Problem Statement: A Creation-First Platform With a Ranking-Driven Consumption Loop
Frames the system as two coupled planes: an on-device editor and a cache-and-ranking-driven distribution engine.
Problem statement
Design a TikTok-like short-video platform. Creators capture clips in-app, edit them with speed ramps, LUT filters, stickers, and transitions, overlay licensed music, and publish. The platform transcodes, screens, and distributes each clip, then serves an endless personalized For You feed driven by engagement signals, with likes, comments, shares, duets, and ephemeral hashtag challenges.
The brief's four functional requirements map to four architectural planes. In-app capture and editing is a client-side GPU rendering problem. Music integration is a licensed-asset and rights-enforcement problem. The personalized feed is a recall-and-rank ML serving problem with a sub-300ms budget. Engagement mechanics are a counter, social-graph, and fan-out problem under extreme skew.
Why short video differs from generic video hosting
Three properties dominate the design. First, consumption starts within minutes of publication: a clip can receive most of its lifetime views in its first hours, so the publish-to-viewable path must be fast and the feed must surface fresh content. Second, the unit is short and rewound: average watch is on the order of the clip length with frequent replays, so start latency and seamless looping matter more than adaptive long-form buffering. Third, distribution is skewed by design: the ranking system intentionally amplifies winners, so one video can jump from hundreds to millions of views per hour, creating hot keys that shatter naive caches.
Public operating baseline
Public figures establish category scale. TikTok is widely reported at roughly 1.6 billion monthly active users in 2024 with tens of millions of uploads per day. Alphabet disclosed that YouTube Shorts exceeded 70 billion daily views in 2024. Instagram is reported above 2 billion monthly actives, with Reels described by Meta as its fastest-growing content format. Kuaishou reports around 400 million daily active users. These are company-reported or widely reported public signals, not our design targets.
For capacity planning this answer assumes a mature product with 500 million MAU, 200 million DAU, 60 minutes of average daily usage, 240 video views per DAU per day, and 30 million uploads per day. Every uncited number in this answer is an explicit design assumption, target, or budget.
The four planes
- Creation plane: capture, edit timeline, effects, audio, drafts, and export on the device.
- Ingest and processing plane: resumable upload, dedupe, transcoding ladder, packaging, thumbnails, and safety screening.
- Distribution plane: object storage, multi-CDN, feed assembly, candidate recall, and ranking.
- Engagement plane: counters, comments, social graph, duets, challenges, and notifications.
A strong answer keeps the planes separate: the editor must work offline, ingest must not block on ranking, ranking must not block on exact counters, and engagement writes must never block video playback.
Key Highlights
- •TikTok is widely reported near 1.6B MAU with tens of millions of daily uploads; Alphabet disclosed 70B daily Shorts views in 2024.
- •The design separates creation, ingest/processing, distribution, and engagement planes with explicit boundaries.
- •Short-form consumption skews to the first hours after publish, making publish-to-viewable latency a first-class SLO.
- •Ranking intentionally creates viral hot keys; the cache layer must be designed for extreme skew, not uniform access.
- •Our assumed scale: 200M DAU, 48B views/day, 30M uploads/day, 3x peak multiplier.
Section Rescue Kit
Buzzwords to use:
Safe statements:
- "I will separate the editor, the ingest pipeline, the feed, and engagement, because each has a different latency and consistency contract."
- "Let me state which published figures are public context and which numbers are my explicit design assumptions."