Fixed #36562

* Use special sorting comparator for tuple arrays which can be created when consolidate_first is called on DataFrames with MultiIndex which contain nan and string values

Author: ssche

doc/source/whatsnew/v1.2.0.rst

@@ -363,7 +363,8 @@ MultiIndex
 ^^^^^^^^^^
 
 - Bug in :meth:`DataFrame.xs` when used with :class:`IndexSlice` raises ``TypeError`` with message ``"Expected label or tuple of labels"`` (:issue:`35301`)
--
+- Bug in :meth:`DataFrame.combine_first` when used with :class:`MultiIndex` containing string and ``NaN`` values raises ``TypeError`` with message ``"'<' not supported between instances of 'float' and 'str'"`` (:issue:`36562`)
+
 
 I/O
 ^^^

pandas/core/algorithms.py

@@ -5,6 +5,7 @@
 from __future__ import annotations
 
 import operator
+import functools
 from textwrap import dedent
 from typing import TYPE_CHECKING, Dict, Optional, Tuple, Union, cast
 from warnings import catch_warnings, simplefilter, warn
@@ -2055,13 +2056,52 @@ def sort_mixed(values):
         strs = np.sort(values[str_pos])
         return np.concatenate([nums, np.asarray(strs, dtype=object)])
 
+    def sort_tuples(values):
+        # sorts tuples with mixed values. can handle nan vs string comparisons.
+        def cmp_func(index_x, index_y):
+            x = values[index_x]
+            y = values[index_y]
+            if x == y:
+                return 0
+            len_x = len(x)
+            len_y = len(y)
+            for i in range(max(len_x, len_y)):
+                # check if the tuples have different lengths (shorter tuples
+                # first)
+                if i >= len_x:
+                    return -1
+                if i >= len_y:
+                    return +1
+                x_i_na = isna(x[i])
+                y_i_na = isna(y[i])
+                # values are the same -> resolve tie with next element
+                if (x_i_na and y_i_na) or (x[i] == y[i]):
+                    continue
+                # check for nan values (sort nan to the end which is consistent
+                # with numpy
+                if x_i_na and not y_i_na:
+                    return +1
+                if not x_i_na and y_i_na:
+                    return -1
+                # normal greater/less than comparison
+                if x[i] < y[i]:
+                    return -1
+                return +1
+            return 0
+
+        ixs = np.arange(len(values))
+        ixs = sorted(ixs, key=functools.cmp_to_key(cmp_func))
+        return values[ixs]
+
     sorter = None
-    if (
-        not is_extension_array_dtype(values)
-        and lib.infer_dtype(values, skipna=False) == "mixed-integer"
-    ):
+
+    ext_arr = is_extension_array_dtype(values)
+    if not ext_arr and lib.infer_dtype(values, skipna=False) == "mixed-integer":
         # unorderable in py3 if mixed str/int
         ordered = sort_mixed(values)
+    elif not ext_arr and values.size and isinstance(values[0], tuple):
+        # 1-D arrays with tuples of potentially mixed type (solves GH36562)
+        ordered = sort_tuples(values)
     else:
         try:
             sorter = values.argsort()

pandas/tests/indexing/multiindex/test_multiindex.py

@@ -91,3 +91,26 @@ def test_multiindex_get_loc_list_raises(self):
         msg = "unhashable type"
         with pytest.raises(TypeError, match=msg):
             idx.get_loc([])
+
+
+def test_combine_first_with_nan_index():
+    mi1 = pd.MultiIndex.from_arrays(
+        [["b", "b", "c", "a", "b", np.nan], [1, 2, 3, 4, 5, 6]],
+        names=["a", "b"]
+    )
+    df = pd.DataFrame({"c": [1, 1, 1, 1, 1, 1]}, index=mi1)
+    mi2 = pd.MultiIndex.from_arrays(
+        [["a", "b", "c", "a", "b", "d"], [1, 1, 1, 1, 1, 1]], names=["a", "b"]
+    )
+    s = pd.Series([1, 2, 3, 4, 5, 6], index=mi2)
+    df_combined = df.combine_first(pd.DataFrame({"col": s}))
+    mi_expected = pd.MultiIndex.from_arrays([
+        ["a", "a", "a", "b", "b", "b", "b", "c", "c", "d", np.nan],
+        [1, 1, 4, 1, 1, 2, 5, 1, 3, 1, 6,]
+    ], names=["a", "b"])
+    assert (df_combined.index == mi_expected).all()
+    exp_col = np.asarray(
+        [1.0, 4.0, np.nan, 2.0, 5.0, np.nan, np.nan, 3.0, np.nan, 6.0, np.nan]
+    )
+    act_col = df_combined['col'].values
+    assert np.allclose(act_col, exp_col, rtol=0, atol=0, equal_nan=True)

pandas/tests/test_sorting.py

@@ -453,3 +453,10 @@ def test_extension_array_codes(self, verify, na_sentinel):
         expected_codes = np.array([0, 2, na_sentinel, 1], dtype=np.intp)
         tm.assert_extension_array_equal(result, expected_values)
         tm.assert_numpy_array_equal(codes, expected_codes)
+
+
+def test_mixed_str_nan():
+    values = np.array(["b", np.nan, "a", "b"], dtype=object)
+    result = safe_sort(values)
+    expected = np.array([np.nan, "a", "b", "b"], dtype=object)
+    tm.assert_numpy_array_equal(result, expected)