GitOps (Argo CD)

Declarative continuous delivery for the inference platform: git is the source of truth, and Argo CD reconciles the cluster to match. This is the CD half of the platform's GitOps and CI setup. The CI half (.github/workflows/ci.yml) lints, schema-validates, and builds every change before it merges.

Applications

applications.yaml defines two Argo CD Applications:

Application	Syncs	Into
inference-serving	vllm-2gpu.yaml + keda-scaledobject-gpu.yaml	inference ns
inference-scaler	the Go external scaler Deployment/Service	inference ns

Both use automated sync with prune and selfHeal, so drift is corrected and deletions in git propagate. A directory include filter keeps the alternative manifests in gpu-node/ (vllm-plain.yaml, kserve-inferenceservice.yaml) out of the synced set, since they're documented options, not the running desired state.

Install & apply

kubectl create namespace argocd
kubectl apply -n argocd -f https://raw.githubusercontent.com/argoproj/argo-cd/stable/manifests/install.yaml
kubectl apply -f gitops/applications.yaml
argocd app list      # or the Argo CD UI

Same WSL2 caveat as KServe/KEDA

Argo CD ships admission/redis/repo-server components and (like cert-manager, KServe, and KEDA) is webhook- and in-cluster-networking-sensitive. On this broken-overlay cluster its controllers should be co-located with the API server (cordon the laptop during install), for the reasons in the troubleshooting log and architecture decisions. On a healthy CNI none of that is needed.

Status

The Application definitions are committed and CI-validated, and Argo CD is installed on the DIY cluster syncing the platform (the bootstrap reconstitutes it, co-locating its controllers with the API server per the caveat above). A Terraform/Helm IaC pass is the remaining delivery work.