Metric anomaly root cause localization

Design a system that, when a top-line metric (e.g. checkout success rate) drops, automatically localizes which dimension is responsible, a specific region, app version, device type, payment provider, or their combination, across a metric with 200 dimensions and millions of dimension-value combinations. Engineers currently slice dashboards by hand for 30 minutes during incidents. The system should surface the responsible slice within ~1 minute of the drop. Design the detection trigger, the dimensional search, and how you avoid spurious explanations.