Multi-armed bandits no longer require sticky bucketing. You can start a bandit without first setting up assignment storage in your SDK, then decide whether returning users keep their original variation. The bandit still shifts traffic based on performance against a single decision metric.
Choose the assignment behavior
Bandits used to list sticky bucketing as a setup prerequisite. Now it's a choice you make for each bandit, based on the experience you're testing:
- Leave sticky bucketing disabled when switching variations is acceptable. Returning users can move between variations as weights change.
- Use sticky bucketing when repeat visits should show the same experience, such as a promotional offer.
Sticky bucketing keeps the experience consistent, but it also keeps users in their first variation. That can limit how much traffic the bandit moves toward better-performing variations.
Align measurement and weight updates
When sticky bucketing is disabled, analysis uses a user's first exposure and its conversion window. A returning user may see a different variation later, while conversions can still count toward the original variation.
Match your conversion window to how long users take to convert, and choose an update cadence that accommodates that window. Your SDK also needs to receive weight updates promptly. Review SDK caches and any CDN or proxy caches against your update cadence.
Getting started
Multi-armed bandits are included in GrowthBook Cloud Pro and Enterprise, and Self-Hosted Enterprise. Bandits use the Bayesian engine.
In the bandit creation flow, choose your assignment attribute and Decision Metric, then enable or disable sticky bucketing for the bandit. If you enable sticky bucketing, also enable it for your organization and configure a sticky-bucketing service and assignment storage in a compatible SDK. Follow the bandit setup guide.
