Make as easy as possible for various features to use S3

I wonder if it wouldn't be a good idea to have a storage abstraction that will either use the filesystem or S3 (or another storage), depending on the configuration. That way, the code doesn't need to care where the data is actually stored.