DEPR: concat ignoring empty objects (#52532)

* DEPR: concat with empty objects

* xfail on 32bit

* missing reason

* Fix AM build

* post-merge fixup

* catch more specifically

* un-xfail

* mypy fixup

* update test

* Fix broken test

* remove duplicate whatsnew entries

* remove unused

Author: jbrockmendel

doc/source/whatsnew/v2.1.0.rst

@@ -253,6 +253,7 @@ Deprecations
 - Deprecated 'fill_method' and 'limit' keywords in :meth:`DataFrame.pct_change`, :meth:`Series.pct_change`, :meth:`DataFrameGroupBy.pct_change`, and :meth:`SeriesGroupBy.pct_change`, explicitly call ``ffill`` or ``bfill`` before calling ``pct_change`` instead (:issue:`53491`)
 - 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 'quantile' keyword in :meth:`Rolling.quantile` and :meth:`Expanding.quantile`, renamed as 'q' instead (:issue:`52550`)
+- 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`)

pandas/core/dtypes/concat.py

@@ -7,10 +7,12 @@
     TYPE_CHECKING,
     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 (
@@ -108,6 +110,17 @@ def concat_compat(
 
     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(),
+            )
 
     if target_dtype is not None:
         to_concat = [astype_array(arr, target_dtype, copy=False) for arr in to_concat]

pandas/core/internals/concat.py

@@ -397,7 +397,9 @@ 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
 
@@ -416,16 +418,14 @@ 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:
@@ -487,17 +487,16 @@ 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
 
 
@@ -553,7 +552,9 @@ 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

@@ -387,9 +387,13 @@ 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)
@@ -403,7 +407,8 @@ 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

@@ -613,7 +613,9 @@ 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

@@ -695,11 +695,14 @@ 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")
@@ -721,11 +724,13 @@ 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"])

pandas/tests/reshape/concat/test_concat.py

@@ -756,8 +756,14 @@ def test_concat_ignore_empty_object_float(empty_dtype, df_dtype):
     df = DataFrame({"foo": [1, 2], "bar": [1, 2]}, dtype=df_dtype)
     empty = DataFrame(columns=["foo", "bar"], dtype=empty_dtype)
 
-    result = concat([empty, df])
-
+    msg = "The behavior of DataFrame concatenation with empty or all-NA entries"
+    warn = None
+    if df_dtype == "datetime64[ns]" or (
+        df_dtype == "float64" and empty_dtype != "float64"
+    ):
+        warn = FutureWarning
+    with tm.assert_produces_warning(warn, match=msg):
+        result = concat([empty, df])
     expected = df
     if df_dtype == "int64":
         # TODO what exact behaviour do we want for integer eventually?
@@ -782,7 +788,7 @@ def test_concat_ignore_all_na_object_float(empty_dtype, df_dtype):
         else:
             df_dtype = "float64"
 
-    msg = "The behavior of DataFrame concatenation with all-NA entries"
+    msg = "The behavior of DataFrame concatenation with empty or all-NA entries"
     warn = None
     if empty_dtype != df_dtype and empty_dtype is not None:
         warn = FutureWarning
@@ -804,7 +810,7 @@ def test_concat_ignore_empty_from_reindex():
 
     aligned = df2.reindex(columns=df1.columns)
 
-    msg = "The behavior of DataFrame concatenation with all-NA entries"
+    msg = "The behavior of DataFrame concatenation with empty or all-NA entries"
     with tm.assert_produces_warning(FutureWarning, match=msg):
         result = concat([df1, aligned], ignore_index=True)
     expected = df1 = DataFrame({"a": [1, 2], "b": [pd.Timestamp("2012-01-01"), pd.NaT]})

pandas/tests/reshape/concat/test_datetimes.py

@@ -578,7 +578,9 @@ def test_concat_float_datetime64(using_array_manager):
 
