Couldn't this mitigated by user-agents introducing random, Poisson distributed delays in all cached responses? The peak of the distribution could be made user configurable to make it further difficult to predict a user-agent.