Using kubelet identity to access ACR OCI charts

[gldraphael]

I created a test cluster exp-aks-02 with the following configuration:

Kubernetes Version: 1.25.6
Authentication and Authorization: Azure AD authentication with Kubernetes RBAC
Network Plugin: Azure CNI

(The cluster does not use the ACR integration.)

I then went ahead and bootstrapped flux, and assigned ACR Pull and Reader permissions to the User Assigned Managed Identity exp-aks-02-agentpool on a ACR instance.

At this point, I expected it to just work, but flux get sources would show this error:

unknown build error: failed to get credential from azure: DefaultAzureCredential: failed to acquire a token.
Attempted credentials:
        EnvironmentCredential: missing environment variable AZURE_TENANT_ID
        ManagedIdentityCredential: no default identity is assigned to this resource
        AzureCLICredential: Azure CLI not found on path

Ideas?

---

Other Observations

Fetching token by specifying the UAI to use

I followed the thread at #898 and concluded the reason this happens is because I have two UAIs (User Assigned managed Identities) attached to this cluster (exp-aks-02-agentpool and aciconnectorlinux-exp-aks-02).

So I tried patching the flux-system kustomization to add AZURE_CLIENT_ID:

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
resources:
  - gotk-components.yaml
  - gotk-sync.yaml
labels:
  - pairs:
      toolkit.fluxcd.io/tenant: sre-team
patches:
  - patch: |
      - op: add
        path: /spec/template/spec/containers/0/args/-
        value: --concurrent=20
      - op: add
        path: /spec/template/spec/containers/0/args/-
        value: --requeue-dependency=5s
    target:
      kind: Deployment
      name: "(kustomize-controller|helm-controller|source-controller)"
  - patch: |
      - op: add
        path: /spec/template/spec/containers/0/env/-
        value:
          name: AZURE_CLIENT_ID
          value: --client-id--
    target:
      kind: Deployment
      name: "(helm-controller|source-controller)"

But I now see this error (which almost feels like a bug):

unknown build error: failed to get credential from azure: error exchanging token: failed to decode the response: invalid character '<' looking for beginning of value

However hitting the token API directly works as long as I include the client_id parameter:

$ kubectl exec -it source-controller-59b5c97495-htrtb -n flux-system -- /bin/sh
$ wget -q -O - "http://169.254.169.254/metadata/identity/oauth2/token?api-version=2018-02-01&resource=https://management.azure.com/&client_id=$AZURE_CLIENT_ID" --header "Metadata: true"
{"access_token":"--redacted--","client_id":"--client-id--","expires_in":"84928","expires_on":"1681412609","ext_expires_in":"86399","not_before":"1681325909","resource":"https://management.azure.com/","token_type":"Bearer"}

akv2k8s works ok

I am able to consume secrets from azure keyvault using the akv2k8s project which appears to use the userAssignedIdentityID value from /etc/kubernetes/azure.json:

apiVersion: v1
kind: Namespace
metadata:
  name: akv2k8s
  labels:
    toolkit.fluxcd.io/tenant: sre-team
---
apiVersion: source.toolkit.fluxcd.io/v1beta2
kind: HelmRepository
metadata:
  name: akv2k8s
  namespace: akv2k8s
spec:
  interval: 60m0s
  url: https://charts.spvapi.no
---
apiVersion: helm.toolkit.fluxcd.io/v2beta1
kind: HelmRelease
metadata:
  name: akv2k8s
  namespace: akv2k8s
spec:
  interval: 60m
  chart:
    spec:
      chart: akv2k8s
      version: "2.*"
      sourceRef:
        kind: HelmRepository
        name: akv2k8s
      interval: 12h
  values:
    global:
      metrics:
        enabled: true
---
apiVersion: spv.no/v1alpha1
kind: AzureKeyVaultSecret
metadata:
  name: test-credentials
  namespace: monitoring
spec:
  vault:
    name: vault-name
    object:
      type: multi-key-value-secret
      name: test-credentials
      contentType: application/x-json
  output:
    secret:
      name: test-credentials

[somtochiama]

Which version of flux are you on? You can run flux version to check?
I want to see if I can reproduce this error on my end, so any more details on how you set up the cluster would be great. We have e2e tests for kubelet identity (but the cluster uses system identity)

[gldraphael]

Flux version returns this:

~ ❯  flux version
flux: v0.41.2
helm-controller: v0.31.2
kustomize-controller: v0.35.1
notification-controller: v0.33.0
source-controller: v0.36.1

I created the cluster from the azure portal but I'm happy to put together a terraform script if that will help.

Edit:

I think to reproduce this, the cluster should use Azure AD, and the cluster should have more than one User Assigned Managed Identity. I'm validating this assumption right now.

[gldraphael]

I created a new cluster with a single User Assigned Managed Identity (UAI):

Node pools
Node pools 1
Enable virtual nodes Disabled

Access
Resource identity: System-assigned managed identity
Local accounts: Disabled
Authentication and Authorization: Azure AD authentication with Kubernetes RBAC
Cluster admin group: Cluster Admin
Encryption type: (Default) Encryption at-rest with a platform-managed key

Networking
Network configuration: Kubenet
Load balancer: Standard
Private cluster: Disabled
Authorized IP ranges: Disabled
Network policy: None

Integrations
Container registry: None
Microsoft Defender for Cloud: Free
Enable Container Logs: Disabled
Alerts: Not enabled
Azure Policy: Disabled

And I see the following error (which is similar to what I saw when I set AZURE_CLIENT_ID in the previous cluster with more than one UAI):

failed to get credential from azure: error exchanging token: failed to decode the response: invalid character '<' looking for beginning of value

Seems like this truly is a bug. Let me know if you have trouble reproducing this.

