Given that a lot of the current activity in Spark Core is in shuffles, I wanted to propose factoring out shuffle implementations in a way that will make experimentation easier. Ideally we will converge on one implementation, but for a while, this could also be used to have several implementations coexist. I'm suggesting this because I aware of at least three efforts to look at shuffle (from Yahoo!, Intel and Databricks). Some of the things people are investigating are:
- Push-based shuffle where data moves directly from mappers to reducers
- Sorting-based instead of hash-based shuffle, to create fewer files (helps a lot with file handles and memory usage on large shuffles)
- External spilling within a key
- Changing the level of parallelism or even algorithm for downstream stages at runtime based on statistics of the map output (this is a thing we had prototyped in the Shark research project but never merged in core)
I've attached a design doc with a proposed interface. It's not too crazy because the interface between shuffles and the rest of the code is already pretty narrow (just some iterators for reading data and a writer interface for writing it). Bigger changes will be needed in the interaction with DAGScheduler and BlockManager for some of the ideas above, but we can handle those separately, and this interface will allow us to experiment with some short-term stuff sooner.
If things go well I'd also like to send a sort-based shuffle implementation for 1.1, but we'll see how the timing on that works out.