One thing I've seen is that a weighted LRU is usually slightly better for unpredictable workloads. i.e. it will never evict high use cache data when there's a flurry of one-off traffic that would normally start evicting a normal LRU cache, and even a 2 random LRU scheme.

This is particularly relevant for databases and keeping rows in memory.

The algorithm is something like "add two weight when an item is used, and decay one from every item every second, evict items under pressure from the lowest weight to the highest."