API: to_timedelta(ints, unit) give requested unit

doc/source/whatsnew/v3.0.0.rst

@@ -380,6 +380,8 @@ In cases with mixed-resolution inputs, the highest resolution is used:
 
 Similarly, the :class:`Timedelta` constructor and :func:`to_timedelta` with a string input now defaults to a microsecond unit, using nanosecond unit only in cases that actually have nanosecond precision.
 
+Moreover, passing an integer to the :class:`Timedelta` constructor or :func:`to_timedelta` along with a ``unit`` will now return an object with that unit when possible, or the closest-supported unit for non-supported units ("W", "D", "h", "m").
+
 .. _whatsnew_300.api_breaking.concat_datetime_sorting:
 
 :func:`concat` no longer ignores ``sort`` when all objects have a :class:`DatetimeIndex`
@@ -1131,6 +1133,7 @@ Timedelta
 - Accuracy improvement in :meth:`Timedelta.to_pytimedelta` to round microseconds consistently for large nanosecond based Timedelta (:issue:`57841`)
 - Bug in :class:`Timedelta` constructor failing to raise when passed an invalid keyword (:issue:`53801`)
 - Bug in :meth:`DataFrame.cumsum` which was raising ``IndexError`` if dtype is ``timedelta64[ns]`` (:issue:`57956`)
+- Bug in adding or subtracting a :class:`Timedelta` object with non-nanosecond unit to a python ``datetime.datetime`` object giving incorrect results; this now works correctly for Timedeltas inside the ``datetime.timedelta`` implementation bounds (:issue:`53643`)
 - Bug in multiplication operations with ``timedelta64`` dtype failing to raise ``TypeError`` when multiplying by ``bool`` objects or dtypes (:issue:`58054`)
 - Bug in multiplication operations with ``timedelta64`` dtype incorrectly raising when multiplying by numpy-nullable dtypes or pyarrow integer dtypes (:issue:`58054`)
 

pandas/_libs/tslibs/timedeltas.pyx

@@ -44,6 +44,7 @@ from pandas._libs.tslibs.conversion cimport (
     cast_from_unit,
 )
 from pandas._libs.tslibs.dtypes cimport (
+    abbrev_to_npy_unit,
     c_DEPR_UNITS,
     get_supported_reso,
     is_supported_unit,
@@ -290,22 +291,24 @@ cpdef int64_t delta_to_nanoseconds(
         ) from err
 
 
-cdef _numeric_to_td64ns(object item, str unit):
+cdef int64_t _numeric_to_td64ns(
+    object item, str unit, NPY_DATETIMEUNIT out_reso=NPY_FR_ns
+):
     # caller is responsible for checking
     #  assert unit not in ["Y", "y", "M"]
     #  assert is_integer_object(item) or is_float_object(item)
     if is_integer_object(item) and item == NPY_NAT:
-        return np.timedelta64(NPY_NAT, "ns")
+        return NPY_NAT
 
     try:
-        item = cast_from_unit(item, unit)
+        ival = cast_from_unit(item, unit, out_reso)
     except OutOfBoundsDatetime as err:
+        abbrev = npy_unit_to_abbrev(out_reso)
         raise OutOfBoundsTimedelta(
-            f"Cannot cast {item} from {unit} to 'ns' without overflow."
+            f"Cannot cast {item} from {unit} to '{abbrev}' without overflow."
         ) from err
 
-    ts = np.timedelta64(item, "ns")
-    return ts
+    return ival
 
 
 # TODO: de-duplicate with DatetimeParseState
@@ -352,16 +355,21 @@ def array_to_timedelta64(
     cdef:
         Py_ssize_t i, n = values.size
         ndarray result = np.empty((<object>values).shape, dtype="m8[ns]")
-        object item, td64ns_obj
+        object item
         int64_t ival
         cnp.broadcast mi = cnp.PyArray_MultiIterNew2(result, values)
         cnp.flatiter it
         str parsed_unit = parse_timedelta_unit(unit or "ns")
-        NPY_DATETIMEUNIT item_reso
+        NPY_DATETIMEUNIT item_reso, int_reso
         ResoState state = ResoState(creso)
         bint infer_reso = creso == NPY_DATETIMEUNIT.NPY_FR_GENERIC
         ndarray iresult = result.view("i8")
 
