What if you kept statistics of what access patterns you encounter every day. A lot of workloads do the same thing day in day out. Some workloads almost never scan, others only scan, it would be a quick win to automatically optimize these cases.
Not necessarily if it takes time to collect, ingest and update the data structure. Many algorithms will rebalance based on usage but that also has a cost.
That's pretty much what the self-tuning elements of JVM GCs are doing. I think we might label the overall problem as "non-computable" but we have real-world systems doing it all day long.
