Kubernetes CRDs & Operators: Extending the API, Episode 1
What Is a CRD? · CRDs You Already Use · CRD Anatomy · Write Your First CRD · CEL Validation · Controller Loop · Build an Operator · CRD Versioning · Admission Webhooks · CRDs in Production

TL;DR

• A Kubernetes CRD (Custom Resource Definition) is how you add new resource types to the Kubernetes API — the same way Deployment and Service exist natively, you can make BackupPolicy or Certificate exist too
  (CRD = the schema/blueprint; Custom Resource = an instance of that schema, just like a Pod is an instance of the Pod schema)
• Every kubectl get crds on a real cluster shows dozens of them — cert-manager, KEDA, Prometheus Operator, Crossplane all ship their own CRDs
• CRDs are served by the same API server as built-in resources — kubectl, RBAC, watches, and events all work identically
• A CRD alone does nothing — a controller watches the custom resources and acts on them; together they form an Operator
• CRDs live in etcd just like Pods and Deployments — they survive API server restarts and cluster upgrades
• You do not need to modify Kubernetes source code or restart the API server to add a CRD

The Big Picture

  HOW KUBERNETES CRDs EXTEND THE API

  ┌──────────────────────────────────────────────────────────────┐
  │  Kubernetes API Server                                       │
  │                                                              │
  │  Built-in resources          Custom resources (via CRD)      │
  │  ─────────────────           ──────────────────────────      │
  │  Pod                         Certificate     (cert-manager)  │
  │  Deployment                  ScaledObject    (KEDA)          │
  │  Service                     ExternalSecret  (ESO)           │
  │  ConfigMap                   BackupPolicy    (your team)     │
  │  ...                         ...                             │
  │                                                              │
  │  All resources: same API, same kubectl, same RBAC, same etcd │
  └──────────────────────────────────────────────────────────────┘
            ▲                          ▲
            │ built in                 │ registered at runtime
            │                         │
         Kubernetes              CustomResourceDefinition
          binary                    (a YAML you apply)

What is a Kubernetes CRD? It is a resource that defines resources — a schema registration that teaches the API server about a new object type you want to use in your cluster.

What Problem CRDs Solve

Kubernetes ships with roughly 50 resource types: Pods, Deployments, Services, ConfigMaps, Secrets, PersistentVolumes, and so on. These cover the general-purpose building blocks for running containerized workloads.

But the moment you operate real infrastructure, you hit the edges. You want to express:

• “This database should have three replicas with point-in-time recovery enabled” — not a Deployment
• “This TLS certificate for api.example.com should renew 30 days before expiry” — not a Secret
• “This queue consumer should scale to zero when the queue is empty” — not a HorizontalPodAutoscaler

Before CRDs (pre-2017), the only options were: use ConfigMaps as a poor substitute (no schema, no validation, no dedicated RBAC), or fork Kubernetes and add the resource natively (impractical for everyone outside the core team).

CRDs, introduced as stable in Kubernetes 1.16, solved this by letting you register a new resource type with the API server at runtime — without touching Kubernetes source code, without restarting the API server, without any special access beyond being able to create cluster-scoped resources.

The Kubernetes API: A Brief Mental Model

Before CRDs make sense, the API model needs to be clear.

  KUBERNETES API STRUCTURE

  apiVersion: apps/v1       ← API group (apps) + version (v1)
  kind: Deployment          ← resource type
  metadata:
    name: web               ← instance name
    namespace: default      ← namespace scope
  spec:
    replicas: 3             ← desired state

Every Kubernetes resource has:
– A group (e.g., apps, batch, networking.k8s.io) — or no group for core resources
– A version (e.g., v1, v1beta1)
– A kind (e.g., Deployment, Pod)
– A scope: namespaced or cluster-wide

The API server is a registry. Each group/version/kind combination maps to a Go struct that knows how to validate, store, and serve that resource type.