+    if unit is None:
+        int_reso = NPY_FR_ns
+    else:
+        int_reso = get_supported_reso(abbrev_to_npy_unit(parsed_unit))
+
     if values.descr.type_num != cnp.NPY_OBJECT:
         # raise here otherwise we segfault below
         raise TypeError("array_to_timedelta64 'values' must have object dtype")
@@ -470,9 +478,19 @@ def array_to_timedelta64(
                     creso = state.creso
                 ival = delta_to_nanoseconds(item, reso=creso)
 
-            elif is_integer_object(item) or is_float_object(item):
-                td64ns_obj = _numeric_to_td64ns(item, parsed_unit)
-                ival = cnp.get_timedelta64_value(td64ns_obj)
+            elif is_integer_object(item):
+                if item == NPY_NAT:
+                    ival = NPY_NAT
+                else:
+                    ival = _numeric_to_td64ns(item, parsed_unit, int_reso)
+                    item_reso = int_reso
+
+                    state.update_creso(item_reso)
+                    if infer_reso:
+                        creso = state.creso
+
+            elif is_float_object(item):
+                ival = _numeric_to_td64ns(item, parsed_unit, NPY_FR_ns)
 
                 item_reso = NPY_FR_ns
                 state.update_creso(item_reso)
@@ -1022,9 +1040,23 @@ cdef _timedelta_from_value_and_reso(cls, int64_t value, NPY_DATETIMEUNIT reso):
     elif reso == NPY_DATETIMEUNIT.NPY_FR_us:
         td_base = _Timedelta.__new__(cls, microseconds=int(value))
     elif reso == NPY_DATETIMEUNIT.NPY_FR_ms:
-        td_base = _Timedelta.__new__(cls, milliseconds=0)
+        if value > -86_399_999_913_600_000 and value < 86_400_000_000_000_000:
+            # i.e. we are in range for pytimedelta. By passing the
+            #  'correct' value here we can
+            #   make pydatetime + Timedelta operations work correctly,
+            #  xref GH#53643
+            td_base = _Timedelta.__new__(cls, milliseconds=value)
+        else:
+            td_base = _Timedelta.__new__(cls, milliseconds=0)
     elif reso == NPY_DATETIMEUNIT.NPY_FR_s:
-        td_base = _Timedelta.__new__(cls, seconds=0)
+        if value > -86_399_999_913_600 and value < 86_400_000_000_000:
+            # i.e. we are in range for pytimedelta. By passing the
+            #  'correct' value here we can
+            #   make pydatetime + Timedelta operations work correctly,
+            #  xref GH#53643
+            td_base = _Timedelta.__new__(cls, seconds=value)
+        else:
+            td_base = _Timedelta.__new__(cls, seconds=0)
     # Other resolutions are disabled but could potentially be implemented here:
     # elif reso == NPY_DATETIMEUNIT.NPY_FR_m:
     #    td_base = _Timedelta.__new__(Timedelta, minutes=int(value))
@@ -2216,7 +2248,18 @@ class Timedelta(_Timedelta):
         elif checknull_with_nat_and_na(value):
             return NaT
 
