TikTok Short Video Platform
TikTok's architecture: short video processing, real-time recommendations, global CDN, and viral content distribution.
What is a short-form video platform?
A short-form video platform helps people upload, discover, and watch a continuous sequence of brief videos. Each swipe is a small systems problem: select an eligible video, start playback quickly, learn from the response, and avoid exposing harmful or unavailable media.
In plain language: the platform must make the next video feel immediate without letting the online feed wait for video encoding, moderation, or long-running data work. This case study designs the two coupled loops behind that promise: an asynchronous media pipeline that makes a video playable, and a low-latency feed path that decides who may see it next.
Core invariant
Only rank and deliver a video after the platform can prove it is eligible for this viewer and can serve a usable rendition within the playback deadline. A relevant video that is not yet safe, playable, or region-allowed is not a valid feed result.
This case study uses a planning envelope of 1B daily viewers, 100M uploads per day, and 2.4M peak feed requests per second. The numbers are illustrative; the reasoning is the transferable part. Review CDNs and caching first if cache misses and edge delivery are new.
The unit of work is a safe, playable next video
The product is not just a video file store. A feed response is a policy decision made under a short deadline. The service must return a small candidate record and playback metadata; the CDN serves the heavy video segments after that decision.
Functional requirements
- Let creators resumably upload a video and receive a clear processing state.
- Produce multiple renditions, thumbnails, captions, metadata, and safety signals.
- Serve a personalized, scrollable feed with likes, follows, comments, shares, and reports.
- Enforce viewer, creator, geography, age, copyright, and moderation restrictions before exposure.
- Learn from watch, skip, completion, report, and follow events without making event ingestion part of playback.
Non-functional requirements
- Return the feed decision within 200 ms P99; reserve most of the budget for retrieval, ranking, and policy rather than video bytes.
- Start video playback within 2 seconds from supported regions through adaptive CDN delivery.
- Keep accepted uploads durable and processing jobs retryable; do not lose a creator's original after an asynchronous handoff.
- Make the read path available during regional or model-service trouble through declared, safe fallbacks.
Scope boundary
This design covers upload processing, recommendation serving, media delivery, and moderation gates. Live streaming, ad auctions, creator payouts, and a full editing suite are separate systems with different latency and consistency contracts.