     if not using_array_manager:
         expected = DataFrame({"A": pd.array(["2000"], dtype="datetime64[ns]")})
-        result = concat([df_time, df_float.iloc[:0]])
+        msg = "The behavior of DataFrame concatenation with empty or all-NA entries"
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            result = concat([df_time, df_float.iloc[:0]])
         tm.assert_frame_equal(result, expected)
     else:
         expected = DataFrame({"A": pd.array(["2000"], dtype="datetime64[ns]")}).astype(

pandas/tests/reshape/concat/test_empty.py

@@ -58,7 +58,9 @@ def test_concat_empty_series(self):
 
         s1 = Series([1, 2, 3], name="x")
         s2 = Series(name="y", dtype="float64")
-        res = concat([s1, s2], axis=0)
+        msg = "The behavior of array concatenation with empty entries is deprecated"
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            res = concat([s1, s2], axis=0)
         # name will be reset
         exp = Series([1, 2, 3])
         tm.assert_series_equal(res, exp)
@@ -238,9 +240,11 @@ def test_concat_inner_join_empty(self):
         df_a = DataFrame({"a": [1, 2]}, index=[0, 1], dtype="int64")
         df_expected = DataFrame({"a": []}, index=RangeIndex(0), dtype="int64")
 
-        for how, expected in [("inner", df_expected), ("outer", df_a)]:
-            result = concat([df_a, df_empty], axis=1, join=how)
-            tm.assert_frame_equal(result, expected)
+        result = concat([df_a, df_empty], axis=1, join="inner")
+        tm.assert_frame_equal(result, df_expected)
+
+        result = concat([df_a, df_empty], axis=1, join="outer")
+        tm.assert_frame_equal(result, df_a)
 
     def test_empty_dtype_coerce(self):
         # xref to #12411

pandas/tests/reshape/concat/test_series.py

@@ -40,7 +40,9 @@ def test_concat_empty_and_non_empty_series_regression(self):
         s2 = Series([], dtype=object)
 
         expected = s1
-        result = concat([s1, s2])
+        msg = "The behavior of array concatenation with empty entries is deprecated"
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            result = concat([s1, s2])
         tm.assert_series_equal(result, expected)
 
     def test_concat_series_axis1(self):

pandas/tests/reshape/merge/test_merge.py

@@ -688,8 +688,13 @@ def test_join_append_timedeltas(self, using_array_manager):
             {"d": [datetime(2013, 11, 5, 5, 56)], "t": [timedelta(0, 22500)]}
         )
         df = DataFrame(columns=list("dt"))
-        df = concat([df, d], ignore_index=True)
-        result = concat([df, d], ignore_index=True)
+        msg = "The behavior of DataFrame concatenation with empty or all-NA entries"
+        warn = FutureWarning
+        if using_array_manager:
+            warn = None
+        with tm.assert_produces_warning(warn, match=msg):
+            df = concat([df, d], ignore_index=True)
+            result = concat([df, d], ignore_index=True)
         expected = DataFrame(
             {
                 "d": [datetime(2013, 11, 5, 5, 56), datetime(2013, 11, 5, 5, 56)],

pandas/tests/series/methods/test_combine_first.py

@@ -63,7 +63,9 @@ def test_combine_first(self):
         # corner case
         ser = Series([1.0, 2, 3], index=[0, 1, 2])
         empty = Series([], index=[], dtype=object)
-        result = ser.combine_first(empty)
+        msg = "The behavior of array concatenation with empty entries is deprecated"
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            result = ser.combine_first(empty)
         ser.index = ser.index.astype("O")
         tm.assert_series_equal(ser, result)
 
@@ -110,7 +112,9 @@ def test_combine_first_timezone_series_with_empty_series(self):
         )
         s1 = Series(range(10), index=time_index)
         s2 = Series(index=time_index)
-        result = s1.combine_first(s2)
+        msg = "The behavior of array concatenation with empty entries is deprecated"
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            result = s1.combine_first(s2)
         tm.assert_series_equal(result, s1)
 
     def test_combine_first_preserves_dtype(self):