Install and operate NVCF self-hosted control-plane and separate compute-plane stacks. Covers Helmfile values and CLI profile installation flows, teardown, values overrides, pull secrets, and troubleshooting. Use for nvcf-self-managed-stack, nvcf-compute-plane-stack, split compute-plane installation, control-plane installation, CLI-generated control-plane profiles, Helmfile, self-managed, or self-hosted deployments. Do NOT use for local k3d environments; use the local k3d development workflow instead.
Instrucciones de origen · Vista previa de solo lectura
name
nvcf-self-managed-installation
description
Install and operate NVCF self-hosted control-plane and separate compute-plane stacks. Covers Helmfile values and CLI profile installation flows, teardown, values overrides, pull secrets, and troubleshooting. Use for nvcf-self-managed-stack, nvcf-compute-plane-stack, split compute-plane installation, control-plane installation, CLI-generated control-plane profiles, Helmfile, self-managed, or self-hosted deployments. Do NOT use for local k3d environments; use the local k3d development workflow instead.
Operational guide for matching NVCF control-plane and compute-plane Helmfile bundles.
Instructions
Use this skill for install, upgrade, or teardown work in matching nvcf-self-managed-stack and nvcf-compute-plane-stack bundles; keep Helmfile values and CLI profile handoffs separate. For functionType: "LLM", read LLM Function Enablement.
Prerequisites
Ask for the extracted nvcf-self-managed-stack path. For split installation,
also require the matching-version nvcf-compute-plane-stack path. Verify both:
ls <control-plane-stack>/helmfile.d/ <control-plane-stack>/environments/ <control-plane-stack>/secrets/ <control-plane-stack>/global.yaml.gotmpl
ls <compute-plane-stack>/helmfile.d/ <compute-plane-stack>/environments/ <compute-plane-stack>/global.yaml.gotmpl
If either directory is missing, download matching bundle versions from NGC:
ngc registry resource download-version <org>/nvcf-self-managed-stack:<version>
ngc registry resource download-version <org>/nvcf-compute-plane-stack:<version>
Control-plane commands below assume the control-plane bundle root. Split-flow
commands and root substitution are in
Split Compute-Plane Installation.
Before You Start
Verify tooling and context before any operation:
helmfile --version # Must be 1.1.x (1.2.0 removed sequential mode)
helm version # Must be >= 3.12
helm plugin list # Must include helm-diff >= 3.11
kubectl version # Client must be within 1 minor version of cluster
Control-plane commands run from the extracted nvcf-self-managed-stack/ root:
cd path/to/nvcf-self-managed-stack
ls helmfile.d/ environments/ secrets/ global.yaml.gotmpl
Identify the environment from environments/<name>.yaml and
secrets/<name>-secrets.yaml; for EKS, use the canonical
CSP End-to-End Example.
How Values Flow
Understanding value precedence prevents the most common configuration mistakes.
environments/base.yaml (defaults)
-> merged with
environments/<env>.yaml (your overrides)
-> consumed by
global.yaml.gotmpl (Go template, constructs per-chart values)
-> consumed by
secrets/<env>-secrets.yaml (sensitive values)
-> overridden by
release inline values: blocks (highest precedence)
Critical: global.yaml.gotmpl only passes through specific keys to each chart (image registries, node selectors, storage, replica counts). Setting an arbitrary chart value in the environment file will be silently ignored if global.yaml.gotmpl does not propagate it.
1. Create namespaces and image pull secrets (if using a private registry)
for ns in cassandra-system nats-system nvcf api-keys ess sis vault-system cert-manager; do
kubectl create namespace "$ns" --dry-run=client -o yaml | kubectl apply -f -
done# Only if pulling from a private registry (e.g., NGC nvcr.io)export NGC_API_KEY="<your-key>"for ns in cassandra-system nats-system nvcf api-keys ess sis vault-system cert-manager; do
kubectl create secret docker-registry nvcr-creds \
--docker-server=nvcr.io \
--docker-username='$oauthtoken' \
--docker-password="$NGC_API_KEY" \
--namespace="$ns" \
--dry-run=client -o yaml | kubectl apply -f -
done
If using pull secrets, you must also configure each helmfile release to reference them. See Image Pull Secrets.
2. Authenticate helm to your chart registry
Both docker login AND helm registry login are required for NGC. Helmfile uses helm (not docker) to pull OCI charts, so the helm credential store must be authenticated separately.
Note: helm registry login keeps only the last login per host (multi-org gotcha).
