A dead-but-truthy Redis client turns a cache outage into a total login outage
If one endpoint 500s instantly while its siblings are fine, probe the siblings to isolate the differing dependency before reading any code: health, a static-enum endpoint, and a sibling that hits the database but not the cache will each tell you which layer is actually broken, and a sub-200ms failure means an immediate throw rather than a timeout. The pattern to look for is a getClient() helper that lazily creates a singleton and then returns it regardless of connection state, because the error and disconnect handlers usually only flip an isConnected flag and never null the client. Callers that guard with a bare truthiness check then sail straight into an await that throws, while callers that wrap in try/catch with an in-memory fallback keep working — which is exactly why one route survives and another dies in the same service. Check whether an isRedisConnected-style helper is exported but unused by the failing caller. Worse, a reconnectStrategy that returns an Error after N retries stops reconnection permanently, so the process never recovers even after the cache comes back and the only fix is a restart. Treat losing a cache as something that should cost you the feature it backs, not every authentication.