Asked about Deploying one AI service, step by step
Walk me through deploying a containerised AI API to production.
An answer that works
Build the image in CI and push it to a registry. Deploy to a managed container service with its own least-privilege identity, secrets read from the secret manager, a health check, minimum and maximum instance counts, and a timeout that fits the model calls. Add logs with request ids, budget alerts, and a rollback path to the previous image.
Then they push further
Users complain the first request in the morning is slow. Why?
The service scaled to zero overnight and the first request waits for a container to start. Keep a minimum of one instance, shrink the image, and move slow start-up work out of the request path.
How do you roll back a bad deploy?
Managed services keep previous revisions. Shift traffic back to the last good revision, which takes seconds, then investigate. That only works if deploys are immutable images with clear versions.
What gets marked up
Strong answers cover secrets, least privilege identity, scaling limits and rollback. Hard-coded keys or no cost controls are marked down.
Reading an answer is not being able to give one. The section that teaches this is Deploying one AI service, step by step, 5 hours of Cloud fundamentals.
Learn this properlyAlso asked about Cloud fundamentals