The stack and the caches span several NGC orgs that need different keys, but they all
live under the one host nvcr.io. A second helm registry login nvcr.io (or
docker login) with a different key silently overwrites the first, so pulls for the
first org then fail with a misleading 403 Access Denied (not a 401: the auth
succeeded, it's the wrong-org key). If you pull charts/images from more than one org,
either re-run the login with the correct key immediately before each pull, or pass
--username '$oauthtoken' --password <key> per helm pull invocation instead of relying
on the stored login. (Pulling everything from a single registry mirror avoids this
entirely.)
2b. (Recommended) Pull from a registry the cluster authenticates to automatically
If you mirror the NVCF images into a registry that the cluster can pull from without an
explicit image pull secret (for example, a registry where the node or workload identity is
granted pull access automatically), point the stack at it and skip the credential steps
above.
Point the stack at the mirror in environments/<env>.yaml:
global:image:registry:<your-registry># e.g. registry.example.comrepository:<mirror-sub-path>helm:sources:# only if you also mirrored the helm chartsregistry:<your-registry>repository:<mirror-sub-path>
The manifest has a mix of OCI and non-OCI artifacts, so convert non-OCI charts before
mirroring. Container images (nvcr.io/...) are already OCI and mirror as-is. Some Helm
charts, though, are published to an https:// Helm repository (for example
helm.ngc.nvidia.com). Those are not OCI and cannot be resolved by global.helm.sources
(OCI) or mirrored directly. Before pointing the stack at your mirror, convert each
https:// chart to OCI: helm repo add/helm pull the .tgz, then
helm push <chart>.tgz oci://<your-registry>/<mirror-sub-path>. Only OCI charts can live
under global.helm.sources.
Skip the credential steps above. When the cluster authenticates to the registry
automatically (the node or workload identity has pull access), the pull needs no
Kubernetes secret, so you do not need: the per-namespace nvcr-creds from step 1
(still create the namespaces, just not the secret), the global.imagePullSecrets entry,
or the docker login/helm registry login nvcr.io from step 2.
One residual credential: nvcf-api still validates the user function image at
function create via the account registry credential, so that credential must point at
whichever registry holds the user image (set it to your mirror if the image is mirrored
there).
3. Set the route domain
global.domain sets the hostname suffix for routes such as api.<domain>. It does not configure the Gateway or its load balancer. Query the exact existing Gateway:
GATEWAY_ADDRESS=$(kubectl get gateway <gateway-name> -n <gateway-namespace> -o jsonpath='{.status.addresses[0].value}')
test -n "$GATEWAY_ADDRESS"
The address may be a hostname or IP. Cloud provider and Gateway configuration determine the load balancer type and address format. The self-managed stack does not select either.
For temporary testing without DNS, use a stable reserved suffix such as global.domain: "nvcf.test", connect to GATEWAY_ADDRESS, and send Host: api.nvcf.test. For production, use your base DNS domain and point its route records to GATEWAY_ADDRESS. If the endpoint later changes, keep global.domain stable and update only DNS or the client destination. See Example 4 in examples.md.
Set BOTH Cassandra passwords in the secrets file. The secrets file documents
DEFAULT_CASSANDRA_PASSWORD, but the cassandra chart has a second credential,
cassandra.serviceRolePassword, that defaults to ch@ng3m3. If you set only
DEFAULT_CASSANDRA_PASSWORD, the service role keeps that default and every DB-backed
service crash-loops with Bad credentials once it tries to authenticate. Set
cassandra.serviceRolePassword to the same value as DEFAULT_CASSANDRA_PASSWORD.
5. Deployment phases
Helmfile deploys in order with dependencies:
Phase
Selector
Services
Wait time
1
release-group=dependencies
NATS, OpenBao, Cassandra
5-10 min
2
release-group=services
API, SIS, ESS, invocation, grpc-proxy, notary, api-keys, optional LLM gateway/router, optional Vanity Gateway, NVCF UI when the stack package includes the addon
5-10 min
3
release-group=ingress
Gateway routes
1-2 min
4
release-group=observability
Observability stack (if enabled)
1-2 min
Monitor in a separate terminal:
kubectl get pods -A -w
6. Enable Vanity Gateway (optional)
Vanity Gateway is disabled by default and is only needed for customer-facing
hostnames or path mappings in front of the standard NVCF API and invocation
routes. It is available only in stack packages that include the Vanity Gateway
addon. If the extracted stack package does not contain a vanity-gateway
release and vanityGateway route values, skip these commands and use a newer
stack package that includes them.
If the stack package includes the addon, enable it in
environments/<env-name>.yaml:
The default route host is vanity.<domain> and the backend is
vanity-gateway.nvcf:8080. Put deployment-specific host and path mappings under
addons.vanityGateway.mappingConfig. If you need custom vanity hosts instead of
vanity.<domain>, use the route hostname overrides supported by the stack
package and create matching DNS records.
