The short answer: count(), as the sum can be partially aggregated on the mappers.
The long answer: http://databricks.gitbooks.io/databricks-spark-knowledge-base/content/best_practices/dont_call_collect_on_a_very_large_rdd.html — FG On Thu, Feb 26, 2015 at 2:28 PM, Emre Sevinc <emre.sev...@gmail.com> wrote: > Hello, > I have a piece of code to force the materialization of RDDs in my Spark > Streaming program, and I'm trying to understand which method is faster and > has less memory consumption: > javaDStream.foreachRDD(new Function<JavaRDD<String>, Void>() { > @Override > public Void call(JavaRDD<String> stringJavaRDD) throws Exception { > //stringJavaRDD.collect(); > // or count? > //stringJavaRDD.count(); > return null; > } > }); > I've checked the source code of Spark at > https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/rdd/RDD.scala, > and see that collect() is defined as: > def collect(): Array[T] = { > val results = sc.runJob(this, (iter: Iterator[T]) => iter.toArray) > Array.concat(results: _*) > } > and count() defined as: > def count(): Long = sc.runJob(this, Utils.getIteratorSize _).sum > Therefore I think calling the count() method is faster and/or consumes less > memory, but I wanted to be sure. > Anyone cares to comment? > -- > Emre Sevinç > http://www.bigindustries.be/