It’s difficult to model relationships. You have to design a hashing strategy manually to avoid hotspots. Most web frameworks are based on sql. The api is extremely verbose and there isn’t the same level of maturity for abstractions which simplify it. Indexes are much more complicated https://docs.aws.amazon.com/amazondynamodb/latest/developerg... In general, the tooling is less mature. You can run a fake dynamodb locally, but not the exact same engine.

There are certainly use cases where nosql is very helpful - heavy writes, as a denormalized cache, simpler non relational data models etc http://highscalability.com/blog/2010/12/6/what-the-heck-are-...