How orin29377053/ray-cluster works
Project type and purpose
• This is an ML-ops/infrastructure project: it automates deployment of a distributed Ray cluster on Minikube via KubeRay, runs Python scripts for parallel compute and YOLOv8 training, and integrates monitoring with Prometheus and Grafana.
File structure
• Top-level artifacts:
• Dockerfile, entrypoint.sh, Makefile.example → build & deploy container image
• k8s/ → Kubernetes manifests for RayCluster (ray-cluster.yaml), Ray service (ray-service.yaml), Prometheus & Grafana config
• dashboards/ & upload-dashboard.sh → Grafana dashboard JSONs plus upload tooling
• scripts/ → Python workloads (parallel_compute.py, unstable_task.py, training.py)
• init/download_dataset.py → data initialization
• Patterns: containerized service → deploy via Helm/kubectl → Ray head & worker pods → Python tasks submitted via Ray job API → metrics scraped by Prometheus → visualized in Grafana.
Key insights from README
• Single Minikube cluster with defined CPU/RAM quotas
• Helm installs for KubeRay operator, Prometheus, Grafana
• RayCluster consists of 1 head node + 3 workers
• Port-forwarding for Ray Dashboard (8265), Grafana (3000), Prometheus (9090)
• Python jobs submitted via ray job submit for compute, stability tests, training with Ray Train/Tune and YOLOv8
• Custom metrics via Ray util metrics, scraped by Prometheus → Grafana