BUG: fix initialization of Series with dict containing NaN as key

closes pandas-dev#18480

Author: toobaz

doc/source/whatsnew/v0.22.0.txt

@@ -205,4 +205,5 @@ Other
 
 - Improved error message when attempting to use a Python keyword as an identifier in a numexpr query (:issue:`18221`)
 - Fixed a bug where creating a Series from an array that contains both tz-naive and tz-aware values will result in a Series whose dtype is tz-aware instead of object (:issue:`16406`)
+- Fixed construction of ``Series`` from ``dict`` containing ``NaN`` as key (:issue:`18480`)
 -

pandas/core/indexes/base.py

@@ -2847,8 +2847,33 @@ def get_indexer_for(self, target, **kwargs):
 
     @Appender(_index_shared_docs['_get_values_from_dict'])
     def _get_values_from_dict(self, data):
-        return lib.fast_multiget(data, self.values,
-                                 default=np.nan)
+        # MultiIndex doesn't support isna()
+        if not data or (not hasattr(self, 'levels') and
+                        not isna(self.values).any()):
+            # lib.fast_multiget doesn't support arbitrary dict subclasses
+            if not data.__class__ is dict:
+                data = dict(data)
+            return lib.fast_multiget(data, self.values,
+                                     default=np.nan)
+        else:
+            # Looking for NaN in dict doesn't work ({np.nan : 1}[float('nan')]
+            # raises KeyError), so we iterate the entire dict, and align
+            keys, values = zip(*compat.iteritems(data))
+            keys = Index(keys)
+            values = np.array(values, dtype='object')
+            if not self.equals(keys):
+                locs = self.get_indexer(keys)
+                order = - np.ones(len(self), dtype=int)
+                order[locs] = np.arange(len(keys))
+                values = values[order]
+
+                nan_idxs = np.where(order == -1)[0]
+                if len(nan_idxs):
+                    if is_integer_dtype(data):
+                        values = values.astype(float)
+                    values[nan_idxs] = np.nan
+            return lib.maybe_convert_objects(values)
+
 
     def _maybe_promote(self, other):
         # A hack, but it works

pandas/tests/series/test_constructors.py

@@ -625,6 +625,26 @@ def test_constructor_dict(self):
         expected.iloc[1] = 1
         assert_series_equal(result, expected)
 
+    def test_constructor_dict_nan_key(self):
+        # GH 18480
+        d = {1: 'a', 2: 'b', np.nan: 'c'}
+        result = Series(d).sort_index()
+        expected = Series(['a', 'b', 'c'], index=[1, 2, np.nan])
+        assert_series_equal(result, expected)
+
+        # Different NaNs:
+        d = {1: 'a', 2: 'b', float('nan'): 'c', float('nan'): 'd'}
+        result = Series(d).sort_values()
+        expected = Series(['a', 'b', 'c', 'd'], index=[1, 2, np.nan, np.nan])
+        assert_series_equal(result, expected)
+
+        # MultiIndex:
+        d = {(1, 1): 'a', (2, 2): 'b', (3, np.nan): 'c'}
+        result = Series(d).sort_values()
+        expected = Series(['a', 'b', 'c'],
+                          index=Index([(1, 1), (2, 2), (3, np.nan)]))
+        assert_series_equal(result, expected)
+
     def test_constructor_dict_datetime64_index(self):
         # GH 9456