// HACKER NEWS — CYBERSECURITY
Show HN: An atlas of system designs with interactive architecture diagrams
Do the work when data is written and reads become instant but can go stale. Do it when data is read and results stay fresh, but every request pays the cost.
Replicas and caches put data closer to readers and survive failures, but every extra copy can lag behind the source of truth. The real question is how stale a read the product can tolerate.
A bigger machine keeps everything simple until it hits a ceiling. Splitting data and traffic across machines removes the ceiling, but adds routing, rebalancing, hot keys and queries that span shards.
When the network splits, a node either refuses requests it cannot guarantee or answers with data that may be stale. Payments and seat inventory lean correct; feeds and like counts lean available.