Github user sethah commented on the issue: https://github.com/apache/spark/pull/19232 Sure, we all agree there is a mechanism for avoiding overhead. However, performance tests are very tricky things, 5% is not a huge improvement, and hard-coding the aggregation depth to `2` limits the utility of using `treeAggregate`. I think the change is probably fine for just the reason that `treeAggregate` shouldn't hurt performance and might speed things up. Still, I don't think there's enough information yet to determine under what circumstances this actually improves the performance, if any.
--- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org