TikTok#

Q: How does TikTok's For-You feed work?

Candidate generation pulls thousands of videos by collaborative filtering and embeddings. A ranker scores them by watch-time, completion, and engagement, then prefetches the top N.

Q: How does TikTok solve the cold-start problem?

New videos are sampled into small audiences to gather engagement signals fast. Strong early metrics expand reach. New users get popular and topic-bootstrapped candidates.

Q: How is the video upload pipeline built?

Upload writes to object storage, a transcode farm produces multiple resolutions and HLS chunks, metadata is updated, and the video enters the recommendation candidate pool.

Q: How is sub-100ms next-video latency achieved?

The client prefetches the next few candidates while the current one plays, uses CDN edges close to the user, and serves the first chunk from a hot in-memory tier.

Q: What ranking signals does TikTok use?

Watch-time ratio, completion rate, replays, likes, shares, comments, follow events, and negative signals like skips and Not Interested feed into a deep ranking model.

Problem statement (interviewer prompt)

Design TikTok: 60-90s vertical video upload + a recommendation-first For-You feed with sub-100ms next-video latency. Cover ranking signals (watch-time, completion, replays), cold-start, audio licensing, and 1B+ MAU.

flowchart LR
  U([Creator])
  UP([Upload + Transcode])
  S3[(Object store)]
  CDN
  CG([Candidate Gen<br/>recommendation-first])
  RANK([Ranker])
  F[For-You Feed]
  V([Viewer])
  U --> UP --> S3 --> CDN
  V --> F --> CG --> RANK --> V
  V -.signals.-> CG

    classDef client fill:#dbeafe,stroke:#1e40af,stroke-width:1px,color:#0f172a;
    classDef edge fill:#cffafe,stroke:#0e7490,stroke-width:1px,color:#0f172a;
    classDef service fill:#fef3c7,stroke:#92400e,stroke-width:1px,color:#0f172a;
    classDef datastore fill:#fee2e2,stroke:#991b1b,stroke-width:1px,color:#0f172a;
    classDef cache fill:#fed7aa,stroke:#9a3412,stroke-width:1px,color:#0f172a;
    classDef queue fill:#ede9fe,stroke:#5b21b6,stroke-width:1px,color:#0f172a;
    classDef compute fill:#d1fae5,stroke:#065f46,stroke-width:1px,color:#0f172a;
    classDef storage fill:#e5e7eb,stroke:#374151,stroke-width:1px,color:#0f172a;
    classDef external fill:#fce7f3,stroke:#9d174d,stroke-width:1px,color:#0f172a;
    classDef obs fill:#f3e8ff,stroke:#6b21a8,stroke-width:1px,color:#0f172a;
    class U,V client;
    class F service;
    class UP,CG,RANK compute;
    class S3 storage;

flowchart TB
  subgraph Creator
    APP([Creator App])
    CAM[Camera + Effects]
  end

  subgraph Upload[Upload Pipeline]
    PRE[Pre-signed URL]
    OBJ[(Origin S3)]
    META[Video Metadata Service]
    TRANS([Transcoder<br/>H.264 + H.265 + AV1<br/>multi-ladder HLS/DASH])
    THUMB[Thumbnail / Cover gen]
    AUD[Audio extract + fingerprint]
    AI([AI moderation<br/>NSFW, violence, copyright, OCR])
    HASH[Perceptual hash<br/>dedup]
  end

  subgraph Distribution
    CDN[Global CDN]
    EDGE[Edge caches]
  end

  subgraph Recommend[Recommendation - the core]
    CG([Candidate Generation<br/>recall: ANN + collaborative])
    RANK1([Coarse ranker<br/>fast DNN])
    RANK2([Fine ranker<br/>heavy DNN multitask])
    RERANK([Diversity + business reranker])
    FRESH[Cold-start exploration]
    NEG[Negative signals dedup]
  end

  subgraph Signals
    WATCH[Watch time / completion]
    REWATCH[Rewatches / loops]
    LIKE[Like / share / save / follow]
    SCROLL[Scroll-past time]
    AUDIO[Audio engagement]
    CMT[Comments / sentiment]
  end

  subgraph Stores
    PDB[(Video metadata)]
    USER[(User profile / interest)]
    EMB[(Embeddings store)]
    GRAPH[(Follow / interest graph)]
    HIST[(Watch history)]
    COMM[(Comments)]
    LIKES[(Likes)]
  end

  subgraph ML[ML Platform]
    FE[Feature store - real-time]
    TR[Training pipelines]
    DEPLOY([Model server fleet])
    AB[A/B exp framework]
  end

  subgraph Social
    DM[Direct Messages]
    LIVE[Live streaming]
    DUET[Duets / Stitches]
  end

