Joe Gregorio's article on sharding counters highlights an example of the counter(heh ;)intuitive techniques that work well on applications with large data sets. "Large" meaning "too big to fit into RAM on a single machine."
One of the reasons sharding works so well is that accessing another computer's RAM over a fast network is orders of magnitude faster than accessing your own disk. (memcached also takes advantage of these constraints)
Here's a presentation with some more details and examples of sharding in appengine:
App Engine Google