Querying is indeed the real issue. I've between reading up on delay tolerant networking; it seems pretty tough to hit the right balance between making metadata generally available and crushing the network with metadata updates. Centralized directories make it easier, though still not trivial, to make content findable on a network.

There have between a bunch of proposals (eg, grokster) out there which have started off as radically decentralized, but eventually took on a layer or three of super peers to manage that metadata transfer, after the fully decentralized model failed under real traffic.

(Good datasets also seem to be a problem here. The available ones are pretty small or have a lot of caveats, which makes it tough to really see how a solution will play out is the real world.)