ML CI/CD versus traditional software CI/CD

WHAT THIS TESTS This checks whether you understand that ML systems change along three axes, code, data, and the model itself, so their delivery pipeline is broader than classic software CI/CD. It rewards naming the extra stages and the reproducibility concerns.

A GOOD ANSWER COVERS A traditional CI/CD pipeline builds code, runs unit and integration tests, and deploys an artifact; the only thing that changes is code. An ML pipeline keeps all of that but adds stages. Continuous integration now also validates data schemas and statistics and runs tests on feature engineering and the training code. It then automatically trains the model on fresh data, evaluates it against quantitative thresholds and against the currently deployed model, and only promotes it if it passes. Continuous delivery deploys not just code but a model artifact and its serving infrastructure. Crucially, ML adds continuous training, an automated retraining loop triggered by schedule, new data, or detected drift, and continuous monitoring of prediction quality in production. Reproducibility matters more because you must version code, data, and the model together to reproduce a result.

COMMON WRONG ANSWERS Describing only code build, test, and deploy with no mention of data validation or model evaluation. Treating the model as a normal binary artifact and ignoring that retraining is part of the lifecycle. Forgetting monitoring and the continuous-training trigger.

LIKELY FOLLOW-UPS What data tests would you run in CI? How do you gate a model promotion? What triggers continuous training?

ONE CONCRETE EXAMPLE A commit to feature code triggers the pipeline: it validates the incoming training data's schema, retrains the model, checks that AUC exceeds 0.80 and beats the live model on a holdout, and if so packages the model and serving image and canaries it to five percent of traffic. Separately, a weekly job and a drift alarm both trigger the same training pipeline, something a traditional code-only pipeline never does.