Design a Music Streaming Service

Understanding the Problem

Design a music streaming service like Spotify that can serve 100M+ songs to millions of concurrent users worldwide.

Functional Requirements:

• Search for songs, artists, and albums
• Play songs with seamless streaming
• Create and manage playlists
• Get personalized recommendations (e.g. Discover Weekly)
• Download songs for offline listening

Non-Functional Requirements:

• Low playback latency: Song playback must start within 200ms of pressing play
• Gapless playback: No silence between consecutive tracks
• Massive scale: Handle 50M concurrent streams
• Global availability: Low-latency streaming worldwide via CDN edge nodes

Estimation

Let's size this system:

• 500M total users, 50M concurrent streams at peak
• 100M songs in the catalog, average ~3 MB per song (compressed) = 300 PB of audio storage
• 200M playlist operations/day — creates, adds, deletes, reorders
• Bandwidth: 50M streams × 160 kbps average = ~8 Tbps aggregate bandwidth
• Metadata: 100M songs × ~5 KB metadata each = ~500 GB (easily fits in memory)

The main challenges are: (1) serving audio at massive scale via CDN, (2) pre-fetching the next track for gapless playback, and (3) building a recommendation engine that keeps users engaged.

Audio Encoding & Chunked Streaming

Music is encoded at different quality levels to adapt to network conditions:

• Low: 96 kbps (Ogg Vorbis) — mobile data saver
• Normal: 160 kbps (Ogg Vorbis) — default quality
• High: 320 kbps (Ogg Vorbis) — premium tier

Spotify uses Ogg Vorbis for streaming; Apple Music uses AAC. Both are lossy but perceptually excellent at 160+ kbps.

Chunked streaming: Songs are split into ~10-second chunks, similar to HLS/DASH for video. This enables:

• Adaptive bitrate: Switch quality mid-song based on bandwidth
• Fast start: Begin playback after buffering just 1-2 chunks (~200ms)
• Pre-fetch: Start loading the next track's first chunks while the current song is still playing (gapless playback)
• Seek: Jump to any point without downloading the entire file

Recommendation Engine

Personalized recommendations are the killer feature that keeps users on the platform. Spotify's Discover Weekly uses a multi-stage pipeline:

Collaborative Filtering:

• "Users who liked song X also liked song Y"
• Matrix factorization on the user-song interaction matrix (billions of plays)
• Works well for popular songs but struggles with new/niche tracks (cold start problem)

Content-Based Filtering:

• Analyze audio features: tempo, key, energy, danceability, acousticness
• Use deep learning models on raw audio spectrograms
• Solves the cold start problem — new songs can be recommended based on their audio features

Hybrid Approach:

• Combine collaborative and content-based signals
• Add contextual features: time of day, listening history, skip patterns
• Re-rank with business rules (boost new releases, licensed content)
• Spotify runs this pipeline weekly for Discover Weekly, daily for Daily Mix

Playlist Management & Offline Sync

Playlists:

• Stored as ordered lists of song IDs in a database
• Support collaborative playlists (multiple editors) with operational transforms or CRDTs for conflict resolution
• 200M operations/day means ~2,300 writes/second — manageable with sharding by user_id

Offline Download:

• Songs are downloaded encrypted (DRM) to the device
• Playback requires a valid license token (checked periodically when online)
• Sync service tracks which songs are cached locally and handles cleanup when storage is low

Royalty & Licensing:

• Every stream is logged for royalty calculations
• Stream events go to a Kafka pipeline for aggregation
• Royalties are calculated per-stream based on licensing agreements (pro-rata or user-centric model)