System Design — Instagram / News Feed

Classic Meta SD prompt. Design Instagram or Facebook News Feed with personalized ranking. Push vs pull fan-out, hot-key handling, and a ranking layer are the canonical decision points. Requirements Functional: post photo/video, follow users, view personalized home feed, like/comment. Scale: ~2B users, ~1B DAU, hundreds of millions of posts/day, fan-out to followers with celebrity-account skew (millions of followers). Decisions to lead with: Fan-out strategy: push (write-time fan-out) for normal users, pull (read-time) for celebrities; hybrid is the expected answer. Storage tiers: hot timeline cache (Redis sorted sets), cold storage (HBase / Cassandra), CDN for media. Ranking layer: candidate gen → light ranker → heavy ranker; the more senior the role, the deeper the personalization discussion expected. Read path latency budget: ~200 ms p95; caching strategy must justify it. Notes E5+ rounds explicitly probe trade-offs (push vs pull, consistency vs availability) and personalization layering. Don't just regurgitate canonical writeups; argue choices. The cleanest hybrid story: precompute feeds for normal users (~2KB / user × 2B users ≈ 4TB cluster-wide), flip a per-edge precomputed=false flag on celebrity follows, and merge their recent posts at read time. Avoids fan-out of one celebrity post to 90M+ feed rows. Cache hot posts in a replicated (not sharded) Redis tier so viral-post traffic distributes across nodes instead of melting a single hot key. Eventual consistency with a ~1 minute staleness budget is acceptable and unlocks all the latency wins. Common drill-down: cache invalidation on unfollow / post delete; thumbnail generation pipeline; abuse detection on the write path. Preparation Memorize the hybrid fan-out template and be able to draw it in 5 minutes. Pre-compute QPS estimates (DAU × posts/user/day × fan-out factor) so you don't burn time on arithmetic. Practice the ranking-layer pitch: features (recency, affinity, engagement), candidate gen, light vs heavy ranker.