How to prevent cascading failures in a distributed system if one of the services in the chain suddenly starts to slow down under load?