NVIDIA Workload Identity Helm chart
Parent chart for one identity shard: SPIRE, Identity Keys, identity injection, the Identity Controller, Kata support, and DPF child-chart resources.
Prerequisites
Configuration
Identity Keys defaults to two replicas. The Identity Controller uses two when enabled. SPIRE deployment and datastore settings pass through to the upstream chart.
The JWT issuer is independent of exchanged SPIFFE IDs, which use
<tenant>.<region>.<service>.<partition>. Identities without a service label use
global.dsx.defaultServiceName.
Provision the host agent outside this chart. The matching
<registry>/<repository>/dsx-spire-agent:<chart-version>
image embeds that plugin at /opt/spire/plugins/dsx-imds-node-attestor-agent.
If the bundled agent is enabled for another node class, constrain it so it
cannot share nodes with the external host agent.
vmSupport.resources.networkSetup and vmSupport.resources.metadata configure
the injected VM sidecars. Defaults are 10m CPU / 32Mi memory for network
setup and 50m / 64Mi for metadata-service. Set both requests and limits
when overriding them. Equal CPU and memory requests and limits preserve the
Guaranteed QoS required for dedicated CPU placement.
VM IMDS networking follows the primary OVN allocation and supports IPv4-only, IPv6-only, and dual-stack guests. See the guest network requirements for supported bindings, addresses, and MAC requirements.
Each Pod selected for token-fetcher injection must set both
spiffe.nvidia.com/inject-dsx-token-fetcher: "true" and its relying-party
audience in spiffe.nvidia.com/token-audience. The
examples cover ordinary Kubernetes
Pods and DPUService deployment.
The fetcher health endpoint and startup probe use
podTokenFetcher.startupProbe.httpGet.port (default 8080). A Pod can override
both with spiffe.nvidia.com/token-fetcher-health-port: "18080" when another
container uses that port. The annotation must contain a valid TCP port number.
Keep the token-fetcher policy in a standalone DPUService. A policy-only
service in a DPUDeployment blocks DPUNodeMaintenance.
Configure the external DPF identity templates before provisioning DPUs. DPF registers DPU Agent and DPUService workload identities. The Identity Controller registers the paired bare-metal host.
identityController.sourceAPIPollInterval sets the discovery retry period for
absent DPU and Launch Layer APIs. It defaults to 1h and must be a positive
Go duration, such as 30m. Installed APIs use watches for source changes.
When dpuService.imds is enabled, the identity controller derives selectors
from the release namespace and dpuService.workload. An explicit
identityController.dpuMetadataSelectors list replaces those defaults. DPF’s
SPIFFE CSI bind-mounts /run/spire, preserving agent.sock for DSX IMDS.
Kata Pods run inside VMs, so SPIFFE CSI cannot carry the Workload API socket across the boundary. The proxy forwards it over vsock for selected Kata Pods, independently of token-fetcher injection.
The host proxy uses hostPID and writes the host cgroup, so Kata support
requires privileged Pod Security Admission, cgroup v2 with the systemd driver,
and sandbox_cgroup_only = true. The default nodeSelector targets
katacontainers.io/kata-runtime=true. List every enabled Kata RuntimeClass.
imagePullSecrets applies to chart-owned and admission-injected DSX images.
Each matched consumer namespace must contain Secrets with the same names.
Native Kubernetes admission policies inject identity sidecars and prevent
application containers from mounting the Workload API or NVIDIA Workload
Identity Keys Proxy sockets. Whenever spire.enabled is true, validation
enforces operator Pod service and role label formats.
Deploy
DPF identity configuration
Setting spec.security.spiffe in DPFOperatorConfig enables DPF’s
csi.spiffe.io driver for DPU workloads.
Configure DPF’s identity templates in the external DPFOperatorConfig before
provisioning DPUs. The local templates must include the tenant and service
segments required by NVIDIA Workload Identity Keys. This fragment shows
operator tenant operator-1, service dsx, region ds9, and partition
nvid.id:
Match the tenant to global.operatorTenant and the service, region, and
partition to global.dsx.defaultServiceName, global.dsx.region, and
global.dsx.partition. DPF supplies the normalized serial,
namespace, and service ID. The DPU Agent’s helper uses tokenExchangeEndpoint
before writing its Kubernetes API token. Its exchanged subject template must
match the token subject because DPF uses it for the DPU Agent’s RoleBinding.
DPUService token-fetcher sidecars use the exchange URL configured in the
identity chart. DPF validates the DPUService exchanged subject template but
does not consume that subject itself.
Set the remaining DPF SPIFFE fields, provide the trust bundle, and configure Kubernetes API authentication as described in the DPF SPIFFE guide.
Secrets
Create referenced Secrets in the release namespace before installing the chart. Keep secret values out of Helm values and source control. Secrets may be managed by any provider, provided they use the data keys and decoded contents below.
Signing keys may be SEC 1 EC, PKCS#1 RSA, or PKCS#8 asymmetric private keys.
Every key must support the algorithm selected by active-kid.json, and
active_kid must identify one key in keyset.json. Supported algorithms are
EdDSA, ES256, ES384, ES512, RS256, RS384, RS512, PS256, PS384,
and PS512.
Identity Keys reloads keyset.json and active-kid.json when either mounted
file changes. Filesystem notifications trigger reloads, and
identityKeys.fileWatchPollInterval sets the polling interval (default 30s).
Invalid or mismatched sources retain the last valid signer and public keys.
It rereads OAuth2 credentials when its cached access token expires.
Vault Secrets Operator example
The following example exports every version of an exportable Vault Transit signing key and transforms Vault values into the three Kubernetes Secret payloads:
Set the chart Secret references to the three destination.name values. Adapt
Vault mounts, paths, authentication, and source field names to the deployment.
Rotation
Signing-key rotation keeps old and new public keys available while issued JWTs remain valid:
- Add the new private key to the exported keyset without changing the active key.
- Wait for Secret synchronization and the new
kidin/.well-known/jwks.json. - Change
active_kidandalg, then wait for exchanged JWTs to use the new key. - Wait for JWTs signed by the old key to expire and downstream JWKS caches to refresh.
- Remove the old private key from the exported keyset. Its public key remains available until its NATS lease expires.
OAuth2 credential rotation does not require a restart. Create the replacement credential at the issuer, update Vault, wait for the mounted Secret to change and for Identity Keys to obtain a token with the replacement, then revoke the old credential.
Install
To install from source, download the locked SPIRE dependency:
Use ./charts/dsx-identity as <oci-chart-reference>.