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

pandas/_libs/lib.pyx

@@ -57,7 +57,7 @@ from tslibs.conversion cimport convert_to_tsobject
 from tslibs.timedeltas cimport convert_to_timedelta64
 from tslibs.timezones cimport get_timezone, tz_compare
 
-from missing cimport (checknull,
+from missing cimport (checknull, isnaobj,
                       is_null_datetime64, is_null_timedelta64, is_null_period)
 
 
@@ -1181,6 +1181,9 @@ def infer_dtype(value: object, skipna: bool=False) -> str:
         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/_libs/missing.pxd

@@ -1,7 +1,10 @@
 # -*- coding: utf-8 -*-
 
+from numpy cimport ndarray, uint8_t
+
 cpdef bint checknull(object val)
 cpdef bint checknull_old(object val)
+cpdef ndarray[uint8_t] isnaobj(ndarray arr)
 
 cdef bint is_null_datetime64(v)
 cdef bint is_null_timedelta64(v)

pandas/_libs/missing.pyx

@@ -124,7 +124,7 @@ cdef inline bint _check_none_nan_inf_neginf(object val):
 
 @cython.wraparound(False)
 @cython.boundscheck(False)
-def isnaobj(ndarray arr):
+cpdef ndarray[uint8_t] isnaobj(ndarray arr):
     """
     Return boolean mask denoting which elements of a 1-D array are na-like,
     according to the criteria defined in `_check_all_nulls`:

pandas/tests/dtypes/test_inference.py

@@ -591,6 +591,22 @@ def test_unicode(self):
         expected = 'unicode' if PY2 else 'string'
         assert result == expected
 
+    @pytest.mark.parametrize('dtype, missing, skipna, expected', [
+        (float, np.nan, False, 'floating'),
+        (float, np.nan, True, 'floating'),
+        (object, np.nan, False, 'floating'),
+        (object, np.nan, True, 'empty'),
+        (object, None, False, 'mixed'),
+        (object, None, True, 'empty')
+    ])
+    @pytest.mark.parametrize('box', [pd.Series, np.array])
+    def test_object_empty(self, box, missing, dtype, skipna, expected):
+        # GH 23421
+        arr = box([missing, missing], 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)]