The load balancer is not equipped to handle mild cases of "login storms", where many users log in at roughly the same time, it will not pick up changing load across machines fast enough and use stale data to distribute the users over the agents. The user sessions also does not produce 100% of their load from moment the session was created, which also can lead to the load balancer distributing the load unevenly.
This has been all but fixed as of bug 4429 where the load balancer now makes its decisions on the number of sessions reported by the sessionstore. There is still a theoretical issue where if many people login at exactly the same time, the number in the sessionstore isn't accurate. The user session is only saved in sessionstore once the agent has reported back that it successfully created a session. However, testing on tl.cendio.se shows that this window is less than a second. This is a massive improvement compared to before where we could potentially have very outdated load information. Considering this bug fixed.
vsmserver.log that confirms that this window is very small: > 2025-04-03 15:48:30 INFO vsmserver.session: User with uid 4036 (samuel) requested a new session > 2025-04-03 15:48:30 INFO vsmserver.session: Session tl.cendio.se:17 created for user samuel