MLOps Is What Happens After the First Deployment
· 5 min read
A model that performed well in validation can degrade in production for reasons that have nothing to do with the model itself — upstream data changes, shifting user behavior, or a schema change three systems away.
Monitoring needs to track more than uptime. Prediction distribution drift, feature drift, and outcome-label delay all need dashboards and alerting, or degradation gets discovered by a customer instead of a data scientist.
Retraining cadence should be a deliberate decision tied to observed drift, not a calendar default. Retraining too rarely lets performance decay; retraining too often on noisy data can actively hurt a stable model.