It depends.

I work with data that is highly denormalized. It works great because of columnstores. The simple structure allows for easy queries and for the use cases of analytics. I can't easily normalize because I work in a MPP database with extremely large datasets. More joins the worse it gets. Thus, I cause redundancy for performance gains oddly enough, especially being I only get one hashing key per table and thus, I replicate data on different hashing keys for that perfomance increase.