Inconsistent behavior with dataframe level arithmetic when using Python's `sum`

[mkleinbort-ic]

Inconsistent behavior with dataframe level arithmetic when using Python's `sum`

[rob-sil]

I think this behavior comes from how polars adds constants to data frames.

Python's sum has a starting value, which defaults to zero. The call sum([a, b]) runs as (0 + a) + b.

When polars goes to add 0 to your first DataFrame, it creates a Series from the constant and infers an i64 type. Apparently, polars allows adding integers to dates, casting the date to an integer using the number of days since the epoch. When you add the second dataframe, addition doesn't crash because i64 + date is allowed, even though date + date is not.

If you change the sum's starting value, you can get different types or different values:

>>> sum(
...     [pl.DataFrame({'x':[1]}).cast({'x':pl.Date}), pl.DataFrame({'x':[2]}).cast({'x':pl.Date})],
...     start=0.0,
... )
┌─────┐
│ x   │
│ --- │
│ f64 │
╞═════╡
│ 3.0 │
└─────┘
>>> sum(
...     [pl.DataFrame({'x':[1]}).cast({'x':pl.Date}), pl.DataFrame({'x':[2]}).cast({'x':pl.Date})],
...     start=10,
... )
┌─────┐
│ x   │
│ --- │
│ i64 │
╞═════╡
│ 13  │
└─────┘

The conversion of dates to integers is pretty weird:

>>> sum([pl.DataFrame({'x':[date(2024, 4, 1)]})])
┌───────┐
│ x     │
│ ---   │
│ i64   │
╞═══════╡
│ 19814 │
└───────┘