How to Design a Video Conferencing System (Zoom Architecture)

Scaling the Video Conferencing Infrastructure

Building one meeting server is nice, but Zoom operates at a massive scale – think millions of meetings and hundreds of millions of participants.

How does the system scale to handle that load while keeping latency low?

Distributed Data Centers

The secret sauce is to deploy servers all over the world.

Zoom has data centers across many regions (reportedly 19 interconnected data centers pre-2020) and uses geolocation to connect each user to the nearest data center.

This dramatically reduces latency – the audio/video only has to travel to a nearby Zoom server, not halfway around the globe, before being distributed to others.

If you’re in London and chatting with someone in New York, ideally Zoom will use a European server for you and a US server for them, and those data centers will relay the call over Zoom’s backbone.

In fact, Zoom’s infrastructure is designed so that it can even deploy on-premise servers for enterprise customers (meeting servers inside a company’s own network) to further reduce latency and comply with security needs.

A central Zone Controller keeps an eye on each cluster’s load and manages connections within that zone.

Cloud Burstability

Zoom primarily uses its own data centers for the heavy lifting of video routing, but it also leverages public cloud providers (like AWS and Oracle Cloud) for flexibility.

Non-real-time services (scheduling meetings, user management, chat, etc.) can run in the cloud, and if Zoom’s servers are nearing capacity (think sudden surge in usage), they can offload some meetings to cloud servers on demand.

This hybrid approach gives Zoom a safety valve to scale quickly.

During peak hours or unexpected spikes (for example, the 2020 pandemic boom), Zoom could spin up extra media servers in AWS/Oracle cloud to handle overflow, then wind them down when demand drops.

The result is a highly elastic system ready to serve hundreds of millions of daily meeting participants.

Load Balancing

At the entrance to Zoom’s backend, load balancers and intelligent routing decide which data center and which server should host your meeting.

The goal is to avoid overloading any single server.

If one region’s servers are full, Zoom can redirect new meetings to another region or the cloud.

Each media router server can only handle so many concurrent streams (since it has to receive and forward for everyone in those meetings), so Zoom runs many of them and distributes meetings among them.

In practice, horizontal scaling (adding more servers) is the name of the game.

As user count grows, Zoom adds more media servers and more datacenters.

This distributed design not only scales capacity but also increases reliability – if one server or data center goes down, others can take over.

The system is built with high availability in mind (calls should continue even if a server fails). Zoom achieves this by having meetings isolated in clusters and maintaining backup paths (remember those HTTP tunnels?) in case the direct path fails.

Maintaining Call Quality at Scale

Scaling to millions of users is pointless if call quality degrades.

A standout feature of Zoom’s design is how it maintains good Quality of Service (QoS) for each call, even on subpar networks or crowded meetings.

Here are some techniques used to keep video and audio clear:

Adaptive Bitrate Streaming

The network conditions between each user and the server can vary wildly – one person might be on gigabit fiber, another on a weak Wi-Fi or cellular data.

Zoom addresses this by using adaptive bitrate and multi-bitrate encoding.

In essence, the Zoom client and server negotiate the video quality on the fly.

Each participant’s app constantly monitors metrics like packet loss, latency, and available bandwidth.

If your connection slows down, Zoom will automatically reduce the video resolution or frame rate to avoid lagging.

It might drop you from HD to SD temporarily, or switch to an audio-only mode if bandwidth is really low.

When your network improves, Zoom raises the quality back up. This dynamic adjustment happens continuously and is why Zoom calls often remain smooth even if someone’s network gets spotty.

Multiple Stream Layers

Modern video codecs and Zoom’s architecture allow sending multiple layers of video quality in one stream (or parallel streams of different quality).

For example, your Zoom app might encode your webcam at 1080p, 720p, and 360p simultaneously.

These layers are sent to the server, and the media router can then deliver the appropriate layer to each participant.

So a user on a smartphone with a slow connection might only get the 360p layer, while a user on a desktop with good Wi-Fi gets 720p or 1080p. This is far more efficient than having the server transcode video down for each user.

Zoom’s Multimedia Router (MMR) essentially helps with this selection, ensuring each client receives a stream quality that matches their device and network, thereby maintaining the best possible quality for everyone. It’s a smart balance between quality and performance.

Forward Error Correction (FEC)

To combat inevitable internet hiccups, Zoom uses forward error correction – a technique where extra redundant data is sent so that the receiver can reconstruct lost packets without needing a retransmit.

For example, Zoom might add a little recovery data every so many packets; if one packet drops, your app can repair it using the FEC data. This happens behind the scenes, and it helps reduce audio glitches or video artifacts when network packets go missing.

Application-Level QoS

Zoom even implements some QoS control at the application level (in the Zoom client itself) rather than relying purely on network routers.

This means the app is smart about prioritizing audio over video (so your voice doesn’t cut out), managing congestion, and adapting encoding settings in real time to match the network conditions of your device.

All these optimizations ensure that even as a call scales up in participants or someone’s network link degrades, the overall meeting experience remains as clear and uninterrupted as possible.

Conclusion

Designing a video conferencing system like Zoom is a balancing act between speed, scalability, and quality.

We need real-time performance (using UDP and WebRTC tricks for low latency), a robust architecture for group calls (media servers with an SFU approach to route streams), and global scaling (distributed servers, cloud fallback) – all while dynamically maintaining high call quality through adaptive techniques.

Zoom’s architecture shows how it’s done: a distributed network of multimedia routers coordinates huge calls, and smart software ensures you see and hear everyone with minimal lag.

For those preparing for a system design interview, “Design Zoom or a video chat system” is a popular question.

If you walk through the points we discussed – from using UDP/WebRTC for real-time streaming, to employing media servers (SFUs) for multiparty calls, to scaling out servers worldwide – you’ll cover the key aspects interviewers look for.

Understanding these concepts not only helps in interviews but also gives an appreciation for the engineering behind everyday tools like Zoom.

FAQs (Frequently Asked Questions)

Q1: How does Zoom achieve low latency in video calls?
Zoom achieves low latency by using UDP for most audio/video data instead of TCP. UDP doesn’t wait for lost packets to be re-sent, so it keeps the conversation flowing with minimal delay. Zoom also connects users to the nearest data center and uses efficient codecs, so the travel time for data is short and real-time interaction feels smooth.

Q2: Does Zoom use peer-to-peer or servers for its video conferencing?
For one-on-one calls, Zoom can use a peer-to-peer connection if possible. However, for group meetings Zoom relies on servers. It uses a Selective Forwarding Unit (SFU) architecture, where each participant sends their stream to a server that then forwards it to others. This server-based approach scales much better than a pure peer-to-peer mesh for large calls.

Q3: How does Zoom maintain video quality on slow networks?
Zoom dynamically adjusts to your network. It uses adaptive bitrate streaming, meaning if your internet is slow or experiencing packet loss, Zoom will reduce the video resolution or frame rate to prevent buffering. It also sends multiple quality layers for each video (high, medium, low) and can drop to a lower layer as needed. Techniques like forward error correction add redundancy to fix lost data without re-sending. All of this helps Zoom calls stay clear and smooth, even on less-than-ideal connections.