A training metric has limited value unless the dataset, feature definitions, code and deployed artifact can be reconstructed. Production monitoring then asks whether the same assumptions still hold for real requests.
In this chapter
- Experiment tracking, model registries and release provenanceConnect training runs, registry artifacts and actual deployments through retained release identities.
- Data drift, concept drift and training-serving skewIdentify the changed quantity, choose valid references and confirm predictive impact.
- Retraining triggers, evaluation gates and rollbackApply label maturity, data checks, persistence, cooldowns and candidate promotion gates.