This was true for a few minutes, but now has settled down.

That suggests to me that there was some huge source of abort_suite golo-proxied commands (probably release builders starting and trying to abort prior suites on that branch en masse?), and thus this sudden spike of abort_suite jobs can eat up a lot of golo proxy "slots".

We need metrics / tracking around how many slots are being used, and for what.

Relatedly, we have sysmon metrics about the server that is acting as the golo proxy. The spikes during which we receive a slew of jobs are clearly visible, and seem to cause load spikes that last around 20 minutes.

https://viceroy.corp.google.com/chromeos/machines?hostname=chromeos-server22&duration=1d&refresh=-1

What are the next steps here?

I'm working on improved metrics here, which will let us track this latency.