I’ve some agents which queues spilled over for some days. There were two reasons: Too much jobs and a very high i/o-load on the server so results were written back to wpt-server slowly. In browsers on agents sometimes I saw periods without any action for more than 10 secs (connected through vnc so these observations should be treated with caution…).
Now I have a changed page-loading-behaviour for some jobs running on these agents. The change occurred around the same time the queues spilled over. I am actually investigating whether I observe an effect of meaqsured pages or of my measurement-environment.
I don’t think the queues or i/o-problem on server could have an impact on measurements on the agents. But you definately have a deeper knowledge of the codebase of agent-components: Can you rule out an impact of server i/o-load or queue-size on measurement results?