Reproducibility and the registry
Track experiments (parameters, metrics, data version, code commit) and promote artefacts through a model registry with stages: staging, production, archived. A model you cannot rebuild exactly is a liability.
Serving patterns
Batch scoring for offline predictions, real-time endpoints behind autoscaling for user-facing latency, and streaming for event pipelines. Shadow deployments and canary rollouts let you compare a new model against production traffic before switching.
Monitoring and drift
Log inputs, outputs, latency and cost. Data drift means the input distribution moved; concept drift means the input–output relationship moved. Alert on both, plus on prediction distribution shifts, and define a retraining trigger up front.