github issue: https://github.com/rancher/rancher/issues/17964
Hi! I found this problem with relation of rancher/lb-service-haproxy:v0.9.3 and 0.9.6 Tested in (Rancher 1.6.20 and 1.6.26)
Once containers in service are stopped there is Draintimeout. And LB does not provide new connections to the container that is being stopped. But once container started again configuration in lb is not reloaded and running container is skipped for newly coming connections. If all containers in service were restarted there is no container left that can accept new connection from LB perspective. And service is left in zombie state it is alive but dead at the same time.
I mentioned that config reloading in lb happens sometimes if there is small delay between container in stopped state and started again. However it happens very infrequently.