YouTube

Video Streaming

Design a video platform where creators upload large files that are processed into many formats, and viewers everywhere start watching in about a second and keep watching without stalls.

Overview

A video platform has two very different halves. On the write side, a creator uploads one large file, sometimes many gigabytes, and the platform turns it into many smaller versions: several resolutions, in more than one codec, cut into short segments. On the read side, viewers ask for those segments millions of times a second, from phones on slow mobile networks and from 4K TVs on fiber, and expect playback to start at once and never stall.

The write side is mostly a compute problem: encoding video is expensive, so it has to run in parallel and in the right order (a watchable version first). The read side is mostly a bandwidth problem: at our assumptions the platform sends around 200 terabits a second at peak, which only works if nearly all of it comes from caches near viewers rather than from the origin.

Questions to ask first

How much is uploaded? — 500 hours of video a minute, about 20 million videos a day
How much is watched? — 1 billion hours a day, twice the average at peak
Largest upload? — 256 GB or 12 hours, whichever is less (YouTube's documented limit)
How soon must a video be watchable? — A low resolution within minutes; higher ones may take hours
Which devices? — Phones, browsers and TVs, on networks from slow mobile to fiber
Live streaming, ads, copyright? — Out of scope

What you will be able to do

Design a resumable upload that survives dropped connections on multi-gigabyte files.
Build a processing pipeline that makes a low-resolution version watchable first.
Explain adaptive bitrate streaming: renditions, segments, manifests and the player's choice.
Estimate delivery bandwidth and explain why caches near viewers, not the origin, carry it.
Shard video metadata and count views without a write per view.

Key concepts

Resumable upload
Transcoding pipeline
Encoding ladder
Adaptive bitrate streaming
Segments and manifests
CDN and ISP caches
Long tail
Sharded metadata
Two-stage recommendations

Functional requirements

Upload a video file, resuming after a dropped connection
Process it into several resolutions and codecs
Play it with quality that adapts to the viewer's network
Show title, channel, view count and thumbnails on a watch page
Show a personalized home page of recommended videos
Search videos by title and description

Non-functional requirements

Playback starts in under 2 seconds at p90 (a target we chose)
Rebuffering under 1% of watch time (a target we chose)
A low-resolution version is watchable within minutes of upload; higher ones may take hours
No acknowledged upload is ever lost
Old, rarely watched videos still play, from slower storage if needed

Interview questions

How do you make a new upload watchable quickly?
How does adaptive bitrate streaming work?
How much bandwidth, and how do you serve it?
How do you handle the long tail of rarely watched videos?
How do you upload a 50 GB file reliably?
How do you count billions of views a day?
How would you build the home page recommendations?
What happens when an encoding worker crashes halfway?