As is often the case, the problem as stated is massively underspecified.

For instance:
- Are your target addresses are correlated, perhaps even strictly sequential (in which case the solution is trivial)?
- How large is the dataset; is it feasible to transfer it to several kernels for parallellisation?
- Is the dataset mutable, or is it feasible to perform an initial, heavy transformation (in which case perfect hashing might outperform probabilistic/Bloom)?

Et cetera. A general database engine is almost certainly not optimal in either case.