Discord Real-Time Messaging Architecture

Discord's gateway servers run Elixir on the BEAM VM, the same platform behind WhatsApp's scale. Each WebSocket connection is an Elixir process consuming ~2KB of memory with its own isolated heap and preemptive scheduling. A single gateway server holds hundreds of thousands of concurrent connections. When a message is sent in a channel, the gateway process for that guild fans out the message to all connected member processes. The BEAM's actor model with message passing makes this fan-out natural — no shared state, no locks, no thread coordination.

Discord initially used Cassandra for message storage but suffered from JVM garbage collection pauses causing multi-second p99 latency spikes — unacceptable for real-time chat. They migrated to ScyllaDB, a C++ reimplementation of Cassandra's architecture that eliminates GC pauses entirely. ScyllaDB uses a shard-per-core architecture: each CPU core owns a specific set of data and runs a single event loop, eliminating cross-core synchronization. The migration reduced p99 read latency from seconds to single-digit milliseconds while handling billions of messages per day.

Messages are stored in ScyllaDB with a partition key of (channel_id, bucket), where buckets are time windows. This ensures that recent messages in a channel — the most common read pattern — are co-located on disk for fast retrieval. Each partition is capped to prevent unbounded growth (hot partitions are a known ScyllaDB/Cassandra antipattern). When a user opens a channel, the client fetches the latest bucket; scrolling up triggers fetches of progressively older buckets. Deleted messages are tombstoned rather than physically removed, with compaction reclaiming space asynchronously.

Discord's presence system tracks the online/offline/idle/DND status of millions of users in real-time. Each user's presence is maintained by their gateway process. When a user's status changes, the update is published to a fan-out system that notifies all guilds the user belongs to, which in turn notify all online members of those guilds. This creates a cascade: a single user going offline can trigger millions of presence updates. Discord batches and rate-limits these updates, prioritizing smaller guilds where presence is more meaningful and sampling updates in mega-servers.

Discord voice channels use WebRTC with Selective Forwarding Units (SFUs) rather than peer-to-peer mesh. Each voice server is an SFU that receives audio/video streams from each participant and selectively forwards them to others — avoiding the N² bandwidth explosion of peer-to-peer in large calls. The SFU decides which streams to forward based on who's speaking (dominant speaker detection via audio energy levels). Voice servers are regionally deployed to minimize latency, and users are routed to the nearest server. Opus codec at 64kbps provides high-quality audio with minimal bandwidth.