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

@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

Messed up the merge 🥲
Fixed

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

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