BUG: .rolling() returns incorrect values when ts index is not nano seconds

doc/source/whatsnew/v2.1.2.rst

@@ -14,6 +14,7 @@ including other versions of pandas.
 Fixed regressions
 ~~~~~~~~~~~~~~~~~
 - Fixed regression in :meth:`DataFrame.join` where result has missing values and dtype is arrow backed string (:issue:`55348`)
+- Fixed regression in :meth:`~DataFrame.rolling` where non-nanosecond index or ``on`` column would produce incorrect results (:issue:`55026`, :issue:`55106`, :issue:`55299`)
 - Fixed regression in :meth:`DataFrame.resample` which was extrapolating back to ``origin`` when ``origin`` was outside its bounds (:issue:`55064`)
 - Fixed regression in :meth:`DataFrame.sort_index` which was not sorting correctly when the index was a sliced :class:`MultiIndex` (:issue:`55379`)
 - Fixed performance regression with wide DataFrames, typically involving methods where all columns were accessed individually (:issue:`55256`, :issue:`55245`)

pandas/core/window/rolling.py

@@ -21,6 +21,7 @@
 
 from pandas._libs.tslibs import (
     BaseOffset,
+    Timedelta,
     to_offset,
 )
 import pandas._libs.window.aggregations as window_aggregations
@@ -112,6 +113,8 @@
     from pandas.core.generic import NDFrame
     from pandas.core.groupby.ops import BaseGrouper
 
+from pandas.core.arrays.datetimelike import dtype_to_unit
+
 
 class BaseWindow(SelectionMixin):
     """Provides utilities for performing windowing operations."""
@@ -1887,7 +1890,12 @@ def _validate(self):
                     self._on.freq.nanos / self._on.freq.n
                 )
             else:
-                self._win_freq_i8 = freq.nanos
+                try:
+                    unit = dtype_to_unit(self._on.dtype)  # type: ignore[arg-type]
+                except TypeError:
+                    # if not a datetime dtype, eg for empty dataframes
+                    unit = "ns"
+                self._win_freq_i8 = Timedelta(freq.nanos).as_unit(unit)._value
 
             # min_periods must be an integer
             if self.min_periods is None:

pandas/tests/window/test_rolling.py

@@ -1950,3 +1950,35 @@ def test_numeric_only_corr_cov_series(kernel, use_arg, numeric_only, dtype):
         op2 = getattr(rolling2, kernel)
         expected = op2(*arg2, numeric_only=numeric_only)
         tm.assert_series_equal(result, expected)
+
+
+@pytest.mark.parametrize("unit", ["s", "ms", "us", "ns"])
+@pytest.mark.parametrize("tz", [None, "UTC", "Europe/Prague"])
+def test_rolling_timedelta_window_non_nanoseconds(unit, tz):
+    # Test Sum, GH#55106
+    df_time = DataFrame(
+        {"A": range(5)}, index=date_range("2013-01-01", freq="1s", periods=5, tz=tz)
+    )
+    sum_in_nanosecs = df_time.rolling("1s").sum()
+    # microseconds / milliseconds should not break the correct rolling
+    df_time.index = df_time.index.as_unit(unit)
+    sum_in_microsecs = df_time.rolling("1s").sum()
+    sum_in_microsecs.index = sum_in_microsecs.index.as_unit("ns")
+    tm.assert_frame_equal(sum_in_nanosecs, sum_in_microsecs)
+
+    # Test max, GH#55026
+    ref_dates = date_range("2023-01-01", "2023-01-10", unit="ns", tz=tz)
+    ref_series = Series(0, index=ref_dates)
+    ref_series.iloc[0] = 1
+    ref_max_series = ref_series.rolling(Timedelta(days=4)).max()
+
+    dates = date_range("2023-01-01", "2023-01-10", unit=unit, tz=tz)
+    series = Series(0, index=dates)
+    series.iloc[0] = 1
+    max_series = series.rolling(Timedelta(days=4)).max()
+
+    ref_df = DataFrame(ref_max_series)
+    df = DataFrame(max_series)
+    df.index = df.index.as_unit("ns")
+
+    tm.assert_frame_equal(ref_df, df)