The autoscaling policies currently operate on a 5 second timescale, which is too slow for 10-100s of ms-latency requests. We should decouple scaling decisions from metadata GC and other, heavier tasks to make the system a little bit lighter on its feet