thats an interesting thought about orderBy and mapPartitions. i guess i could emulate a groupBy with secondary sort using those two. however isn't using an orderBy expensive since it is a total sort? i mean a groupBy with secondary sort is also a total sort under the hood, but its on (hashCode(key), secondarySortColumn) which is easier to distribute and therefore can be implemented more efficiently.
On Thu, Nov 3, 2016 at 8:59 PM, Michael Armbrust <mich...@databricks.com> wrote: > It is still unclear to me why we should remember all these tricks (or add >> lots of extra little functions) when this elegantly can be expressed in a >> reduce operation with a simple one line lamba function. >> > I think you can do that too. KeyValueGroupedDataset has a reduceGroups > function. This probably won't be as fast though because you end up > creating objects where as the version I gave will get codgened to operate > on binary data the whole way though. > >> The same applies to these Window functions. I had to read it 3 times to >> understand what it all means. Maybe it makes sense for someone who has been >> forced to use such limited tools in sql for many years but that's not >> necessary what we should aim for. Why can I not just have the sortBy and >> then an Iterator[X] => Iterator[Y] to express what I want to do? >> > We also have orderBy and mapPartitions. > >> All these functions (rank etc.) can be trivially expressed in this, plus >> I can add other operations if needed, instead of being locked in like this >> Window framework. >> > I agree that window functions would probably not be my first choice for > many problems, but for people coming from SQL it was a very popular > feature. My real goal is to give as many paradigms as possible in a single > unified framework. Let people pick the right mode of expression for any > given job :) >