Drift Detection
Where today stopped looking like the baseline — for metrics feeding alerts and features feeding models
Features tracked
0
Drifting
0
Numeric (KS)
0
Categorical (Chi-square)
0
How drift is measured
Two tests, chosen by what the feature is:
- Numeric— Kolmogorov–Smirnov against the baseline distribution
- Categorical— Chi-square over the category mix
Drift is a warning, not a failure. A p95 that shifted because traffic doubled is real drift and entirely fine; a feature that shifted because an upstream schema changed is not. The evidence column is there so you can tell them apart.
Significant drift needing review
Tracked features
Compared against a rolling 14-day baseline, recomputed nightly per organization
No matching results
