Question

Your service runs on a platform that scales to zero when idle (Knative/Cloud Run style). A monitoring synthetic and the first real users after an idle period get 10–20 second responses or timeouts, while warm responses are ~150ms. The platform logs show the request waiting on a new pod being created: a large container image is pulled, the runtime starts, and only then does the app boot and connect to its dependencies. Traffic is spiky — long idle gaps then sudden bursts. How do you triage and cut the first-request latency?