---
title: 'YouTube Upload Architecture: How It Works'
source: 'https://youtube.com/watch?v=hMTOYedA818'
video_id: 'hMTOYedA818'
date: 2026-08-04
duration_sec: 66
---

# YouTube Upload Architecture: How It Works

> Source: [YouTube Upload Architecture: How It Works](https://youtube.com/watch?v=hMTOYedA818)

## Summary

This video explains the architecture behind YouTube's video upload pipeline, detailing how a 10 GB file is processed from client upload to global CDN distribution. It covers chunked uploads via pre-signed URLs, a DAG-based processing pipeline with parallel transcoding, and the final push to edge nodes for low-latency viewing.

### Key Points

- **Chunked Upload with Pre-Signed URLs** [00:01] — The client splits the file into chunks and uploads them to blob storage using pre-signed URLs. This allows retries on individual chunks rather than the entire file.
- **DAG Fan-Out and Transcoding** [00:16] — After upload, the pipeline fans out into a DAG. The heaviest branch is transcoding, where the video is segmented and a job is queued for each segment at every target resolution (360p to 4K).
- **Parallel Processing Workers** [00:28] — Workers pull jobs from the queue and process transcoding in parallel. Other branches handle integrity checks, metadata extraction, and copyright matching.
- **Central Jobs Table and CDN Distribution** [00:44] — A central jobs table tracks completion across all branches. Once processing finishes, the video is pushed to CDN edge nodes globally, so viewers in Tokyo don't pull bytes from Virginia.

### Conclusion

YouTube's upload pipeline is a masterclass in distributed systems, using chunked uploads, parallel processing, and global CDN caching to handle massive files efficiently.

## Transcript

back-end after user clicks upload? Well, first, the client splits the file into to blob storage via what's called a pre-signed URL. This way, any dropped connection will retry just that one chunk, not the whole 10 GB file. Once
uploaded, the pipeline then fans out into a DAG. The heaviest branch in this DAG is transcoding, where the video is put into segments and a job is queued for each segment at every target resolution, all the way from 360p up to
4K. Workers then pull off the queue and process the transcoding in parallel. branches are doing things like integrity checks, extracting metadata, generating content for copyright matches. There's also a central jobs table that tracks
the completion across every branch. When all that processing finishes, the video pushed out to CDN edge nodes all over the world. This is so that a viewer in Tokyo isn't having to pull bytes all the way from a data center in Virginia. Read
our full breakdown on how to design YouTube at hellointerview.com.