Ubers Journey To Ray On Kubernetes Ray Setup

Uber Engineering moved machine-learning job placement out of individual engineers' configurations and into a shared controller. Engineers specify the work and required resources; the controller filters available clusters for compatibility, then scores suitable choices using factors such as current load. This addresses an earlier arrangement in which fixed cluster choices left some clusters heavily loaded and others underused.

The controller also follows jobs through completion and releases their resources. Idle detection handles cases where a client crashes or fails to clean up. The setup account explains placement across clusters; the companion resource-management article covers sharing capacity within them.