Additionally, you hit (in passing) on CAP theory and the rather large implications of choosing A over C, and didn't mention how well AWS or your implementation using AWS handles P (which is a corner case albeit, but when we're talking large scales that makes all corner cases much more common...).
For financial transactions, eventually-consistent is often not acceptable, rather than merely less-than ideal. Similar for inventory management (which often goes hand-in-hand). I worked on a very large project where mongodb had been chosen. Extracting the consistency semantics we needed out of it via the limitations of the protocol it provides was a difficult task requiring very careful planning. There is no perfect solution to HA--trade-offs exist no matter what approach you take. /* PLUG: http://plug.org, #utah on irc.freenode.net Unsubscribe: http://plug.org/mailman/options/plug Don't fear the penguin. */