-        elif is_integer_object(value) or is_float_object(value):
+        elif is_integer_object(value):
+            # unit=None is de-facto 'ns'
+            if value != NPY_NAT:
+                unit = parse_timedelta_unit(unit)
+                if unit != "ns":
+                    # Return with the closest-to-supported unit by going through
+                    #  the timedelta64 path
+                    td = np.timedelta64(value, unit)
+                    return cls(td)
+                value = _numeric_to_td64ns(value, unit)
+
+        elif is_float_object(value):
             # unit=None is de-facto 'ns'
             unit = parse_timedelta_unit(unit)
             value = _numeric_to_td64ns(value, unit)

pandas/core/arrays/datetimelike.py

@@ -1656,7 +1656,7 @@ def mean(self, *, skipna: bool = True, axis: AxisInt | None = 0):
         >>> tdelta_idx = pd.to_timedelta([1, 2, 3], unit="D")
         >>> tdelta_idx
         TimedeltaIndex(['1 days', '2 days', '3 days'],
-                        dtype='timedelta64[ns]', freq=None)
+                        dtype='timedelta64[s]', freq=None)
         >>> tdelta_idx.mean()
         Timedelta('2 days 00:00:00')
         """

pandas/core/arrays/timedeltas.py

@@ -847,7 +847,7 @@ def to_pytimedelta(self) -> npt.NDArray[np.object_]:
         >>> tdelta_idx = pd.to_timedelta([1, 2, 3], unit="D")
         >>> tdelta_idx
         TimedeltaIndex(['1 days', '2 days', '3 days'],
-                        dtype='timedelta64[ns]', freq=None)
+                        dtype='timedelta64[s]', freq=None)
         >>> tdelta_idx.to_pytimedelta()
         array([datetime.timedelta(days=1), datetime.timedelta(days=2),
                datetime.timedelta(days=3)], dtype=object)
@@ -880,7 +880,7 @@ def to_pytimedelta(self) -> npt.NDArray[np.object_]:
     0   1 days
     1   2 days
     2   3 days
-    dtype: timedelta64[ns]
+    dtype: timedelta64[s]
     >>> ser.dt.days
     0    1
     1    2
@@ -915,7 +915,7 @@ def to_pytimedelta(self) -> npt.NDArray[np.object_]:
     0   0 days 00:00:01
     1   0 days 00:00:02
     2   0 days 00:00:03
-    dtype: timedelta64[ns]
+    dtype: timedelta64[s]
     >>> ser.dt.seconds
     0    1
     1    2
@@ -927,7 +927,7 @@ def to_pytimedelta(self) -> npt.NDArray[np.object_]:
     >>> tdelta_idx = pd.to_timedelta([1, 2, 3], unit='s')
     >>> tdelta_idx
     TimedeltaIndex(['0 days 00:00:01', '0 days 00:00:02', '0 days 00:00:03'],
-                   dtype='timedelta64[ns]', freq=None)
+                   dtype='timedelta64[s]', freq=None)
     >>> tdelta_idx.seconds
     Index([1, 2, 3], dtype='int32')"""
     )
@@ -955,7 +955,7 @@ def to_pytimedelta(self) -> npt.NDArray[np.object_]:
     0   0 days 00:00:00.000001
     1   0 days 00:00:00.000002
     2   0 days 00:00:00.000003
-    dtype: timedelta64[ns]
+    dtype: timedelta64[us]
     >>> ser.dt.microseconds
     0    1
     1    2
@@ -968,7 +968,7 @@ def to_pytimedelta(self) -> npt.NDArray[np.object_]:
     >>> tdelta_idx
     TimedeltaIndex(['0 days 00:00:00.000001', '0 days 00:00:00.000002',
                     '0 days 00:00:00.000003'],
-                   dtype='timedelta64[ns]', freq=None)
+                   dtype='timedelta64[us]', freq=None)
     >>> tdelta_idx.microseconds
     Index([1, 2, 3], dtype='int32')"""
     )
@@ -1205,7 +1205,9 @@ def _ints_to_td64ns(data, unit: str = "ns") -> tuple[np.ndarray, bool]:
         dtype_str = f"timedelta64[{unit}]"
         data = data.view(dtype_str)
 
-        data = astype_overflowsafe(data, dtype=TD64NS_DTYPE)
+        new_dtype = get_supported_dtype(data.dtype)
+        if new_dtype != data.dtype:
+            data = astype_overflowsafe(data, dtype=new_dtype)
 
