An interactive study companion for Martin Kleppmann's classic on reliable, scalable, and maintainable systems.
Billions of users expect near-instant delivery, offline support, and privacy — on unreliable mobile networks.
Chat apps are connection-heavy. A user with one WebSocket can receive thousands of messages per day; the hard part is routing each message to the right connection set with minimal latency.
WhatsApp famously runs Erlang/BEAM for massive connection counts. Message IDs are client-generated UUIDs so retries do not duplicate chats. Signal protocol handles E2E on devices.
When the receiver is offline, the server persists ciphertext and triggers APNs/FCM. On reconnect, the client syncs from its last acked sequence — no message loss without unbounded server memory.
Millions of concurrent WebSocket connections per machine with preemptive scheduling and hot code reload. Node.js or Go work at smaller scale; BEAM is proven for telco-grade connection fan-in.
Bi-directional, low-overhead channel for instant delivery and typing indicators. Long polling wastes bandwidth; SSE is server→client only.
Append-heavy writes at massive QPS with partition by chat_id. PostgreSQL struggles with write fan-out to billions of messages; the store optimizes for log-like access, not complex joins.
Mobile OSes kill background apps — you cannot hold a WebSocket forever on a phone. Push wakes the app to pull pending messages when offline.
Ephemeral state with TTL: online status, typing indicators, connection routing hints. Loss on restart is acceptable — users reconnect. Wrong tool for durable message history.
Servers route ciphertext they cannot read — privacy promise. Trade-off: server-side search and moderation become harder.
Full-text search across all chats is expensive and often not a product requirement. WhatsApp optimizes delivery, not global search.