Serverless scale-to-zero compute costs and cold-start concurrency

Low-traffic workloads often incur unnecessary costs when maintaining active compute instances. Serverless container platforms offer a scale-to-zero feature to eliminate per-hour charges for unused capacity, shifting the billing model to request counts and actual resource consumption (vCPU and memory seconds).

While this reduces idle spend, the transition from zero to active introduces cold-start latency. There is uncertainty regarding how the platform handles a burst of concurrent requests that arrive while the first instance is still initializing.

• Does the platform queue subsequent requests during the initial cold start, or does it trigger multiple parallel cold starts?
• Is the initialization time for the first instance billed as part of the request duration for all concurrent requests waiting in the queue?