Skip to content
Quantoros

What are the key considerations when deciding whether to retrain a model on a schedule vs. trigger-based retraining?

ML System DesignLevel: FoundationML system design and MLOps

What are the key considerations when deciding whether to retrain a model on a schedule vs. trigger-based retraining?

Hints
  • Start with quality, load and latency requirements.
  • Trace data to prediction; identify a failure point and a validation method.

Answer

Scheduled retraining (e.g., weekly/monthly) is simpler to operate and predictable, appropriate when data/concept drift happens gradually and the cost of a slightly stale model is low. Trigger-based retraining (kicked off when monitored drift/performance metrics cross a threshold) responds faster to sudden shifts (e.g., a major behavior change from an external event) but requires robust automated monitoring and can be riskier if triggers fire on noise rather than genuine drift. Many production systems use a hybrid: a baseline schedule plus alerting that can trigger an off-cycle retrain when performance degrades unexpectedly. Also weigh the cost of retraining (compute, data pipeline complexity, validation/approval process) against the expected benefit of freshness.

Source and licence: Ishan Dutta — Awesome Data Science Interview Q&A · MIT

Related questions