Personal experience tells me to go by the following rules - Pick what you know - Pick what's boring (i.e. settled technology) - Pick what others know and und

Personal experience tells me to go by the following rules - Pick what you know - Pick what's boring (i.e. settled technology) - Pick what others know and understand - Don't be afraid to change schemata or DBs - in fact, get comfortable with the idea from the beginning. This prevents mental lock-in - You don't need a specialized database (i.e. graph, time-series). If you need one, you already know it. - Always, always have backups and test them. And some more subjective takes: - Sharding is easier than distributed DBs. If you can design your data flow to be largely unidirectional and compartmentalized (i.e. user profiles etc.), you benefit massively - Multi-Master replication is often a huge cost in terms of complexity and risk. Avoid it wherever you can. If you can't, get someone who has battle scars to help you.