Design review questions that catch failures before code exists: what breaks, how easy it is to roll back, how far the damage spreads, and how to ask business people about consistency without hearing "make it 100% safe". Ends with a checklist you can paste into your next design doc.
Instagram hit the connection wall, the vertical scaling wall and the ID wall on Postgres, and cleared all three without switching databases. Here is what broke, in what order, plus the logical sharding and Snowflake ID patterns you can copy before your own database runs out of room.
Message ordering in distributed systems does not come from a FIFO queue. What actually keeps events sane: idempotency as the real safety net, partition keys when order truly matters, optimistic locking as a version guard, and the traps of multiple consumers.
OpenAI serves over 800 million weekly ChatGPT users without database sharding. Inside the architecture: a single primary PostgreSQL, cascading read replicas, aggressive caching against thundering herds, and connection pooling pushed to its limits.
Snapshot databases only show the current mess, not who knocked over the chair. Event sourcing explained through an investigation lens: anatomy of an event, state recreation as time travel, why audit tables are a dead end, and the honest cost of eventual consistency.
Business archetypes prove your 'unique' domain is usually a store selling something unusual. Map complex requirements onto universal building blocks: product catalogue, orders, entitlements, inventory for services, party and billing, and stop reinventing the wheel.