Oh goody, the tried and true, 'fast but inflexible' idea of using just serialized objects instead of relational datastores has a new name:Object Prevalence. To be fair, there's a consistent API to define the way the serialization is used, but it IS just serialization.
The tradeoffs are the usual ones: No cross object queries. No real transaction support (well, you can use a transactional datastore that is non-relational (i.e. what people do using Berkeley DB as an object store from perl via one of the many 'persist to hash' modules on CPAN) and the two tier nature of this way of hacking storage means that it is good and fast if you don't need to scale it or query it in as yet unthought of ways.
What WOULD be nice would be an upgrade path to relational data and replicated objects and transactions - but wait, that's what EJB's are supposed to do, isn't it?