[somtochiama]

Hey, Sorry for the long wait. I just tested this on the latest version and it worked okay:

fleet-infra git:(main) flux -v
flux version 2.0.0-rc.2

I created an AKS cluster with the following properties (as stated in the previous comment)

I assigned an AcrPull role to the cluster's managed identity and it reconciled successfully.
Next, I added a second managed identity to the cluster and it failed to reconcile (which is expected):

► annotating OCIRepository podinfo in flux-system namespace
✔ OCIRepository annotated
◎ waiting for OCIRepository reconciliation
✗ OCIRepository reconciliation failed: 'failed to get credential from azure: DefaultAzureCredential: failed to acquire a token.
Attempted credentials:
        EnvironmentCredential: missing environment variable AZURE_TENANT_ID
        WorkloadIdentityCredential: missing environment variables for workload identity. Check webhook and pod configuration
        ManagedIdentityCredential: no default identity is assigned to this resource
        AzureCLICredential: Azure CLI not found on path

Then I added the AZURE_CLIENT_ID env variable to the source-controller pod and it reconciled successfully.

Can you try upgrading to 2.0.0-rc.2

[gldraphael]

Thanks for testing this out @somtochiama

I just tested it with v2.0.0-rc.3 but still see the same error unfortunately:

failed to get credential from azure: error exchanging token: failed to decode the response: invalid character '<' looking for beginning of value

I will try again on Monday just to be certain.

[gldraphael]

I am still seeing the same error. I see it when I add the following source:

apiVersion: source.toolkit.fluxcd.io/v1beta2
kind: HelmRepository
metadata:
  name: myacr
  namespace: experiments
spec:
  type: oci
  provider: azure
  url: oci://myacr.azurecr.io
  interval: 5m

Are you able to reproduce this?

[somtochiama]

I was testing using OCIRepository instead of HelmRepository. I will try again today

[somtochiama]

Hey @gldraphael ,

I have been able to reproduce this. Can you try specifying the repository in the URL i.e

spec:
  type: oci
  provider: azure
  url: oci://myacr.azurecr.io/<repo-name>

[gldraphael]

Well, that kinda works, but not quite. My chart is at oci://myacr.azurecr.io/clippy. Not at oci://myacr.azurecr.io/charts/clippy.

Earlier, I tried:

apiVersion: source.toolkit.fluxcd.io/v1beta2
kind: HelmRepository
metadata:
  name: myacr
  namespace: experiments
spec:
  type: oci
  provider: azure
  url: oci://myacr.azurecr.io
  interval: 5m
---
apiVersion: helm.toolkit.fluxcd.io/v2beta1
kind: HelmRelease
metadata:
  name: clippy
  namespace: experiments
spec:
  releaseName: clippy
  chart:
    spec:
      chart: clippy
      sourceRef:
        kind: HelmRepository
        name: myacr
      version: 1.0.1
  interval: 50m
  install:
    remediation:
      retries: 3
  values: {}

And that shows the error I reported earlier.

Now, I tried the following:

apiVersion: source.toolkit.fluxcd.io/v1beta2
kind: HelmRepository
metadata:
  name: clippy
  namespace: experiments
spec:
  type: oci
  provider: azure
  url: oci://myacr.azurecr.io/clippy
  interval: 5m
---
apiVersion: helm.toolkit.fluxcd.io/v2beta1
kind: HelmRelease
metadata:
  name: clippy
  namespace: experiments
spec:
  releaseName: clippy
  chart:
    spec:
      chart: clippy
      sourceRef:
        kind: HelmRepository
        name: clippy
        namespace: experiments
      version: 1.0.1
  interval: 50m
  install:
    remediation:
      retries: 3
  values: {}

I see no errors on the HelmRepository anymore, but the HelmChart shows the following error:

 ~/flux/flux get source chart experiments-clippy -n experiments
NAME                    REVISION        SUSPENDED       READY   MESSAGE

experiments-clippy                      False           False   chart pull error: failed to download chart for remote reference: failed to get 'oci://myacr.azurecr.io/clippy/clippy:1.0.1': myacr.azurecr.io/clippy/clippy:1.0.1: not found

It appears to be trying to get the chart from the wrong place: myacr.azurecr.io/clippy/clippy:1.0.1 instead of myacr.azurecr.io/clippy:1.0.1

I think a possible workaround may be to move my chart to myacr.azurecr.io/charts/clippy:1.0.1.

What I do not understand is why I no longer see any error on the HelmRepository when I use oci://myacr.azurecr.io/clippy as opposed to oci://myacr.azurecr.io. Does that URL always expect a base path after the origin?

[somtochiama]

I think a possible workaround may be to move my chart to myacr.azurecr.io/charts/clippy:1.0.1.

Yes, you would have to use this as a workaround while I get this fixed.

The HelmRepository should work with the repository root address but right now there's a bug that prevents it from doing so. When exchanging the token, it makes a request to index.docker.io due to some defaulting in a library we use.
Thanks for reporting this!

[gldraphael]

Ah! Feel free let me know if you'd like me to test anything.
Appreciate your patience here!