I'd argue that most of the time it is better to write normalized data and use some form of permanently existing database views (rather than actual table) to read from pseudo-denormalized data. That way you combine the best of both worlds.
That really, really depends on whether you’re driving something that gets read often enough that query complexity ruins your product’s ability to predictably deliver on query deadlines — whether that’s “load a webpage” deadline or a “submit half a million payments to the bank before it closes” deadline.
I've started doing this (materialized views denormalizing data) at work more recently and it's been immensely helpful. Surprised more people don't do this.
