Accelerating experimentation at DoorDash with a multi-armed bandit platform

DoorDash built a Multi-Armed Bandit (MAB) platform to overcome traditional A/B testing limitations by dynamically allocating traffic to better-performing variants during experiments. The platform uses Thompson sampling with Bayesian inference to balance exploration and exploitation, reducing opportunity costs and accelerating product iteration. Key improvements include modeling treatment effects rather than absolute metric values to avoid Simpson's paradox. The system integrates reward computation, arm allocation, and automated feedback loops to minimize regret while surfacing insights faster than fixed-duration experiments.