reduceByKey or reduceByKeyLocally which should be preferred

Question

Below are the def 

def reduceByKey(partitioner: Partitioner, func: (V, V) ⇒ V): RDD[(K, V)]

def reduceByKeyLocally(func: (V, V) ⇒ V): Map[K, V]

Both are almost similar

Answer 1

Yes, they both merge the values using an associative reduce function. reduceByKeyLocally returns the result to Master as a Map.

Now talking from a project perspective, reduceByKey the data is distributed among the cluster as it is represented as RDD. reduceByKeyLocally merges all the output to a Single Master (machine) as a Map. This completely defeats the usage of a distributed data, which is necessary while working at a large scale.