DEPR: concat ignoring all-NA columns (#52613)

* DEPR: concat ignoring all-NA columns

* silence warning

* use code-block

* fix duplicate whatsnew entry

* remove duplicate whatsnew entry

* Fix duplicates in whatsnew

Author: jbrockmendel

doc/source/whatsnew/v1.4.0.rst

@@ -279,11 +279,11 @@ if one of the DataFrames was empty or had all-NA values, its dtype was
 *sometimes* ignored when finding the concatenated dtype.  These are now
 consistently *not* ignored (:issue:`43507`).
 
-.. ipython:: python
+.. code-block:: ipython
 
-    df1 = pd.DataFrame({"bar": [pd.Timestamp("2013-01-01")]}, index=range(1))
-    df2 = pd.DataFrame({"bar": np.nan}, index=range(1, 2))
-    res = pd.concat([df1, df2])
+    In [3]: df1 = pd.DataFrame({"bar": [pd.Timestamp("2013-01-01")]}, index=range(1))
+    In [4]: df2 = pd.DataFrame({"bar": np.nan}, index=range(1, 2))
+    In [5]: res = pd.concat([df1, df2])
 
 Previously, the float-dtype in ``df2`` would be ignored so the result dtype
 would be ``datetime64[ns]``. As a result, the ``np.nan`` would be cast to
@@ -293,8 +293,8 @@ would be ``datetime64[ns]``. As a result, the ``np.nan`` would be cast to
 
 .. code-block:: ipython
 
-    In [4]: res
-    Out[4]:
+    In [6]: res
+    Out[6]:
              bar
     0 2013-01-01
     1        NaT
@@ -306,8 +306,8 @@ object, the ``np.nan`` is retained.
 
 .. code-block:: ipython
 
-    In [4]: res
-    Out[4]:
+    In [6]: res
+    Out[6]:
                        bar
     0  2013-01-01 00:00:00
     1                  NaN

doc/source/whatsnew/v2.1.0.rst

@@ -233,6 +233,7 @@ Deprecations
 - Deprecated :meth:`DataFrame.applymap`. Use the new :meth:`DataFrame.map` method instead (:issue:`52353`)
 - Deprecated :meth:`DataFrame.swapaxes` and :meth:`Series.swapaxes`, use :meth:`DataFrame.transpose` or :meth:`Series.transpose` instead (:issue:`51946`)
 - Deprecated ``freq`` parameter in :class:`PeriodArray` constructor, pass ``dtype`` instead (:issue:`52462`)
+- Deprecated behavior of :func:`concat` when :class:`DataFrame` has columns that are all-NA, in a future version these will not be discarded when determining the resulting dtype (:issue:`40893`)
 - Deprecated behavior of :meth:`Series.dt.to_pydatetime`, in a future version this will return a :class:`Series` containing python ``datetime`` objects instead of an ``ndarray`` of datetimes; this matches the behavior of other :meth:`Series.dt` properties (:issue:`20306`)
 - Deprecated logical operations (``|``, ``&``, ``^``) between pandas objects and dtype-less sequences (e.g. ``list``, ``tuple``), wrap a sequence in a :class:`Series` or numpy array before operating instead (:issue:`51521`)
 - Deprecated making :meth:`Series.apply` return a :class:`DataFrame` when the passed-in callable returns a :class:`Series` object. In the future this will return a :class:`Series` whose values are themselves :class:`Series`. This pattern was very slow and it's recommended to use alternative methods to archive the same goal (:issue:`52116`)

pandas/core/internals/concat.py

@@ -5,6 +5,7 @@
     TYPE_CHECKING,
     Sequence,
 )
+import warnings
 
 import numpy as np
 
@@ -15,6 +16,7 @@
 )
 from pandas._libs.missing import NA
 from pandas.util._decorators import cache_readonly
+from pandas.util._exceptions import find_stack_level
 
 from pandas.core.dtypes.astype import astype_array
 from pandas.core.dtypes.cast import (
@@ -439,6 +441,19 @@ def is_na(self) -> bool:
                 return False
             return all(isna_all(row) for row in values)
 
+    @cache_readonly
+    def is_na_without_isna_all(self) -> bool:
+        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:
         values: ArrayLike
 
@@ -487,7 +502,7 @@ def _concatenate_join_units(join_units: list[JoinUnit], copy: bool) -> ArrayLike
     """
     Concatenate values from several join units along axis=1.
     """
-    empty_dtype = _get_empty_dtype(join_units)
+    empty_dtype, empty_dtype_future = _get_empty_dtype(join_units)
 
     has_none_blocks = any(unit.block.dtype.kind == "V" for unit in join_units)
     upcasted_na = _dtype_to_na_value(empty_dtype, has_none_blocks)
@@ -526,6 +541,19 @@ def _concatenate_join_units(join_units: list[JoinUnit], copy: bool) -> ArrayLike
     else:
         concat_values = concat_compat(to_concat, axis=1)
 
