Straight To The Database

An endpoint takes caller-supplied product IDs. Cache hit rate on that endpoint is near zero, database QPS tracks request QPS almost exactly, and logs show most of the requested IDs do not exist in the catalog. Which fix targets this failure directly?

Switch the eviction policy from LRU to LFU so the frequently requested product keys survive a burst of one-off lookups

Put a Bloom filter of existing IDs in front and short-circuit on a miss, and cache the negative result briefly

Raise the TTL on cached entries so the popular products stay resident and fewer real lookups reach the catalog table

Add cache replicas so the read fan-out spreads over more nodes and no single shard absorbs the whole request rate