It's an asynchronous read model or a projection and it's now just part of your application.
For instance if I look up a value and pass it to three pure functions to ask about it, versus having the 3 functions look up the value and try to share a cache entry.
This is doubly wrong because the final question may see a cache eviction and answer a different question due to changes in the data. Now I may have a situation where True:False:True was not a corner case we cover because it's logically impossible, but it's happening in the code because the last one changed from False to True due to concurrent read and write.
The latter is also harder to test, because it requires mocks instead of manufactured inputs. And at this point the code is using the cache as global, mutable state, instead of semi-local state. Three strikes, you're out.
It's a subtle word change, but can communicate the slightly difference architecture and break the reader away from an assumption.
If your site comes crashing down because the cache went down, it's not a cache anymore and you should use a different mindset in approaching it.
The answer varies depending on a whole lot of relevant numbers that can vary widely, but an example of an answer I have in one of my systems is that I store all the cache queries more persistently than the cache answers. This particular system gets hammered with requests that are mostly the same as last time at certain times but spends a lot of time idle. The numbers work out that if I'm trying to compute the cache during one of my crunches, I need a huge amount of resources for just a couple of minutes, but if I restart the service between these crunches, there's plenty of time to recompute the entire cache before it is necessary with no additional resources. But that's just one particular example of not really being a "cache" anymore suitable in my very particular case.