DEPR: concat ignoring empty objects

doc/source/whatsnew/v2.1.0.rst

@@ -218,22 +218,29 @@ Deprecations
 ~~~~~~~~~~~~
 - Deprecated 'broadcast_axis' keyword in :meth:`Series.align` and :meth:`DataFrame.align`, upcast before calling ``align`` with ``left = DataFrame({col: left for col in right.columns}, index=right.index)`` (:issue:`51856`)
 - Deprecated 'method', 'limit', and 'fill_axis' keywords in :meth:`DataFrame.align` and :meth:`Series.align`, explicitly call ``fillna`` on the alignment results instead (:issue:`51856`)
+- Deprecated :func:`concat` behavior when any of the objects being concatenated have length 0; in the past the dtypes of empty objects were ignored when determining the resulting dtype, in a future version they will not (:issue:`39122`)
 - Deprecated :meth:`.DataFrameGroupBy.apply` and methods on the objects returned by :meth:`.DataFrameGroupBy.resample` operating on the grouping column(s); select the columns to operate on after groupby to either explicitly include or exclude the groupings and avoid the ``FutureWarning`` (:issue:`7155`)
 - Deprecated :meth:`.Groupby.all` and :meth:`.GroupBy.any` with datetime64 or :class:`PeriodDtype` values, matching the :class:`Series` and :class:`DataFrame` deprecations (:issue:`34479`)
 - Deprecated :meth:`Categorical.to_list`, use ``obj.tolist()`` instead (:issue:`51254`)
 - Deprecated :meth:`DataFrame._data` and :meth:`Series._data`, use public APIs instead (:issue:`33333`)
 - Deprecated :meth:`DataFrameGroupBy.dtypes`, check ``dtypes`` on the underlying object instead (:issue:`51045`)
+- Deprecated :meth:`DataFrameGroupBy.dtypes`, check ``dtypes`` on the underlying object instead (:issue:`51045`)
 - Deprecated ``axis=1`` in :meth:`DataFrame.ewm`, :meth:`DataFrame.rolling`, :meth:`DataFrame.expanding`, transpose before calling the method instead (:issue:`51778`)
 - Deprecated ``axis=1`` in :meth:`DataFrame.groupby` and in :class:`Grouper` constructor, do ``frame.T.groupby(...)`` instead (:issue:`51203`)
+- Deprecated ``axis=1`` in :meth:`DataFrame.groupby` and in :class:`Grouper` constructor, do ``frame.T.groupby(...)`` instead (:issue:`51203`)
 - Deprecated accepting slices in :meth:`DataFrame.take`, call ``obj[slicer]`` or pass a sequence of integers instead (:issue:`51539`)
 - Deprecated explicit support for subclassing :class:`Index` (:issue:`45289`)
+- Deprecated explicit support for subclassing :class:`Index` (:issue:`45289`)
 - Deprecated passing a :class:`DataFrame` to :meth:`DataFrame.from_records`, use :meth:`DataFrame.set_index` or :meth:`DataFrame.drop` instead (:issue:`51353`)
 - Deprecated silently dropping unrecognized timezones when parsing strings to datetimes (:issue:`18702`)
 - Deprecated the ``axis`` keyword in :meth:`DataFrame.ewm`, :meth:`Series.ewm`, :meth:`DataFrame.rolling`, :meth:`Series.rolling`, :meth:`DataFrame.expanding`, :meth:`Series.expanding` (:issue:`51778`)
 - Deprecated the ``axis`` keyword in :meth:`DataFrame.resample`, :meth:`Series.resample` (:issue:`51778`)
 - Deprecated the behavior of :func:`concat` with both ``len(keys) != len(objs)``, in a future version this will raise instead of truncating to the shorter of the two sequences (:issue:`43485`)
+- Deprecated the behavior of :func:`concat` with both ``len(keys) != len(objs)``, in a future version this will raise instead of truncating to the shorter of the two sequences (:issue:`43485`)
+- Deprecated the default of ``observed=False`` in :meth:`DataFrame.groupby` and :meth:`Series.groupby`; this will default to ``True`` in a future version (:issue:`43999`)
 - Deprecated the default of ``observed=False`` in :meth:`DataFrame.groupby` and :meth:`Series.groupby`; this will default to ``True`` in a future version (:issue:`43999`)
 - Deprecating pinning ``group.name`` to each group in :meth:`SeriesGroupBy.aggregate` aggregations; if your operation requires utilizing the groupby keys, iterate over the groupby object instead (:issue:`41090`)
