hard20 pts · first answer countsmonitoringdrift

Drift alerts firing all the time

You run a Kolmogorov-Smirnov test on each feature comparing millions of production rows with training data. Almost every feature shows p < 0.001 every day, though model quality is stable. What is the best fix?

A.Ignore all drift alerts permanentlyB.With huge samples, tiny harmless shifts become statistically significant; alert on effect-size measures (such as PSI or distance thresholds) tuned to impact, and prioritise important featuresC.Switch to comparing only 10 rows per dayD.Retrain the model every hour

← Same seed, different results Hundreds of low-traffic models →