Skip to content
anythingai.space
MLOpsintermediate14 min · 9 steps

Monitoring, Drift and Degradation

Models fail silently: inputs shift, the world changes, accuracy slides. Measure against a baseline, detect drift, and alert someone who can act.

accuracy81%
p9542ms
max PSI0.31
transactionshealthACCURACY ON LATER-LABELLED TRANSACTIONS (ILLUSTRATIVE)month 194%month 389%month 681% — nobody alertedLive trafficFraud model v3silently wrongSystem dashboardup · 42 ms · 0 errors
requestsignaldataresponseerror
metrics are illustrative
1 / 9
ProblemStep 1 of 9

Every dashboard is green — and the model is wrong

A fraud model has been live for six months. The service is up, latency is fine, error rate is zero. Yet fraud losses are rising: the model's predictions are getting worse, and nothing measures that.

ML systems fail differently from ordinary software. They rarely crash; they quietly become wrong as the world moves away from the training data.

Watch outUptime is not correctness. You need model-level monitoring on top of system monitoring.

Lesson 6 of 7 in MLOps