A cache changes the system
Caching is often described as a database optimization. Under load, it becomes a concurrency, freshness and failure-control mechanism. That broader role changes what teams need to measure.
A useful design begins with the workload: access distribution, tolerance for stale data, miss cost and invalidation triggers. Cache technology comes later.
Control the miss path
A high hit rate can hide a dangerous miss path. Request coalescing, bounded concurrency and jittered expiry prevent popular keys from turning routine expiration into a database incident.
Negative results may also deserve short-lived caching when the underlying lookup is expensive. The policy should reflect business semantics, not a universal time-to-live.
Observe the pressure transfer
Track hit rate by workload, miss latency, origin amplification, eviction and stale serves. These signals explain whether the cache is protecting the system or merely postponing saturation.