I would argue we should keep the track_features
in place until someone presents a compelling reason otherwise.
If no one needs them, does it make sense to deprecate them?
We only weighed them down initially because there was no __cuda
dependency initially.
We only weighed them down initially because there was no
__cuda
dependency initially.
That was my understanding of the history behind this as well. I'm planning to try to remove this for v18 and we can see how things go.
Also just found that I had opened an issue for this even longer ago already (#1169) where there was another voice in favour of doing this.
libarrow
&pyarrow
have had a track_features weighing down the CUDA builds ever since those got introduced in https://github.com/conda-forge/arrow-cpp-feedstock/commit/1e57bc08f594170075e20c99db44f081e0518f63 & https://github.com/conda-forge/arrow-cpp-feedstock/commit/99a1e8df43adaa1f4671ea7c77e56f2c3ce17171I don't see why we need to do that - people with the right drivers installed (and thus the right
__cuda
virtual package) should just get the CUDA builds? We do have run-dependence on__cuda
already https://github.com/conda-forge/arrow-cpp-feedstock/blob/909b79ce4a04dcc031fe93c534ef03ad3f2e7dfc/recipe/meta.yaml#L208 so clients without CUDA support wouldn't pointlessly download the bigger binaries.Am I missing something else?
CC @conda-forge/arrow-cpp @conda-forge/core