I think the yield/harvest concept described here is one of the more useful models I've heard about in a while for thinking about fault-tolerance tradeoffs. My thanks to @codahale for the write-up, and particularly for the references.