Comment 1 Tobias 2025-04-08 09:59:49 CEST

The load balancer has been simplified (bug 4429) to base user distribution solely on current number of ThinLinc users. Acquiring this number requires requesting the session store for all public session data -- a request that one should be careful of repeating with high frequency.

Considering these attempts are likely separated in the order of 10s of seconds apart, depending on the agent problem, this kind of session store requests should not be too frequent to be an issue.

Comment 2 Tobias 2025-04-08 11:08:28 CEST

If feedback from the agent is quick, i.e. for whatever reason session startup fails within a second or so, requesting an updated list of agents to try is likely not warranted and may approach dangerous levels of session store requests.

By virtue of that, a grace period within which the agent list is considered sufficiently fresh would be good. That way, we'd still be addressing the problem with long timeouts but avoid unpalatable side effects with short timeouts.

Comment 5 Tobias 2025-04-08 16:17:39 CEST

>MUST:
>* Suggested agent list must update in between session attempts
✅︎ Yes, given we're past the short grace period. 
>* Previously failed agents must be discarded
✅︎ They are.
>* A grace period must strike a good balance between not overloading the session store and sufficiently updated agent lists
✅︎ It does. A grace period of 5 seconds was added, which should be sufficiently long to never overload the session store but short enough to catch any rapid agent user distribution changes.

Comment 6 Samuel Mannehed 2025-04-10 13:20:59 CEST

Reopening due to some code-review comments.

Comment 13 Tobias 2025-04-16 14:53:32 CEST

(In reply to Samuel Mannehed from comment #6)
> Reopening due to some code-review comments.

Reported issues have been addressed.

Comment 14 Samuel Mannehed 2025-04-22 12:59:03 CEST

Looks good. Tested on a cluster with 3 CentOS 8 agents. This is a nice little improvement.