Network services with heavy load will cause "connection reset" from time to time. Especially those with big payloads. When packets with sequence number out-of-window arrived k8s node, conntrack marked them as INVALID. kube-proxy will ignore them, without rewriting DNAT. The packet goes back the original pod, who doesn't recognize the packet because of the wrong source ip, end up RSTing the connection.