> Or maybe it's useful to get started off with a low overhead, easy to implement DB like Mongo and then as you grow larger, spin off uses that it doesn't serve

> Or maybe it's useful to get started off with a low overhead, easy to implement DB like Mongo and then as you grow larger, spin off uses that it doesn't serve well to other more specialized and complicated DBs? Not really, converting to relational data is quite a bit of work. Actually the reverse is the correct approach. You start with normalized data, when there's a bottleneck you start denormalizing it, if that's still not enough you move /subset/ of data to NoSQL database. > One of the biggest problem with relational DBs is that once you decide on a schema, if it's the wrong one, you're gonna be in a lot of pain. Not really from my experience all migrations were done through SQL. Also if multiple people (who understand relational databases) come with a schema they pretty much will arrive to the same normalized result.