I think a balance is needed. Completely denormalised isn’t a good idea because if you have a single table with a large row, but a single column receives most the updates and it updates a lot, you’re going to have tons of dead tuples and write churn/overhead whenever updating a row in that table.

But I agree that some people go too far with normalisation. When done reasonably, with awareness of access patterns and application behaviour, I think it’s important though.