I can see DBs for aggregators, for instance, so that you can version events without having to keep all revisions live, and do complex queries/computation on other hardware, and just serve up the results as vector embeddings.
Also allows you to store events off-relay and then pass them to the relay when only certain npubs send a request. Tighter access-control and a protective layer against hacking. You could also move the data one layer back and have multiple relays pulling from one source, and control what they are each allowed to pull and for whom. Then a relay is like a query server that only runs a subset of queries, which allows for query optimization.
It could get pretty wild 🤯 .
