[Python] Adopt a more pythonic method for UDF registration

[westonpace]

Describe the enhancement requested

It's currently quite complicated to register a UDF. In a recent PR we reviewed a few different suggestions. We should adopt one of these.

Component(s)

Python

[westonpace]

I think this suggestion, from @pitrou , is rather compelling:

udf = pc.register_scalar_function(
    name="y=x*2",
    arg_names=["x"],
    doc = {"summary": "multiply by two function",
           "description": "test multiply function"}
    )
for in_type in [pa.int64(), pa.float64()]:
    udf.add_kernel([in_type], in_type, impl)

It could also open the way to a decorator-like syntax:

udf = pc.register_scalar_function(
    name="y=x*2",
    arg_names=["x"],
    doc = {"summary": "multiply by two function",
           "description": "test multiply function"}
    )

@udf.kernel([pa.float64()], pa.float64())
@udf.kernel([pa.int64()], pa.int64())
def impl(ctx, x):
    # ...

[dangotbanned]

We could definitely benefit from this and/or #31612 in https://github.com/narwhals-dev/narwhals

Related

• feat(expr-ir): Support group_by, utilize pyarrow.acero narwhals-dev/narwhals#3143
• [enh]: More permissive group_by().agg(...) for pyarrow narwhals-dev/narwhals#2385