[tracker1]

Usually the bigger reasons are easier scaling. By taking a different approach to data, it's often easier to scale across data centers etc. The second being the shape of your data. If most requests are simple key/value queries, a much simpler model can work better, more so if you can keep all your data for a given query together (document dbs in particular), this can have performance benefits over a typically normalized database.

Another is when you need read and more specifically write performance that a single system cannot keep up with. When you hit these boundaries, it gets interesting. Sometimes it's just easier to design for such a system up front.

If it's an internal application SQL first is probably fine, if it's SaaS you may want to look at alternatives.