Design Spotify: music streaming, search and playlists

Design a music streaming service like Spotify for hundreds of millions of listeners and a catalog of over 100 million tracks. Users search for music, build playlists, get recommendations, play music smoothly over poor networks and download tracks for offline listening.

Requirements

• Functional: catalog of artists, albums and tracks; search; play with seeking; playlists that users create and share; recommendations; offline downloads; play history.
• Non-functional: playback starts in a fraction of a second; adapts audio quality to the network without stopping; serves hundreds of millions of users worldwide; licensing rules by country; offline files protected from copying.

Clarifying questions

• How many daily users and concurrent streams at peak?
• Which audio qualities, and how large are files?
• How fresh must recommendations and play counts be?
• Which licensing rules must be enforced, and where?
• Do we cover podcasts and video, or music only?

What a strong answer covers

• Audio storage: each track encoded at several bitrates, split into segments, kept in object storage and served through a CDN; popular tracks cached at the edge.
• Playback: the client fetches segments, buffers ahead, switches bitrate with network conditions, and prefetches the next track.
• Metadata and search: a catalog service and a search index with typo tolerance and popularity ranking.
• Playlists: a data model for very large and shared playlists, with concurrent edits.
• Play events into a stream for history, royalties and recommendations; recommendations computed offline and refreshed.
• Offline downloads with encrypted files and license checks; country rights checks at play time.

Follow-ups

• A huge artist releases an album at midnight worldwide.
• A user on a train keeps losing connection.
• Count plays accurately enough to pay royalties.
• Collaborative playlists edited by many people at once.