fix: Preserve pandas column name attribute

[FBruzzesi]

What type of PR is this? (check all applicable)

• 💾 Refactor
• ✨ Feature
• 🐛 Bug Fix
• 🔧 Optimization
• 📝 Documentation
• ✅ Test
• 🐳 Other

Related issues

• Closes bug: select and with_columns don't always preserve column index name for pandas-like #1483

Checklist

• Code follows style guide (ruff)
• Tests added
• Documented the changes

[FBruzzesi]

This might be one step ahead of dask#11874

[FBruzzesi]

Mental sanity for symmetry with other pandas-like and order of the spec 😅

[FBruzzesi]

I wanted to concatenate two methods here, mostly for the sake of it.
Might be worth reparametrizing it

[FBruzzesi]

Dask select and with_columns have no issues since we use .assign method for both!

[MarcoGorelli]

thanks! i'll do a little test to check there's no accidental / surprise overhead

[FBruzzesi]

@MarcoGorelli I tried a simple script to check rename_axis overhead (not in narwhals), and it seems to be a pretty low overhead operation:

import timeit
from pprint import pprint

import numpy as np
import pandas as pd


def benchmark_rename_axis_overhead(n_rows: int, n_cols: int, repetitions: int = 1000) -> float:
    df = pd.DataFrame(np.random.randn(n_rows, n_cols))

    return timeit.timeit(
        lambda: df.rename_axis(columns="new_name", copy=False),
        number=repetitions
    ) / repetitions


TOTAL_SIZE = 10_000_000

results = {
    (int(TOTAL_SIZE/10**i), 10**i): benchmark_rename_axis_overhead(n_rows=int(TOTAL_SIZE/10**i), n_cols=10**i)
    for i in range(1, 6)
}

pprint(results)

{(100, 100000): 1.2109115000015436e-05,
 (1000, 10000): 1.679945200001498e-05,
 (10000, 1000): 1.2669480999988991e-05,
 (100000, 100): 1.331965500000365e-05,
 (1000000, 10): 2.9313745999985485e-05}

Let me know what your thoughts are and if we need to check something more extensively

[FBruzzesi]

Messed up the merge 🥲
Fixed

[MarcoGorelli]

thanks!

Let me know what your thoughts are and if we need to check something more extensively

i'm still slightly concerned about there being extra copies in old versions, i'll take another look

[FBruzzesi]

thanks!

Let me know what your thoughts are and if we need to check something more extensively

i'm still slightly concerned about there being extra copies in old versions, i'll take another look

No rush! I keep merging main to avoid having to deal with a large conflicting diff if that happens to be the case

[MarcoGorelli]

🤔 i'm wondering if we could/should just set df.columns.name = self._native_columns_name

else we need to trust pandas' copy-on-write / copy=False working properly, which i'm not sure i do

[FBruzzesi]

I trust your judgement here - but should I would be concerned as a pandas user?

[MarcoGorelli]

thanks @FBruzzesi !

I don't trust pandas' rename_axis, nor do I trust that copy-on-write will take effect properly, we would need more comprehensive benchmarks to be sure

but, in select and with_columns, we create new objects, so it should be safe to just use .columns.name = ... there