cgn-operator — Kubernetes controller

kube-rs reconciler that watches Cognitora CRDs and mirrors desired state into etcd (/cognitora/*) so cgn-router and cgn-agent behave the same as in bare-metal profiles — no second control plane.

Overview

The operator translates declarative Kubernetes objects into the same routing policies and node expectations that scripts/run/up.sh profiles use. Helm installs the Deployment; CRDs ship under deploy/kubernetes/crds/.

CRDs

CRD	File	Purpose
`InferenceCluster`	`inferencecluster.yaml`	Cluster identity, etcd endpoints, TLS/auth bootstrap hints
`ModelPool`	`modelpool.yaml`	Model deployments (engine, tensor parallelism, resources)
`RoutingPolicy`	`routingpolicy.yaml`	Score weights, admission, cascade / disagg toggles

YAML shapes mirror the examples in Kubernetes guide.

Features

Single reconciliation loop per CRD kind — edits propagate to etcd for live router reload
Works alongside the packaged Helm chart (helm install … ./deploy/kubernetes/helm/cognitora or published OCI chart)
No Go runtime — pure Rust/Kubernetes dependency chain via kube-rs

Architecture

kubectl apply → Kubernetes API → cgn-operator watch → etcd (/cognitora/routing/policy, …) → cgn-router watches.

Example (`ModelPool` + `RoutingPolicy`)

apiVersion: cognitora.dev/v1alpha1
kind: ModelPool
metadata:
  name: llama3-70b
  namespace: cognitora
spec:
  model: meta-llama/Meta-Llama-3-70B-Instruct
  tensorParallel: 4
  dtype: bfloat16
  replicas: 2
---
apiVersion: cognitora.dev/v1alpha1
kind: RoutingPolicy
metadata:
  name: default
  namespace: cognitora
spec:
  scoreWeights:
    kv: 0.55
    load: 0.25
    power: 0.10
    capacity: 0.10
  admission:
    maxQueue: 8192
    ttftSloMs: 600

Field names follow deploy/kubernetes/crds/; use kubectl explain modelpool.spec after kubectl apply -f deploy/kubernetes/crds/.

Dependencies

Kubernetes 1.28+
etcd reachable from the cluster (same requirement as bare metal)
cgn-router / cgn-agent DaemonSets or Deployments charted to talk to that etcd