Update dask-cudf wheel name

[raydouglass]

A recent change in pypa/setuptools#4159 may have caused our dask-cudf wheels to be published as dask_cudf-cu12 instead of dask-cudf-cu12.

Additionally, cudf_kafka wheels would have this issue, but 1) we do not publish wheels for cudf_kafka and 2) the conda packages are published as cudf_kafka (with underscore), so be a larger refactor later on.

[bdice]

Thanks @raydouglass. Currently there is no intention to publish wheels for cudf_kafka, because the wheel would have to include libcudf. I experimented a bit with this while working on #14292. I think we're safe with just fixing dask-cudf.

[charlesbluca]

Looking at the failures, seems like we might want to reconsider how we construct the sed commands updating ${package_dir}/${package_name}/_version.py?

sed: can't read python/dask_cudf/dask-cudf/_version.py: No such file or directory

[ajschmidt8]

Do these lines also need to be changed?

• https://github.com/raydouglass/cudf/blob/f58baee9a87ac389d6766ec1bdbaa63be7863493/python/cudf_kafka/pyproject.toml#L55
• https://github.com/raydouglass/cudf/blob/f58baee9a87ac389d6766ec1bdbaa63be7863493/python/custreamz/pyproject.toml#L75
• https://github.com/raydouglass/cudf/blob/f58baee9a87ac389d6766ec1bdbaa63be7863493/python/dask_cudf/pyproject.toml#L78

[bdice]

Do these lines also need to be changed?

• raydouglass/cudf@f58baee/python/cudf_kafka/pyproject.toml#L55

• raydouglass/cudf@f58baee/python/custreamz/pyproject.toml#L75

• raydouglass/cudf@f58baee/python/dask_cudf/pyproject.toml#L78

No changes are needed there. Those are rules for telling isort how to sort imports. The import name is still dask_cudf because Python module names can't have dashes (only underscores). The package name is dask-cudf but the import name (module name) is dask_cudf.

[raydouglass]

Adding back DO NOT MERGE because I want to examine the built wheel before merging.

[bdice]

I think the better fix here would be to rename the directory python/dask_cudf to python/dask-cudf, to align with the package name. However, that is more breaking. It would also affect cudf_kafka. It's probably good to separate that discussion from this PR since its scope is considerably larger.

[bdice]

I don't have strong feelings about this. I just don't think there's a good reason to favor a string replacement in a script (more of a workaround) over the principle of matching the directory structure to the package names. The change would come with some nonzero cost and potential for breakage, though, so I wouldn't do it lightly.

[raydouglass]

FWIW, the change here changes python/dask_cudf/dask-cudf => python/dask_cudf/dask_cudf basically fixing the module name.

But I think you're right that it should be python/$package_name/$module_name and therefore python/dask-cudf/dask_cudf.

[ajschmidt8]

Do these lines also need to be changed?

• raydouglass/cudf@f58baee/python/cudf_kafka/pyproject.toml#L55
• raydouglass/cudf@f58baee/python/custreamz/pyproject.toml#L75
• raydouglass/cudf@f58baee/python/dask_cudf/pyproject.toml#L78

No changes are needed there. Those are rules for telling isort how to sort imports. The import name is still dask_cudf because Python module names can't have dashes (only underscores). The package name is dask-cudf but the import name (module name) is dask_cudf.

Good catch. I didn't see those lines were in the isort section when I did my global search. There are a lot of dask_cudf strings in this repository.

[raydouglass]

The wheel's contents including METADATA/Name looks correct now, so this is ready to merge.

[bdice]

@raydouglass Feel free to admin merge if you'd like. Tests are blocked by #14723.