System Design: Leaderboard / Category Ranking

Design a per-category leaderboard service — for each category, show the top-K most-clicked posts in near-real time. The standard Redis-sorted-set solution is the expected starting point; the round goes deep on category cardinality, write rate, and consistency. Requirements Design a leaderboard service that returns the top-K posts, pins, or users for a category. Address the query API, real-time score updates, deterministic tie-breaking, millions of categories, hot categories, and historical or rolling windows. Notes Use one Redis sorted set per category: ZINCRBY for each engagement and ZREVRANGE 0 K-1 for reads. Writes are O(log n); returning K members is O(log n + K). Memory is one entry per active (category, item) pair. Shard categories by a stable hash so one category's update and read path reaches one owner. Asynchronous stream consumption makes the leaderboard eventually consistent; state the allowed lag. Keep click count as the sorted-set score. For a deterministic secondary order, maintain a separate first-achieved timestamp or monotonic sequence and resolve equal-score groups in the service, or use a bounded composite encoding only after proving the maximum count and tie-field ranges fit Redis's exact-integer double range. Do not combine raw Unix timestamps with ZINCRBY click increments. Rolling windows can use minute buckets and merge the active buckets at query time, or maintain a separate windowed aggregate with expiration events. Explain the write/read/memory trade-off. For a hot category, batch increments at stream consumers or route that category to a dedicated owner. Splitting one category across writers requires an explicit merge and consistency contract. Preparation Draw the event stream, category owner, Redis cluster, and read path. Test a tie between two items and a one-click difference to prove the secondary key can never overtake the primary count. Walk a three-bucket rolling-window query and the expiration of the oldest bucket. Quantify batching lag and hot-category write reduction.