fix: prevent exponential memory growth in UnionArray (#3119)

Author: jpivarski

src/awkward/contents/unionarray.py

@@ -431,7 +431,7 @@ def simplified(
             ]
 
         if len(contents) == 1:
-            next = contents[0]._carry(index, True)
+            next = contents[0]._carry(index, False)
             return next.copy(parameters=parameters_union(next._parameters, parameters))
 
         else:
@@ -702,7 +702,7 @@ def project(self, index):
         nextcarry = ak.index.Index64(
             tmpcarry.data[: lenout[0]], nplike=self._backend.index_nplike
         )
-        return self._contents[index]._carry(nextcarry, True)
+        return self._contents[index]._carry(nextcarry, False)
 
     @staticmethod
     def regular_index(

tests/test_2713_from_buffers_allow_noncanonical.py

@@ -120,9 +120,10 @@ def test_union_simplification():
     projected = ak.from_buffers(
         projected_form, length, container, allow_noncanonical_form=True
     )
+
     assert projected.layout.form.to_dict(verbose=False) == {
-        "class": "IndexedArray",
-        "index": "i64",
-        "content": {"class": "RecordArray", "fields": ["x"], "contents": ["int64"]},
+        "class": "RecordArray",
+        "fields": ["x"],
+        "contents": ["int64"],
     }
     assert ak.almost_equal(array[["x"]], projected)

tests/test_3118_prevent_exponential_memory_growth_in_unionarray.py

@@ -0,0 +1,34 @@
+# BSD 3-Clause License; see https://github.com/scikit-hep/awkward/blob/main/LICENSE
+
+from __future__ import annotations
+
+import awkward as ak
+
+
+def test():
+    one_a = ak.Array([{"x": 1, "y": 2}], with_name="T")
+    one_b = ak.Array([{"x": 1, "y": 2}], with_name="T")
+    two_a = ak.Array([{"x": 1, "z": 3}], with_name="T")
+    two_b = ak.Array([{"x": 1, "z": 3}], with_name="T")
+    three = ak.Array([{"x": 4}, {"x": 4}], with_name="T")
+
+    first = ak.zip({"a": one_a, "b": one_b})
+    second = ak.zip({"a": two_a, "b": two_b})
+
+    cat = ak.concatenate([first, second], axis=0)
+
+    cat["another"] = three
+
+    def check(layout):
+        if hasattr(layout, "contents"):
+            for x in layout.contents:
+                check(x)
+        elif hasattr(layout, "content"):
+            check(layout.content)
+        else:
+            assert layout.length <= 2
+
+    for _ in range(5):
+        check(cat.layout)
+
+        cat["another", "w"] = three.x