Github user mridulm commented on the issue: https://github.com/apache/spark/pull/16574 Couple of points : a) Can recomputation be expensive ? Unfortunately, yes if not used properly. For better or for worse, this has been the implementation in spark since early days - pre-0.5; and the costs are known. Particularly given Apache spark's ability to cache/checkpoint data, the assumption is that shuffle is more expensive. This might not hold anymore actually, given improvements since 1.0 - but only redoing benchmarks will give a better picture. b) If we were to do a shuffle for cartesian, I would implement it differently - take a look at how Apache Pig has implemented it for a more efficient way to do it. (Btw, I dont think the impl in the PR actually works, but I have not looked at it in detail).
--- If your project is set up for it, you can reply to this email and have your reply appear on GitHub as well. If your project does not have this feature enabled and wishes so, or if the feature is enabled but not working, please contact infrastructure at infrastruct...@apache.org or file a JIRA ticket with INFRA. --- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org