This is really interesting, thanks. I think we’ve definitely seen issues where readinessProbes failing due to an upstream dependency have caused problems — but I’m not yet convinced that removing all dependency checks is the best answer.

If the application can’t do anything useful (other than return an error page), what is the point of sending traffic to it? In the situation when all pods become unavailable could the error handling be done further up the stack? This way if the issue is only with a single pod, it does get taken out of load-balancing and your users won’t get routed to a pod that we already know is going to return an error, when there are others that could of successfully handled it.