Several workloads impact the PostgreSQL buffer cache and the Operating System page cache in different ways.
Here is an attempt to track the real-life cases where the *_cost GUC are abused to make a workload perform well. Some hypothetical scenarios are also described to illustrate possible corner-case situation or common workloads.
Cached data may not be re-used vs Cached data may be re-used.
Query trash cache vs Query build cache vs Query use cache
Server is not dedicated to PostgrSQL vs server is dedicated.
Add Kevin case if possible.
One query to trash them all. (but the kernel is not stupid)
Day vs Night workload
Small server, small usage.