The way I see it there are only two ways to perfectly invalidate a cache. The first is to only cache the results of pure functions whose arguments can be perfe

The way I see it there are only two ways to perfectly invalidate a cache. The first is to only cache the results of pure functions whose arguments can be perfectly equality-checked. But if your arguments aren't trivial to compare, this can incur some overhead even on cache hits. It also doesn't work for a stateful system, because not all relevant information gets passed in as arguments. So the second way is to track every possible mutation to any data in the system, associate each piece of data with all caches that depend on it, and "push" invalidations to those caches when something changes. In the context of a front-end app, this is what MobX does. Salsa is a Rust library that does something similar and is used by rust-analyzer. Broadly this falls under the term "reactivity", though I've seen the word used to describe several different related concepts. Most of the invalidation strategies people use in practice are imperfect and ad-hoc and depend on domain knowledge. Often you get a version of #2, but done manually instead of automatically: "I know that when X happens, we need to bust cache Y, so I'll manually code that behavior". This is the "hard" version that the adage refers to.