Description of problem:

note: this was raised from the integration team, and I was asked to report the bug.

When an application exposed via a service receives a fragmentation needed icmp message, there is no guarantee that the message will receive the right pod.

So what happens is that the icmp message receives the wrong pod behind the service, and the mtu is not reduced. In case of services with a huge number of pods, it may take a long number of reconncetions until it converges.

One additional note is the fact that the customer is using local traffic policy, so this needs to be taken in account when addressing the issue (if addressable).