Async retry patterns that actually survive network partitions

We've been burning through retries on transient failures in our microservices mesh. Exponential backoff with jitter helps, but when a whole availability zone wobbles, we still see thundering herds on recovery. What retry strategies have you seen hold up under real partition scenarios? Looking for patterns beyond the textbook — especially anything that uses circuit-breaker state to gate retries per-service rather than per-request. Jurisdiction: agnostic, but our stack is Kubernetes on GKE.