+    if empty_dtype != empty_dtype_future:
+        if empty_dtype == concat_values.dtype:
+            # 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.",
+                FutureWarning,
+                stacklevel=find_stack_level(),
+            )
+
     return concat_values
 
 
@@ -552,7 +580,7 @@ def _dtype_to_na_value(dtype: DtypeObj, has_none_blocks: bool):
     raise NotImplementedError
 
 
-def _get_empty_dtype(join_units: Sequence[JoinUnit]) -> DtypeObj:
+def _get_empty_dtype(join_units: Sequence[JoinUnit]) -> tuple[DtypeObj, DtypeObj]:
     """
     Return dtype and N/A values to use when concatenating specified units.
 
@@ -564,11 +592,11 @@ def _get_empty_dtype(join_units: Sequence[JoinUnit]) -> DtypeObj:
     """
     if len(join_units) == 1:
         blk = join_units[0].block
-        return blk.dtype
+        return blk.dtype, blk.dtype
 
     if lib.dtypes_all_equal([ju.block.dtype for ju in join_units]):
         empty_dtype = join_units[0].block.dtype
-        return empty_dtype
+        return empty_dtype, empty_dtype
 
     has_none_blocks = any(unit.block.dtype.kind == "V" for unit in join_units)
 
@@ -581,7 +609,23 @@ def _get_empty_dtype(join_units: Sequence[JoinUnit]) -> DtypeObj:
     dtype = find_common_type(dtypes)
     if has_none_blocks:
         dtype = ensure_dtype_can_hold_na(dtype)
-    return dtype
+
+    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
+        ]
+        if not len(dtypes_future):
+            dtypes_future = [
+                unit.block.dtype for unit in join_units if unit.block.dtype.kind != "V"
+            ]
+
+        if len(dtypes) != len(dtypes_future):
+            dtype_future = find_common_type(dtypes_future)
+            if has_none_blocks:
+                dtype_future = ensure_dtype_can_hold_na(dtype_future)
+
+    return dtype, dtype_future
 
 
 def _is_uniform_join_units(join_units: list[JoinUnit]) -> bool:

pandas/tests/reshape/concat/test_concat.py

@@ -747,7 +747,9 @@ def test_concat_ignore_empty_object_float(empty_dtype, df_dtype):
     # https://github.com/pandas-dev/pandas/issues/45637
     df = DataFrame({"foo": [1, 2], "bar": [1, 2]}, dtype=df_dtype)
     empty = DataFrame(columns=["foo", "bar"], dtype=empty_dtype)
+
     result = concat([empty, df])
+
     expected = df
     if df_dtype == "int64":
         # TODO what exact behaviour do we want for integer eventually?
@@ -764,14 +766,24 @@ def test_concat_ignore_empty_object_float(empty_dtype, df_dtype):
 def test_concat_ignore_all_na_object_float(empty_dtype, df_dtype):
     df = DataFrame({"foo": [1, 2], "bar": [1, 2]}, dtype=df_dtype)
     empty = DataFrame({"foo": [np.nan], "bar": [np.nan]}, dtype=empty_dtype)
-    result = concat([empty, df], ignore_index=True)
 
     if df_dtype == "int64":
         # TODO what exact behaviour do we want for integer eventually?
         if empty_dtype == "object":
             df_dtype = "object"
         else:
             df_dtype = "float64"
+
+    msg = "The behavior of DataFrame concatenation with all-NA entries"
+    warn = None
+    if empty_dtype != df_dtype and empty_dtype is not None:
+        warn = FutureWarning
+    elif df_dtype == "datetime64[ns]":
+        warn = FutureWarning
+
+    with tm.assert_produces_warning(warn, match=msg):
+        result = concat([empty, df], ignore_index=True)
+
     expected = DataFrame({"foo": [None, 1, 2], "bar": [None, 1, 2]}, dtype=df_dtype)
     tm.assert_frame_equal(result, expected)
 
@@ -782,7 +794,11 @@ def test_concat_ignore_empty_from_reindex():
     df1 = DataFrame({"a": [1], "b": [pd.Timestamp("2012-01-01")]})
     df2 = DataFrame({"a": [2]})
 
-    result = concat([df1, df2.reindex(columns=df1.columns)], ignore_index=True)
+    aligned = df2.reindex(columns=df1.columns)
+
+    msg = "The behavior of DataFrame concatenation with 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]})
     tm.assert_frame_equal(result, expected)