转自: http://stackoverflow.com/questions/34529953/why-is-the-fold-action-necessary-in-spark
Empty RDD
It cannot be substituted when RDD is empty:
val rdd = sc.emptyRDD[Int]
rdd.reduce(_ + _)
// java.lang.UnsupportedOperationException: empty collection at
// org.apache.spark.rdd.RDD\$\$anonfun$reduce$1\$\$anonfun$apply$ ...
rdd.fold(0)(_ + _)
// Int = 0
You can of course combine reduce with condition on isEmpty but it is rather ugly.
Mutable buffer
Another use case for fold is aggregation with mutable buffer. Consider following RDD:
Lets say we want a sum of all elements. A naive solution is to simply reduce with +:
import breeze.linalg.DenseVector
val rdd = sc.parallelize(Array.fill(100)(DenseVector(1)), 8)
rdd.reduce(_ + _)
Unfortunately it creates a new vector for each element. Since object creation and subsequent garbage collection is expensive it could be
better to use a mutable object. It is not possible with reduce (immutability of RDD doesn't imply immutability of the elements), but can be
achieved with fold as follows:
rdd.fold(DenseVector(0))((acc, x) => acc += x)
Zero element is used here as mutable buffer initialized once per partition leaving actual data untouched.
acc = op(obj, acc), why this operation order is used instead of acc = op(acc, obj) , See SPARK-6416 and SPARK-7683