fix: use public pandas APIs where possible (#60)

* refactor: use public pandas APIs where possible

* no need to override take

* backport take implementation

* move remaining private pandas methods to backports

* add note about _validate_scalar to docstring

* comment why we can't use public mixin

db_dtypes/__init__.py

@@ -22,13 +22,7 @@
 import numpy
 import packaging.version
 import pandas
-import pandas.compat.numpy.function
-import pandas.core.algorithms
-import pandas.core.arrays
-import pandas.core.dtypes.base
-import pandas.core.dtypes.dtypes
-import pandas.core.dtypes.generic
-import pandas.core.nanops
+import pandas.api.extensions
 import pyarrow
 import pyarrow.compute
 
@@ -44,7 +38,7 @@
 pandas_release = packaging.version.parse(pandas.__version__).release
 
 
-@pandas.core.dtypes.dtypes.register_extension_dtype
+@pandas.api.extensions.register_extension_dtype
 class TimeDtype(core.BaseDatetimeDtype):
     """
     Extension dtype for time data.
@@ -113,7 +107,7 @@ def _datetime(
                 .as_py()
             )
 
-        if scalar is None:
+        if pandas.isna(scalar):
             return None
         if isinstance(scalar, datetime.time):
             return pandas.Timestamp(
@@ -194,7 +188,7 @@ def __arrow_array__(self, type=None):
         )
 
 
-@pandas.core.dtypes.dtypes.register_extension_dtype
+@pandas.api.extensions.register_extension_dtype
 class DateDtype(core.BaseDatetimeDtype):
     """
     Extension dtype for time data.
@@ -238,7 +232,7 @@ def _datetime(
         if isinstance(scalar, (pyarrow.Date32Scalar, pyarrow.Date64Scalar)):
             scalar = scalar.as_py()
 
-        if scalar is None:
+        if pandas.isna(scalar):
             return None
         elif isinstance(scalar, datetime.date):
             return pandas.Timestamp(

db_dtypes/core.py

@@ -12,20 +12,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any, Optional, Sequence
+from typing import Optional
 
 import numpy
 import pandas
-from pandas._libs import NaT
+from pandas import NaT
 import pandas.api.extensions
-import pandas.compat.numpy.function
-import pandas.core.algorithms
-import pandas.core.arrays
-import pandas.core.dtypes.base
-from pandas.core.dtypes.common import is_dtype_equal, is_list_like, pandas_dtype
-import pandas.core.dtypes.dtypes
-import pandas.core.dtypes.generic
-import pandas.core.nanops
+from pandas.api.types import is_dtype_equal, is_list_like, pandas_dtype
 
 from db_dtypes import pandas_backports
 
@@ -107,42 +100,11 @@ def isna(self):
         return pandas.isna(self._ndarray)
 
     def _validate_scalar(self, value):
-        if pandas.isna(value):
-            return None
-
-        if not isinstance(value, self.dtype.type):
-            raise ValueError(value)
-
-        return value
-
-    def take(
-        self,
-        indices: Sequence[int],
-        *,
-        allow_fill: bool = False,
-        fill_value: Any = None,
-    ):
-        indices = numpy.asarray(indices, dtype=numpy.intp)
-        data = self._ndarray
-        if allow_fill:
-            fill_value = self._validate_scalar(fill_value)
-            fill_value = (
-                numpy.datetime64() if fill_value is None else self._datetime(fill_value)
-            )
-            if (indices < -1).any():
-                raise ValueError(
-                    "take called with negative indexes other than -1,"
-                    " when a fill value is provided."
-                )
-        out = data.take(indices)
-        if allow_fill:
-            out[indices == -1] = fill_value
-
-        return self.__class__(out)
-
-    # TODO: provide implementations of dropna, fillna, unique,
-    # factorize, argsort, searchsoeted for better performance over
-    # abstract implementations.
+        """
+        Validate and convert a scalar value to datetime64[ns] for storage in
+        backing NumPy array.
+        """
+        return self._datetime(value)
 
     def any(
         self,
@@ -152,10 +114,8 @@ def any(
         keepdims: bool = False,
         skipna: bool = True,
     ):
-        pandas.compat.numpy.function.validate_any(
-            (), {"out": out, "keepdims": keepdims}
-        )
-        result = pandas.core.nanops.nanany(self._ndarray, axis=axis, skipna=skipna)
+        pandas_backports.numpy_validate_any((), {"out": out, "keepdims": keepdims})
+        result = pandas_backports.nanany(self._ndarray, axis=axis, skipna=skipna)
         return result
 
