Mh, my current company does this, and it's a real pain and a massive innovation showstopper. It's nice as long as you have a single application and a single database. That's easy to do on premise.

Then you need search, so you either build the search yourself, or build on solr/elasticsearch or anything else building on lucene. Then you need a cache, and suddenly you need to worry about cluster setups, choose a cache implementation. Then you end up with a use case which would be easily solved with a persistent queue, or neo4j or whatever, but your on-premise solutions block you.

It sounds simple to have n different backend solutions, and to offer a binary choice - everything built-in and everything not-built in. But usually you end up with about n! * #on-premise different setups and they all end up with their own classes of bugs and difficulties to setup.