         # the astype conversion makes a copy, so we can avoid re-copying later
         copy_made = True

pandas/core/indexes/accessors.py

@@ -489,7 +489,7 @@ def to_pytimedelta(self) -> np.ndarray:
         2   2 days
         3   3 days
         4   4 days
-        dtype: timedelta64[ns]
+        dtype: timedelta64[s]
 
         >>> s.dt.to_pytimedelta()
         array([datetime.timedelta(0), datetime.timedelta(days=1),
@@ -535,7 +535,7 @@ def components(self) -> DataFrame:
         2   0 days 00:00:02
         3   0 days 00:00:03
         4   0 days 00:00:04
-        dtype: timedelta64[ns]
+        dtype: timedelta64[s]
         >>> s.dt.components
            days  hours  minutes  seconds  milliseconds  microseconds  nanoseconds
         0     0      0        0        0             0             0            0

pandas/core/indexes/datetimelike.py

@@ -132,7 +132,7 @@ def mean(self, *, skipna: bool = True, axis: int | None = 0):
         >>> tdelta_idx = pd.to_timedelta([1, 2, 3], unit="D")
         >>> tdelta_idx
         TimedeltaIndex(['1 days', '2 days', '3 days'],
-                        dtype='timedelta64[ns]', freq=None)
+                        dtype='timedelta64[s]', freq=None)
         >>> tdelta_idx.mean()
         Timedelta('2 days 00:00:00')
         """

pandas/core/tools/timedeltas.py

@@ -171,10 +171,10 @@ def to_timedelta(
     >>> pd.to_timedelta(np.arange(5), unit="s")
     TimedeltaIndex(['0 days 00:00:00', '0 days 00:00:01', '0 days 00:00:02',
                     '0 days 00:00:03', '0 days 00:00:04'],
-                   dtype='timedelta64[ns]', freq=None)
+                   dtype='timedelta64[s]', freq=None)
     >>> pd.to_timedelta(np.arange(5), unit="D")
     TimedeltaIndex(['0 days', '1 days', '2 days', '3 days', '4 days'],
-                   dtype='timedelta64[ns]', freq=None)
+                   dtype='timedelta64[s]', freq=None)
     """
     if unit is not None:
         unit = parse_timedelta_unit(unit)

pandas/tests/arithmetic/test_timedelta64.py

@@ -10,7 +10,6 @@
 
 from pandas._libs.tslibs import timezones
 from pandas.compat import WASM
-from pandas.errors import OutOfBoundsDatetime
 import pandas.util._test_decorators as td
 
 import pandas as pd
@@ -728,17 +727,14 @@ def test_tdi_add_overflow(self):
         # See GH#14068
         # preliminary test scalar analogue of vectorized tests below
         # TODO: Make raised error message more informative and test
-        with pytest.raises(OutOfBoundsDatetime, match="10155196800000000000"):
-            pd.to_timedelta(106580, "D") + Timestamp("2000")
-        with pytest.raises(OutOfBoundsDatetime, match="10155196800000000000"):
-            Timestamp("2000") + pd.to_timedelta(106580, "D")
 
         _NaT = NaT._value + 1
+        td = pd.to_timedelta([106580], "D").as_unit("ns")
         msg = "Overflow in int64 addition"
         with pytest.raises(OverflowError, match=msg):
-            pd.to_timedelta([106580], "D") + Timestamp("2000")
+            td + Timestamp("2000")
         with pytest.raises(OverflowError, match=msg):
-            Timestamp("2000") + pd.to_timedelta([106580], "D")
+            Timestamp("2000") + td
         with pytest.raises(OverflowError, match=msg):
             pd.to_timedelta([_NaT]) - Timedelta("1 days")
         with pytest.raises(OverflowError, match=msg):

pandas/tests/frame/indexing/test_mask.py

@@ -127,14 +127,15 @@ def test_mask_where_dtype_timedelta():
     # https://github.com/pandas-dev/pandas/issues/39548
     df = DataFrame([Timedelta(i, unit="D") for i in range(5)])
 
