when do you denormalize your database

hey all, been thinking about database design lately. we're running into some query performance issues on a read-heavy service, and denormalization keeps popping up as a potential solution. ive seen it used effectively in some places, but im always a bit hesitant to introduce redundancy. what are your thoughts? when have you found denormalization to be a good trade-off versus the added complexity of managing duplicate data? are there specific types of queries or data access patterns where its almost always a win?