Design Zoom

Problem framing: an SFU video meeting platform

Zoom is a cloud video-conferencing platform, and the fact that defines the architecture is the media topology: a meeting is a session on a selective forwarding unit (SFU). Each participant sends one uplink stream to the SFU and the SFU forwards the right streams to each other participant — it does not mix (that is an MCU, too costly at scale) and it does not require everyone to send to everyone (that is a P2P mesh, which collapses past a handful). With simulcast, each sender publishes several quality layers, and the SFU forwards the best layer each receiver can handle, so one person on weak wifi never drags the whole meeting down.

So the spine is: join and signaling to bring a participant onto an SFU, simulcast plus per-receiver layer selection driven by the active speaker and each downlink, screen-share as a second high-resolution track, breakout rooms as moves between SFU sessions, cloud recording via a server-side compositor, host controls, and cascading SFUs to span regions and large meetings. State the SFU-forwarding framing up front, because the whole design — capacity, adaptation, recording, scale to thousands — follows from the server forwarding selected layers rather than mixing or meshing.