     def all(
@@ -166,22 +126,20 @@ def all(
         keepdims: bool = False,
         skipna: bool = True,
     ):
-        pandas.compat.numpy.function.validate_all(
-            (), {"out": out, "keepdims": keepdims}
-        )
-        result = pandas.core.nanops.nanall(self._ndarray, axis=axis, skipna=skipna)
+        pandas_backports.numpy_validate_all((), {"out": out, "keepdims": keepdims})
+        result = pandas_backports.nanall(self._ndarray, axis=axis, skipna=skipna)
         return result
 
     def min(self, *, axis: Optional[int] = None, skipna: bool = True, **kwargs):
-        pandas.compat.numpy.function.validate_min((), kwargs)
-        result = pandas.core.nanops.nanmin(
+        pandas_backports.numpy_validate_min((), kwargs)
+        result = pandas_backports.nanmin(
             values=self._ndarray, axis=axis, mask=self.isna(), skipna=skipna
         )
         return self._box_func(result)
 
     def max(self, *, axis: Optional[int] = None, skipna: bool = True, **kwargs):
-        pandas.compat.numpy.function.validate_max((), kwargs)
-        result = pandas.core.nanops.nanmax(
+        pandas_backports.numpy_validate_max((), kwargs)
+        result = pandas_backports.nanmax(
             values=self._ndarray, axis=axis, mask=self.isna(), skipna=skipna
         )
         return self._box_func(result)
@@ -197,11 +155,9 @@ def median(
             keepdims: bool = False,
             skipna: bool = True,
         ):
-            pandas.compat.numpy.function.validate_median(
+            pandas_backports.numpy_validate_median(
                 (),
                 {"out": out, "overwrite_input": overwrite_input, "keepdims": keepdims},
             )
-            result = pandas.core.nanops.nanmedian(
-                self._ndarray, axis=axis, skipna=skipna
-            )
+            result = pandas_backports.nanmedian(self._ndarray, axis=axis, skipna=skipna)
             return self._box_func(result)

db_dtypes/pandas_backports.py

@@ -20,15 +20,32 @@
 """
 
 import operator
+from typing import Any
 
 import numpy
 import packaging.version
 import pandas
-from pandas._libs.lib import is_integer
+from pandas.api.types import is_integer
+import pandas.compat.numpy.function
+import pandas.core.nanops
 
 
 pandas_release = packaging.version.parse(pandas.__version__).release
 
+# Create aliases for private methods in case they move in a future version.
+nanall = pandas.core.nanops.nanall
+nanany = pandas.core.nanops.nanany
+nanmax = pandas.core.nanops.nanmax
+nanmin = pandas.core.nanops.nanmin
+numpy_validate_all = pandas.compat.numpy.function.validate_all
+numpy_validate_any = pandas.compat.numpy.function.validate_any
+numpy_validate_max = pandas.compat.numpy.function.validate_max
+numpy_validate_min = pandas.compat.numpy.function.validate_min
+
+if pandas_release >= (1, 2):
+    nanmedian = pandas.core.nanops.nanmedian
+    numpy_validate_median = pandas.compat.numpy.function.validate_median
+
 
 def import_default(module_name, force=False, default=None):
     """
@@ -55,6 +72,10 @@ def import_default(module_name, force=False, default=None):
     return getattr(module, name, default)
 
 
+# pandas.core.arraylike.OpsMixin is private, but the related public API
+# "ExtensionScalarOpsMixin" is not sufficient for adding dates to times.
+# It results in unsupported operand type(s) for +: 'datetime.time' and
+# 'datetime.date'
 @import_default("pandas.core.arraylike")
 class OpsMixin:
     def _cmp_method(self, other, op):  # pragma: NO COVER
@@ -81,6 +102,8 @@ def __ge__(self, other):
     __add__ = __radd__ = __sub__ = lambda self, other: NotImplemented
 
 
+# TODO: use public API once pandas 1.5 / 2.x is released.
+# See: https://github.com/pandas-dev/pandas/pull/45544
 @import_default("pandas.core.arrays._mixins", pandas_release < (1, 3))
 class NDArrayBackedExtensionArray(pandas.core.arrays.base.ExtensionArray):
 
@@ -130,6 +153,28 @@ def copy(self):
     def repeat(self, n):
         return self.__class__(self._ndarray.repeat(n), self._dtype)
 
+    def take(
+        self,
+        indices,
+        *,
+        allow_fill: bool = False,
+        fill_value: Any = None,
+        axis: int = 0,
+    ):
+        from pandas.core.algorithms import take
+
+        if allow_fill:
+            fill_value = self._validate_scalar(fill_value)
+
+        new_data = take(
+            self._ndarray,
+            indices,
+            allow_fill=allow_fill,
+            fill_value=fill_value,
+            axis=axis,
+        )
+        return self._from_backing_data(new_data)
+
     @classmethod
     def _concat_same_type(cls, to_concat, axis=0):
         dtypes = {str(x.dtype) for x in to_concat}