Don't kill dependencies — use backoff

Every retry reserves CPU and memory on the receiver. If the receiver is already close to out-of-memory, aggressive constant retries can finish it off. Use a backoff strategy (exponential backoff is the common default) so retries back off when problems persist.