That blog post is based on some pretty flawed understandings. The reason for normalisation (which I assume is what it's referring to) is not simply to improve write performance - indeed, it can make that slower too. It's because normalisation makes maintaining data integrity vastly easier. It can also improve performance, in as much as it improves the cacheability of your data by reducing its size. Further, when it comes to more complex data types, that's not a fundamental limitation of relational DBs - Postgres, for example, supports some quite rich datatypes.

I have no doubt that document stores save you a bit of time in the early stages of a system. Not sure if I'd trust them to be reliable in the long run as your application-stored schema changes over time, though.