Trigger on evidence, not just the calendar. Monitor input drift, prediction distribution, and where labels allow it, live performance on a rolling window. Retrain when a metric crosses a threshold you defined in advance, or on a cadence matched to how fast the domain moves. Always validate the candidate against the incumbent on recent data before promoting it.
Why interviewers ask this
Interviewers want to know you treat retraining as a controlled release rather than a cron job that overwrites production. The important parts are a trigger defined ahead of time, an understanding of label latency, and never promoting a retrained model without comparing it to the current one on fresh data. Automatic retraining with no gate is a common and expensive mistake.
How to structure your answer
- Separate scheduled retraining from triggered retraining.
- Name what you monitor, and mention label delay as the constraint.
- Insist the candidate is compared to the incumbent before promotion.
- Mention feedback loops where the model shapes its own training data.
Example answer
I use both a schedule and triggers. The schedule is set by how fast the domain moves: a fraud model I would retrain weekly, a demand forecast monthly, an image classifier on a stable taxonomy maybe twice a year. Triggers come from monitoring: population stability on the key inputs, the prediction distribution shifting, and where labels arrive fast enough, actual performance on a rolling window. The constraint people underestimate is label latency. If churn labels take sixty days, I cannot measure real degradation for two months, so I lean harder on input drift as an early warning and I say plainly that it is a proxy. The part I will not skip is the gate. A retrained model is a release candidate, not an upgrade. It gets evaluated against the incumbent on the most recent held out period, and if it does not win we keep the old one, because fresher training data does not automatically make a better model. And I watch for feedback loops: if the model decides who sees an offer, the next training set only holds outcomes for people it already liked, so I keep a small random holdout to keep the data honest.
Walking into this interview soon? GhostPilot listens to your live call, spots the question the moment it is asked, and puts a structured answer on your screen in real time. Try it on your next mock, or grab a $29 Session Pass, no subscription, for the real thing.
See how it worksFollow-up questions to expect
- What statistic would you use to detect input drift, and what threshold?
- How do you handle retraining when labels are heavily delayed?
- What would make you retrain from scratch instead of continuing training?
Related machine learning engineer questions
Your interviewer will ask their own version of this. Paste your actual job description into the free Question Predictor and get the 20 questions that role is most likely to ask, with what each one is really probing.
Predict my questions