Recent Like Count and Top Posts in a Sliding Window

Design the last-minute like-count backend for a social platform: exact count for one post, global top posts over the last minute, and region-based top posts over the same window. Requirements Design a social-platform service supporting three read APIs: Given post_id, return that post's like count in the past 60 seconds. Return the globally most-liked posts in the past 60 seconds. Return the most-liked posts per region in the past 60 seconds. Assume peak write scale around 1M likes/sec and average write scale around 100K-300K likes/sec. Clarify whether the 60-second window is sliding or tumbling, whether exactness is required, how region is derived, and whether unlike events subtract from the count. One viable split: Exact point query: Redis hash buckets keyed by post and second, e.g. likes:{post_id} with fields for recent seconds. Sum the latest 60 buckets at read time. Top-K: stream likes through Kafka into a Flink sliding-window job, aggregate by post and region, then publish precomputed snapshots into Redis sorted sets such as trending:global and trending:{region}. Write path: client → Like Service auth/dedup → Kafka → Redis bucket incrementer and Flink consumer group. Read path: point query reads bucket hash; top-K reads sorted-set snapshots with no per-read heavy computation. Notes Exact counts and approximate trending can use different paths. The point query is user-visible and should be closer to exact; top-K is a product ranking surface and can tolerate approximation. Deduplicate repeated likes by the same user before publishing count events. If unlike is supported, the event stream must carry negative deltas and late-event handling. Region partitioning needs a product decision: IP geolocation, registered country, or current device region all produce different semantics. The hot-key problem is real for celebrity posts. Discuss partitioning by post_id, local aggregation, and snapshot write frequency. Preparation Practice drawing the dual-path design: exact point counters plus streaming top-K. Be ready to explain sliding-window bucketing, TTL cleanup, and why top-K should be precomputed. Prepare trade-offs for exact vs approximate counts, Flink vs Redis-only designs, and global vs region partition keys.