Deployment & Drift
Training ends; the world does not. A deployed model is a snapshot of the past, and reality drifts away from it — customers change habits, language picks up new slang, fraudsters adapt, products launch. Nothing crashes and no error fires; accuracy just quietly erodes, which makes staleness the most dangerous failure mode in production ML. The discipline is MLOps: log predictions, compare incoming data against the training distribution (data drift) and outcomes against predictions (concept drift), alarm at a threshold, retrain on fresh data, and roll out the new version safely — shadow tests and canary releases before full traffic.
Compare deploy-and-forget against monitor-and-retrain over two years. Same model, same world — the only difference is whether anyone is watching the decay and resetting it.
Check yourself
A fraud-detection model's accuracy falls from 92% to 70% in a year with zero code changes and zero bugs. What changed, and why did no alert fire?
Go deeper (free): Chip Huyen — Data distribution shifts & monitoring ↗