Github user shivaram commented on a diff in the pull request:
https://github.com/apache/spark/pull/7461#discussion_r38497342
--- Diff: core/src/main/scala/org/apache/spark/rdd/RDD.scala ---
@@ -1080,6 +1080,21 @@ abstract class RDD[T: ClassTag](
var partiallyAggregated = mapPartitions(it =>
Iterator(aggregatePartition(it)))
var numPartitions = partiallyAggregated.partitions.length
val scale = math.max(math.ceil(math.pow(numPartitions, 1.0 /
depth)).toInt, 2)
+
+ // Do one level of aggregation based on executorId before starting
the tree
+ // NOTE: exclude the driver from list of executors
+ val numExecutors = math.max(context.getExecutorStorageStatus.length
- 1, 1)
+ partiallyAggregated = partiallyAggregated.mapPartitionsWithIndex {
case (idx, iter) =>
+ def isAllDigits(x: String) = x forall Character.isDigit
+ val execId = SparkEnv.get.executorId
+ if (isAllDigits(execId)) {
+ iter.map((execId.toInt, _))
+ } else {
+ iter.map((execId.hashCode, _))
--- End diff --
Ah I can actually reproduce it -- so the problem is that the keyspace get
un-balanced when we use hashCode on Strings. For example if you start with 100
exec ids and want to get 50 partitions out, some partitions will get 3 ids
while some will get 1
```
(0 until 100).map(_.toString).map(_.hashCode %
50).groupBy(identity).map(_._2.size)
```
while with integers you get a uniform distribution
```
(0 until 100).map(_.hashCode % 50).groupBy(identity).map(_._2.size)
```
So you could end up with unbalanced trees which is not desirable.
@andrewor14 Can we guarantee that execId will be an integer ? Is there any
other integer we can get a handle on otherwise ?
---
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 [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
