Before you begin, you need a task with applied routing and a candidate model that is available in Studio.
1
Add the candidate
Open the task’s Routing page, select Add model, and choose the candidate. Start it at weight
0 if you want to save the configuration before serving traffic.2
Set rollout weights
Give the candidate a small positive weight and keep the current model in the configuration. Weights are relative: weights
1 and 9 send approximately 10% and 90% of traffic.For conversational applications, include wandb.thread_id with requests so every conversation remains on one model during the rollout.3
Save the routing configuration
Choose the task version that should receive the change and select Save routing. Wait until the deployment is Applied at the current revision before relying on the new configuration.
4
Verify the rollout
Send representative requests through the proxy. In Analytics, confirm that the candidate receives the intended traffic share and does not increase the error rate.Use evaluation results and application metrics to decide whether to increase the candidate’s weight. Set its weight back to
0 to stop new weighted traffic without removing it from the configuration.Routing and version behavior
Learn about sticky routing, parameter overrides, versions, and rollback.