Distribuire su un runtime per container serverless

A serverless container runtime runs your image without you managing any VMs. You hand over a container, and the platform handles the rest. ☁️

Google offers Cloud Run and AWS offers App Runner. Both pull your image from a registry and expose it behind a managed HTTPS URL.

You point the runtime at your pushed image and it creates a service. With Cloud Run, one deploy command turns an image into a live URL.

The platform sends traffic to a port it chooses, exposed as the PORT environment variable. Your app must bind to that value, not a hardcoded one.

Inside a container, bind the server to 0.0.0.0, not 127.0.0.1. Otherwise the platform cannot route outside traffic into your process.

Instances start and stop at any time, so anything saved to local disk can vanish. Keep the service stateless and store data externally.

Serverless billing is usually per request and per active second. When no one calls your model, idle time can cost almost nothing. 💸

If an idle service scaled to zero, the next request waits for a container to boot. That delay is a cold start, common with big model files.

To dodge cold starts, set a minimum instances floor. The platform keeps that many containers warm and ready to answer instantly.

The runtime streams your container output to its logs console. Tail them right after a deploy to confirm the model loaded cleanly.

That is the whole loop: push an image, run one deploy, and the platform returns a public URL serving your model over HTTPS.

Let us check why your first request can feel slow.

You deployed an image to a serverless runtime, bound to PORT on 0.0.0.0, stayed stateless, and tamed cold starts. Your model is live. 🎉