spark-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Reynold Xin <r...@databricks.com>
Subject Re: Implementing rdd.scanLeft()
Date Thu, 05 Jun 2014 17:47:47 GMT
I think the main concern is this would require scanning the data twice, and
maybe the user should be aware of it ...


On Thu, Jun 5, 2014 at 10:29 AM, Andrew Ash <andrew@andrewash.com> wrote:

> I have a use case that would greatly benefit from RDDs having a .scanLeft()
> method.  Are the project developers interested in adding this to the public
> API?
>
>
> Looking through past message traffic, this has come up a few times.  The
> recommendation from the list before has been to implement a parallel prefix
> scan.
>
> http://comments.gmane.org/gmane.comp.lang.scala.spark.user/1880
> https://groups.google.com/forum/#!topic/spark-users/ts-FdB50ltY
>
> The algorithm Reynold sketched in the first link leads to this working
> implementation:
>
> val vector = sc.parallelize(1 to 20, 3)
>
> val sums = 0 +: vector.mapPartitionsWithIndex{ case(partition, iter) =>
> Iterator(iter.sum) }.collect.scanLeft(0)(_+_).drop(1)
>
> val prefixScan = vector.mapPartitionsWithIndex { case(partition, iter) =>
>   val base = sums(partition)
>   println(partition, base)
>   iter.scanLeft(base)(_+_).drop(1)
> }.collect
>
>
> I'd love to have that replaced with this:
>
> val vector = sc.parallelize(1 to 20, 3)
> val cumSum: RDD[Int] = vector.scanLeft(0)(_+_)
>
>
> Any thoughts on whether this contribution would be accepted?  What pitfalls
> exist that I should be thinking about?
>
> Thanks!
> Andrew
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message