After confirming the package includes the vanity-gateway release, preview and
apply the service and route:
Verify the route only when the addon is present and enabled:
kubectl get deploy,svc -n nvcf -l app.kubernetes.io/name=vanity-gateway
kubectl get httproute -A | grep -i vanity
curl -H "Host: vanity.<domain>""http://<gateway-address>/health"
Do not enable Vanity Gateway for standard API, API Keys, invocation, LLM
invocation, or gRPC traffic. Those routes are provided by the base gateway
routes.
7. Enable NVCF UI (optional)
NVCF UI is an optional, customer-facing admin-panel UI. It is disabled by
default and available only in stack packages that include the NVCF UI addon. If
your extracted stack package does not contain a nvcf-ui release and nvcfUi
route values, skip this section. The NVCF UI admin panel is currently
unauthenticated; do not expose it to the public internet.
For enablement, namespace and pull secret setup, apply, and verification steps,
see NVCF UI Addon.
8. Install the separate compute-plane stack
The control-plane stack does not install compute-plane components. After it is healthy, choose exactly one compute-plane handoff:
Helmfile values flow: author separate environment files, then register and
install with the compute-plane Makefile.
CLI profile flow: use the profile generated by the CLI control-plane install,
then run the complete CLI register and install commands.
Do not mix these handoffs. Follow Split Compute-Plane Installation
for commands, EKS routing, pull secrets, verification, and teardown order.
Clean Teardown
Destroy every compute plane before the control plane, using the teardown for
the same handoff; see Split Compute-Plane Installation.
Scope: Only destroy releases managed by this control-plane helmfile stack. The NVCF releases are: nats, openbao-server, cassandra, api-keys, sis, api, invocation-service, grpc-proxy, ess-api, notary-service, optional vanity-gateway, nvcf-ui, admin-issuer-proxy, and ingress; cert-manager is included only when certManager.enabled: true. The control-plane namespace inventory also includes cert-manager, but default namespace cleanup preserves it. Namespace preservation does not preserve the Helm release. External cert-manager must use certManager.enabled: false; otherwise Helmfile treats the release as stack-managed. Do not delete other helm releases or namespaces on the cluster.
Before control-plane destroy, identify whether cert-manager is external. For an external installation, require effective certManager.enabled: false and abort if it is true. For an intended stack-managed release with effective certManager.enabled: true, verify matching stack Helm release metadata and abort on missing or mismatched evidence. Preserve the namespace by default; only plan its removal after verifying namespace provenance and getting explicit confirmation.
Standard teardown
Run from inside the nvcf-self-managed-stack/ directory:
HELMFILE_ENV=<env-name> helmfile destroy
Delete namespaces
for ns in cassandra-system nats-system nvcf api-keys ess sis nvcf-ui vault-system; do
kubectl delete namespace "$ns" --ignore-not-found
done
Verify clean
kubectl get ns | grep -E '(cassandra|nats|vault|nvcf|api-keys|ess|sis)'# Should return empty
Overriding Helm Chart Values
Environment file (limited)
Works only for keys that global.yaml.gotmpl propagates (e.g., cassandra.replicaCount, global.storageClass, global.image).
Release values block (any chart value)
Edit the release in helmfile.d/*.yaml.gotmpl and add a values: block. In the examples below, <private-values> refers to the secrets/ directory at the helmfile stack root.
YAML merge gotcha: When a release uses <<: *dependency or inherit, specifying values: replaces the template's values list. You must re-include global.yaml.gotmpl and the secrets file.
There are three distinct credential types. Do not conflate them, and mind the
format rule below (two of them fail even with the correct key if the format is wrong):
Control Plane Pull Secrets
API Bootstrap Registry Creds
Sidecar Image-Pull Secret
Purpose
K8s pulls NVCF service images
nvcf-api validates the user function image at function create
nvcf-api pulls the platform sidecars injected into every worker pod
NVCF_API_SIDECARS_IMAGE_PULL_SECRET in <env>-secrets.yaml -> OpenBao services/nvcf-api/kv/sidecars/image-pull-secret (field secret)
Format
standard .dockerconfigjson (kubectl builds it for you)
base64("$oauthtoken:<key>")
base64("$oauthtoken:<key>")
Note: the account-bootstrap cred and the sidecar secret are NOT a
dockerconfigjson. The secrets template ships the placeholder
REPLACE_WITH_BASE64_DOCKER_CREDENTIAL, which reads as "base64 of a dockerconfigjson."
