pandas groupby + agg(dict) is ~5.32× slower than polars group_by + agg for three aggregates per group

Same three aggregates on the same groups.

Result

Spread over 3 runs: pandas groupby + agg(dict) CV 8.7% / polars group_by + agg CV 3.7% / ratio CV 9.5%

The two calls

df.groupby('g').agg(v_sum=('v','sum'), v_mean=('v','mean'), k_max=('k','max'))

df.group_by('g').agg(pl.col('v').sum(), pl.col('v').mean(), pl.col('k').max())

Conditions

1,000 groups, 400,000 rows.

Where this number does not apply

Absolute times here are small, so the ratio moves more between runs than the large operations do.

Reproduce

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