Github user pwendell commented on the pull request:
https://github.com/apache/spark/pull/136#issuecomment-37732233
I don't think we typically run jobs inside of getPartitions - so this
changes some semantics of calling that function. For instance a lot of the
other RDD constructors immediately access the partitions of their parents when
constructed. This would change the lazy evaluation model of Spark.
The difficulties here are coming from the fact that sliding really _isn't_
a parallelizable problem. In the limit case where the partitions size is small
and/or the window size is close to the number of partitions, this effectively
requires all-to-all communication.
I wonder if it's better implemented using a shuffle. This is more expensive
but it might be the only way. Curious what @mateiz, @rxin, @aarondav and others
think on this.
---
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.
---
