fix(rust, python): treat null columns as zero in sum_horizontal

[edavisau]

Fixes #13113

• Null columns are treated as additive identity in the presence of non-null columns. They can be skipped in the computation
• If all columns are Null, return Null

I have also changed the behaviour of single columns sums. This seems more consistent to multi-column sums. Example:

    df = pl.DataFrame(
        {
            "A": [5, None, 3, None],
            "B": [5, 3, None, None],
        }
    )

As reference, when summing multiple columns
sum_horizontal("A", "B") = [10, 3, 3, 0]

Before
sum_horizontal("A") = [5, None, 3, None]
After
sum_horizontal("A") = [5, 0, 3, 0]

[stinodego]

Thanks for the PR!

If all columns are Null, return Null

I don't think this is correct. Null should just be treated as the identity always.

[edavisau]

@stinodego In this case I was referring to the actual Null data type, as opposed to all values being null i.e.

df = pl.DataFrame([
    pl.Series("A", [None, None, None], pl.Null),
    pl.Series("B", [None, None, None], pl.Null),
    pl.Series("C", [None, None, None], pl.Null),
])

If you replaced pl.Null with pl.Int32 above you'll get 0 in the result, similarly if you replaced with pl.Utf8 you'll get empty strings in the result. In addition, if 2 of the columns were pl.Null, and one was pl.Int32, you'd still get a pl.Int32 result. The all-pl.Null situation seemed too ambiguous though. Do you have another suggestion for this case?

Thanks!

[stinodego]

Ah, I hadn't understood that part. Then it's probably fine!

[ritchie46]

Thanks!