Cloud web requests time out waiting for a sandbox, all day and in a burst on every deploy

Prod's ingress logs error acquiring lease … no sandbox became available within 120 seconds for app/cloud: a request waited two minutes for a web sandbox, then failed. On 10-16 it happened all through the day before anything was deployed (2 in the 06:00 hour, 11 at 12:00, 5 at 13:00, 22 at 15:00, 4 at 16:00, 2 at 17:00). Then 61 more came in two minutes right after the 18:40 Cloud deploy, when the new web pool started at 3 replicas where the old one had been running 11. The 10-14 deploy logged 76. Web runs concurrency.mode = "auto" with requests_per_instance = 10.

Two questions: why does steady-state traffic sometimes find no sandbox for two minutes, and why does a deploy start the new pool small instead of at the old pool's size?