A CRD registers a new entry in that registry. You supply the group, version, kind, and schema. The API server handles everything else — serving it via REST, storing it in etcd, exposing it to kubectl.

What a CRD Looks Like

Here is the smallest possible CRD — it creates a new BackupPolicy resource type in the storage.example.com API group:

apiVersion: apiextensions.k8s.io/v1
kind: CustomResourceDefinition
metadata:
  name: backuppolicies.storage.example.com
spec:
  group: storage.example.com
  versions:
    - name: v1alpha1
      served: true
      storage: true
      schema:
        openAPIV3Schema:
          type: object
          properties:
            spec:
              type: object
              properties:
                schedule:
                  type: string
                retentionDays:
                  type: integer
  scope: Namespaced
  names:
    plural: backuppolicies
    singular: backuppolicy
    kind: BackupPolicy
    shortNames:
      - bp

Apply it:

kubectl apply -f backuppolicy-crd.yaml

Now create an instance:

apiVersion: storage.example.com/v1alpha1
kind: BackupPolicy
metadata:
  name: nightly
  namespace: default
spec:
  schedule: "0 2 * * *"
  retentionDays: 30

kubectl apply -f nightly-backup.yaml
kubectl get backuppolicies
kubectl get bp            # shortName works
kubectl describe bp nightly

The API server validates the spec against the schema, stores it in etcd, and returns it via all the standard API endpoints — all without a single line of custom code.

CRD vs Built-In Resource: What Is Different?

Not much, deliberately.

The last row is the critical one. When you create a Deployment, the deployment controller immediately starts managing ReplicaSets. When you create a BackupPolicy, nothing happens — until you write and deploy a controller that watches BackupPolicy objects and acts on them.

That controller + the CRD is what people call an Operator.

A Real Cluster: What You Actually See

Run this on any cluster running cert-manager, Prometheus Operator, or any other tooling:

kubectl get crds

Sample output (abbreviated):

NAME                                                  CREATED AT
certificates.cert-manager.io                          2024-11-01T08:12:00Z
certificaterequests.cert-manager.io                   2024-11-01T08:12:00Z
issuers.cert-manager.io                               2024-11-01T08:12:00Z
clusterissuers.cert-manager.io                        2024-11-01T08:12:00Z
scaledobjects.keda.sh                                 2024-11-01T08:13:00Z
scaledjobs.keda.sh                                    2024-11-01T08:13:00Z
externalsecrets.external-secrets.io                   2024-11-01T08:14:00Z
prometheuses.monitoring.coreos.com                    2024-11-01T08:15:00Z
servicemonitors.monitoring.coreos.com                 2024-11-01T08:15:00Z

Every tool that ships as a CRD-based system registers its resource types here first. The count often surprises engineers: a production cluster with a typical toolchain easily has 40–80 CRDs.

Check how many are on your cluster:

kubectl get crds --no-headers | wc -l

How the API Server Handles a CRD

When you apply a CRD, the API server does three things:

  CRD REGISTRATION FLOW

  kubectl apply -f my-crd.yaml
          │
          ▼
  1. API server validates the CRD manifest
     (is the schema valid OpenAPI v3? are names correct?)
          │
          ▼
  2. CRD stored in etcd
     (under /registry/apiextensions.k8s.io/customresourcedefinitions/)
          │
          ▼
  3. New REST endpoints activated immediately:
     GET  /apis/storage.example.com/v1alpha1/namespaces/{ns}/backuppolicies
     POST /apis/storage.example.com/v1alpha1/namespaces/{ns}/backuppolicies
     ...

From this point, any kubectl get backuppolicies or API call to those endpoints is handled exactly like a built-in resource call — the API server serves it from etcd, applies RBAC, runs admission webhooks, and returns standard JSON.

No restart required. The new endpoints appear within seconds.

