Github user zsxwing commented on a diff in the pull request:
https://github.com/apache/spark/pull/19763#discussion_r152087573
--- Diff: core/src/main/scala/org/apache/spark/MapOutputTracker.scala ---
@@ -472,16 +475,48 @@ private[spark] class MapOutputTrackerMaster(
shuffleStatuses.get(shuffleId).map(_.findMissingPartitions())
}
+ /**
+ * To equally divide n elements into m buckets, basically each bucket
should have n/m elements,
+ * for the remaining n%m elements, add one more element to the first n%m
buckets each.
+ */
+ def equallyDivide(numElements: Int, numBuckets: Int): Iterator[Seq[Int]]
= {
+ val elementsPerBucket = numElements / numBuckets
+ val remaining = numElements % numBuckets
+ if (remaining == 0) {
+ 0.until(numElements).grouped(elementsPerBucket)
+ } else {
+ val splitPoint = (elementsPerBucket + 1) * remaining
+ 0.to(splitPoint).grouped(elementsPerBucket + 1) ++
--- End diff --
`grouped` is expensive here. I saw it generates Vector rather than `Range`:
```
scala> (1 to 100).grouped(10).foreach(g => println(g.getClass))
class scala.collection.immutable.Vector
class scala.collection.immutable.Vector
class scala.collection.immutable.Vector
class scala.collection.immutable.Vector
class scala.collection.immutable.Vector
class scala.collection.immutable.Vector
class scala.collection.immutable.Vector
class scala.collection.immutable.Vector
class scala.collection.immutable.Vector
class scala.collection.immutable.Vector
```
It means we need to generate all of numbers between 0 and `numElements`.
Could you implement a special `grouped` for Range instead?
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