-    expected = DataFrame(np.full(5, np.nan, dtype="timedelta64[ns]"))
+    expected = DataFrame(np.full(5, np.nan, dtype="timedelta64[s]"))
     tm.assert_frame_equal(df.mask(df.notna()), expected)
 
     expected = DataFrame(
         [np.nan, np.nan, np.nan, Timedelta("3 day"), Timedelta("4 day")],
-        dtype="m8[ns]",
+        dtype="m8[s]",
     )
-    tm.assert_frame_equal(df.where(df > Timedelta(2, unit="D")), expected)
+    result = df.where(df > Timedelta(2, unit="D"))
+    tm.assert_frame_equal(result, expected)
 
 
 def test_mask_return_dtype():

pandas/tests/frame/indexing/test_setitem.py

@@ -1000,8 +1000,8 @@ def test_loc_expansion_with_timedelta_type(self):
             index=Index([0]),
             columns=(["a", "b", "c"]),
         )
-        expected["a"] = expected["a"].astype("m8[ns]")
-        expected["b"] = expected["b"].astype("m8[ns]")
+        expected["a"] = expected["a"].astype("m8[s]")
+        expected["b"] = expected["b"].astype("m8[s]")
         tm.assert_frame_equal(result, expected)
 
     def test_setitem_tuple_key_in_empty_frame(self):

pandas/tests/frame/test_constructors.py

@@ -893,7 +893,7 @@ def create_data(constructor):
         [
             (lambda x: np.timedelta64(x, "D"), "m8[s]"),
             (lambda x: timedelta(days=x), "m8[us]"),
-            (lambda x: Timedelta(x, "D"), "m8[ns]"),
+            (lambda x: Timedelta(x, "D"), "m8[s]"),
             (lambda x: Timedelta(x, "D").as_unit("s"), "m8[s]"),
         ],
     )

pandas/tests/groupby/methods/test_quantile.py

@@ -363,15 +363,14 @@ def test_groupby_quantile_allNA_column(dtype):
 
 def test_groupby_timedelta_quantile():
     # GH: 29485
-    df = DataFrame(
-        {"value": pd.to_timedelta(np.arange(4), unit="s"), "group": [1, 1, 2, 2]}
-    )
+    tdi = pd.to_timedelta(np.arange(4), unit="s").as_unit("us")
+    df = DataFrame({"value": tdi, "group": [1, 1, 2, 2]})
     result = df.groupby("group").quantile(0.99)
     expected = DataFrame(
         {
             "value": [
-                pd.Timedelta("0 days 00:00:00.990000").as_unit("ns"),
-                pd.Timedelta("0 days 00:00:02.990000").as_unit("ns"),
+                pd.Timedelta("0 days 00:00:00.990000"),
+                pd.Timedelta("0 days 00:00:02.990000"),
             ]
         },
         index=Index([1, 2], name="group"),

pandas/tests/indexes/timedeltas/methods/test_shift.py

@@ -49,7 +49,7 @@ def test_tdi_shift_int(self):
                 "5 days 01:00:00",
             ],
             freq="D",
-            dtype="m8[ns]",
+            dtype="m8[s]",
         )
         tm.assert_index_equal(result, expected)
 
@@ -67,7 +67,7 @@ def test_tdi_shift_nonstandard_freq(self):
                 "10 days 01:00:03",
             ],
             freq="D",
-            dtype="m8[ns]",
+            dtype="m8[s]",
         )
         tm.assert_index_equal(result, expected)
 

pandas/tests/indexes/timedeltas/test_constructors.py

@@ -267,4 +267,5 @@ def test_unit_deprecated(self, unit, unit_depr):
 
         with tm.assert_produces_warning(Pandas4Warning, match=msg):
             tdi = to_timedelta([1, 2], unit=unit_depr)
-        tm.assert_index_equal(tdi, expected.as_unit("ns"))
+        exp_unit = unit if unit in ["s", "ms", "us"] else "s"
+        tm.assert_index_equal(tdi, expected.as_unit(exp_unit))