upon further investigation, it looks like this does happen on the internal endpoint too - just a few seconds of "connection refused" then resumes working normally, which i suppose explains the 503 from the ingress, but i am still confused as to why it would happen in a deployment with multiple replicas. or even a deployment with a single replica, as i had assumed it would wait for the new pod to start and become ready before killing the old. this is pushing the boundaries of my knowledge of kubernetes, but it seems like the old pods are being killed before traffic stops being routed to them or something?

here is what i see in the events log in yatai for two updates of the deployment. all i'm doing is changing the version of the bento. the failed readiness checks with "connection refused" appearing after the "stopping container" messages feels like a clue but i am not sure what to make of it.