Over the weekend I was having a look at Disco, an open source map-reduce framework, built atop Erlang/OTP that allows users to write mapper/reducer jobs in Python. Disco does not mandate any Erlang knowledge on part of the user, who can use all the expressiveness of Python to implement map/reduce jobs.
One more addition to the stack of using Erlang as middleware.
As a programmer, you can concentrate on composing map/reduce jobs using all the expressiveness of Python. Disco master receives jobs from the clients, adds them to the job queue, and makes them run on Erlang powered clusters. Each node of the cluster runs the usual supervisor-worker hierarchies of Erlang processes that fires up the concurrent processing of all client jobs. Server crash does not affect job execution, new servers can be added on the fly and high availability can be ensured through a multiple Disco master configuration.
As I had mentioned before, Erlang as middleware is catching up ..