That 0.8 second P99 Latency Cliff in Production Wasnt Supposed to Happen

A Redis-backed config layer pushed P99 latency to 700 ms in a high-throughput matchmaking service.

Every request synchronously hit Redis, and cache flushes plus gRPC fetches created stampedes and latency spikes.

Scaling Redis, disabling flushes, and using replicas or cluster mode did not solve the consistency and performance issues.

The team replaced the hot-path design with ConfigEdge, using a Git-backed control plane and node-local file reads for config delivery.