It's a question about GAE behind the scene.
ok, it can be a silly question, but I want figure out how google can support geographic distributed datacenters and what strategies used in memcached in this case.
In a logic view perspective, it a "one" (and consistent) memcached server by app? Let me explain: one of most knowed cases to use memcached is a "Page View Count" sample. Well, if memcached is distributed/replicated (that is, the key:value will be distributed and replicated in n+1 machines) I can't guarantee a consistent value (or something like this), because the count value will be divide by n+1.
Another case that a distributed/replicated memcached can result in troubles: I would use some not updated entity value. ex: step 1. req webserver1+memcached1 (do some queries and cache entity1 in memcached1) step 2. req webserver2+memcached2 (update entity1 and invalidate it in memcached2) step 3. req webserver1+memcached1 (get entity1 from memcached1, but it's a old data, because it was updated in datastore, invalidated in memcached2 but NOT in memcached1!!)