Architecture Blueprint

Design Spotify / Music Streaming Service

High-scale music streaming architecture delivering sub-200ms audio playback start, adaptive audio bitrate chunking (Ogg Vorbis / AAC), offline caching, and personalized music discovery.

1. System Requirements

Functional

Search, stream, and play millions of licensed audio tracks with instant playback start.
Support adaptive audio bitrate switching (96kbps, 160kbps, 320kbps) based on network conditions.
Provide playlist management, user library persistence, and offline track download capability.

Non-Functional / SLA

Sub-200ms audio playback start time (Time to First Audio Frame) to eliminate playback lag.
High streaming availability (99.99%) with zero mid-song audio buffering.
Scalable global audio distribution optimizing CDN bandwidth and licensing egress costs.

2. Scale & Resource Estimates

Throughput: 600 Million Monthly Active Users. 50 Million concurrent listeners during peak hours.

Storage & RAM: 100 Million tracks * 5 MB average per track = 500 Terabytes raw audio. Encoded into 3 bitrates = 1.5 Petabytes total storage.

3. System API Endpoints

GET /api/v1/tracks/{track_id}/stream?bitrate=320
Response (200 OK):
{
  "stream_url": "https://cdn.spotify.com/audio/chunks/trk-9901-320k.m3u8",
  "track_id": "trk-9901",
  "duration_ms": 214000,
  "codec": "vorbis"
}

POST /api/v1/users/me/playback/play
Request: { track_id: "trk-9901", position_ms: 0, device_id: "dev-macbook-pro" }
Response: { status: "playing" }

4. Database Schema Design

Table: tracks (PostgreSQL Metadata Store)
- track_id (UUID - Primary Key)
- title (VARCHAR(255))
- artist_id (UUID - Foreign Key)
- album_id (UUID)
- duration_ms (INTEGER)
- isrc_code (VARCHAR(12) - Unique)

Table: playback_history (Cassandra - Time-Series Partitioned by user_id)
- user_id (UUID - Partition Key)
- played_at (TimeUUID - Clustering Key DESC)
- track_id (UUID)
- completion_percentage (NUMERIC(5,2))

Table: user_playlists (PostgreSQL)
- playlist_id (UUID - Primary Key)
- user_id (UUID)
- track_ids (UUID[] - Ordered Array)
- is_collaborative (BOOLEAN)

5. Interactive High-Level Architecture

graph TD Client["Spotify Mobile / Desktop App"] -->|Audio Byte-Range Request| EdgeCDN["Global CDN Edge PoP: Audio Chunks"] EdgeCDN -->|Cache Miss| AudioStorage[("S3 / GCS Audio Chunk Store")] Client -->|Playback Telemetry| Gateway["API Gateway: Envoy"] Gateway --> TelemetryQueue[("Kafka Telemetry Stream")] TelemetryQueue --> Flink["Apache Flink Royalty and Session Processor"] Flink --> PlaybackDB[("Cassandra Playback Store")] Client -->|Search and Metadata| MetadataService["Track and Playlist Service"] MetadataService --> MetaDB[("PostgreSQL Master-Replica")] MetadataService --> RecEngine[("Milvus Vector DB: Music Recs")]

6. Scalability Deep Dive & Bottlenecks

Spotify avoids sending monolithic 5 MB audio files over the wire. During ingestion, raw audio masters are transcoded into multiple bitrates (96 kbps HE-AAC, 160 kbps Ogg Vorbis, 320 kbps Vorbis) and diced into 10-second encrypted chunks. When a user presses play, the client issues an HTTP Range request for the first 500 KB to buffer the first few seconds instantly, while asynchronously pre-fetching remaining chunks in the background. Content-addressed chunk hashes are cached across hundreds of global CDN Edge Points of Presence (PoPs).

7. Production Scaling Strategies

Initial Chunk Aggressive Prefetch: When a user starts or skips a track, the client requests chunk 0 with high priority. Once playback begins, background workers prefetch chunk 0 of the *next* queued song to guarantee zero-latency gapless playback.

CDN Byte-Range Optimization: Audio chunks are immutable and content-addressed via SHA-256. They carry `Cache-Control: public, max-age=31536000, immutable`, offloading 98% of streaming traffic to edge CDNs.

Cassandra Playback Ring: User listening histories and royalty play counts write to an append-only Cassandra cluster partitioned by `user_id`. Cassandra's LSM-tree architecture absorbs 150,000+ playback metric writes/sec without locking.

Vector Search for Discover Weekly: Songs and listener listening sessions are encoded into 256-dimensional vector embeddings using two-tower collaborative filtering models. Vector databases (Milvus/Qdrant) query similar embeddings for personalized discovery.

Offline Sync via Widevine / FairPlay DRM: Downloaded tracks are encrypted on disk with device-bound AES-128 keys. Periodic heartbeat license renewals check subscription status without re-downloading music assets.

Spotify Connect WebSocket Sync: Real-time playback synchronization across phones, laptops, and smart speakers is maintained via persistent WebSocket connections to an event bus running Erlang/Elixir.

8. Runnable Production Starter Code

Scaffold

System Design Architectures