Cold start and what to monitor

A new customer with no history breaks collaborative filtering. There are three reliable answers. Lean on content features in the item tower so brand-new items still embed sensibly. Fall back to a popularity prior for brand-new customers. Or run a contextual bandit that explores intelligently. We will cover the first two now.

Four signals catch most production regressions: API p99 latency, retrieval cache miss rate, ranking score distribution drift, and model-version freshness. Any one of them moving by more than a few percent is worth a page.