Wrong groupby_rolling result when period=2d and offset=-3d

[SaneBow]

Polars version checks

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of Polars.

Issue description

When period=2d and offset=-3d, the groupby_rolling gives unexpected result. Possibly due to incorrect boundary handling.
Strangely, when offset is set to some other values like offset=-4d or offset=-5d, then results are correct.

Reproducible example

import polars as pl
from datetime import datetime
dcol = pl.date_range(datetime(2021,1,1), datetime(2021,1,4), '1d', eager=True)
vcol = np.arange(1, len(dcol)+1)
df = pl.DataFrame({'datetime': dcol, 'v': vcol})
df

┌─────────────────────┬─────┐
│ datetime            ┆ v   │
│ ---                 ┆ --- │
│ datetime[μs]        ┆ i64 │
╞═════════════════════╪═════╡
│ 2021-01-01 00:00:00 ┆ 1   │
│ 2021-01-02 00:00:00 ┆ 2   │
│ 2021-01-03 00:00:00 ┆ 3   │
│ 2021-01-04 00:00:00 ┆ 4   │
└─────────────────────┴─────┘

df.groupby_rolling('datetime', period='2d', offset='-3d').agg(pl.col('v'))

Wrong output:

┌─────────────────────┬───────────┐
│ datetime            ┆ v         │
│ ---                 ┆ ---       │
│ datetime[μs]        ┆ list[i64] │
╞═════════════════════╪═══════════╡
│ 2021-01-01 00:00:00 ┆ [1]       │
│ 2021-01-02 00:00:00 ┆ [1, 2]    │
│ 2021-01-03 00:00:00 ┆ [1, 2, 3] │
│ 2021-01-04 00:00:00 ┆ [2, 3, 4] │
└─────────────────────┴───────────┘

Expected behavior

┌─────────────────────┬───────────┐
│ datetime            ┆ v         │
│ ---                 ┆ ---       │
│ datetime[μs]        ┆ list[i64] │
╞═════════════════════╪═══════════╡
│ 2021-01-01 00:00:00 ┆ []        │
│ 2021-01-02 00:00:00 ┆ [1]       │
│ 2021-01-03 00:00:00 ┆ [1, 2]    │
│ 2021-01-04 00:00:00 ┆ [2, 3]    │
└─────────────────────┴───────────┘

Installed versions

--------Version info---------
Polars:      0.18.0
Index type:  UInt32
Platform:    Linux-5.4.0-149-generic-x86_64-with-glibc2.31
Python:      3.10.11 (main, Apr 20 2023, 19:02:41) [GCC 11.2.0]

----Optional dependencies----
numpy:       1.24.3
pandas:      1.5.3
pyarrow:     12.0.0
connectorx:  <not installed>
deltalake:   <not installed>
fsspec:      2023.5.0
matplotlib:  3.6.3
xlsx2csv:    <not installed>
xlsxwriter:  <not installed>

[MarcoGorelli]

Thanks @SaneBow for the report

In this case, the windows will be

• (2021-01-01 -3d, 2021-01-01 -3d + 2d]
• (2021-01-02 -3d, 2021-01-02 -3d + 2d]
• (2021-01-03 -3d, 2021-01-03 -3d + 2d]
• (2021-01-04 -3d, 2021-01-04 -3d + 2d]

i.e.

• (2020-12-29, 2020-12-31], values: []
• (2020-12-30, 2021-01-01], values: [1]
• (2020-12-31, 2021-01-02], values: [1, 2]
• (2021-01-01, 2021-01-03], values: [2, 3]

So, I agree, the output does look incorrect. Will take a look

[MarcoGorelli]

The same example with offset='3d' is also incorrect:

In [30]: import polars as pl
    ...: from datetime import datetime
    ...: dcol = pl.date_range(datetime(2021,1,1), datetime(2021,1,4), '1d', eager=True)
    ...: vcol = np.arange(1, len(dcol)+1)
    ...: df = pl.DataFrame({'datetime': dcol, 'v': vcol})
    ...: df.groupby_rolling('datetime', period='2d', offset='-3d', closed='left').agg(pl.col('v'))
Out[30]:
shape: (4, 2)
┌─────────────────────┬───────────┐
│ datetime            ┆ v         │
│ ---                 ┆ ---       │
│ datetime[μs]        ┆ list[i64] │
╞═════════════════════╪═══════════╡
│ 2021-01-01 00:00:00 ┆ []        │
│ 2021-01-02 00:00:00 ┆ []        │
│ 2021-01-03 00:00:00 ┆ []        │
│ 2021-01-04 00:00:00 ┆ []        │
└─────────────────────┴───────────┘

Was expecting:

shape: (4, 2)
┌─────────────────────┬───────────┐
│ datetime            ┆ v         │
│ ---                 ┆ ---       │
│ datetime[μs]        ┆ list[i64] │
╞═════════════════════╪═══════════╡
│ 2021-01-01 00:00:00 ┆ []        │
│ 2021-01-02 00:00:00 ┆ []        │
│ 2021-01-03 00:00:00 ┆ [1]       │
│ 2021-01-04 00:00:00 ┆ [1, 2]    │
└─────────────────────┴───────────┘