Question

Every rolling deploy of a service causes a brief but sharp burst of client-facing 5xx errors and 'connection reset' / 'upstream closed connection' errors at the load balancer — lasting 10-20 seconds per batch of pods replaced — even though the new pods are healthy and the steady state is clean. The errors line up exactly with old pods terminating. The deployment replaces pods in batches; the LB health check passes for new pods quickly. The app receives SIGTERM and exits in ~1s. There's no connection draining configured and no preStop hook. Triage and explain the durable fix.