Bytes-in-flight higher than receiver window in frozen client connections

Question

I am dealing with sort of a "ghost issue". We have an endpoint URL that some people can use at all times with no issues but others have a frozen connection on the client side (checked with multiple client codes written in Java, Python C#) that never finishes while server can see reset connection. I have captured traffic over network that works and it's perfectly synchronized and I can see when the sender gets close to the receiver window size it backs down while the network that fails the Bytes in flight go 3-4x over The calculated window size. To make things more confusing we have 2 different endpoints with almost identical set up and there is a one that causes described issue and another that can handle it all. (The only difference between endpoints is the VM type used on the cluster). It appears that the autoscaling is advertised in both cases and initial packets look identical to me. Here is a capture of working network The failing network capture The failing network capture