Module 6 · The Model Lifecycle: Train, Evaluate, Deploy, Monitor
Deployment, Monitoring & Model Drift
65 min
Learning objectives
- Explain what deployment means and the basics of MLOps
- Describe why deployed models degrade over time (model drift)
- Outline how monitoring and retraining keep a model healthy
From notebook to the real world
A model that scores well in testing is worthless until it's deployed — wired into a real system where it serves predictions to users or other software. Deployment brings a new set of concerns: speed, reliability, cost, and versioning. The discipline that handles all of this is called MLOps.
Deployment — Integrating a trained model into a live system (an app, website, or pipeline) so it produces predictions on real, incoming data.
MLOps — The set of practices and tools for deploying, monitoring, versioning, and updating ML models reliably in production — the ML equivalent of DevOps.
Building the model is often a minority of the work. Keeping it running, monitored, and up to date in production is where most ongoing effort goes.
Why models quietly get worse
A trained model captures patterns from a snapshot of the past. The world keeps moving. As live data drifts away from the training data, the model's predictions get less accurate — even though the model itself hasn't changed a single line. This silent decay is called model drift.
Model drift — The gradual decline in a deployed model's accuracy as real-world data diverges from the data it was trained on. Includes data drift (the inputs change) and concept drift (the relationship between inputs and outcomes changes).
Analogy
It's like a map of a fast-growing city. The map was perfect the day it was printed, but new roads open and old ones close. The map didn't change — the city did — and slowly it starts leading you astray.
Example — Drift in action
A retailer's demand-forecasting model was trained on years of stable shopping patterns. When buying behavior shifted suddenly, its forecasts went badly wrong — the model was unchanged, but the world it described was not. Performance recovered only after the team retrained on recent data.
Monitoring and retraining
- Monitor live performance metrics (accuracy, precision/recall) against ground truth as it becomes known.
- Watch input data for shifts in distribution that signal drift even before outcomes are confirmed.
- Set alert thresholds so a drop triggers human review rather than going unnoticed.
- Retrain on fresh data on a schedule or when drift is detected, then re-evaluate before redeploying.
Watch out
A deployed model with no monitoring is a liability. It can degrade silently for months, making steadily worse decisions while everyone assumes it still works as well as it did on launch day.
The lifecycle closes here: monitoring feeds drift signals back into new data and retraining, which produces a new model version — and the loop begins again.
Knowledge check
Quick practice — not part of your exam score.
What is 'model drift'?
Which best describes the role of MLOps?
A model's accuracy has been falling for months and no one noticed. What was most likely missing?
Sign in to track your progress and mark lessons complete.
Sign in