So, let's say I've hit machine A in a cluster of 50 machines. What chance is there of me hitting machine J in the same cluster before the session data is replicated to that machine? .1 seconds? .01 seconds? I still have a hard time believing that *all* memory data (figure 500 users at 1k of session data each = 500k) can be replicated across a few dozen machines without there ever being the chance of data not being in sync. It seems like a lot of gyrations to do to get around a central database (which would also help survive power cuts, and which also gives an easier central place to do analysis of user session data from).
BTW, the reason the *nuke clones are not scaleable is less to do with hitting the database on every request as it is that they don't know when to use the database and when not to.