Having worked on a framework for efficient data resolution for the last couple of months I greatly enjoyed this talk by Simon Marlow from Facebook (
Haxl: Making Concurrency Unreasonably Easy).
The problem space is quite different from ours and the solution isn't viable for us, but it's none the less inspiring.
TL;DR Facebook's Haxi framework for Haskell coupled with a compiler pass allows your data fetching to be cached and parallelised transparently. The compiler converts your Haskell to use Applicative operators, from there on Haxi is allowed to parallelise the data fetches as it desires.
It certainly reduces the level of complexity for such a system quite substantially when you have a language that supports asynchronies calls (not PHP :D)