LRU inversion: the problem with not caring about it is that it’s not a visible problem until it very suddenly is. Your system will not gradually degrade but very suddenly and unpredictably hit a wall that it cannot get itself over.

All this talk just confirms my feelings that there is a general lack of understanding of actual modern workloads.

RAM (normal w/wo zram) doesn't get full, then stay full forever in real workloads. Not only is that not realistic at the "opened apps"/"running processes" level, it's not real at the heap allocation level within tasks within processes. And this is much more pronounced with code written in modern languages like Rust and some styles of C++. Modern heap allocators batch and cache (primarily to help with performance). But still, A LOT of memory is getting allocated and deallocated all the time, even from the kernel's PoV.

LRU itself is an imperfect approximation, not a goal. In the setup described in my other comment (fast SSD swap storages only used sparingly most of the time), so called LRU inversion gets auto-cancelled relatively quickly, as free space in RAM(+zram) gets available all the time, and some "LRU-hot" pages in SSD swap turn out to be actually cold, and those ones are the only ones that actually stay there.

This is why, I would imagine a lot of fake scenarios, and "benchmarks" based on them, may fail to replicate the practical reality of many (overall system) use-cases.

More tangentially, the oversized concern for file caching pages also points to specific aligned use-cases in mind, as if everyone is running DB-centric workloads or something.