bug: invalid values returned by .first().over(w) or .last().over(w) when using the pandas backend

Minimal reproducer:

import ibis
import pandas as pd

df = pd.DataFrame(
    {

        "g": ["a", "a", "a", "a", "a"],
        "x": [0, 1, 2, 3, 4],
        "y": [3, 2, 0, 1, 1],
    }
)
df.to_parquet("test.parquet")
t_pandas = ibis.pandas.connect({"t": df}).table("t")
t_duckdb = ibis.duckdb.connect().register("test.parquet", table_name="t")


def simple_window_ops(t):
    w = ibis.window(
        group_by=t.g,
        order_by=[t.x, t.y],
        preceding=1,
        following=0,
    )
    return t.mutate(
        x_first=t.x.first().over(w),
        x_last=t.x.last().over(w),
        y_first=t.y.first().over(w),
        y_last=t.y.last().over(w),
    )

Then pandas does not seem to take the preceding and following window boundaries into account:

>>> print(simple_window_ops(t_pandas).execute())
   g  x  y  x_first  x_last  y_first  y_last
0  a  0  3        0       4        3       1
1  a  1  2        0       4        3       1
2  a  2  0        0       4        3       1
3  a  3  1        0       4        3       1
4  a  4  1        0       4        3       1

while duckdb works as expected:

>>> print(simple_window_ops(t_duckdb).execute())
   g  x  y  x_first  x_last  y_first  y_last
0  a  0  3        0       0        3       3
1  a  1  2        0       1        3       2
2  a  2  0        1       2        2       0
3  a  3  1        2       3        0       1
4  a  4  1        3       4        1       1

[saulpw]

Thanks for reporting this, @ogrisel. This is definitely two bugs. We’ll try to get pandas to return valid results, or at the very least, raise a meaningful error if it can’t.

[ogrisel]

BTW, what about the second problem documented in https://github.com/ibis-project/ibis/issues/4676#issuecomment-1283756388

Do you want me to open a dedicated issue or both problems are likely to be solved by the same PR?