+- Deprecating pinning ``group.name`` to each group in :meth:`SeriesGroupBy.aggregate` aggregations; if your operation requires utilizing the groupby keys, iterate over the groupby object instead (:issue:`41090`)
 - Deprecated the 'axis' keyword in :meth:`.GroupBy.idxmax`, :meth:`.GroupBy.idxmin`, :meth:`.GroupBy.fillna`, :meth:`.GroupBy.take`, :meth:`.GroupBy.skew`, :meth:`.GroupBy.rank`, :meth:`.GroupBy.cumprod`, :meth:`.GroupBy.cumsum`, :meth:`.GroupBy.cummax`, :meth:`.GroupBy.cummin`, :meth:`.GroupBy.pct_change`, :meth:`GroupBy.diff`, :meth:`.GroupBy.shift`, and :meth:`DataFrameGroupBy.corrwith`; for ``axis=1`` operate on the underlying :class:`DataFrame` instead (:issue:`50405`, :issue:`51046`)
 - Deprecated :class:`.DataFrameGroupBy` with ``as_index=False`` not including groupings in the result when they are not columns of the DataFrame (:issue:`49519`)
 - Deprecated :func:`is_categorical_dtype`, use ``isinstance(obj.dtype, pd.CategoricalDtype)`` instead (:issue:`52527`)

pandas/core/dtypes/concat.py

@@ -8,24 +8,21 @@
     Sequence,
     cast,
 )
+import warnings
 
 import numpy as np
 
 from pandas._libs import lib
+from pandas.util._exceptions import find_stack_level
 
 from pandas.core.dtypes.astype import astype_array
 from pandas.core.dtypes.cast import (
     common_dtype_categorical_compat,
     find_common_type,
 )
-from pandas.core.dtypes.dtypes import (
-    CategoricalDtype,
-    DatetimeTZDtype,
-    ExtensionDtype,
-)
+from pandas.core.dtypes.dtypes import CategoricalDtype
 from pandas.core.dtypes.generic import (
     ABCCategoricalIndex,
-    ABCExtensionArray,
     ABCSeries,
 )
 
@@ -41,6 +38,9 @@
     )
 
 
+_dtype_obj = np.dtype(object)
+
+
 def _is_nonempty(x, axis) -> bool:
     # filter empty arrays
     # 1-d dtypes always are included here
@@ -100,45 +100,63 @@ def concat_compat(
     # Creating an empty array directly is tempting, but the winnings would be
     # marginal given that it would still require shape & dtype calculation and
     # np.concatenate which has them both implemented is compiled.
+    orig = to_concat
     non_empties = [x for x in to_concat if _is_nonempty(x, axis)]
     if non_empties and axis == 0 and not ea_compat_axis:
         # ea_compat_axis see GH#39574
         to_concat = non_empties
 
-    dtypes = {obj.dtype for obj in to_concat}
-    kinds = {obj.dtype.kind for obj in to_concat}
-    contains_datetime = any(
-        isinstance(dtype, (np.dtype, DatetimeTZDtype)) and dtype.kind in "mM"
-        for dtype in dtypes
-    ) or any(isinstance(obj, ABCExtensionArray) and obj.ndim > 1 for obj in to_concat)
+    any_ea, kinds, target_dtype = _get_result_dtype(to_concat, non_empties)
+
+    if len(to_concat) < len(orig):
+        _, _, alt_dtype = _get_result_dtype(orig, non_empties)
+        if alt_dtype != target_dtype:
+            # GH#39122
+            warnings.warn(
+                "The behavior of array concatenation with empty entries is "
+                "deprecated. In a future version, this will no longer exclude "
+                "empty items when determining the result dtype. "
+                "To retain the old behavior, exclude the empty entries before "
+                "the concat operation.",
+                FutureWarning,
+                stacklevel=find_stack_level(),
+            )
 
-    all_empty = not len(non_empties)
-    single_dtype = len(dtypes) == 1
-    any_ea = any(isinstance(x, ExtensionDtype) for x in dtypes)
+    if target_dtype is not None:
+        to_concat = [astype_array(arr, target_dtype, copy=False) for arr in to_concat]
+
+    if not isinstance(to_concat[0], np.ndarray):
+        # i.e. isinstance(to_concat[0], ExtensionArray)
+        to_concat_eas = cast("Sequence[ExtensionArray]", to_concat)
+        cls = type(to_concat[0])
+        return cls._concat_same_type(to_concat_eas)
+    else:
+        to_concat_arrs = cast("Sequence[np.ndarray]", to_concat)
+        result = np.concatenate(to_concat_arrs, axis=axis)
+
+        if not any_ea and "b" in kinds and result.dtype.kind in "iuf":
