Whenever you're ready, walk me through how you'd approach this.
Design a backend service that returns the top-K most viewed YouTube videos in a rolling 24-hour window at planet-scale traffic, including how you would shard counters and reconcile approximate vs exact counts.
Discuss the serving API, view-event ingestion, time-bucketed rolling-window aggregation, hot-video mitigation, regional replication, caching, and consistency trade-offs. Explain how the system produces candidates cheaply, verifies exact counts when necessary, and handles late, duplicated, or missing events.