The current implementation is entirely in-memory and synchronized out of band. Finally, for reads, the checks are done in parallel with the API call, so overhead is negligible