+            # GH#39817 cast to object instead of casting bools to numeric
+            result = result.astype(object, copy=False)
+    return result
 
-    if contains_datetime:
-        return _concat_datetime(to_concat, axis=axis)
 
+def _get_result_dtype(to_concat: Sequence[ArrayLike], non_empties: Sequence[ArrayLike]):
+    target_dtype = None
+
+    dtypes = {obj.dtype for obj in to_concat}
+    kinds = {obj.dtype.kind for obj in to_concat}
+
+    any_ea = any(not isinstance(x, np.ndarray) for x in to_concat)
     if any_ea:
+        # i.e. any ExtensionArrays
+
         # we ignore axis here, as internally concatting with EAs is always
         # for axis=0
-        if not single_dtype:
+        if len(dtypes) != 1:
             target_dtype = find_common_type([x.dtype for x in to_concat])
             target_dtype = common_dtype_categorical_compat(to_concat, target_dtype)
-            to_concat = [
-                astype_array(arr, target_dtype, copy=False) for arr in to_concat
-            ]
-
-        if isinstance(to_concat[0], ABCExtensionArray):
-            # TODO: what about EA-backed Index?
-            to_concat_eas = cast("Sequence[ExtensionArray]", to_concat)
-            cls = type(to_concat[0])
-            return cls._concat_same_type(to_concat_eas)
-        else:
-            to_concat_arrs = cast("Sequence[np.ndarray]", to_concat)
-            return np.concatenate(to_concat_arrs)
 
-    elif all_empty:
+    elif not len(non_empties):
         # we have all empties, but may need to coerce the result dtype to
         # object if we have non-numeric type operands (numpy would otherwise
         # cast this to float)
@@ -148,17 +166,16 @@ def concat_compat(
                 pass
             else:
                 # coerce to object
-                to_concat = [x.astype("object") for x in to_concat]
+                target_dtype = np.dtype(object)
                 kinds = {"o"}
+    else:
+        # Argument 1 to "list" has incompatible type "Set[Union[ExtensionDtype,
+        # Any]]"; expected "Iterable[Union[dtype[Any], None, Type[Any],
+        # _SupportsDType[dtype[Any]], str, Tuple[Any, Union[SupportsIndex,
+        # Sequence[SupportsIndex]]], List[Any], _DTypeDict, Tuple[Any, Any]]]"
+        target_dtype = np.find_common_type(list(dtypes), [])  # type: ignore[arg-type]
 
-    # error: Argument 1 to "concatenate" has incompatible type
-    # "Sequence[Union[ExtensionArray, ndarray[Any, Any]]]"; expected
-    # "Union[_SupportsArray[dtype[Any]], _NestedSequence[_SupportsArray[dtype[Any]]]]"
-    result: np.ndarray = np.concatenate(to_concat, axis=axis)  # type: ignore[arg-type]
-    if "b" in kinds and result.dtype.kind in "iuf":
-        # GH#39817 cast to object instead of casting bools to numeric
-        result = result.astype(object, copy=False)
-    return result
+    return any_ea, kinds, target_dtype
 
 
 def union_categoricals(
@@ -320,45 +337,3 @@ def _maybe_unwrap(x):
 
     dtype = CategoricalDtype(categories=categories, ordered=ordered)
     return Categorical._simple_new(new_codes, dtype=dtype)
-
-
-def _concatenate_2d(to_concat: Sequence[np.ndarray], axis: AxisInt) -> np.ndarray:
-    # coerce to 2d if needed & concatenate
-    if axis == 1:
-        to_concat = [np.atleast_2d(x) for x in to_concat]
-    return np.concatenate(to_concat, axis=axis)
-
-
-def _concat_datetime(to_concat: Sequence[ArrayLike], axis: AxisInt = 0) -> ArrayLike:
-    """
-    provide concatenation of an datetimelike array of arrays each of which is a
-    single M8[ns], datetime64[ns, tz] or m8[ns] dtype
-
-    Parameters
-    ----------
-    to_concat : sequence of arrays
-    axis : axis to provide concatenation
-
-    Returns
-    -------
-    a single array, preserving the combined dtypes
-    """
-    from pandas.core.construction import ensure_wrapped_if_datetimelike
-
-    to_concat = [ensure_wrapped_if_datetimelike(x) for x in to_concat]
-
-    single_dtype = lib.dtypes_all_equal([x.dtype for x in to_concat])
-
-    # multiple types, need to coerce to object
