Models go stale
The world changes, so a frozen model slowly degrades. Retraining cadence is how often you refresh it on new data.
Schedule versus trigger
- Scheduled retrain on a fixed cycle, such as weekly
- Triggered retrain when a monitor detects drift or a metric drop
Triggered retraining reacts faster but needs reliable monitoring. Scheduled is simpler and predictable.
What drives the right cadence
- Drift speed fast moving domains need frequent retrains
- Data volume enough fresh labeled data must accumulate
- Cost each retrain consumes compute and engineering attention
- Risk every new model can regress, so each needs validation
Automate carefully
A fully automated pipeline still needs a gate that compares the candidate to the live model and only promotes a winner.
Key idea
Match retraining cadence to how fast the world drifts, and always gate promotion on a validation that beats the current model.