What triggers automatic full retraining in an ML pipeline?

WHAT THIS TESTS: This question checks whether you understand the operational boundary between experimentation and production MLOps. Interviewers want to see that you can distinguish event-driven triggers from time-driven triggers and that you know why blindly retraining is as dangerous as never retraining. At the senior level, they are listening for cost awareness, data versioning discipline, and integration with CI/CD.

A GOOD ANSWER COVERS: A good answer hits four categories in order. First, data triggers: new labeled data arriving in a storage bucket or warehouse crosses a volume or freshness threshold, kicking off a run. Second, code triggers: a merge to the training pipeline, feature engineering logic, or model architecture in Git should rebuild and retrain via CI/CD. Third, model performance triggers: production monitoring detects statistical drift, data quality anomalies, or accuracy decay below a service level objective, which triggers remediation retraining. Fourth, schedule or policy triggers: a cron job, fiscal quarter, or compliance window mandates retraining regardless of other signals. You should also mention that these are often gated by human approval in regulated environments.

COMMON WRONG ANSWERS: A red flag is answering with only manual triggers like a scientist pressing a button. Another is conflating retraining with batch prediction or inference pipelines. Saying you retrain daily just because it feels safe without mentioning data volume or cost is also weak. Finally, ignoring rollback strategy or failing to mention that retraining requires a validation gate before promotion shows immature ops thinking.

LIKELY FOLLOW-UPS: Expect the interviewer to ask how you prevent a runaway pipeline from retraining on bad data, how you version datasets alongside code, or how you decide whether to retrain from scratch versus warm-start from the previous checkpoint. They may also probe cost: how much does a full retraining run cost in dollars or GPU hours, and how do you throttle it.

ONE CONCRETE EXAMPLE: At a previous company, our computer vision pipeline triggered full retraining in three ways. First, when the weekly labeled image dump exceeded fifty thousand new samples in GCS. Second, when a pull request merged into the main branch of our training repository, Cloud Build ran integration tests and then kicked off a training job on Vertex AI. Third, if the production model's F1 score on a held-out live slice dropped below zero point eight five for two consecutive hours, a PagerDuty alert fired and an Airflow DAG triggered retraining after on-call approval. Scheduled retraining happened monthly for compliance baseline resets even if no other trigger fired.