The only thing that really matters here is the write load and the distribution of your reads. The read load can easily be handled by a memecache-like cacheing

The only thing that really matters here is the write load and the distribution of your reads. The read load can easily be handled by a memecache-like cacheing layer unless it is completely random and you don't have enough RAM. The write load will determine how fast keys get expired out of the cache layer (plus the associated read load this creates) and what sort of i/o throughput your persistence engine will require. To be completely honest, if memcache, memcachedb, redis, and tokyo cabinet are not getting the job done then you are probably doing things very, very wrong...