DEPR: datetimelike.astype(int_other_than_i8) return requested dtype (pandas-dev#45574)

Author: jbrockmendel

doc/source/whatsnew/v1.5.0.rst

@@ -238,6 +238,7 @@ Other Deprecations
 - Deprecated behavior of :meth:`DatetimeIndex.intersection` and :meth:`DatetimeIndex.symmetric_difference` (``union`` behavior was already deprecated in version 1.3.0) with mixed time zones; in a future version both will be cast to UTC instead of object dtype (:issue:`39328`, :issue:`45357`)
 - Deprecated :meth:`DataFrame.iteritems`, :meth:`Series.iteritems`, :meth:`HDFStore.iteritems` in favor of :meth:`DataFrame.items`, :meth:`Series.items`, :meth:`HDFStore.items`  (:issue:`45321`)
 - Deprecated :meth:`Series.is_monotonic` and :meth:`Index.is_monotonic` in favor of :meth:`Series.is_monotonic_increasing` and :meth:`Index.is_monotonic_increasing` (:issue:`45422`, :issue:`21335`)
+- Deprecated behavior of :meth:`DatetimeIndex.astype`, :meth:`TimedeltaIndex.astype`, :meth:`PeriodIndex.astype` when converting to an integer dtype other than ``int64``. In a future version, these will convert to exactly the specified dtype (instead of always ``int64``) and will raise if the conversion overflows (:issue:`45034`)
 - Deprecated the ``__array_wrap__`` method of DataFrame and Series, rely on standard numpy ufuncs instead (:issue:`45451`)
 - Deprecated the behavior of :meth:`Series.fillna` and :meth:`DataFrame.fillna` with ``timedelta64[ns]`` dtype and incompatible fill value; in a future version this will cast to a common dtype (usually object) instead of raising, matching the behavior of other dtypes (:issue:`45746`)
 - Deprecated the ``warn`` parameter in :func:`infer_freq` (:issue:`45947`)

pandas/core/arrays/datetimelike.py

@@ -446,6 +446,36 @@ def astype(self, dtype, copy: bool = True):
             if is_unsigned_integer_dtype(dtype):
                 # Again, we ignore int32 vs. int64
                 values = values.view("uint64")
+                if dtype != np.uint64:
+                    # GH#45034
+                    warnings.warn(
+                        f"The behavior of .astype from {self.dtype} to {dtype} is "
+                        "deprecated. In a future version, this astype will return "
+                        "exactly the specified dtype instead of uint64, and will "
+                        "raise if that conversion overflows.",
+                        FutureWarning,
+                        stacklevel=find_stack_level(),
+                    )
+                elif (self.asi8 < 0).any():
+                    # GH#45034
+                    warnings.warn(
+                        f"The behavior of .astype from {self.dtype} to {dtype} is "
+                        "deprecated. In a future version, this astype will "
+                        "raise if the conversion overflows, as it did in this "
+                        "case with negative int64 values.",
+                        FutureWarning,
+                        stacklevel=find_stack_level(),
+                    )
+            elif dtype != np.int64:
+                # GH#45034
+                warnings.warn(
+                    f"The behavior of .astype from {self.dtype} to {dtype} is "
+                    "deprecated. In a future version, this astype will return "
+                    "exactly the specified dtype instead of int64, and will "
+                    "raise if that conversion overflows.",
+                    FutureWarning,
+                    stacklevel=find_stack_level(),
+                )
 
             if copy:
                 values = values.copy()

pandas/tests/arrays/period/test_astype.py

@@ -9,18 +9,29 @@
 
 
 @pytest.mark.parametrize("dtype", [int, np.int32, np.int64, "uint32", "uint64"])
-def test_astype(dtype):
+def test_astype_int(dtype):
     # We choose to ignore the sign and size of integers for
     # Period/Datetime/Timedelta astype
     arr = period_array(["2000", "2001", None], freq="D")
-    result = arr.astype(dtype)
 
     if np.dtype(dtype).kind == "u":
         expected_dtype = np.dtype("uint64")
+        warn1 = FutureWarning
     else:
         expected_dtype = np.dtype("int64")
-
-    expected = arr.astype(expected_dtype)
+        warn1 = None
+
+    msg_overflow = "will raise if the conversion overflows"
+    with tm.assert_produces_warning(warn1, match=msg_overflow):
+        expected = arr.astype(expected_dtype)
+
+    warn = None if dtype == expected_dtype else FutureWarning
+    msg = " will return exactly the specified dtype"
+    if warn is None and warn1 is not None:
+        warn = warn1
+        msg = msg_overflow
+    with tm.assert_produces_warning(warn, match=msg):
+        result = arr.astype(dtype)
 
