TikTok

Short-Form Video Platform

Design a short-video app where people upload clips and scroll an endless, personalized feed chosen by a recommendation system rather than by whom they follow.

Overview

In this app, people record short videos and everyone else scrolls through them one at a time. The main screen, the For You feed, is not built from accounts you follow. A recommendation system picks each video from the whole catalog based on what you have watched, finished, replayed, liked or skipped. Following still exists, but it is a second tab.

That changes the design compared with a follow-based feed such as Twitter. There is no fan-out of posts to followers. Instead there are three heavy systems: a video pipeline that turns every upload into several playable versions, a delivery system that starts a video instantly on any network, and a recommendation system that ranks millions of videos for each viewer and updates itself from their behavior within minutes.

Questions to ask first

How many users and views? — 100 million daily users, about 100 videos each: 10 billion views a day
How many uploads, how long? — 5 million a day, 30 seconds on average
Which part should we go deep on? — The For You feed first, then upload and playback
How fast must a video start? — Under about 300 ms after the swipe (a target we chose)
How fresh must recommendations be? — They should react to this session's behavior within seconds or minutes
Must videos be checked before they are shown? — Yes: automated checks before a video can be recommended

What you will be able to do

Size the video pipeline, storage and delivery bandwidth from views and uploads.
Design the upload pipeline so every step can be retried safely.
Explain adaptive bitrate playback and why the next videos are prefetched.
Walk through the recommendation funnel: candidates, ranking, re-ranking.
Describe how watch events reach the model within minutes, and how new users and new videos get a fair start.

Key concepts

Interest graph
Candidate generation
Ranking
Watch-time signals
Online training
Embeddings
Cold start
Transcoding
Adaptive bitrate
CDN prefetch

Functional requirements

Upload a short video with a caption and a sound; it becomes watchable once processed and checked
Scroll a personalized For You feed, one video at a time, without an end
Scroll a Following feed of recent videos from followed accounts
Like, comment, share, follow; show counts on each video
Record what each viewer watched, for how long, and what they skipped

Non-functional requirements

A video starts playing within about 300 ms of the swipe (a target we chose)
A feed page of 10 videos is chosen within about 150 ms at p99 (a target we chose)
Watch behavior changes recommendations within minutes
Playback and the feed stay up when the recommender is degraded: fall back to simpler candidates
No video is recommended before it passes automated checks

Interview questions

How does the For You feed choose videos?
How does the feed react to what I just watched?
How do you make a video start instantly on swipe?
Walk through an upload.
How do new videos get a chance?
How do you count likes and views on a viral video?
What happens if the recommender goes down?