I have an issue related to how your load balancer distributes traffic across replicas.

When I scale a service to 50 replicas and start sending requests, it looks like the load balancer uses some kind of resource‑based routing. It seems to send traffic to one replica until it reaches a high CPU/RAM usage threshold, and only then switches to another replica.

For workloads like browser spawning, this behavior is extremely inefficient. Spawning a browser takes time, so when I send, for example, 50 simultaneous requests to create browser instances, almost all of them are routed to the first replica. As a result, that replica becomes overloaded and crashes. This problem combines with the resource allocation issues I described in my other post.

If the load balancer used a simple round‑robin strategy, it would distribute each incoming request evenly across all replicas regardless of their momentary resource usage. This would completely solve the issue for my case.

I tried writing my own load balancer with round‑robin logic, but I couldn’t get it to work correctly in your environment. I also found a template with a custom round‑robin load balancer, but the repository was removed.

Do you have any recommendations or supported approaches for implementing proper round‑robin request distribution on Railway? Any guidance would help a lot.