BUG: DataFrame from dict with non-nano Timedelta

pandas/_libs/tslibs/timedeltas.pxd

@@ -22,6 +22,7 @@ cdef class _Timedelta(timedelta):
 
     cpdef timedelta to_pytimedelta(_Timedelta self)
     cdef bint _has_ns(self)
+    cdef bint _is_in_pytimedelta_bounds(self)
     cdef _ensure_components(_Timedelta self)
     cdef inline bint _compare_mismatched_resos(self, _Timedelta other, op)
     cdef _Timedelta _as_creso(self, NPY_DATETIMEUNIT reso, bint round_ok=*)

pandas/_libs/tslibs/timedeltas.pyx

@@ -1093,8 +1093,27 @@ cdef class _Timedelta(timedelta):
             #  non-invariant behavior.
             #  see GH#44504
             return hash(self.value)
-        else:
+        elif self._is_in_pytimedelta_bounds() and (
+            self._creso == NPY_FR_ns or self._creso == NPY_DATETIMEUNIT.NPY_FR_us
+        ):
+            # If we can defer to timedelta.__hash__, do so, as that
+            #  ensures the hash is invariant to our _reso.
+            # We can only defer for ns and us, as for these two resos we
+            #  call _Timedelta.__new__ with the correct input in
+            #  _timedelta_from_value_and_reso; so timedelta.__hash__
+            #  will be correct
             return timedelta.__hash__(self)
+        else:
+            # We want to ensure that two equivalent Timedelta objects
+            #  have the same hash.  So we try downcasting to the next-lowest
+            #  resolution.
+            try:
+                obj = (<_Timedelta>self)._as_creso(<NPY_DATETIMEUNIT>(self._creso + 1))
+            except OverflowError:
+                # Doesn't fit, so we're off the hook
+                return hash(self.value)
+            else:
+                return hash(obj)
 
     def __richcmp__(_Timedelta self, object other, int op):
         cdef:
@@ -1152,6 +1171,13 @@ cdef class _Timedelta(timedelta):
         else:
             raise NotImplementedError(self._creso)
 
+    cdef bint _is_in_pytimedelta_bounds(self):
+        """
+        Check if we are within the bounds of datetime.timedelta.
+        """
+        self._ensure_components()
+        return -999999999 <= self._d and self._d <= 999999999
+
     cdef _ensure_components(_Timedelta self):
         """
         compute the components

pandas/tests/frame/test_constructors.py

@@ -846,30 +846,19 @@ def create_data(constructor):
         tm.assert_frame_equal(result_Timestamp, expected)
 
     @pytest.mark.parametrize(
-        "klass",
+        "klass,name",
         [
-            pytest.param(
-                np.timedelta64,
-                marks=pytest.mark.xfail(
-                    reason="hash mismatch (GH#44504) causes lib.fast_multiget "
-                    "to mess up on dict lookups with equal Timedeltas with "
-                    "mismatched resos"
-                ),
-            ),
-            timedelta,
-            Timedelta,
+            (lambda x: np.timedelta64(x, "D"), "timedelta64"),
+            (lambda x: timedelta(days=x), "pytimedelta"),
+            (lambda x: Timedelta(x, "D"), "Timedelta[ns]"),
+            (lambda x: Timedelta(x, "D").as_unit("s"), "Timedelta[s]"),
         ],
     )
-    def test_constructor_dict_timedelta64_index(self, klass):
+    def test_constructor_dict_timedelta64_index(self, klass, name):
         # GH 10160
         td_as_int = [1, 2, 3, 4]
 
-        if klass is timedelta:
-            constructor = lambda x: timedelta(days=x)
-        else:
-            constructor = lambda x: klass(x, "D")
-
-        data = {i: {constructor(s): 2 * i} for i, s in enumerate(td_as_int)}
+        data = {i: {klass(s): 2 * i} for i, s in enumerate(td_as_int)}
 
         expected = DataFrame(
             [
@@ -881,7 +870,13 @@ def test_constructor_dict_timedelta64_index(self, klass):
             index=[Timedelta(td, "D") for td in td_as_int],
         )
 
-        result = DataFrame(data)
+        if name == "Timedelta[s]":
+            # TODO(2.0): passing index here shouldn't be necessary, is for now
+            #  otherwise we raise in _extract_index
+            result = DataFrame(data, index=expected.index)
+        else:
+            result = DataFrame(data)
+
         tm.assert_frame_equal(result, expected)
 
     def test_constructor_period_dict(self):

pandas/tests/libs/test_lib.py

@@ -2,6 +2,7 @@
 import pytest
 
 from pandas._libs import (
+    Timedelta,
     lib,
     writers as libwriters,
 )
@@ -42,6 +43,34 @@ def test_fast_unique_multiple_list_gen_sort(self):
         out = lib.fast_unique_multiple_list_gen(gen, sort=False)
         tm.assert_numpy_array_equal(np.array(out), expected)
 
+    def test_fast_multiget_timedelta_resos(self):
+        # This will become relevant for test_constructor_dict_timedelta64_index
+        #  once Timedelta constructor preserves reso when passed a
+        #  np.timedelta64 object
+        td = Timedelta(days=1)
+
+        mapping1 = {td: 1}
+        mapping2 = {td.as_unit("s"): 1}
+
+        oindex = Index([td * n for n in range(3)])._values.astype(object)
+
+        expected = lib.fast_multiget(mapping1, oindex)
+        result = lib.fast_multiget(mapping2, oindex)
+        tm.assert_numpy_array_equal(result, expected)
+
+        # case that can't be cast to td64ns
+        td = Timedelta(np.timedelta64(400, "Y"))
+        assert hash(td) == hash(td.as_unit("ms"))
+        assert hash(td) == hash(td.as_unit("us"))
+        mapping1 = {td: 1}
+        mapping2 = {td.as_unit("ms"): 1}
+
+        oindex = Index([td * n for n in range(3)])._values.astype(object)
+
+        expected = lib.fast_multiget(mapping1, oindex)
+        result = lib.fast_multiget(mapping2, oindex)
+        tm.assert_numpy_array_equal(result, expected)
+
 
 class TestIndexing:
     def test_maybe_indices_to_slice_left_edge(self):