The Difference Between CRD and CR

Two terms that are easily confused:

• CRD (CustomResourceDefinition) — the schema/blueprint. There is one CRD per resource type. certificates.cert-manager.io is a CRD.
• CR (Custom Resource) — an instance of a CRD. Every Certificate object you create is a custom resource. You can have thousands of CRs per CRD.

  CRD (one)          →  Custom Resource (many)
  ─────────             ─────────────────────
  certificates          web-tls           (namespace: production)
  .cert-manager.io      api-tls           (namespace: production)
                        admin-tls         (namespace: staging)
                        ...

The CRD is applied once (usually by the tool’s Helm chart). Custom resources are created by your users, your CI pipeline, or your GitOps system throughout the life of the cluster.

Where CRDs Fit in the Kubernetes Extension Model

CRDs are one of three ways to extend Kubernetes:

  KUBERNETES EXTENSION MECHANISMS

  1. CRDs + Controllers (Operators)
     Add new resource types + behavior
     → cert-manager, KEDA, Argo CD, Crossplane
     Used for: domain-specific abstractions, infrastructure management

  2. Admission Webhooks
     Intercept API requests to validate or mutate objects
     → OPA/Gatekeeper, Kyverno, Istio injection
     Used for: policy enforcement, sidecar injection, defaulting

  3. API Aggregation (AA)
     Register a fully separate API server behind the main API server
     → metrics-server, custom autoscalers
     Used for: when you need non-CRUD semantics (e.g. exec, attach, streaming)

For 95% of use cases, CRDs + controllers are the right mechanism. API aggregation is complex and only warranted for non-standard API semantics. Admission webhooks are complementary to CRDs, not an alternative.

Common Mistakes

Confusing the CRD with the controller. The CRD is just a schema registration — it does not execute code. If you apply a CRD but do not deploy its controller, creating custom resources will succeed (the API server accepts them) but nothing will happen. This catches many people the first time they try to use cert-manager by only applying the CRDs without installing the cert-manager controller.

Assuming CRD deletion is safe. Deleting a CRD deletes all custom resources of that type from etcd. There is no “are you sure?” prompt. If you delete the certificates.cert-manager.io CRD, every Certificate object in every namespace is gone.

Treating CRDs as ConfigMap replacements. Some teams store configuration in CRDs purely to get schema validation. This works, but without a controller, the custom resources are inert data. If you only need configuration storage with validation, a CRD is viable — just be explicit that there is no reconciliation loop.

Quick Reference

# List all CRDs in the cluster
kubectl get crds

# Inspect a specific CRD's schema
kubectl get crd certificates.cert-manager.io -o yaml

# List all custom resources of a type
kubectl get certificates -A

# Get details on a specific custom resource
kubectl describe certificate web-tls -n production

# Delete a CRD (WARNING: deletes all instances)
kubectl delete crd backuppolicies.storage.example.com

# Check if a CRD is established (ready to use)
kubectl get crd backuppolicies.storage.example.com \
  -o jsonpath='{.status.conditions[?(@.type=="Established")].status}'
# Returns: True

Key Takeaways

• A Kubernetes CRD registers a new resource type with the API server — no source code changes, no restart required
• Custom resources behave identically to built-in resources: kubectl, RBAC, watches, etcd, admission webhooks all work the same way
• The CRD is just the schema; a controller gives custom resources behavior — together they form an Operator
• Every production cluster running modern tooling already uses dozens of CRDs
• Deleting a CRD deletes all its instances — treat CRDs as production-critical objects

What’s Next

EP02: CRDs You Already Use makes this concrete before we go deeper — we walk through cert-manager’s Certificate, KEDA’s ScaledObject, and External Secrets’ ExternalSecret as working examples, so you understand what a well-designed CRD looks like from a user’s perspective before you design your own.

Get EP02 in your inbox when it publishes → subscribe at linuxcent.com

The post What Is a Kubernetes CRD? How Custom Resources Extend the API appeared first on Linuxcent.