when doent data denormalization become a problem, perhaps?

ive been thinking a lot about data modeling lately, specifically around denormalization. we often denormalize to improve read performance, especially when dealing with complex joins or when a specific query pattern is critical. however, ive seen situations where excessive denormalization leads to update anomalies and makes the data harder to maintain. at what point do the trade-offs of denormalization start to outweigh the benefits? what are your experiences with managing denormalized data and avoiding these pitfalls?