K8s Crd From Typed Schema

k8s CRDs from a typed schema

Generating CRDs from typed schemas (zod / TypeBox / Valibot / hand-written JSON Schema) looks straightforward — convert to OpenAPI v3, embed as openAPIV3Schema, done. It isn't. CRDs use a subset of OpenAPI v3 plus k8s-specific extensions ("Structural Schema"), and several common JSON-Schema features are silently incompatible. Plus there are two non-schema gotchas (status subresource, metadata prohibition) that don't show up until you try to use the CRD.

This skill is the punch list of what to fix before kubectl apply -f crd.yaml succeeds and the operator actually works.

When to use

• Building an operator / controller and writing your own CRDs.
• Generating CRDs from a typed source rather than hand-authoring YAML.
• Hitting kubectl errors like unknown field "$schema" / must not have "oneOf" / must not specify anything other than name and generateName / cannot have both "additionalProperties" and "properties".
• Operator works once but then reconciles forever in a loop after the first status write.

Workflow

1. Convert typed schema → JSON Schema (OpenAPI v3 dialect)

For zod:

tsCopy code
import { zodToJsonSchema } from 'zod-to-json-schema';

const json = zodToJsonSchema(specSchema, {
  target: 'openApi3',
  $refStrategy: 'none',  // CRDs forbid $refs across the document
});

$refStrategy: 'none' is mandatory — k8s' Structural Schema does not resolve cross-document $ref. Inline everything.

For TypeBox: pass the schema through as-is, then run the adapter below.

2. Run the Structural-Schema adapter

The Structural Schema dialect rejects several JSON Schema features. Strip / rewrite them:

A reference adapter (TypeScript, ~40 lines) lives at examples/adapter.ts in this skill (or copy from k1c/src/cli/export-crds.ts).

3. Wrap the spec schema in the CRD envelope

yamlCopy code
apiVersion: apiextensions.k8s.io/v1
kind: CustomResourceDefinition
metadata:
  name: <plural>.<group>     # e.g. r2buckets.cloudflare.k1c.io
spec:
  group: <group>
  scope: Namespaced
  names:
    kind: <Kind>
    singular: <kind>
    plural: <plural>
    listKind: <Kind>List
  versions:
    - name: v1alpha1
      served: true
      storage: true
      schema:
        openAPIV3Schema:
          type: object
          properties:
            apiVersion: { type: string }
            kind:       { type: string }
            spec:       <your adapted schema here>
            status:
              type: object
              x-kubernetes-preserve-unknown-fields: true
          required: [spec]
      subresources:
        status: {}            # <-- mandatory if your operator writes status

4. Compute the plural correctly

spec.names.plural ends up in the kubectl URL (/apis/<group>/<version>/<plural>). Don't kind.toLowerCase() + 's' — handle the regular-verb cases:

tsCopy code
function pluralize(kind: string): string {
  const lower = kind.toLowerCase();
  if (lower.endsWith('s') || lower.endsWith('x') || lower.endsWith('z')) return `${lower}es`;
  if (lower.endsWith('y')) return `${lower.slice(0, -1)}ies`;
  return `${lower}s`;
}

This matches the inflection Cloudflare and most k8s API authors use. Irregular kinds (Endpoints, Quota → quotas) need a manual override map.

5. Verify with kubectl in dry-run mode

shCopy code
kubectl apply --dry-run=server -f crd.yaml
kubectl apply --dry-run=server -f sample-cr.yaml   # after CRD is registered

Server-side dry-run catches structural-schema violations that local dry-run misses (e.g. additionalProperties + properties is only enforced server-side).

Pitfalls

Pitfall 1: declaring metadata in openAPIV3Schema.properties

yamlCopy code
# ❌ kubectl rejects this:
openAPIV3Schema:
  type: object
  properties:
    metadata: { type: object, properties: { name: { type: string } } }
    spec: ...
# error: must not specify anything other than name and generateName for metadata

Fix: don't list metadata at all. The apiserver auto-applies the standard ObjectMeta schema. Only declare apiVersion / kind / spec / status.

Pitfall 2: missing subresources.status: {} causes infinite reconcile loops

Without this, every kubectl patch --subresource=status call:

1. Bumps .metadata.generation.
2. Wakes up the watch.
3. Operator re-reconciles → patches status again.
4. Goto 1.

Symptom: operator looks healthy but logs show reconcile loop lines firing every few hundred ms forever. CPU climbs slowly. Apiserver QPS budget burns.

Fix: always include subresources: { status: {} } in every served version.

Pitfall 3: oneOf from discriminated unions

z.discriminatedUnion(...) and Type.Union(...) lower to oneOf. Structural schema rejects this outside x-kubernetes-validations (CEL). Two fixes:

1. Permissive escape: replace the node with x-kubernetes-preserve-unknown-fields: true (loses validation).
2. Strict CEL: emit x-kubernetes-validations: [{ rule: "self.type == 'a' && has(self.payload)" }]. Only apiserver ≥ 1.25 supports this.

The adapter in step 2 picks option 1 by default. Bump to option 2 only when the validation is load-bearing.

Pitfall 4: structural schema requires type on every node

z.unknown() lowers to an empty {} JSON Schema. Structural schema rejects nodes without type (or enum, or the escape hatch). Stamp x-kubernetes-preserve-unknown-fields: true on every empty leaf during adaptation.

Pitfall 5: only one schema can be the storage version

versions[N].storage: true must be exactly one entry per CRD across all served versions. Multiple storage: true → apiserver refuses the CRD on registration. When adding v1beta1 → v1, flip exactly one over.

Verifying without a real cluster

Stand up kind in CI and run:

shCopy code
kubectl apply -f crd.yaml                              # registers the CRD
kubectl apply --dry-run=server -f every-example.yaml   # validates each CR against the new schema

This is exactly what k1c does in .github/workflows/k8s-validate.yml — see that workflow for a reference job.

Why this matters

CRDs are a contract between humans, the apiserver, and your operator. The Structural Schema rules are not arbitrary — they exist so the apiserver can do server-side validation, defaulting, and pruning. But the rules aren't documented in one place: the official Structural Schema docs cover the dialect, the subresource docs cover /status, and the metadata prohibition only shows up as a runtime apiserver error. This skill collapses all four into one workflow.

Related skills

• cloudflare-deploy — for Cloudflare-specific resource shapes if you're operating on CF (independent of CRD generation).
• apm-usage — for shipping the generated CRD bundle as a skill artifact.