The point is that the query I mentioned requires joins. You can of course get the same information from key value pairs, it will just require a number of scans over all your data, which doesn't scale if you need the queries to be fast. On the RDBMS side, there has been more than three decades of research on optimizing patterns like this. You don't want to try and reinvent that.

If you can know for sure from the start that you'll never need queries like this, then of course something like Mongo will be awesome. But requirements change, hence this article.