gke-metadata-server

GA is here!

gke-metadata-server

release

A GKE Metadata Server emulator for making it easier to use GCP Workload Identity Federation inside non-GKE Kubernetes clusters, e.g. KinD, on-prem, managed Kubernetes from other clouds, or even self-managed clusters deployed in GCE VMs. This implementation tries to mimic the gke-metadata-server DaemonSet deployed automatically by Google in the kube-system namespace of GKE clusters that have the feature Workload Identity Federation for GKE enabled. See how the GKE Metadata Server works.

Usage

Steps:

Configure GCP Workload Identity Federation for Kubernetes.
Deploy gke-metadata-server in the cluster using the Workload Identity Provider full name, obtained after step 1.
Add the label iam.gke.io/gke-metadata-server-enabled: "true" to the Nodes you want to have the emulator Pods running on. Only Pods running on these Nodes will be able to use the emulator. The most recommended way to do this is by using Node pooling/grouping features from your Kubernetes provider. Alternatively, set requireNodeLabel: false to skip this label requirement and run the emulator on all nodes.
See ./testdata/pod.yaml for an example of how to configure your Pods and their ServiceAccounts.
(Optional but highly recommended) Verify gke-metadata-server’s artifact signatures to make sure you are deploying authentic artifacts distributed by this project.

Configure GCP Workload Identity Federation for Kubernetes

Steps:

Create a Workload Identity Pool and Provider pair for the cluster.
Grant Kubernetes ServiceAccounts permission to impersonate Google Service Accounts.

Official docs and examples are here.

More examples for all the configuration described in this section are available here: ./terraform/test.tf. This is where we provision the infrastructure required for testing this project in CI and development.

Create a Workload Identity Pool and Provider pair for the cluster

Before granting ServiceAccounts of a Kubernetes cluster permission to impersonate Google Service Accounts, a Workload Identity Pool and Provider pair must be created for the cluster in Google Cloud Platform.

A Pool is a namespace for Subjects and Providers, where Subjects represent the identities to whom permission to impersonate Google Service Accounts are granted, i.e. the Kubernetes ServiceAccounts in this case.

A Provider stores the OpenID Connect configuration parameters retrieved from the cluster for allowing Google to verify ServiceAccount Tokens issued by the cluster. Specifically for the creation of a Provider, see an example in the ./Makefile, target create-or-update-provider.

The ServiceAccount Tokens issued by Kubernetes are JWT tokens. Each such token is exchanged for an OAuth 2.0 Access Token that impersonates a Google Service Account. During a token exchange, the Subject is mapped from the sub claim of the JWT (achieved in the Makefile example by --attribute-mapping=google.subject=assertion.sub). In this case the Subject has the following format:

system:serviceaccount:{k8s_namespace}:{k8s_sa_name}

The Subject total length cannot exceed 127 characters (docs).

Attention 1: If you update the private keys of the ServiceAccount Issuer of the cluster you must update the Provider with the new OpenID Connect configuration, otherwise service will be disrupted.

Attention 2: Please make sure not to specify any audiences when creating the Provider. The emulator uses the default audience when issuing the ServiceAccount Tokens. The default audience contains the full name of the Provider, which is a strong restriction:

//iam.googleapis.com/{provider_full_name}

where {provider_full_name} has the form:

{pool_full_name}/providers/{provider_short_name}

and {pool_full_name} has the form:

projects/{gcp_project_number}/locations/global/workloadIdentityPools/{pool_short_name}

Grant Kubernetes ServiceAccounts permission to impersonate Google Service Accounts

For allowing the Kubernetes ServiceAccount {k8s_sa_name} from the namespace {k8s_namespace} to impersonate the Google Service Account {gcp_service_account}@{gcp_project_id}.iam.gserviceaccount.com, grant the IAM Role roles/iam.workloadIdentityUser on this Google Service Account to the following principal:

principal://iam.googleapis.com/{pool_full_name}/subject/system:serviceaccount:{k8s_namespace}:{k8s_sa_name}

This principal will be reflected as a Subject in the Google Cloud Console webpage of the Pool.

Alternatively, grant direct resource access to the Kubernetes ServiceAccount

Workload Identity Federation for Kubernetes allows you to directly grant Kubernetes ServiceAccounts access to Google resources, without the need to impersonate a Google Service Account. This is done by granting the given IAM Roles directly to principals of the form described above for impersonation. See docs.

Some specific GCP services do not support this method. See docs.

The GET /computeMetadata/v1/instance/service-accounts/default/identity API is not supported by this method. If you plan to use this API, you must use the impersonation method described above.

Deploy `gke-metadata-server` in your cluster

Using the Helm Chart

A Helm Chart is available in the following Helm OCI Repository:

ghcr.io/matheuscscp/gke-metadata-server-helm:{version} (GitHub Container Registry)

See the Helm Values API in the file ./helm/gke-metadata-server/values.yaml. Make sure to specify at least the full name of the Workload Identity Provider.

Using the Timoni Module

If you prefer something newer and strongly typed, a Timoni Module is available in the following Timoni OCI Repository:

ghcr.io/matheuscscp/gke-metadata-server-timoni:{version} (GitHub Container Registry)

See the Timoni Values API in the files ./timoni/gke-metadata-server/templates/config.tpl.cue and ./timoni/gke-metadata-server/templates/settings.cue. Make sure to specify at least the full name of the Workload Identity Provider.

Using only the container image (with your own Kubernetes manifests)

Alternatively, you can write your own Kubernetes manifests and consume only the container image:

ghcr.io/matheuscscp/gke-metadata-server:{version} (GitHub Container Registry)

Verify the OCI artifacts

For verifying the OCI artifacts above, use the cosign CLI tool.

