Numerical stability of Expr.mean in GroupBy context (GroupBy.mean()) seems not only worse than the pandas version, but also worse than Expr.mean in select context.
import numpy as np
import pandas as pd
import polars as pl
df = pd.DataFrame({'data':[10_00_00_00]*100_00_00, 'group':[1,2]*50_00_00}, dtype=np.int32)
print(df.groupby('group').mean())
print(pl.from_pandas(df).groupby('group').agg(pl.col('data').mean()))
print(pl.from_pandas(df).groupby('group').mean())
"""
data
group
1 10000000.0
2 10000000.0
shape: (2, 2)
┌───────┬─────────────┐
│ group ┆ data │
│ --- ┆ --- │
│ i32 ┆ f64 │
╞═══════╪═════════════╡
│ 2 ┆ 1316.134912 │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌╌╌╌╌╌╌┤
│ 1 ┆ 1316.134912 │
└───────┴─────────────┘
shape: (2, 2)
┌───────┬─────────────┐
│ group ┆ data │
│ --- ┆ --- │
│ i32 ┆ f64 │
╞═══════╪═════════════╡
│ 2 ┆ 1316.134912 │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌╌╌╌╌╌╌┤
│ 1 ┆ 1316.134912 │
└───────┴─────────────┘
"""
But Expr.mean in select context does works:
print(df.mean())
print(pl.from_pandas(df).select(pl.col('data').mean()))
print(pl.from_pandas(df).mean())
"""
data 10000000.0
group 1.5
dtype: float64
shape: (1, 1)
┌──────┐
│ data │
│ --- │
│ f64 │
╞══════╡
│ 1e7 │
└──────┘
shape: (1, 2)
┌──────┬───────┐
│ data ┆ group │
│ --- ┆ --- │
│ f64 ┆ f64 │
╞══════╪═══════╡
│ 1e7 ┆ 1.5 │
└──────┴───────┘
"""
I am on Windows with polars==0.14.18. I know int64 is the solution, but it is a little bit confusing, is it a bug?
