I think the argument for denormalisation is something like this: If you have a DB that is mostly read from (and more importantly rows are rarely updated), with lots of data and many concurrent requests, denormalisation is a way to improve performance by avoiding the overhead of a join. But as you point out, the price is added complexity when updating.

It's a useful trick, but maybe not as useful as some think. It is, after all, just another way to optimise performance.