Spotify separates personalization and experimentation into distinct tech stacks rather than combining them. Personalization systems (recommendation engines, contextual bandits, ML models) live in the ML platform for low-latency inference and rich feature access, while experimentation tools evaluate these systems through A/B tests. This separation avoids technical debt, enables better infrastructure optimization for each domain, and allows hundreds of teams to run thousands of experiments simultaneously. Multi-armed bandits are avoided for experimentation due to single-metric optimization limitations and complexity with multiple objectives.

10m read timeFrom engineering.atspotify.com
Post cover image
Table of contents
IntroductionWhat is personalization?The overlap between personalization and experimentationWhy we separate experimentation and personalization at SpotifyThe one-dimensional focus of multi-armed banditsHow we run personalization experiments efficientlyWrapping up

Sort: