diff --git a/core/src/main/scala/org/apache/spark/rdd/RDD.scala b/core/src/main/scala/org/apache/spark/rdd/RDD.scala
index 032939b49a708..71c733bafb6ff 100644
--- a/core/src/main/scala/org/apache/spark/rdd/RDD.scala
+++ b/core/src/main/scala/org/apache/spark/rdd/RDD.scala
@@ -1061,7 +1061,7 @@ abstract class RDD[T: ClassTag](
    * Aggregates the elements of this RDD in a multi-level tree pattern.
    *
    * @param depth suggested depth of the tree (default: 2)
-   * @see [[org.apache.spark.rdd.RDD#aggregate]]
+   * @see [[org.apache.spark.rdd.RDD#aggregate]] These two methods have identical semantics.
    */
   def treeAggregate[U: ClassTag](zeroValue: U)(
       seqOp: (U, T) => U,
@@ -1075,7 +1075,7 @@ abstract class RDD[T: ClassTag](
       val cleanCombOp = context.clean(combOp)
       val aggregatePartition =
         (it: Iterator[T]) => it.aggregate(zeroValue)(cleanSeqOp, cleanCombOp)
-      var partiallyAggregated = mapPartitions(it => Iterator(aggregatePartition(it)))
+      var partiallyAggregated: RDD[U] = mapPartitions(it => Iterator(aggregatePartition(it)))
       var numPartitions = partiallyAggregated.partitions.length
       val scale = math.max(math.ceil(math.pow(numPartitions, 1.0 / depth)).toInt, 2)
       // If creating an extra level doesn't help reduce
@@ -1087,9 +1087,10 @@ abstract class RDD[T: ClassTag](
         val curNumPartitions = numPartitions
         partiallyAggregated = partiallyAggregated.mapPartitionsWithIndex {
           (i, iter) => iter.map((i % curNumPartitions, _))
-        }.reduceByKey(new HashPartitioner(curNumPartitions), cleanCombOp).values
+        }.foldByKey(zeroValue, new HashPartitioner(curNumPartitions))(cleanCombOp).values
       }
-      partiallyAggregated.reduce(cleanCombOp)
+      val copiedZeroValue = Utils.clone(zeroValue, sc.env.closureSerializer.newInstance())
+      partiallyAggregated.fold(copiedZeroValue)(cleanCombOp)
     }
   }
 
diff --git a/core/src/test/scala/org/apache/spark/rdd/RDDSuite.scala b/core/src/test/scala/org/apache/spark/rdd/RDDSuite.scala
index 24daedab2090f..bff87b8eb663d 100644
--- a/core/src/test/scala/org/apache/spark/rdd/RDDSuite.scala
+++ b/core/src/test/scala/org/apache/spark/rdd/RDDSuite.scala
@@ -159,6 +159,22 @@ class RDDSuite extends SparkFunSuite with SharedSparkContext {
     assert(ser.serialize(union.partitions.head).limit() < 2000)
   }
 
+  test("fold") {
+    val rdd = sc.makeRDD(-1000 until 1000, 10)
+    def op: (Int, Int) => Int = (c: Int, x: Int) => c + x
+    val sum = rdd.fold(0)(op)
+    assert(sum === -1000)
+  }
+
+  test("fold with op modifying first arg") {
+    val rdd = sc.makeRDD(-1000 until 1000, 10).map(x => Array(x))
+    def op: (Array[Int], Array[Int]) => Array[Int] = { (c: Array[Int], x: Array[Int]) =>
+      c(0) += x(0)
+      c
+    }
+    val sum = rdd.fold(Array(0))(op)
+  }
+
   test("aggregate") {
     val pairs = sc.makeRDD(Array(("a", 1), ("b", 2), ("a", 2), ("c", 5), ("a", 3)))
     type StringMap = HashMap[String, Int]
@@ -185,7 +201,19 @@ class RDDSuite extends SparkFunSuite with SharedSparkContext {
     def combOp: (Long, Long) => Long = (c1: Long, c2: Long) => c1 + c2
     for (depth <- 1 until 10) {
       val sum = rdd.treeAggregate(0L)(seqOp, combOp, depth)
-      assert(sum === -1000L)
+      assert(sum === -1000)
+    }
+  }
+
+  test("treeAggregate with ops modifying first args") {
+    val rdd = sc.makeRDD(-1000 until 1000, 10).map(x => Array(x))
+    def op: (Array[Int], Array[Int]) => Array[Int] = { (c: Array[Int], x: Array[Int]) =>
+      c(0) += x(0)
+      c
+    }
+    for (depth <- 1 until 10) {
+      val sum = rdd.treeAggregate(Array(0))(op, op, depth)
+      assert(sum(0) === -1000)
     }
   }