How would you design a Redis-based caching layer to avoid cache stampede (thundering herd) problems under high load?