Instrumenting Kueue Scheduling for ML Training - Amy Chen, CoreWeave & Gabriel Saba, Google

This talk focuses on Kueue, a job scheduler designed for managing batch and machine learning workloads. The speakers, Amy Chen from CoreWeave and Gabriel Saba from Google, examine the challenges associated with scheduling at scale, particularly opaque failures that can arise. They provide insights into how they enhanced observability within Kueue’s scheduler by translating complex scheduling logic into clear signals, while also exploring a fair-sharing quota bug and discussing metrics and logging improvements. Attendees gain practical knowledge to track the Kueue workload lifecycle, empowering them to resolve the challenge of pending workloads effectively.