Re:Announcing Spark 1.6
by rxin on Monday January 04, 2016 @04:22PM
(#10838740)
Despite our attempts at warning people, a lot of users still use groupByKey in RDDs. Hopefully over time this won't be a problem as the engine should be able to figure out more intelligently and do the proper rewrite (of course, we won't be able to do it 100%).
[ Parent
| on HN ]
Many people blindly point to the docs to say "don't use groupBy, prefer reduce because it's faster..." Are there better examples that illustrate the fundamental differences between the two operations? Surely there is still a need for both operations
Re:Announcing Spark 1.6
by IvanVergiliev on Monday January 04, 2016 @05:55PM
(#10839398)
Reduce can perform reductions on locally on each machine before shuffling the data. This decreases the memory as well as the network overhead.
If you need all the elements for a given key - e.g. to display them to a user or save them to a DB, perhaps you should use groupBy. If you're going to perform some form of a reduce after that though, it's likely sub-optimal.
databricks has a page that describes the pitfalls: https://databricks.gitbooks.io/databricks-spark-knowledge-ba...
I don't know if the OutOfMemory exception can still occur in recent versions of Spark, but the performance impact of groupByKey is very real.