I've always believed that replication/scaling should be reframed as an application level concern, at a minimum to serve as a though experiment regarding what you actually need from your data providers.

The exact semantics regarding how "lively" and "safe" each domain type must be almost certainly varies type to type.

I think a lot of developers come into this thinking one big SQLite database. And for many cases that is the most correct path. But you can do a database per domain type or service, and customize the replication and performance considerations at this smaller grain. You can still do ETL magic on the other side to build composite views of everything for reporting purposes. SQLite is amazing at this kind of thing. It can really help decompose a monster problem if you're willing to throw away dogma and focus on the actual problem on hand.