This sounds a lot like optimistic vs. pessimistic concurrency control, but over the set of all transactions in flight. You proceed assuming none of them are going to fail, but if any of them do, you're really screwed--you have to abort all of them and roll back to the last valid state to accept any more. Still, if we can stop rolling our own half-assed transactions over a set of feature-poor data stores, we'll avoid a lot of ugly problems.