Unverified Commit d8af5ea7 authored byBrowse files
Ingress rollouts should specify minReadySeconds
Ingress components may run directly behind load balancers and thus need to delay deployment rollout long enough for the load balancers to see the new process. Without minReadySeconds, the upgrade process will immediately delete the old pod as soon as the new pod is ready, which means that the LB may briefly see no pods, which then triggers on some LB the behavior of including all hosts (generally if all hosts are unhealthy the LB will allow requests to go to any host as opposed to failing closed). In the future we may wish to allow minReadySeconds to be customized by customers since not all load balancers will take less than 30s to converge. However, 30s is a reasonable start.