Trace one request before drawing a fleet. A browser or mobile client resolves a name, reaches a public endpoint and sends a request to a service. A load balancer distributes requests among eligible backends. Health checks stop new traffic from reaching a failed instance, but they don't make a failed request disappear. Timeouts and retries still need a policy.
Replicas help only when another instance can serve the request. If all session state lives on one server, routing to a healthy replacement can lose the session. Move shared state to a suitable store or use a signed client token when the requirements allow it. These exercises use fictional read services, so independent API replicas are suitable. Follow the direction of requests, and distinguish replicated compute from replicated durable data.