  APP --> PRE --> OBJ
  APP --> META --> PDB
  META --> TRANS --> CDN
  TRANS --> THUMB
  TRANS --> AUD
  TRANS --> AI --> HASH
  Viewer[Viewer] --> CDN
  Viewer --> RANK1
  CG --> RANK1 --> RANK2 --> RERANK --> Viewer
  Signals --> FE
  FE --> CG
  FE --> RANK1
  FE --> RANK2
  EMB --> CG
  HIST --> CG
  GRAPH --> CG
  AB --> DEPLOY
  DEPLOY --- RANK1
  DEPLOY --- RANK2
  Social --- Viewer

    classDef client fill:#dbeafe,stroke:#1e40af,stroke-width:1px,color:#0f172a;
    classDef edge fill:#cffafe,stroke:#0e7490,stroke-width:1px,color:#0f172a;
    classDef service fill:#fef3c7,stroke:#92400e,stroke-width:1px,color:#0f172a;
    classDef datastore fill:#fee2e2,stroke:#991b1b,stroke-width:1px,color:#0f172a;
    classDef cache fill:#fed7aa,stroke:#9a3412,stroke-width:1px,color:#0f172a;
    classDef queue fill:#ede9fe,stroke:#5b21b6,stroke-width:1px,color:#0f172a;
    classDef compute fill:#d1fae5,stroke:#065f46,stroke-width:1px,color:#0f172a;
    classDef storage fill:#e5e7eb,stroke:#374151,stroke-width:1px,color:#0f172a;
    classDef external fill:#fce7f3,stroke:#9d174d,stroke-width:1px,color:#0f172a;
    classDef obs fill:#f3e8ff,stroke:#6b21a8,stroke-width:1px,color:#0f172a;
    class APP,Viewer client;
    class CDN edge;
    class CAM,PRE,META,THUMB,AUD,HASH,FRESH,NEG,WATCH,REWATCH,LIKE,SCROLL,AUDIO,CMT,FE,TR,AB,DM,LIVE,DUET service;
    class PDB,USER,EMB,HIST,COMM,LIKES datastore;
    class EDGE cache;
    class TRANS,AI,CG,RANK1,RANK2,RERANK,DEPLOY compute;
    class OBJ storage;

What makes TikTok different#

Feed is recommendation-first: most videos shown aren't from followed accounts.
Heavy reliance on watch-time and completion signals; like is a weaker signal.
Recall stage uses two-tower embeddings (user × video) with ANN (HNSW/ScaNN).
Fine ranker is a multi-task DNN predicting many heads (like, share, follow, complete) combined.

Cold start#

New videos enter exploration bucket; if early CTR good, broaden distribution.
New users get popular by region/language until profile builds.

Audio + copyright#

Audio fingerprinting identifies music; license / mute / monetization.
DMCA pipeline.

Real-time#

Watch event → Kafka → feature store → next session feed.
Latency budget for feature push: < 1 min for recent watch to influence next call.

Scale#

1B+ MAU, hundreds of millions of new videos / day.
Multi-region storage; edge caches for cold-start regional bias.

Glossary & fundamentals#

Concepts referenced in this design. Each row links to its canonical page; the tag column shows whether it is a high-level (HLD) or low-level (LLD) concept.

Tag	Concept	What it is	Page
`HLD`	CDN	edge caching for static assets	cdn
`HLD`	Pub/Sub & message brokers	topics, consumer groups, delivery semantics	pub-sub-pattern
`HLD`	Search internals	inverted index, BM25, embeddings, ANN	search-internals
`HLD`	Multi-region & DR	RTO / RPO, active-active, failover	multi-region-dr

Quick reference#

Functional#

Short-video upload + edit + publish.
Personalized "For You" feed.
Follow / DM / comments / duets / stitches.
Live streams.
Music / sound library.

Non-functional#

p99 video first frame < 1 s.
1B+ MAU, multi-region.
Real-time-ish personalization (< 1 min).

Capacity#

Uploads: 100M+ videos/day, avg 30 s × ~5 MB encoded = ~500 PB/yr storage with multi-ladder.
Views: trillions/yr.

Schema highlights#

videos(id, author, audio_id, ts, duration, hashtags[], language, region, status)
watch_events(user, video, ts, watch_ms, completion, skipped)
embeddings(user_vec, video_vec) in Faiss/ScaNN
audio_fingerprints(audio_id, hash)

Trade-offs#

Recommendation-first loosens dependency on social graph → cold-start friendlier.
Heavy ML stack = big infra cost; gains in engagement justify it.
Watch-time signal rewards retention but can incentivize manipulative content; needs safety brake.
Region segregation for compliance / latency.

Refs#

ByteDance engineering posts, "Monolith" (TikTok's recommendation system paper), Facebook's two-tower retrieval paper, FAISS/HNSW papers, ByteByteGo "Design TikTok", Alex Xu Vol 2.

FAQ#

How does TikTok's For-You feed work?#

Candidate generation pulls thousands of videos by collaborative filtering and embeddings. A ranker scores them by watch-time, completion, and engagement, then prefetches the top N.

How does TikTok solve the cold-start problem?#

New videos are sampled into small audiences to gather engagement signals fast. Strong early metrics expand reach. New users get popular and topic-bootstrapped candidates.

How is the video upload pipeline built?#

Upload writes to object storage, a transcode farm produces multiple resolutions and HLS chunks, metadata is updated, and the video enters the recommendation candidate pool.

How is sub-100ms next-video latency achieved?#

The client prefetches the next few candidates while the current one plays, uses CDN edges close to the user, and serves the first chunk from a hot in-memory tier.

What ranking signals does TikTok use?#

Watch-time ratio, completion rate, replays, likes, shares, comments, follow events, and negative signals like skips and Not Interested feed into a deep ranking model.

Video walkthrough

Google System Design Interview: Design TikTok (with ex-Google EM) : via Exponent