Purdue Analysis Facility

GitOps source of truth for the Purdue Analysis Facility — a Kubernetes-based interactive analysis platform for high energy physics research at CMS experiment.

Everything the cluster runs is declared here and reconciled by Flux.

User documentation: analysis-facility.physics.purdue.edu Admin documentation: https://purdue-cms-tier2.gitlab.io/documentation

Platform at a glance

Orchestration	Kubernetes on the Geddes cluster; Flux CD (roots)
Sessions	JupyterHub — JupyterLab and code-server interfaces, CILogon auth
Scale-out	Dask Gateway
Training	Ray on KubeRay — GPU training sent from notebooks, one cluster per user (gateway)
User environment	the purdue-af image, pixi environments
Data	CVMFS, XRootD, XCache, EOS, Depot NFS; ServiceX for columnar delivery
Inference	SuperSONIC
Observability	Prometheus, Grafana, Loki, Pyroscope, Alloy + purpose-built exporters
Agents	MCP server exposing AF-specific tools to any MCP client

Component status

Whether each component on the cluster is running what is on main (which ref each root deploys; ).

Core —

Experimental

Images — which tag each one ships on: RELEASING.md.

Reading the badges:

• deployed — no drift, the cluster has it
• awaiting release — validated, and only a release stands between it and the cluster
• validating — CI has not finished on those commits yet
• failed CI — do not release; the drift is broken
• the trailing number is how many commits it has moved since it was deployed
• a leading X.Y.Z is the deployed version, for images on a version stream

Recomputed hourly and after every CI run on main by component-status.yml; the badge data lives on the status branch, so keeping it current never touches main.

How a change reaches the cluster, version rules and rollback: RELEASING.md.