I've been using Drools for what I consider "small" semantic databases, right now in the range of 1M-10M facts but that could get bigger.

One thing I notice is that performance is dependent on the order that I put patterns in the RHS; if the high selectivity patterns go first I am happy with the results but put them in the wrong order and gee wiz, I won't wait long enough for it to try all the matches that fail.

I see Drools as competitive with products such as triple stores, deductive databases and many of them have this feature