system-design

Estimate average QPS as DAU multiplied by actions per user divided by 86,400 seconds, then multiply by 2-5x for peak. Storage equals records per day times record size times retention period, rounded to order of magnitude.

Use base62-encoded auto-increment IDs stored in a key-value store, with cache-aside Redis for hot URLs. Choose 302 redirects for analytics or 301 for maximum caching performance, and estimate roughly 116 write QPS per 100M DAU.

Choose SQL when you need ACID transactions, complex joins, and a stable schema. Choose NoSQL for flexible schemas, horizontal write scaling, or simple key lookups, matching the model (key-value, document, wide-column, graph) to your access pattern.

Shard only after exhausting vertical scaling, read replicas, caching, and query optimization. Sharding adds major operational complexity, so it is justified mainly for write-heavy workloads or data too large for one server.

Fanout-on-write pre-computes feeds at post time for fast reads but is expensive for celebrity accounts. Fanout-on-read assembles feeds at read time, keeping writes cheap but reads slow; a hybrid approach pushes for normal users and pulls for celebrities.

Estimation targets order of magnitude, not precision, so results can be off by 2-10x. Common mistakes include ignoring peak traffic multipliers, media storage, replication overhead, and future growth rates.