-    if not single_dtype:
-        # ensure_wrapped_if_datetimelike ensures that astype(object) wraps
-        #  in Timestamp/Timedelta
-        return _concatenate_2d([x.astype(object) for x in to_concat], axis=axis)
-
-    # error: Unexpected keyword argument "axis" for "_concat_same_type" of
-    # "ExtensionArray"
-    to_concat_eas = cast("list[ExtensionArray]", to_concat)
-    result = type(to_concat_eas[0])._concat_same_type(  # type: ignore[call-arg]
-        to_concat_eas, axis=axis
-    )
-    return result

pandas/core/internals/concat.py

@@ -475,7 +475,9 @@ def is_na(self) -> bool:
 
         values = blk.values
         if values.size == 0:
+            # GH#39122 this case will return False once deprecation is enforced
             return True
+
         if isinstance(values.dtype, SparseDtype):
             return False
 
@@ -494,16 +496,14 @@ def is_na(self) -> bool:
             return all(isna_all(row) for row in values)
 
     @cache_readonly
-    def is_na_without_isna_all(self) -> bool:
+    def is_na_after_size_and_isna_all_deprecation(self) -> bool:
+        """
+        Will self.is_na be True after values.size == 0 deprecation and isna_all
+        deprecation are enforced?
+        """
         blk = self.block
         if blk.dtype.kind == "V":
             return True
-        if not blk._can_hold_na:
-            return False
-
-        values = blk.values
-        if values.size == 0:
-            return True
         return False
 
     def get_reindexed_values(self, empty_dtype: DtypeObj, upcasted_na) -> ArrayLike:
@@ -565,17 +565,16 @@ def _concatenate_join_units(join_units: list[JoinUnit], copy: bool) -> ArrayLike
 
     if empty_dtype != empty_dtype_future:
         if empty_dtype == concat_values.dtype:
-            # GH#40893
+            # GH#39122, GH#40893
             warnings.warn(
-                "The behavior of DataFrame concatenation with all-NA entries is "
-                "deprecated. In a future version, this will no longer exclude "
-                "all-NA columns when determining the result dtypes. "
-                "To retain the old behavior, cast the all-NA columns to the "
-                "desired dtype before the concat operation.",
+                "The behavior of DataFrame concatenation with empty or all-NA "
+                "entries is deprecated. In a future version, this will no longer "
+                "exclude empty or all-NA columns when determining the result dtypes. "
+                "To retain the old behavior, exclude the relevant entries before "
+                "the concat operation.",
                 FutureWarning,
                 stacklevel=find_stack_level(),
             )
-
     return concat_values
 
 
@@ -631,7 +630,9 @@ def _get_empty_dtype(join_units: Sequence[JoinUnit]) -> tuple[DtypeObj, DtypeObj
     dtype_future = dtype
     if len(dtypes) != len(join_units):
         dtypes_future = [
-            unit.block.dtype for unit in join_units if not unit.is_na_without_isna_all
+            unit.block.dtype
+            for unit in join_units
+            if not unit.is_na_after_size_and_isna_all_deprecation
         ]
         if not len(dtypes_future):
             dtypes_future = [

pandas/tests/dtypes/test_concat.py

@@ -12,8 +12,11 @@ def test_concat_mismatched_categoricals_with_empty():
     ser1 = Series(["a", "b", "c"], dtype="category")
     ser2 = Series([], dtype="category")
 
-    result = _concat.concat_compat([ser1._values, ser2._values])
-    expected = pd.concat([ser1, ser2])._values
+    msg = "The behavior of array concatenation with empty entries is deprecated"
+    with tm.assert_produces_warning(FutureWarning, match=msg):
+        result = _concat.concat_compat([ser1._values, ser2._values])
+    with tm.assert_produces_warning(FutureWarning, match=msg):
+        expected = pd.concat([ser1, ser2])._values
     tm.assert_categorical_equal(result, expected)
 
 

pandas/tests/groupby/test_groupby.py

@@ -374,9 +374,13 @@ def f3(x):
 
