pandas pivot_table is ~1.08× slower than polars pivot for pivoting to a wide table

This is the case where the two libraries are closest.

Result

Spread over 3 runs: pandas pivot_table CV 4.7% / polars pivot CV 2.9% / ratio CV 1.8%

The two calls

df.pivot_table(index='g', columns='k', values='v', aggfunc='sum')

df.pivot(on='k', index='g', values='v', aggregate_function='sum')

Conditions

1,000 group keys by 1,000 column keys, summed.

Where this number does not apply

A dense output is required on both sides; sparse pivots are a different path.

Reproduce

cd bench && python run.py --run 1 --out ../data/run1.json