← Back to Resources

Model Retraining Strategies

Practical Resources - AI Engineering

Once you accept that models can go stale (see Data and Concept Drift), the natural next question is: when, and how, should you actually retrain? There's no single right answer — the right strategy depends on how fast your data changes, how costly retraining is, and how costly a stale model is.


Common retraining triggers:

What "retraining" can mean in practice:

Before deploying a retrained model, always re-evaluate it properly — against the same held-out evaluation set (or an updated equivalent) used for the previous version, not just against whatever new data triggered the retrain. This connects directly to A/B Testing and Canary Deployments: a retrained model should generally go through the same careful rollout process as any other new model version, not skip straight to full production traffic just because it's "the same model, just updated."

Why is this important? Without a deliberate retraining strategy, teams tend to drift (no pun intended) into one of two bad defaults: never retraining until something breaks visibly, or retraining reflexively without checking whether the new version is actually better. A clear, agreed strategy avoids both.

Where to go deeper: Google Cloud's MLOps continuous delivery and automation pipelines guide covers retraining triggers as part of a broader automated pipeline design.