YouTube#

Q: How does YouTube handle 500 hours of video uploaded per minute?

Uploads use resumable chunked transfer to origin storage, then a transcoder fleet produces a per title ladder of resolutions and codecs. Work is fanned out across thousands of workers in parallel.

Q: How does YouTube serve video globally with low start time?

Encoded segments are pushed to a global CDN. Players use adaptive bitrate streaming over HLS or DASH, picking a rung based on bandwidth and buffer to keep p99 playback start under 1.5 seconds.

Q: How does YouTube detect copyrighted content?

Content ID fingerprints audio and video on upload and matches against a reference database of rights holder assets. Matches trigger block, monetize, or track actions defined by the rights holder.

Q: How does YouTube rank videos in recommendations?

A two stage system uses candidate generation from collaborative and content embeddings, then a deep ranker scores candidates using watch history, freshness, and dwell time signals.

Q: What storage does YouTube use for video files?

Originals and transcoded ladders live in Google's distributed object storage. Metadata sits in Spanner-like stores while transient state lives in caches like Memcached and Bigtable.

Problem statement (interviewer prompt)

Design YouTube: resumable video upload from creators, multi-resolution + multi-codec transcoding, global CDN delivery with adaptive bitrate, recommendation-driven discovery, comments, live streaming, and copyright (Content ID). Handle 500h of video uploaded per minute.

flowchart LR
  U([Creator])
  UP[Upload Service]
  S3[(Origin Storage)]
  TR([Transcoder])
  CDN
  META[(Metadata DB)]
  V([Viewer])
  REC([Recommendations])
  U --> UP --> S3 --> TR --> CDN
  UP --> META
  V --> CDN
  V --> REC
  REC --> META

    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 UP service;
    class META datastore;
    class TR,REC compute;
    class S3 storage;

flowchart TB
  subgraph Creator
    APP([Creator app / Studio])
  end

  subgraph Upload[Upload Pipeline]
    PRE[Resumable upload<br/>chunked - tus / GCS]
    OBJ[(Origin: Colossus / S3)]
    META[(Video metadata)]
    QUE[[Transcode jobs queue]]
  end

  subgraph Transcode[Transcode Farm]
    SPLIT[Chunked by GoP /<br/>shot boundary]
    ENC([Encoders<br/>H.264 / VP9 / AV1 ladders])
    AUD([Audio encoder<br/>AAC / Opus])
    SUB([Auto captions / ASR])
    THUMB[Thumbnail extractor]
    PACK[HLS / DASH packager]
    DRM[DRM: Widevine / FairPlay / PlayReady]
  end

  subgraph Distribution
    EDGE[YouTube CDN<br/>Google Edge / ISP peering]
    GGC[Google Global Cache<br/>inside ISP]
    ABR([Adaptive Bitrate selector<br/>client side])
  end

  subgraph Discovery
    HOME[Home feed]
    SEARCH[Search]
    REL[Related / Up Next]
    SUBS[Subscriptions feed]
    SHORTS[Shorts]
  end

  subgraph Reco[Recommendation]
    CG([Candidate Gen<br/>two-tower + collaborative])
    RANK([Ranker DNN<br/>watch time objective])
    RR([Reranker - diversity, safety])
    SIG[Signals: watch, like, skip, search]
  end

  subgraph Engagement
    LIKE[Like / dislike]
    COMM[Comments]
    SUBN[Subscriptions]
    NOTIF[Notifications]
    PLAY[Playlists]
  end

  subgraph Ads
    ADSVR[Ad server]
    AUC[Auction / bidding]
    BR[Brand safety]
  end

  subgraph Live[Live Streaming]
    RTMP([Ingest RTMP / SRT])
    PKG[Live packager LL-HLS]
    DVR[DVR window]
  end

  subgraph Safety
    CP[Content ID]
    MOD[Moderation: ML + human]
    AGE[Age gating]
  end

