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.