Using the file system as a persistence engine for a database

I'm currently designing a sort of object oriented database engine in Java. I'd like to have a plug-in design for all the components, so that I can choose to implement them as I want. Now I'm thinking about the persistence engine, and I'd like to approach the problem with a different perspective; let's see what are the features I'd like to obtain from this component: first of all I'd like to implement it easily, then I'd like it to be enough performing to be usable not only for example purpose; I want caching, expandability, eventually replication and multi-volume. I've started thinking about an “ad hoc” file format, but its design is not so trivial: deleting records could cause an inefficient use of the space, and eventually fragmentation; and caching could be a nice problem too. Having to deal with this problems ain't that easy, which means that this approach could cause a big amount of time spent. So is there any thing able to perform this sort of operations efficiently? Sure: the file system! Modern file systems (have a look at Sun's ZFS ) do offer a lot, and they are virtually free, in the sense that any modern OS offers a set of features that include caching, mirroring, space saving features such as compression, and even security features such as journalling, error recovery and file content encryption.
So the choice is done, I'll think about a persistence engine based on the file system and directory structure (I could implement inheritance via subdirectory organization, and object reference with simple urls or links).