Problem
The self-managed stack can enable PKI for the LLM request router and render a
cert-manager Certificate that references the nvcf-openbao-pki
ClusterIssuer. The stack also installs cert-manager by default and provisions
the corresponding OpenBao signing role, but it does not currently create that
ClusterIssuer.
As a result, the Helm render can succeed while the request-router certificate
remains unissued at runtime because its referenced issuer does not exist.
Expected behavior
When self-managed LLM PKI is enabled, the stack should:
- consume a versioned
helm-nvcf-pki chart maintained under deploy/helm;
- create the
nvcf-openbao-pki ClusterIssuer;
- configure it for the current OpenBao service-issuing PKI role;
- authenticate using the cert-manager ServiceAccount identity;
- depend on OpenBao and, when the stack manages cert-manager, on that release;
- continue to support an externally managed cert-manager installation;
- preserve the existing PKI-disabled behavior;
- fail rendering when required issuer configuration is missing; and
- include render tests for enabled, disabled, external-controller, and invalid
configurations.
Security considerations
- Do not embed credentials, tokens, environment-specific endpoints, or private
deployment values in the chart.
- Keep the issuer name distinct from other certificate authorities.
- Keep the OpenBao signing path and authentication role least-privileged.
- Ensure test output and committed fixtures contain no secrets.
Coordination
PR #483 adds a separate compute-plane webhook PKI. The self-managed LLM issuer
should use distinct resource names and avoid creating conflicting ownership of
shared cert-manager resources.
Acceptance criteria
- A rendered PKI-enabled self-managed stack contains exactly one
ClusterIssuer named nvcf-openbao-pki.
- Its signing path matches the current OpenBao service-issuing role used by the
LLM PKI migration.
- The issuer release depends on OpenBao and additionally depends on cert-manager
when the stack owns the cert-manager release.
- The external cert-manager mode renders without depending on a disabled Helm
release.
- The request-router
Certificate references the same issuer name.
- The default PKI-disabled path does not render the issuer.
- The chart is published before a self-managed stack version that pins it is
promoted.
- No sensitive values are introduced into public files, fixtures, logs, or
generated artifacts.
Problem
The self-managed stack can enable PKI for the LLM request router and render a
cert-manager
Certificatethat references thenvcf-openbao-pkiClusterIssuer. The stack also installs cert-manager by default and provisions
the corresponding OpenBao signing role, but it does not currently create that
ClusterIssuer.
As a result, the Helm render can succeed while the request-router certificate
remains unissued at runtime because its referenced issuer does not exist.
Expected behavior
When self-managed LLM PKI is enabled, the stack should:
helm-nvcf-pkichart maintained underdeploy/helm;nvcf-openbao-pkiClusterIssuer;configurations.
Security considerations
deployment values in the chart.
Coordination
PR #483 adds a separate compute-plane webhook PKI. The self-managed LLM issuer
should use distinct resource names and avoid creating conflicting ownership of
shared cert-manager resources.
Acceptance criteria
ClusterIssuernamednvcf-openbao-pki.LLM PKI migration.
when the stack owns the cert-manager release.
release.
Certificatereferences the same issuer name.promoted.
generated artifacts.