Introduction
Once a model is live, it quietly starts to age. Customers change, products change, sensors get replaced, upstream systems get updated. Unlike a crashed server, a degrading model gives no error message — it just gets worse predictions, sometimes for months before anyone notices.
Model monitoring is the early-warning system. It watches four things:
| Pillar | Question | Example signals |
|---|---|---|
| service health | is the system working? | latency, error rate, uptime, memory |
| data quality & drift | do inputs look like the training data? | missing values, ranges, distributions |
| predictions | do outputs look normal? | share of each class, average score |
| performance | is it still accurate? | accuracy, RMSE, precision/recall (once labels arrive) |