Question

Your ad-ranking service runs an online/offline feature-skew monitor that compares the live feature vector logged at serving time against the same feature recomputed by the offline training pipeline. At 09:00 the monitor fires: for the feature 'ctr_30d_smoothed', the online and offline values now diverge for ~40% of impressions (online values are systematically lower), where yesterday they matched within tolerance. The model serves 200s, latency is normal, and revenue-per-impression has started drifting down. Dashboards: the skew started right after an overnight offline pipeline change that switched the smoothing prior from a fixed constant to a category-level Bayesian prior; the online serving service still uses the old fixed-constant smoothing. No serving deploy. How do you triage and respond?