Connection loss after approx 30 Seconds

Hey all, while testing / integrating travelping UPF with open5gs I've discovered odd behaviour I could not understand / find the root cause for it.

We use open5gs with kubernetes. The SMF is running in a pod and listening on 0.0.0.0 for gpt(c/u) and pfcp. I used a LoadBalancer for PFCP and GTP(u) with Advertise address to be sent to UPF by SMF. Right after the SMF association I can connect a UE and it will successfully create a UPF session and I can ping the DN. SMF Association and UE Sessions seem normal when using sh upf session / association. Approx 30 seconds after the association the ping stops working and the UE Sessions gets removed. I cannot re-connect since UPF reports Cause 72 (no association). I slightly modifed open5gs SMF to use the GTPU advertise address for FAR3 Rule IP (without this fix It wont worked at all, rejected with cause 73 with reason "FAR Rule invalid (3)" since it contained 0.0.0.0 as IP).

I assume there is some kind of "timeout / heartbeat problem" which might be caused by the "Natting / Kubernetes layer" between SMF and UPF? How could I best debug it and what could possibly fix this problem? Is a scenario with natting / different Subnets between SMF and UPF possible at all at the moment?

EDIT: after the 30 seconds the gNB (UERANSIM) is still connected with UPF but the UPF drops packet from UE (drop packet counter on n3 goes up)

travelping / upg-vpp

Connection loss after approx 30 Seconds #352