Worked on the jeejeelee/vllm repository to enhance observability and performance in the scheduler by introducing a detailed Waiting Requests Breakdown metric. This feature categorizes waiting requests by capacity and deferred reasons, enabling more precise monitoring and debugging of scheduling behavior. The implementation involved backend development and metrics instrumentation using Python, with a focus on exposing labeled metrics through core updates. The work supported business goals such as reducing mean time to resolution and improving scheduling efficiency. Collaboration was demonstrated through co-authored commits, and the changes improved the system’s ability to localize issues and plan for capacity more effectively.