System Design — Trending Hashtags

Design a backend system that detects and serves trending hashtags from Facebook posts. The system must surface trends while the real-world event is happening, target about 1-minute latency, consider the last 24 hours of posts with recency weighting, reward popularity across the community, and distinguish genuinely novel spikes from long-running background chatter. Requirements Ingest Facebook posts and extract hashtags continuously. Surface a hashtag as trending while the real-world event is still active; target end-to-end freshness around 1 minute. Score posts from the last 24 hours, with more recent posts carrying more weight. Popularity must reflect broad community interest, not only a small cluster posting repeatedly. Novelty is required: a trend should be about something new, or at least a sharp intensity increase versus its prior baseline. Serve the ranked trend list to product surfaces with low read latency. Notes The hard part is not counting hashtags; it is combining recency, popularity, and novelty without letting stale evergreen tags dominate. Be ready to define a baseline window for each hashtag, a decay function for recent posts, and an anti-spam / dedup layer so one coordinated group cannot manufacture a trend. The 1-minute freshness target pushes toward streaming aggregation plus approximate top-K structures, with batch recomputation used for correction rather than the primary path. Preparation Practice a streaming counter design with sliding 1-minute, 1-hour, and 24-hour windows. Prepare a trend score formula that includes current rate, baseline rate, recency decay, and unique-user / unique-community normalization. Drill the operational deep dive: late events, hashtag extraction quality, bot filtering, regional trends, and serving-cache invalidation.