MLOps Engineer interview question
How do you design CI/CD for machine learning?
CI validates code, data contracts, feature logic, reproducible training, model thresholds, security and serving contracts. CD promotes one signed model bundle progressively through environments with approvals, monitoring and rollback gates.
Direct answer
What a strong answer covers
How it works
Continuous integration for ML validates more than code: unit tests for feature logic, data contract tests against a fixture, a small reproducible training run that must hit a metric floor, model behaviour tests on a golden set, dependency and secret scanning, and a contract test of the serving container. Continuous delivery then treats the trained model, its preprocessing, and its schema as one signed bundle that moves through environments the same way an application does: automated promotion to staging, an approval gate with the evaluation report attached, a canary in production with automatic rollback on service guardrails, and full rollout when the delayed quality metrics clear.
Example
A pull request that changes feature code triggers unit tests, a training smoke run on 5,000 rows, and a golden-set evaluation; a metric below 0.88 blocks the merge. A nightly job trains on fresh data, registers a candidate, and opens a promotion request with the evaluation diff against production. Approval by the model owner deploys to 5 percent of traffic, and a job promotes to 100 percent after 48 hours if conversion has not dropped.
Common mistakes
- Running only code tests and letting a bad model pass CI.
- Retraining as part of the deploy step so what ships was never evaluated.
- Promoting a model with no approval trail and no attached evaluation.
A useful answer structure
Answer from your own work.
Do not memorise a script: use only examples you can explain honestly in follow-up questions.
Open the desktop workspace and rehearse a clear answer in your own words.