Thinking out loud. Suppose you wanted to experiment with a P2P version of a centralized service, especially a popular design like del.icio.us or flickr. Maybe you can't stand the uptime of a feed indexing service or want to experiment with some extensions to a social bookmark service. Trust me, this is all completely hypothetical. What's the design and implementation process look like? What guarantees can you make and what do you have to give up? I think, and this is mostly speculation, the majority of the academic P2P work probably relies on at least a few reliable, resilient, super nodes providing a backbone.
Alternatively, how far can you push towards P2P purity in implementing one of these popular, large scale Web services?