I'm curious about this too--clj-spark didn't work for me so I'm currently prototyping a job using the Java bindings in Clojure. If I end up wrapping the Java bindings in a useful way I would consider putting together some kind of release if there's community interest in that.
However, shouldn't there be a much more flexible approach where you just send your functions to an execution server (just like an agent)? You might want to define some keywords to refer to previously used functions or data.
Then again, such an agent would be pretty much a REPL, so you might just want to ssh to a REPL that does load balancing and has sub-repls (on other machines) that fail over.