System requirements

Functional:

1. Upload videos with title, description
2. Search videos by keywords, creation time
3. Watch selected videos online, support different resolutions

Non-Functional:

1. Low latency especially when watch videos
2. High availability
3. Consistency can be relaxed a bit, eventual consistency across regions should be enough.

Capacity estimation

500M DAU, watch 10 videos per day, 100M upload 1 video per day

Traffic

write: 100M / 24 / 3600 = 1.2k TPS

read: 500M * 10 / 24 / 3600 = 5800k TPS

Storage

Suppose a video is 500MB in size

500MB * 100M = 50PB/day

API design

1. upload(title, description, user_id, video_id, chunk_id)
2. stream(video_id, user_id)
3. get_all_videos(author_id)

Database design

The actual videos will be stored in object storage.

We can use NoSQL to store the metadata of the videos

• user_id (partition key)
• created_at (sort key)
• title
• description
• link
• tags

High-level design

• App server - handles user requests, breaks down videos into chunks
• Upload queue - message queue that contains all the chunked video upload requests
• Encoder - workers that encode videos into different format/resolution
• Processed queue - message queue that contains all the encoded video chunks
• Combine server - update metadata of the video chunks, saves chunks into object storage and CDN
• Search service - create index on video metadata, such as title, tags

Request flows

Upload

The requests are sent to the write app server, app server saves the metadata to database, and queued the request. A set of encoders pull from the queue, process the chunks into different resolutions, push them to another queue. The combine server read from the second queue, saves the video chunks to S3, all the chunks for the same video is saved to the same file directory, update the metadata in database, and upload the chunks to CDN.

Read

The client fetch the chunked videos from CDN.

Search

The combine server will save the data to search database as well for indexing and search.

Detailed component design

Encoder

1. Encoder concurrently pull from the request queue and encode the videos into different format (mp4, mov etc) and different resolutions (1040p, 720p etc)

CDN

1. store hot movies on CDN
2. change resolution based on users' network condition on the fly

Trade offs/Tech choices

1. Add the processed queue instead of having encoders do all the things after encoding videos because the TPS is pretty high, encoders might fall behind. Keep functionality centralized for each micro service.

Failure scenarios/bottlenecks

Try to discuss as many failure scenarios/bottlenecks as possible.

Future improvements

What are some future improvements you would make? How would you mitigate the failure scenario(s) you described above?