Content Popularity System

Begin by defining popularity: views, unique users, weighted likes, watch time, or a decayed score, plus whether results are global or personalized and how fresh they must be. Write engagement as immutable events with an idempotency key, aggregate them by content ID and time bucket, and maintain a materialized ranking such as a top-K heap or sorted store for reads. Use event time and a lateness policy so delayed events do not silently corrupt windows. A decay function or sliding window prevents old viral content from dominating forever. Partition by content ID but account for hot items through sharding or local aggregation. The read path should never recompute all content rankings. Exact per-event update cost depends on the ranking store, typically O(log K) for a bounded top-K structure.