Design Short-Video Upload & Delivery

This looks like the YouTube upload design at a smaller size, and that resemblance is where the round goes wrong. The bytes-in, transcode, deliver skeleton is the same, so a candidate who stops there has drawn a generic video platform. The part that makes it TikTok sits at the end of the pipeline. On TikTok almost nobody follows a search bar to a video. People watch the For You feed, the personalized stream that autoplays one clip after another. That feed is powered by a recommendation model that learns from engagement in near real time. So the upload pipeline is not just storage and playback. It is the front door to that model. Its job is to get a brand-new clip transcoded, screened for policy, and turned into a vector, fast enough that the feed can start testing it on a small audience while it is still fresh.

Hold onto that framing, because it reshapes two decisions later. First, publish latency and time-to-first-view matter more than raw throughput. A clip that takes an hour to become watchable has already missed the window where it could go viral. Second, a moderation gate sits on the critical path in a way it does not for a passive video library. A clip that autoplays to strangers within minutes cannot be reviewed lazily after the fact. Everything downstream, the queue design, the codec choices, the delivery prefetch, follows from wanting a fresh upload to enter the loop quickly and safely.

The TikTok angle

The distinctive part of this design is not the transcode ladder, it is how fast a fresh upload reaches the For You feed. ByteDance built TikTok around a recommendation model that trains online and reflects engagement from minutes ago. So the interviewer probes the coupling. How does a clip get transcoded, screened, and embedded quickly enough that the feed can test it while it is still new. Draw a pipeline that dead-ends at a blob store and you have described a video archive.