Lesson 9 · 9 min
CI/CD for ML pipelines
Pipelines that ship models like code: tested, versioned, reviewable, rollback-able.
What's different about ML CI/CD
A web-app pipeline tests: types, unit tests, e2e. An ML pipeline adds:
- Data validation — schema, value ranges, freshness, no silent column rename.
- Model evaluation — pass-or-fail on a held-out probe set + drift / regression check vs current production model.
- Smoke inference — load the new artifact in a real serving stack, send 50 sample queries, verify shape + latency.
- Artifact promotion — once all checks pass, mark the new model
canary(notproduction). Production tag flips only after canary observation period.