When would a read-through cache just be too complicated to bother with?

Lately, I've really been thinking about different ways to cache things, and read-through caches always seem to be a strong choice. It's pretty cool how the cache can just grab data from the source if it doesn't have it, which makes your application code a lot simpler. But, it does make things a bit more complicated with extra steps and places where stuff could go wrong. So, when do you think the good parts of a read-through cache, like easier client code and keeping data consistent, get overshadowed by how much harder it makes setting up, watching, and fixing errors? Are there times you'd totally steer clear of a read-through cache, even though it has some nice perks?