Well structured relational databases are an awesome tool. I have built and used hundreds. But they are a bit of cargo cult architecture. Deciding to use one means adding lots of code. It adds a lot of complexity around testing. All of it multiplies. It's another big Artifact to maintain and you may have to decide how you will scale it, replicate it, make it highly-available, etc. NOSQL was the first crack in the seemingly unassailable assumption that you need to use a DB. Now I think a few of us are questioning whether to use a datastore at all. After all, once you've dropped SQL for some functions, why stuff them into yet another opaque datastore? You might end up having to use a relational database anyway at some point and then you're stuck with another piece of junk and all its conceptual overhead and extra code. Start with files and introduce a relational database only when necessary.
Formats are straight forward these days. Just write out JSON, XML, HTML5 w/ microformats or even Yaml. The art is in the "schema" or how you store the files. Choose wisely with some knowledge of your requirements and it seems like it can work well.
Files on disk mean you can use all your great UNIX tooling to do all sorts of complex operations that would take you many many hours of skilled development to do with a database. Then there are all the possible things you could do with Git or (imagine!) ZFS! Versioning everything for free? Keep an activity log within a hierarchy. What kind of cool stuff can you do with that? I aim to find out.