The point is that in many cases the operation passed to reduceByKey aggregates data into much smaller size, say + and * for integer. String concatenation doesn’t actually “shrink” data, thus in your case, rdd.reduceByKey(_ ++ _) and rdd.groupByKey suffer similar performance issue. In general, don’t do these unless you have to.
And in Konstantin’s case, I guess he knows what he’s doing. At least we can’t know whether we can help to optimize without further information about the "business logic” is provided. On Aug 7, 2014, at 10:22 PM, chutium <teng....@gmail.com> wrote: > a long time ago, in Spark Summit 2013, Patrick Wendell said in his talk about > performance > (http://spark-summit.org/talk/wendell-understanding-the-performance-of-spark-applications/) > > that, reduceByKey will be more efficient than groupByKey... he mentioned > groupByKey "copies all data over network". > > is that still true? which one should we choice? because actually we can > replace all of groupByKey with reduceByKey > > for example, if we want to use groupByKey on a RDD[ String, String ], to get > a RDD[ String, Seq[String] ], > > we can also do it with reduceByKey: > at first, map RDD[ String, String ] to RDD[ String, Seq[String] ] > then, reduceByKey(_ ++ _) on this RDD[ String, Seq[String] ] > > > > -- > View this message in context: > http://apache-spark-user-list.1001560.n3.nabble.com/reduceByKey-to-get-all-associated-values-tp11645p11652.html > Sent from the Apache Spark User List mailing list archive at Nabble.com. > > --------------------------------------------------------------------- > To unsubscribe, e-mail: user-unsubscr...@spark.apache.org > For additional commands, e-mail: user-h...@spark.apache.org > --------------------------------------------------------------------- To unsubscribe, e-mail: user-unsubscr...@spark.apache.org For additional commands, e-mail: user-h...@spark.apache.org