     df2 = DataFrame({"a": [3, 2, 2, 2], "b": range(4), "c": range(5, 9)})
 
+    depr_msg = "The behavior of array concatenation with empty entries is deprecated"
+
     # correct result
-    result1 = df.groupby("a").apply(f1)
-    result2 = df2.groupby("a").apply(f1)
+    with tm.assert_produces_warning(FutureWarning, match=depr_msg):
+        result1 = df.groupby("a").apply(f1)
+    with tm.assert_produces_warning(FutureWarning, match=depr_msg):
+        result2 = df2.groupby("a").apply(f1)
     tm.assert_frame_equal(result1, result2)
 
     # should fail (not the same number of levels)
@@ -390,7 +394,8 @@ def f3(x):
     with pytest.raises(AssertionError, match=msg):
         df.groupby("a").apply(f3)
     with pytest.raises(AssertionError, match=msg):
-        df2.groupby("a").apply(f3)
+        with tm.assert_produces_warning(FutureWarning, match=depr_msg):
+            df2.groupby("a").apply(f3)
 
 
 def test_attr_wrapper(ts):

pandas/tests/indexes/test_base.py

@@ -616,7 +616,9 @@ def test_append_empty_preserve_name(self, name, expected):
         left = Index([], name="foo")
         right = Index([1, 2, 3], name=name)
 
-        result = left.append(right)
+        msg = "The behavior of array concatenation with empty entries is deprecated"
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            result = left.append(right)
         assert result.name == expected
 
     @pytest.mark.parametrize(

pandas/tests/reshape/concat/test_append.py

@@ -162,7 +162,9 @@ def test_append_preserve_index_name(self):
         df2 = DataFrame(data=[[1, 4, 7], [2, 5, 8], [3, 6, 9]], columns=["A", "B", "C"])
         df2 = df2.set_index(["A"])
 
-        result = df1._append(df2)
+        msg = "The behavior of array concatenation with empty entries is deprecated"
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            result = df1._append(df2)
         assert result.index.name == "A"
 
     indexes_can_append = [

pandas/tests/reshape/concat/test_append_common.py

@@ -693,11 +693,14 @@ def test_concat_categorical_empty(self):
         s1 = Series([], dtype="category")
         s2 = Series([1, 2], dtype="category")
 
-        tm.assert_series_equal(pd.concat([s1, s2], ignore_index=True), s2)
-        tm.assert_series_equal(s1._append(s2, ignore_index=True), s2)
+        msg = "The behavior of array concatenation with empty entries is deprecated"
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            tm.assert_series_equal(pd.concat([s1, s2], ignore_index=True), s2)
+            tm.assert_series_equal(s1._append(s2, ignore_index=True), s2)
 
-        tm.assert_series_equal(pd.concat([s2, s1], ignore_index=True), s2)
-        tm.assert_series_equal(s2._append(s1, ignore_index=True), s2)
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            tm.assert_series_equal(pd.concat([s2, s1], ignore_index=True), s2)
+            tm.assert_series_equal(s2._append(s1, ignore_index=True), s2)
 
         s1 = Series([], dtype="category")
         s2 = Series([], dtype="category")
@@ -719,11 +722,13 @@ def test_concat_categorical_empty(self):
 
         # empty Series is ignored
         exp = Series([np.nan, np.nan])
-        tm.assert_series_equal(pd.concat([s1, s2], ignore_index=True), exp)
-        tm.assert_series_equal(s1._append(s2, ignore_index=True), exp)
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            tm.assert_series_equal(pd.concat([s1, s2], ignore_index=True), exp)
+            tm.assert_series_equal(s1._append(s2, ignore_index=True), exp)
 
-        tm.assert_series_equal(pd.concat([s2, s1], ignore_index=True), exp)
-        tm.assert_series_equal(s2._append(s1, ignore_index=True), exp)
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            tm.assert_series_equal(pd.concat([s2, s1], ignore_index=True), exp)
+            tm.assert_series_equal(s2._append(s1, ignore_index=True), exp)
 
     def test_categorical_concat_append(self):
         cat = Categorical(["a", "b"], categories=["a", "b"])