Design Facebook Messenger

Messenger at Meta scale

We are designing Facebook Messenger: real-time one-to-one and group chat for billions of users, integrated with the Meta social graph. The fact that defines the architecture is per-thread ordering at planet scale. Within a single conversation, every participant must see messages in the same order, but there is no requirement to order messages across different conversations, so the system is a sea of independent, totally-ordered threads rather than one global log. This is what makes Messenger shardable: each conversation is its own small ordered world.

Unlike a privacy-first messenger, Messenger is server-readable by default (the server stores and can process message content for search, integrations, and rich features), with secret conversations as an opt-in end-to-end-encrypted mode. The spine is therefore: a send path that assigns each message a per-thread sequence idempotently, a fan-out that delivers to connected devices over WebSocket and pushes to offline ones, an inbox-sync path that catches a reconnecting device up via cursors, and presence and receipts on top. State the framing up front, because per-thread ordering, idempotent send, and sync are the load-bearing ideas everything else hangs on.