Case Study: Churn Model, End-to-End MLOps
"Walk me through what happens to a model after a data scientist trains it" is one of the most revealing questions an MLOps interview can ask, because the honest answer for most teams two years ago was "someone emails a .pkl file to whoever runs the serving box." The gap between that answer and a mature answer is not one big tool — it's a sequence of specific, boring disciplines, each of which exists because a specific incident kept happening without it. This case study is a single narrative that exercises all of them, in order, on one running example: a subscription-product churn-prediction model.
This case study applies the track's subjects rather than re-teaching them. Where a stage uses a specific practice — containerized environments, CI gates, experiment tracking and registries, canary releases, drift monitoring, model lifecycle governance — the relevant subject is named, and the theory is not re-derived here. This subject's job is to show how those pieces click together into one coherent path a real model actually travels, and, at each stage, to name the specific failure that stage prevents — because in an interview, "we'd add a registry" is a weaker answer than "we'd add a registry, because without one nobody can say with confidence which version is live, which is exactly what caused last quarter's incident."
Read it as the answer you would give in 40–45 minutes, walking the lifecycle map from the MLOps Model Lifecycle & Governance subject one stage at a time against a concrete model, then use the follow-up questions near the end to pressure-test yourself.
The Setup
A subscription product wants to predict which customers will cancel within the next 30 days, so a retention team can proactively reach out with an offer. Assume:
| Quantity | Assumption |
|---|---|
| Active subscribers | ~2 million |
| Baseline monthly churn | ~4% |
| Model use | Score all active subscribers weekly; retention team acts on the top 5% |
| Financial impact | Medium-to-high — retention discounts cost real money per contact, and a bad model wastes retention-team capacity on customers who weren't going to churn while missing ones who were |
| Regulatory exposure | None directly (not credit/healthcare), but customer-facing decisions still warrant a real approval trail |
State early: "the retention discount this model triggers is a real cost per contact, so both a false positive (wasted discount) and a false negative (a preventable cancellation) have a measurable dollar cost — that's the frame the whole lifecycle below is built to protect." This is the same risk-tiering instinct the MLOps Model Lifecycle & Governance subject asks you to apply: not every model needs the same rigor, and this one sits in a deliberate middle tier — real financial impact, but no regulator asking for its record six months later.