Entrypoints in conda recipe

[jacobtomlinson]

@jameslamb we are specifying the entrypoints in the pyproject.toml and the recipe.yaml. Given that the recipe just calls pip install do we need to specify them in both places?

rapids-cli/conda/recipes/rapids-cli/recipe.yaml

Lines 26 to 35 in 6e8d487

	entry_points:
	- driver = rapids_cli.doctor.checks.cuda_driver:check_driver_compatibility
	- gpu = rapids_cli.doctor.checks.gpu:gpu_check
	- gpu_compute_capability = rapids_cli.doctor.checks.gpu:check_gpu_compute_capability
	- cuda = rapids_cli.doctor.checks.cuda_driver:cuda_check
	- driver_compatibility = rapids_cli.doctor.checks.cuda_driver:check_driver_compatibility
	- sdd_nvme = rapids_cli.doctor.checks.sdd_nvme:check_sdd_nvme
	- memory_to_gpu_ratio = rapids_cli.doctor.checks.memory:check_memory_to_gpu_ratio
	- nvlink_status = rapids_cli.doctor.checks.nvlink:check_nvlink_status
	- os = rapids_cli.doctor.checks.os:detect_os

The problem is that if we need to modify the entrypoints then we have to remember to do it in two places.

[jameslamb]

Good question! I think the entry_points do also need to be specified in the rattler-build recipe.

Given that the recipe just calls pip install do we need to specify them in both places?

This pip install is used during the build, you're right

rapids-cli/conda/recipes/rapids-cli/recipe.yaml

Line 24 in 6e8d487

python -m pip install -v .

But when you conda install rapids-cli, conda is not literally running pip install again... it's just downloading and unpacking the package.

I think that's why the entrypoints need to be specified in the conda package's metadata, so that they can be registered in the conda environment at install time.

To test this, you could do something like:

1. remove those entrypoints from the recipe
2. add a call to one of the entrypoints in https://github.com/rapidsai/rapids-cli/blob/main/ci/test_python.sh