[crashedsnow]

It doesn't "exactly" work that way. The instance doesn't disappear after the request is served. It hangs around in case there's another request (for a while), but you only pay while the request is active. In practice "cold starts" are single-digit percentage of most common workloads (# of requests that are cold). Also FWIW it's not just scale to 0. Scaling any number of instances up will hit cold starts, the difference is this happens automatically versus fixed cluster sizes which have to either be pre-provisioned for peak, or tend toward much slower scale up times.

Disclosure: I work for GCP

[unraveller]

These averages that always get dropped to alleviate all concern tell me that serverless container cold starts will never be solved.

No platform will sweat to improve the start time of an environment 10x, not when it only affects <10% of executions that tend toward commodity prices.

Individual end-users must occasionally endure the waking up of successive microservices and devs must risk a bad first impression.