Summary
The full deployment cycle for an AI application. Secrets, scaling, instrumentation, rollback. The lab that turns a prototype into something on-call for.
Objectives
- 01Deploy an AI application to a real host.
- 02Handle secrets and scaling correctly.
- 03Instrument latency, cost, and errors.
- 04Document and rehearse the rollback path.
Key Ideas
- A system is only deployed if you can also un-deploy it.
- The rollback rehearsal is where you find out the plan does not work.
Lab
Take an AI application you have built (from an earlier lab or your own work) and deploy it as a real service: authenticated, scaled to handle at least 10 requests per second, with secrets managed via a vault, logs shipped to a searchable backend, and dashboards for latency, cost, and error rate. Write a runbook: how to deploy, how to roll back, how to respond to a spike in error rate. Rehearse the rollback: intentionally break a deployment, revert, document the incident.
Deliverables
- — Deployed, authenticated, scaled service.
- — Secret management not in plain text or Git.
- — Dashboards for latency, cost, error rate.
- — Runbook and one written incident report from a rehearsed rollback.
Rubric
- — Production-grade — no plaintext secrets, no unmonitored errors.
- — Instrumentation — the three dashboards exist and update.
- — Operability — the runbook works when read cold.
- — Rollback — the rehearsal happened and was documented.