When does eventual consistency start to feel like a necessary evil instead of a design choice?

Everyone talks about how useful eventual consistency is for making distributed systems bigger, especially if you want them to be super available and handle network problems. But I'm really interested in when you actually *have* to use it. When do you choose eventual consistency not because it's the best idea, but because trying to get strong consistency at a huge scale just gets too complicated or pricey? Are there particular kinds of data or tasks where you've noticed it's just better to go with eventually consistent systems, even if that means more headaches dealing with old data or conflicts?