Concurrency control for one thing, if we stored the information in flat files we'd need to build an infrastructure including AOLserver-level locking to keep things consistent. A busy site supporting a bunch of locales - as GP Planet will grow to become - will have more than one translator working at a time on a subsite.
If the information were in flat files you'd still need to load the NSV cache at start-up, so that shouldn't be an issue.
The general problem of dumping and restoring data of this sort is something we should provide a simple solution for. XML-based, presumably? Straight DB dumps are a problem for any table that includes sequence-generated keys (makes it difficult to merge data from multiple instances) ...
