MLOps

MLOps covers model registry, versioned deployment, canary rollout, rollback, drift monitoring and retraining pipelines.

The registry and rollback path are what let a bad model version be undone in minutes instead of overnight.

The discipline it most resembles is release engineering. Everything that makes web deployment safe, versioning, staged rollout, fast rollback, monitoring, applies to models, with the added wrinkle that a model can degrade without any code changing at all.

The piece teams most often skip is the registry, and it is the piece they most regret skipping. Without a record of which model version is serving which traffic, trained on what data, a bad deployment cannot be reversed quickly because nobody can say with confidence what the previous good state was. That single artefact turns a long evening into a two-minute rollback.

For AI systems specifically, the monitoring surface is wider than in conventional deployment. You are watching not only latency and error rates but input distribution, output distribution and cost per call, a model can be perfectly healthy by infrastructure metrics while quietly producing worse answers on a shifting population of inputs.