API: fix corner case of lib.infer_dtype (pandas-dev#23422)

Author: h-vetinari

doc/source/whatsnew/v0.24.0.txt

@@ -246,6 +246,8 @@ Backwards incompatible API changes
 - :meth:`Series.str.cat` will now raise if `others` is a `set` (:issue:`23009`)
 - The `.str`-accessor will perform more rigorous type checking for inputs. Previously, some types that were never intended to be used
   "worked" purely due to limitations of dtype checking -- e.g. ``bytes``, which is now disabled except for :meth:`Series.str.decode` (:issue:`23011`, :issue:`23163`)
+- The method `pandas._libs.lib.infer_dtype` now returns `'empty'` rather than (sometimes) the dtype of the array,
+  in case the array only consists of missing values and `skipna=True` (:issue:`23421`)
 
 .. _whatsnew_0240.api_breaking.deps:
 

pandas/_libs/lib.pyx

@@ -59,6 +59,7 @@ from tslibs.timezones cimport get_timezone, tz_compare
 
 from missing cimport (checknull,
                       is_null_datetime64, is_null_timedelta64, is_null_period)
+from missing import isnaobj
 
 
 # constants that will be compared to potentially arbitrarily large
@@ -1171,6 +1172,9 @@ def infer_dtype(object value, bint skipna=False):
         values = construct_1d_object_array_from_listlike(value)
 
     values = getattr(values, 'values', values)
+    if skipna:
+        values = values[~isnaobj(values)]
+
     val = _try_infer_map(values)
     if val is not None:
         return val

pandas/tests/dtypes/test_inference.py

@@ -591,6 +591,19 @@ def test_unicode(self):
         expected = 'unicode' if PY2 else 'string'
         assert result == expected
 
+    @pytest.mark.parametrize('dtype, skipna, expected', [
+        (float, False, 'floating'),
+        (float, True, 'floating'),
+        (object, False, 'floating'),
+        (object, True, 'empty')
+    ])
+    def test_object_empty(self, dtype, skipna, expected):
+        # GH 23421
+        arr = pd.Series([np.nan, np.nan], dtype=dtype)
+
+        result = lib.infer_dtype(arr, skipna=skipna)
+        assert result == expected
+
     def test_datetime(self):
 
         dates = [datetime(2012, 1, x) for x in range(1, 20)]