  APP --> PRE --> OBJ
  PRE --> QUE
  QUE --> SPLIT --> ENC --> PACK --> DRM --> EDGE
  ENC --> AUD
  SPLIT --> THUMB
  SPLIT --> SUB
  Viewer --> EDGE
  Viewer --> ABR
  Viewer --> Discovery
  Discovery --> Reco
  Reco --> CG --> RANK --> RR --> Viewer
  SIG --> Reco
  Engagement --> SIG
  Ads --- Viewer
  Live --> EDGE
  Safety --- ENC
  Safety --- Discovery
  CP --- OBJ

    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,ABR client;
    class EDGE edge;
    class SPLIT,THUMB,PACK,DRM,GGC,HOME,SEARCH,REL,SUBS,SHORTS,SIG,LIKE,COMM,SUBN,NOTIF,PLAY,ADSVR,AUC,BR,PKG,DVR,CP,MOD,AGE service;
    class META datastore;
    class QUE queue;
    class ENC,AUD,SUB,CG,RANK,RR,RTMP compute;
    class PRE,OBJ storage;

Transcoding ladders (typical)#

144p, 240p, 360p, 480p, 720p, 1080p, 1440p, 2160p (4K), HDR
Codecs: H.264 (baseline), VP9 (Premium / 4K), AV1 (modern devices).
Audio: AAC + Opus.
Output: HLS (Apple) + DASH (rest) segments, typically 2-6 s.

Storage#

Origin in Colossus (Google) or S3 (other CDNs).
Many copies per ladder × codec; popular videos pre-pushed to edge.
Cold videos kept on lower tier.

CDN strategy#

Google Edge nodes globally.
Google Global Cache appliances inside large ISPs to save peering cost.
Anycast routing to nearest POP.

Live streaming#

Ingest RTMP (legacy) or SRT/WebRTC; transcode on the fly.
Low-Latency HLS (LL-HLS) for ~2-3 s glass-to-glass.

Content ID#

Audio + video fingerprint match against rightsholder library.
Auto-claim → revenue share / block / track.

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`	Load balancer / GSLB	L4/L7 traffic distribution and failover	load-balancer
`HLD`	CDN	edge caching for static assets	cdn
`LLD`	REST API design	verbs, statuses, pagination, errors	rest-api-design
`LLD`	Behavioural patterns	Strategy, Observer, State, Command, Chain	behavioral-patterns

Quick reference#

Functional#

Resumable video upload.
Multi-resolution + codec transcode.
Global delivery via CDN with ABR.
Discovery: home, search, related, subscriptions, shorts.
Engagement: likes, comments, subs, notifications, playlists.
Live streaming + DVR.
Ads, Premium (ad-free), monetization for creators.
Content ID copyright management.

Non-functional#

2.5B MAU; 500h video / minute uploaded.
p99 first frame < 2 s globally.
99.99% playback availability.

Capacity#

500h/min uploads × 60 min × 24h = 720k h/day.
Avg 1 GB/h raw → 720 PB/day raw → many EB/yr.
Transcoded ladders push that 3-5× to storage; offset by aggressive cold-tiering.
View traffic: tens of exabytes/year egress.

Schema highlights#

videos(id, channel_id, title, desc, ts, lang, ladders[], status)
engagements(video_id, user_id, type, ts) for watch/like
subs(viewer, channel)
embeddings(video_id, vec)

Trade-offs#

Resumable chunked upload mandatory for large/spotty creator uploads.
VP9 / AV1 saves egress but costs CPU encoding time; tier roll-out.
Pre-push vs on-demand to edge: predict by signals; cold videos pull on first miss.
Watch-time objective pushed dwell up but moderation backlash → balance with safety metrics.

Refs#

"Deep Neural Networks for YouTube Recommendations" (Covington et al., RecSys '16).
"Vitess" architecture (YouTube's MySQL scaler).
Google's CDN / Edge engineering posts; Open Connect (Netflix) for comparison.
ByteByteGo "Design YouTube", Alex Xu Vol 2.

FAQ#

How does YouTube handle 500 hours of video uploaded per minute?#