PERF: MultiIndex.argsort / MultiIndex.sort_values (#48406)

* perf improvement in MultiIndex.argsort and MultiIndex.sort_values

* whatsnew

* add asv for Int64

Author: lukemanley

asv_bench/benchmarks/multiindex_object.py

@@ -8,6 +8,7 @@
     MultiIndex,
     RangeIndex,
     Series,
+    array,
     date_range,
 )
 
@@ -176,6 +177,20 @@ def time_sortlevel_one(self):
         self.mi.sortlevel(1)
 
 
+class SortValues:
+
+    params = ["int64", "Int64"]
+    param_names = ["dtype"]
+
+    def setup(self, dtype):
+        a = array(np.tile(np.arange(100), 1000), dtype=dtype)
+        b = array(np.tile(np.arange(1000), 100), dtype=dtype)
+        self.mi = MultiIndex.from_arrays([a, b])
+
+    def time_sort_values(self, dtype):
+        self.mi.sort_values()
+
+
 class Values:
     def setup_cache(self):
 

doc/source/whatsnew/v1.6.0.rst

@@ -101,6 +101,7 @@ Deprecations
 Performance improvements
 ~~~~~~~~~~~~~~~~~~~~~~~~
 - Performance improvement in :meth:`.GroupBy.median` for nullable dtypes (:issue:`37493`)
+- Performance improvement in :meth:`MultiIndex.argsort` and :meth:`MultiIndex.sort_values` (:issue:`48406`)
 - Performance improvement in :meth:`.GroupBy.mean` and :meth:`.GroupBy.var` for extension array dtypes (:issue:`37493`)
 - Performance improvement for :meth:`MultiIndex.unique` (:issue:`48335`)
 -

pandas/core/indexes/multi.py

@@ -2225,6 +2225,10 @@ def append(self, other):
             return Index._with_infer(new_tuples)
 
     def argsort(self, *args, **kwargs) -> npt.NDArray[np.intp]:
+        if len(args) == 0 and len(kwargs) == 0:
+            # np.lexsort is significantly faster than self._values.argsort()
+            values = [self._get_level_values(i) for i in reversed(range(self.nlevels))]
+            return np.lexsort(values)
         return self._values.argsort(*args, **kwargs)
 
     @Appender(_index_shared_docs["repeat"] % _index_doc_kwargs)

pandas/tests/indexes/multi/test_setops.py

@@ -530,11 +530,6 @@ def test_union_duplicates(index, request):
     if index.empty or isinstance(index, (IntervalIndex, CategoricalIndex)):
         # No duplicates in empty indexes
         return
-    if index.dtype.kind == "c":
-        mark = pytest.mark.xfail(
-            reason="sort_values() call raises bc complex objects are not comparable"
-        )
-        request.node.add_marker(mark)
 
     values = index.unique().values.tolist()
     mi1 = MultiIndex.from_arrays([values, [1] * len(values)])