[FEA] Add narwhals tests to our test suite

[vyasr]

Is your feature request related to a problem? Please describe.
@MarcoGorelli has reported a number of issues to cudf stemming from narwhals. narwhals is sussing out a number of issues that we don't currently see anywhere else except likely in the pandas test suite, where our pass rate remains too low to be used as an absolute pass/fail indicator in CI.

Describe the solution you'd like
We would benefit from running the narwhals test suite as part of cudf's CI similarly to how we run the pandas test suite.

Describe alternatives you've considered
None

Additional context
The narwhals tests look like they take <5 min, so they shouldn't be too expensive, but I would still recommend only running them on a single CUDA ver/Python ver/arch combination to limit resource usage.

[MarcoGorelli]

Thanks for the ping! Glad to hear our test suite has been useful

You can run the Narwhals test suite in ~60-70 seconds: https://www.kaggle.com/code/marcogorelli/testing-cudf-in-narwhals?scriptVersionId=215469980

git clone https://github.com/narwhals-dev/narwhals --depth=1
cd narwhals
pip install -U -e ".[dev]"
pytest tests

So long as cuDF is installed, it will include cuDF in the parametrised tests

[vyasr]

Awesome. From a quick look at the configuration it looks like the test suite automatically selects which narwhals backends to test. Is there a way to select some explicitly? If not, would you be open to adding some custom flags to your pytest_addoption implementation to allow for selecting the pandas tests, polars tests, etc? The reason that I ask is that if your test suite is that fast, we might even want to consider running it for each of our front-ends separately, i.e. run the cuDF tests, then run the pandas tests with cudf.pandas enabled, then run the polars test suite with GPU support via cuDF enabled.

[MarcoGorelli]

Definitely in-scope, we'll look into how to provide this

[MarcoGorelli]

Alright, done!

Here's an example of how you could run the tests: https://www.kaggle.com/code/marcogorelli/testing-cudf-in-narwhals?scriptVersionId=216065554

• cuDF tests: pytest --constructors=cudf (~35 seconds)
• polars[gpu] tests: NARWHALS_POLARS_GPU=1 pytest --constructors=polars[lazy] (there's currently two failures, which I've opened issues about) (~14 seconds)
• cudf.pandas tests: NARWHALS_DEFAULT_CONSTRUCTORS=pandas python -m cudf.pandas -m pytest (many of these fail due to [BUG] using python -m cudf.pandas and calling hasattr converts NA to NaN #17666 ) (will check the timings once [BUG] using python -m cudf.pandas and calling hasattr converts NA to NaN #17666 is addressed, but it's probably similar)