Fix GH-61477: Prevent spurious sort warning in concat with unorderable MultiIndex

[Neer-Pathak]

Fix GH-61477: Stop Spurious Warning When concat(..., sort=False) on Mixed-Type MultiIndex

Overview

When you do something like:

pd.concat([df1, df2], axis=1, sort=False)

and your two DataFrames have MultiIndex columns that mix tuples and integers, pandas used to try to sort those labels under the hood. Since Python cannot compare tuple < int, you’d see:

RuntimeWarning: '<' not supported between instances of 'int' and 'tuple'; sort order is undefined for incomparable objects with multilevel columns

This warning is confusing, and worse, you explicitly asked not to sort (sort=False), so pandas should never even try.

What Changed

1. Short-circuit Index.union when sort=False
   Before: Even with sort=False, pandas would call its normal union logic, which might attempt to compare labels.

Now: If you pass sort=False, we simply concatenate the two index arrays with:

np.concatenate([self._values, other._values])

and wrap that in a new Index. No comparisons, no warnings, and your original order is preserved.

2. Guard sorting in MultiIndex._union
   Before: pandas would call result.sort_values() when sort wasn’t False, and if labels were unorderable it would warn you.

Now: We only call sort_values() when sort is truthy (True), and we wrap it in a try/except TypeError that silently falls back to the existing order on failure. No warning is emitted.

3. New Regression Test
   A pytest test reproduces the original bug scenario, concatenating two small DataFrames with mixed-type MultiIndex columns and sort=False. The test asserts:

No RuntimeWarning is raised

Column order is exactly “first DataFrame’s columns, then second DataFrame’s columns”

Respects sort=False: If a user explicitly disables sorting, pandas won’t try.

Silences spurious warnings: No more confusing messages about comparing tuples to ints.

Keeps existing behavior for sort=True: You still get a sort or a real error if the labels truly can’t be ordered.

For testing we can try

import numpy as np, pandas as pd

left = pd.DataFrame(
    np.random.rand(5, 2),
    columns=pd.MultiIndex.from_tuples([("A", 1), ("B", (2, 3))])
)
right = pd.DataFrame(
    np.random.rand(5, 1),
    columns=pd.MultiIndex.from_tuples([("C", 4)])
)

# No warning, order preserved:
out = pd.concat([left, right], axis=1, sort=False)
print(out.columns)  # [("A", 1), ("B", (2, 3)), ("C", 4)]

# Sorting still works if requested:
sorted_out = pd.concat([left, right], axis=1, sort=True)
print(sorted_out.columns)  # sorted order or TypeError if impossible