     assert result.dtype == expected_dtype
     tm.assert_numpy_array_equal(result, expected)

pandas/tests/arrays/test_datetimes.py

@@ -77,14 +77,20 @@ def test_astype_copies(self, dtype, other):
     @pytest.mark.parametrize("dtype", [int, np.int32, np.int64, "uint32", "uint64"])
     def test_astype_int(self, dtype):
         arr = DatetimeArray._from_sequence([pd.Timestamp("2000"), pd.Timestamp("2001")])
-        result = arr.astype(dtype)
 
         if np.dtype(dtype).kind == "u":
             expected_dtype = np.dtype("uint64")
         else:
             expected_dtype = np.dtype("int64")
         expected = arr.astype(expected_dtype)
 
+        warn = None
+        if dtype != expected_dtype:
+            warn = FutureWarning
+        msg = " will return exactly the specified dtype"
+        with tm.assert_produces_warning(warn, match=msg):
+            result = arr.astype(dtype)
+
         assert result.dtype == expected_dtype
         tm.assert_numpy_array_equal(result, expected)
 

pandas/tests/arrays/test_timedeltas.py

@@ -11,14 +11,20 @@ class TestTimedeltaArray:
     @pytest.mark.parametrize("dtype", [int, np.int32, np.int64, "uint32", "uint64"])
     def test_astype_int(self, dtype):
         arr = TimedeltaArray._from_sequence([Timedelta("1H"), Timedelta("2H")])
-        result = arr.astype(dtype)
 
         if np.dtype(dtype).kind == "u":
             expected_dtype = np.dtype("uint64")
         else:
             expected_dtype = np.dtype("int64")
         expected = arr.astype(expected_dtype)
 
+        warn = None
+        if dtype != expected_dtype:
+            warn = FutureWarning
+        msg = " will return exactly the specified dtype"
+        with tm.assert_produces_warning(warn, match=msg):
+            result = arr.astype(dtype)
+
         assert result.dtype == expected_dtype
         tm.assert_numpy_array_equal(result, expected)
 

pandas/tests/indexes/datetimes/methods/test_astype.py

@@ -52,7 +52,11 @@ def test_astype_uint(self):
             name="idx",
         )
         tm.assert_index_equal(arr.astype("uint64"), expected)
-        tm.assert_index_equal(arr.astype("uint32"), expected)
+
+        msg = "will return exactly the specified dtype instead of uint64"
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            res = arr.astype("uint32")
+        tm.assert_index_equal(res, expected)
 
     def test_astype_with_tz(self):
 

pandas/tests/indexes/interval/test_astype.py

@@ -205,10 +205,18 @@ def index(self, request):
     @pytest.mark.parametrize("subtype", ["int64", "uint64"])
     def test_subtype_integer(self, index, subtype):
         dtype = IntervalDtype(subtype, "right")
-        result = index.astype(dtype)
-        expected = IntervalIndex.from_arrays(
-            index.left.astype(subtype), index.right.astype(subtype), closed=index.closed
-        )
+
+        warn = None
+        if index.isna().any() and subtype == "uint64":
+            warn = FutureWarning
+        msg = "In a future version, this astype will raise if the conversion overflows"
+
+        with tm.assert_produces_warning(warn, match=msg):
+            result = index.astype(dtype)
+            new_left = index.left.astype(subtype)
+            new_right = index.right.astype(subtype)
+
+        expected = IntervalIndex.from_arrays(new_left, new_right, closed=index.closed)
         tm.assert_index_equal(result, expected)
 
     def test_subtype_float(self, index):

pandas/tests/indexes/period/methods/test_astype.py

@@ -58,7 +58,11 @@ def test_astype_uint(self):
         arr = period_range("2000", periods=2, name="idx")
         expected = UInt64Index(np.array([10957, 10958], dtype="uint64"), name="idx")
         tm.assert_index_equal(arr.astype("uint64"), expected)
-        tm.assert_index_equal(arr.astype("uint32"), expected)
+
+        msg = "will return exactly the specified dtype instead of uint64"
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            res = arr.astype("uint32")
+        tm.assert_index_equal(res, expected)
 
     def test_astype_object(self):
         idx = PeriodIndex([], freq="M")

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

@@ -79,7 +79,11 @@ def test_astype_uint(self):
             np.array([3600000000000, 90000000000000], dtype="uint64")
         )
         tm.assert_index_equal(arr.astype("uint64"), expected)
-        tm.assert_index_equal(arr.astype("uint32"), expected)
+
+        msg = "will return exactly the specified dtype instead of uint64"
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            res = arr.astype("uint32")
+        tm.assert_index_equal(res, expected)
 
     def test_astype_timedelta64(self):
         # GH 13149, GH 13209