Architecture & System Design › System Design Fundamentals

Designing a Chat System

Real-time messaging, presence and message storage.

Backend Engineer · Senior Frontend Engineer · Senior

Also known as: chat system design, messaging system, design chat

Designing a chat system delivers messages between users in real time with history: connection handling (WebSocket/long-poll), message routing and fan-out, ordered persistent history, presence and receipts — at scales from team chat to billions of messages a day. The core tension is real-time delivery versus durable, ordered history.

send → gateway → router → recipients' connections (+ persist to history store)
offline user → history catch-up on reconnect (cursor-based)

Key decisions: protocol (WebSocket for real-time, with fallback), 1:1 vs group fan-out (groups multiply delivery work), ordering (per-conversation sequence numbers), storage (recent in fast store, archive in cheap), presence and receipts (ephemeral state, separate from messages), and media (uploads via object storage, not inline).

The classic mistakes:

No ordering story. Concurrent sends, retries and reconnects scramble conversations without per-conversation sequencing. Sequence at write; reconcile on read.
History as an afterthought. Real-time without durable ordered history breaks on every reconnect and device switch. Persist first-class; replay by cursor.
Group fan-out naively. A 10k-member group turns one send into 10k deliveries; viral groups need fan-out batching and read-path optimisation (see hot partition).
Presence coupled to messages. Online indicators on the message path slow delivery and fail together. Separate ephemeral presence from durable messaging.
Unbounded history growth. Infinite scrollback without archiving and retention policy balloons storage. Tier history; expire per policy.
Missing abuse controls. Spam, harassment and illegal content need rate limits, reporting, blocking and scanning from the start — not after the incident.
Encryption promises. End-to-end encryption changes everything (history, search, moderation, multi-device). Decide the trust model before the architecture, not after.

Why it teaches: real-time transport, fan-out, ordering, presence, storage tiering and abuse — chat compresses distributed messaging into one relatable system. Sequence, persist, fan out deliberately.