A model that scored well offline can quietly go stale in production. The world keeps moving — user behaviour shifts, upstream systems change, seasons turn — and the data a model sees in production drifts away from the data it was trained on. This scene streams synthetic "requests" (small particles) from the left into a central model block, then bins their feature values into two live histograms: a fixed training baseline (blue) and the current production distribution (orange), so you can watch the gap between them grow.
Teams often track the Population Stability Index (PSI) in production: a PSI below 0.1 is considered stable, 0.1–0.25 signals moderate drift worth investigating, and above 0.25 usually triggers an automatic alert or retraining pipeline — the same three-zone logic this gauge visualizes.
A stream of production requests flows into a deployed model while a live histogram of its incoming data drifts away from the fixed training baseline, driving a drift-score gauge that pages an alert and — if auto-retrain is armed — closes the feedback loop.
The gap between a static training-time baseline (blue bars) and a shifting live distribution (orange bars) is measured every frame into a simplified PSI-style drift score, exactly the kind of statistic real monitoring pipelines track per feature.
Raise the drift rate or live noise to push the orange histogram away from the blue one. Watch the gauge climb — cross the dashed threshold ring and the beacon alerts, triggering an automatic retrain pulse back to the model if auto-retrain is armed.
A Population Stability Index (PSI) below 0.1 is considered stable in most production ML systems; teams often set automatic retraining triggers around the 0.25 mark, the same three-zone green/amber/red logic this gauge visualizes.