Improve performance of intersection of MultiIndex

Author: Jean-Francois Zinque

asv_bench/benchmarks/multiindex_object.py

@@ -160,4 +160,30 @@ def time_equals_non_object_index(self):
         self.mi_large_slow.equals(self.idx_non_object)
 
 
+class SetOperations:
+
+    params = [
+        ("monotonic", "non_monotonic"),
+        ("intersection", "union", "symmetric_difference"),
+    ]
+    param_names = ["index_structure", "method"]
+
+    def setup(self, index_structure, method):
+        N = 10 ** 5
+        values = np.arange(N)
+
+        if index_structure == "monotonic":
+            mi = MultiIndex.from_arrays([values, values])
+        else:
+            mi = MultiIndex.from_arrays(
+                [np.concatenate([values[::2], values[1::2]]), values]
+            )
+
+        self.left = mi
+        self.right = mi[:-1]
+
+    def time_operation(self, index_structure, method):
+        getattr(self.left, method)(self.right)
+
+
 from .pandas_vb_common import setup  # noqa: F401 isort:skip

pandas/core/indexes/multi.py

@@ -3314,10 +3314,16 @@ def intersection(self, other, sort=False):
         if self.equals(other):
             return self
 
-        uniq_other = set(other.values)
+        lvals = self._ndarray_values
+        rvals = other._ndarray_values
+
+        if self.is_monotonic and other.is_monotonic:
+            return self._inner_indexer(lvals, rvals)[0]
+
+        other_uniq = set(rvals)
         seen = set()
         uniq_tuples = [
-            x for x in self.values if x in uniq_other and not (x in seen or seen.add(x))
+            x for x in lvals if x in other_uniq and not (x in seen or seen.add(x))
         ]
 
         if sort is None: