DoorDash Adopts Multi‑Armed Bandits For Experimentation
DoorDash engineers Caixia Huang and Alex Weinstein adopt a multi-armed bandits (MAB) approach to optimize product experiments, using Thompson sampling to adaptively allocate traffic and reduce opportunity cost. They report MAB accelerates learning and lowers regret compared with fixed-split A/B tests but complicates metric inference and can create inconsistent user experiences. DoorDash plans contextual bandits, Bayesian optimization, and sticky user assignment to mitigate limitations.
Key Points
- 1Implement bandits: DoorDash uses Thompson sampling MAB to adaptively allocate experiment traffic.
- 2Reduce regret: adaptive allocation accelerates learning and cuts opportunity cost versus fixed A/B splits.
- 3Adopt practices: practitioners must design richer reward metrics and address user consistency and delayed feedback.
Scoring Rationale
Practical, credible company implementation yields actionable guidance, but limited novelty beyond established multi-armed bandit techniques.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
