Hello,
It is great to hear that you are experimenting setting up servers and it is mostly working!
To run a computation there are multiple requests and connections between the client and the server. It is essential that all of them are routed to the same server. Our public setup uses load balancing between containers even on a single machine to improve resilience - if one container crashes, others are available while the crashed one reboots. For HA-Proxy you can replicate configuration from here:
for other load balancers the idea is to do something similar - extract the session/kernel ID from the URL or header and use it for routing.
Hope this helps, please let me know about your progress!
Andrey