Cold-start economics, isolate reuse, and the case for per-region warm capacity in modern nodejs systems.
Serverless was sold as the answer to over-provisioning. In practice, for latency-sensitive node-js workloads, it trades a known cost for an unknown one: cold starts at exactly the wrong moment.
A node js handler that takes 12ms warm typically takes 1.5–4 seconds cold. That tax lands on whichever request is unlucky enough to arrive first. For B2C traffic that means real user-visible latency on the request the user notices most: the first one.
The fix isn't a different runtime — it's keeping a small floor of warm nodejs workers per region, sized to your p99 arrival rate. Adensabais's scheduler treats the floor as a constraint and only scales above it.
Where the workload allows, reuse a single V8 isolate across many requests. The marginal CPU cost is near zero and the latency floor drops by an order of magnitude.