What events trigger automatic model retraining beyond code changes?

What's really being asked

The interviewer wants to know if you understand that ML pipelines are fundamentally different from traditional software CI/CD because the code is only one of several moving parts. They are checking whether you think in terms of event-driven architecture for retraining, recognizing that models decay due to data and environment changes, not just code bugs. They also want to see if you can distinguish between proactive triggers like schedules, reactive triggers like drift, and external triggers like upstream data completion.

The full answer

A good answer hits six categories of triggers in order of relevance. First, data events such as schema changes, feature drift, or new labeled data landing in a warehouse or feature store. Second, time-based triggers including cron schedules, batch windows, or business cycles like end-of-quarter retraining. Third, model performance events such as accuracy degradation, prediction distribution shift, or concept drift detected by a monitoring layer. Fourth, upstream pipeline completion where ETL or data ingestion jobs finishing successfully publish an event. Fifth, infrastructure or dependency changes like new library versions, base image updates, or GPU driver changes. Sixth, manual or business-driven triggers such as regulatory requirements, A/B test resets, or ad-hoc experiments. Mentioning idempotency and conditional checks, for example only retraining if drift exceeds a threshold, shows seniority.

The mistakes people make

The biggest red flag is treating the ML pipeline like a standard app deployment and listing only git push, pull requests, or code merges. Another weak pattern is listing vague buzzwords like data changes without specifying schema drift versus volume thresholds. Saying the model stops working without explaining how you detect that in production also signals shallow monitoring knowledge. Forgetting cost controls, such as triggering expensive retraining on every minor data update, suggests you have not operated large-scale training pipelines.

What usually comes next

Expect the interviewer to ask how you prevent a feedback loop where a bad retraining trigger deploys a worse model. They may ask how you gate retraining with a validation set or shadow deployment. They might probe cost optimization, such as using spot instances or only retraining when drift exceeds a statistical threshold. Another common follow-up is how you version data and models together so that every training run is reproducible.

A concrete example

Imagine a fraud detection model. A robust pipeline triggers retraining on four events. One, a daily cron job at 2 AM after the previous day's transactions are fully ingested. Two, a monitoring alert fires when the KS statistic between production feature distributions and the training set exceeds 0.1. Three, the upstream data pipeline publishes a message confirming that new labeled fraud cases have been verified and loaded into the warehouse. Four, on the first of each month, a scheduled job retrains on the full quarterly window to capture seasonal patterns. Each trigger includes a human approval gate for production promotion if training cost exceeds five hundred dollars.