Comment 5 for bug 1803629

The panics in metrics are due to the system state being closed but the worker to gather the metrics is not stopped. The stateworker which registers the prometheus collector should unregister it when the worker dies. This is only done using a defer in the worker loop and it seems possible on a heavily loaded system that the state tracker Done() could be called before the defer executes, so there's a window where the collector could use a closed state pool.