Ashok Malhotra from Oracle gave an interesting talk about his vision for connecting RDF(SPARQL) to relational databases. His approach is quite different from the most basic use of things like D2RQ and SquirrelRDF, which maps the tables in an RDB to classes in OWL. In the simplest form of this, abstractions over the base tables can be described in OWL, leaving the problem of generating appropriate queries to the inference engine.
In contrast to this approach, Malhotra suggests that each class be mapped to the database by writing a special purpose query. This is a knowledge-intensive approach, but the effort you spend here will be payed back in improved performance later on, since you can define abstractions in optimizable SQL.
It isn't clear to me whether there is still a need for OWL inferencing at this point - if you can define your classes as abstractions specified in SQL, do you really need to describe the relations between those classes and other abstractions in OWL? Malhotra agrees that this is an important issue, but it is still an open issue in his approach.
Another issue we haven't discussed is the decidability, provability and explanation capability of such a solution. One of the benefits of using OWL (at least, OWL-DL) is that you have a decidable language and you can prove T-Box assertions about your model. While I think this might be formally possible in SQL, it certainly isn't a normal thing to do, and it can't be done for queries that cross databases. But how important is this?
At the moment, it seems to me that this approach is quite a bit at odds with the approaches taken by D2RQ and SquirrelRDF, but perhaps that is just my own narrow-minded view, and there is a synthesis of the two approaches.