It isn't. Both values are consumed as a raw docker auth string: nvcf-api takes the
value verbatim and drops it into the auth field of the dockerconfig it builds. So the
value must be base64("$oauthtoken:<key>"), e.g. printf '$oauthtoken:%s' "$KEY" | openssl base64 -A,
not base64(<dockerconfigjson>).
Wrong account-bootstrap value: function create returns
400 "must be base64 encoded username:password format".
Wrong sidecar value: every worker's sidecar pull gets a malformed credential and
403s, regardless of which key is inside. The symptom looks like a bad or wrong-org
key, but the key is fine and the encoding is the bug. To repair an already-installed
stack without a full re-sync:
AUTH=$(printf'$oauthtoken:%s'"$KEY" | openssl base64 -A)
bao kv put services/nvcf-api/kv/sidecars/image-pull-secret secret="$AUTH"# if NVIDIA Cloud Tasks is enabled, repeat for services/nvct-api/kv/sidecars/image-pull-secret
kubectl rollout restart -n nvcf deploy/nvcf-api # picks up the new sidecar cred
Note (multi-org): when pulling from NGC (not a mirror), the user function image and
the platform sidecars often live in different nvcr.io orgs, each needing a different
key, so the account-bootstrap cred (function image) and the sidecar secret (sidecars)
are frequently two different keys. Mirroring everything into a single registry the
cluster authenticates to automatically collapses this to one credential path.
Configuring with Kyverno (recommended)
Use a Kyverno mutating admission policy to automatically inject imagePullSecrets into all pods in NVCF namespaces. This works uniformly for all charts -- no per-chart configuration or helmfile modifications needed.
kubectl get pods -A -o wide # All pods
helm list -A # All helm releases
kubectl get events -n <ns> --sort-by='.lastTimestamp'# Recent events
Common failure patterns
Symptom
Cause
Fix
ImagePullBackOff + 401 Unauthorized
Missing or wrong pull secret
Check secret exists, check SA has imagePullSecrets
Init:0/1 stuck on service pods
Vault-agent waiting for OpenBao
Check OpenBao pods + migration job status
Init:0/1 + vault-agent-init shows auth/jwt/login -> 400 ... no known key successfully validated the token signature
OpenBao JWT auth configured with a static pubkey instead of the cluster's live JWKS -- the base default openbao.migrations.issuerDiscovery.enabled: false. Common on AKS / any OIDC-issuer cluster
Set openbao.migrations.issuerDiscovery.enabled: true, delete job openbao-server-migrations, re-sync openbao-server. See references/debugging.md "Init:0/1 Stuck" fix #4.
OOMKilled on Cassandra
Default resources too small
Override cassandra.resources via values block
DB-backed services crash-loop Bad credentials connecting to Cassandra
Secrets file set only DEFAULT_CASSANDRA_PASSWORD; cassandra.serviceRolePassword still defaults to ch@ng3m3
Set cassandra.serviceRolePassword in the secrets file equal to DEFAULT_CASSANDRA_PASSWORD, then re-sync cassandra + the DB-backed services
function / account calls return 404 Unknown client_id
The account-bootstrap post-install hook never ran: the api release install failed, or was repaired with live patches instead of a clean re-sync, so the hook (and the account) was skipped
Services fail to read vault secrets; secrets.json not found
Vault path hardcoded to /home/app/vault/ in _helpers.tpl; the runtime resolves the mounted path relative to the working directory and drops the leading /
Override podAnnotations and set JAVA_TOOL_OPTIONS: "-Duser.dir=/" in release values
NATS connection fails at startup; placement tag mismatch
NATS server tags hardcoded to dc:ncp; app derives tag from AWS_REGION (e.g., us-gov-west-1)
Set AWS_REGION=ncp and NVCF_AWS_REGION=ncp in env config
Verify the account bootstrap ran
The default account is created by a post-install hook on the api release
(job/nvcf-api-account-bootstrap in nvcf). Because it is a post-install hook, it only
fires on a successful api install. If that install failed and you repaired it with live
patches (rather than a clean re-sync), the hook never runs, no account is created, and
every subsequent function/account call returns a cryptic 404 Unknown client_id with
nothing obviously wrong in the running pods.
If the job is absent or 0, re-run it by re-syncing the api release so the hook fires
again (delete the prior job first if it lingers in a completed/failed state and blocks the
hook):
kubectl delete job -n nvcf nvcf-api-account-bootstrap --ignore-not-found
HELMFILE_ENV=<env> helmfile --selector name=api sync
kubectl logs -n nvcf job/nvcf-api-account-bootstrap # confirm the account was created
