What problem does it solve?
PyTorch CI data is scattered across GitHub Actions, HUD, ClickHouse, and Prometheus, making it hard to diagnose failing jobs, slow runners, or queue backlogs without manually stitching dashboards together.
Core Features & Use Cases
- Workflow and Job Analytics: Query GitHub webhook data in ClickHouse to find the most failing workflow jobs, test pass/fail counts, or per-file test run history.
- Runner and Queue Metrics: Query Prometheus to inspect runner queue depth, jobs running per cluster, and infrastructure health over time ranges.
- Use Case: Investigate why a pull request is stuck by checking current queue depth per runner type and correlating with recent failure rates for the affected workflow.
Quick Start
Use the ci-metrics skill to list the top 10 most failing workflow jobs on pytorch/pytorch main in the last two weeks.