Orkestra

A runtime for Kubernetes operators.

Declare. Run.

Quick Start · Docs · Discussions · Early Access Slack

⚠️ Pre‑alpha — actively building. APIs may change. Feedback shapes what ships next.

---

You have a CRD. Kubernetes stores it, validates it, and serves it.

The only missing piece is something that watches it and acts on it.

Traditionally that means Go: informers, workqueues, reconcile loops, code generation, Dockerfiles, Helm charts — a software project per operator. Most engineers never start. Teams that do spend weeks before the first CR reconciles.

Orkestra removes that entirely.

---

Declare

apiVersion: orkestra.orkspace.io/v1
kind: Katalog
metadata:
  name: website-operator
spec:
  crds:
    website:
      crdFile: ./crd.yaml
      crFiles: [./cr.yaml]
      operatorBox:
        onCreate:
          deployments:
            - name: "{{ .metadata.name }}"
              image: "{{ .spec.image }}"
              replicas: "{{ .spec.replicas }}"
              reconcile: true
          services:
            - name: "{{ .metadata.name }}-svc"
              port: 80
              targetPort: "{{ .spec.port }}"
              reconcile: true

That is the whole operator.

Run

ork run

Orkestra reads the Katalog, applies the CRD and CR, starts the operator, creates the Deployment and Service, sets owner references on both, writes status, emits Kubernetes events, corrects drift, and exposes health, metrics, and a control center.

Not a single line of Go.

Your CRD is enough. The rest is just a Katalog.

---

What every CRD gets

Every CRD declared in a Katalog becomes a complete, isolated operator. Nothing to configure.

Informer	Watches your exact GVK. In-memory cache. Zero API calls on read.
Workqueue	Per-CRD. Rate-limited. Deduplicated. Isolated from every other CRD.
Worker pool	Configurable concurrency. A panic in one CRD does not affect any other.
Drift correction	reconcile: true — desired state is enforced on every cycle.
Owner references	Child resources deleted when the CR is deleted. No onDelete logic needed.
Finalizers	CRs protected from dirty deletion automatically.
Events	Every reconcile is a traceable Kubernetes event.
Leader election	One active instance. Followers hold warm caches. Failover in under 15s.
Status	Ready condition + your own status fields written after every reconcile.
Health API	/katalog/{crd}/health, /katalog/{crd}/cr, /metrics — per CRD.
Prometheus metrics	Reconcile totals, queue depth, error rate — labeled by GVK.
Deletion protection	Orkestra and everything it manages cannot be accidentally kubectl delete.
Control Center	Realtime visibility per CRD, per Katalog, across instances. Auto-generated operator docs — overview, reconcile mode, child resources, kubectl reference, access control.

---

Getting started

# Install (macOS)
brew install orkspace/tap/ork orkspace/tap/orkcc

# Install (Linux)
curl -sSL https://get.orkestra.sh | bash

# Initialize and run
ork init --pack beginner
cd beginner/01-hello-website
ork run

No cluster? Add --dev to create a temporary kind cluster. Requires Docker.

ork init with no arguments scaffolds a katalog.yaml, crd.yaml, and cr.yaml in the current directory — like terraform init. The --pack flag pulls a collection of runnable examples.

→ Learning to Orkestrate — the guided path from first operator to full platform. Every capability has a runnable example.

---

Control Center

In another terminal:

ork control

→ localhost:8081 · username:password → orkestra

Six Runtimes. 75 CRDs. One Control Center.

Live deployment: cc.orkestra.sh

---

Numbers

	Traditional (75 operators)	Orkestra
Processes	75	6 runtimes + 1 control center
Memory	3.75 GB – 15 GB	~79 MB per runtime (measured)
CRDs under management	75	75
First operator	3–6 weeks	Under 1 hour
Lines of Go	400+ per operator	0
Adding a new CRD	Days to weeks	Minutes

79 MB is a live measurement from a 10-CRD runtime (process_resident_memory_bytes from the /metrics endpoint — raw scrape). The memory reduction works because Orkestra pays the cost of client-go, leader election, and health servers once per runtime. Per-CRD cost is a goroutine pool and an in-memory cache. Isolation works the same way kube-controller-manager isolates Deployment, StatefulSet, and Job controllers — dedicated informer, queue, and worker pool per CRD. A panic in one is caught by safeReconcile; the others keep running. The Control Center aggregates all runtimes into a single dashboard.

---

What Orkestra is not

CRD generation is a starting point, not the source of truth. ork generate crd scaffolds a base CRD from your Katalog. You own the final schema — add validation, printer columns, and version history to it. crdFile just points to whatever CRD file you maintain.

It does not replace Go for complex logic. Hooks and constructors exist for exactly this reason. ~90% of operators are declarative structure; ~10% need code. Orkestra handles the 90% and gives the 10% a clean interface.

External infrastructure providers are in development. For AWS, MongoDB, or cloud DNS alongside Kubernetes resources, use Crossplane for external infrastructure and Orkestra for the application layer. The two complement each other.

It does not auto-sync from Git. Configuration is resolved at startup and locked in. Katalogs define long-lived API contracts; silently reloading them is dangerous. Use a deployment pipeline like any other runtime change.

---

Documentation

Why Orkestra	What Orkestra is, how it works, and why it’s different
Trust and Failure Model	What happens when things go wrong
Getting Started	First operator in under an hour
Learning to Orkestrate	Every capability, as a runnable example
Katalog Reference	Complete field reference
Orkestra Registry	OCI distribution for operators
Security	Deletion protection, RBAC, admission

---

Community

Issues · Discussions · Contributing

---

Apache 2.0 — see LICENSE