Verify the container image

For verifying the container image of a given release, download the digest file container-digest.txt attached to the Github Release and use it with cosign:

VERSION="x.y.z" # replace with the release version
cosign verify ghcr.io/matheuscscp/gke-metadata-server@$(cat container-digest.txt) \
  --certificate-oidc-issuer=https://token.actions.githubusercontent.com \
  --certificate-identity=https://github.com/matheuscscp/gke-metadata-server/.github/workflows/release.yml@refs/tags/v${VERSION}

Automatic verification

If you are using Sigstore’s Policy Controller for enforcing policies you can automate the image verification using Keyless Authorities.

If you are using Kyverno for enforcing policies you can automate the image verification using Keyless Verification.

Verify the Helm Chart

For verifying the OCI Helm Chart of a given release, download the digest file helm-digest.txt attached to the Github Release and use it with cosign:

VERSION="x.y.z" # replace with the release version
cosign verify ghcr.io/matheuscscp/gke-metadata-server-helm@$(cat helm-digest.txt) \
  --certificate-oidc-issuer=https://token.actions.githubusercontent.com \
  --certificate-identity=https://github.com/matheuscscp/gke-metadata-server/.github/workflows/release.yml@refs/tags/v${VERSION}

Automatic verification

If you are using a Flux OCIRepository for deploying Helm Charts you can automate the verification using Keyless Verification.

Verify the Timoni Module

For verifying the OCI Timoni Module of a given release, download the digest file timoni-digest.txt attached to the Github Release and use it with cosign:

VERSION="x.y.z" # replace with the release version
cosign verify ghcr.io/matheuscscp/gke-metadata-server-timoni@$(cat timoni-digest.txt) \
  --certificate-oidc-issuer=https://token.actions.githubusercontent.com \
  --certificate-identity=https://github.com/matheuscscp/gke-metadata-server/.github/workflows/release.yml@refs/tags/v${VERSION}

Automatic verification

If you are using the Timoni CLI to deploy Modules/Bundles you can automate the verification using Keyless Verification.

Further documentation

Routing Modes

The Google libraries attempt to retrieve tokens from the hard-coded endpoint 169.254.169.254:80. In order to make this work, the emulator must intercept the traffic from the client Pods to this endpoint. To solve this problem, the emulator supports three routing modes: eBPF, Loopback, and None. See details of each mode below. In all modes the unencrypted communication between the emulator Pod and a client Pod never leaves the Node where they are both running on. This is exactly how the native GKE implementation works.

The annotation or label node.gke-metadata-server.matheuscscp.io/routingMode can be added to a Node in order to configure the routing mode of the emulator in that particular Node. The default is eBPF.

If you are using similar tools or workload identity features of managed Kubernetes services from other clouds, this configuration may have a direct conflict with other such tools or features. It’s a recurrent practice among cloud providers using this hard-coded IP address and port to implement workload identity. If your use case requires, the recommended solution is to use Node pooling/grouping features from your Kubernetes provider to isolate the workloads that need to use the emulator as mentioned in the Usage section. Remember to add the label iam.gke.io/gke-metadata-server-enabled: "true" to such Nodes (or set requireNodeLabel: false) and appropriate Node selectors/tolerations to your Pods.

`eBPF`

In this routing mode the emulator hooks an eBPF program to the connect() syscall. The eBPF program looks at the destination address and, if it matches the hard-coded address mentioned above, modifies the address to the actual address of the emulator.

This mode also attaches a cgroup/sock_ops program that records the cgroup ID of every TCP-connecting task into a 4-tuple-keyed map at TCP_CONNECT_CB. The emulator consults that map for hostNetwork pod requests and resolves the cgroup ID to a pod UID by walking /sys/fs/cgroup, giving hostNetwork pods per-pod identity (see Pods running on the host network).

This mode coexists with other eBPF tools that attach to cgroup/connect4 — notably Cilium, including with kubeProxyReplacement enabled (which makes Cilium attach its own cgroup/connect4 for service-VIP redirection). Both programs are loaded under the same hook via the kernel’s multi-attach API (BPF_LINK_CREATE), and act on disjoint destinations: Cilium rewrites ClusterIP-range destinations, this emulator rewrites 169.254.169.254:80 only. Either ordering between the two programs produces the same result. CI validates this coexistence on every PR (e2e (cilium-kpr) matrix entry).

The advantage of this mode is that the emulator can bind to any port, not just port 80. This is because we modify the destination address of connections targeting our endpoint of interest, so we have the freedom to choose any port.

Note that in this mode the emulator will proxy connections that are not targeting Google metadata to the 169.254.169.254:80 endpoint, i.e. any connections whose first few bytes do not match GET /computeMetadata/v1. This is to allow other environments that expose services on this endpoint to continue working properly. For example, AWS EKS clusters expose the AWS Instance Metadata Service on this endpoint. The emulator works fine in such environments without disrupting the native metadata service due to this proxying feature. The emulator will not work with environments exposing a service in this endpoint that also identifies Pods by their IP addresses. This is because the source IP address the service will see will be the Node IP address, and not the IP address of the client Pod. This is not the case for AWS EKS, since the AWS Instance Metadata Service does not identify Pods by their IP addresses, so the emulator works fine in AWS EKS clusters.

`Loopback`

In this routing mode the emulator adds the hard-coded address mentioned above to the lo interface of the Node with the label lo:gke-md-sv and the host scope. This way the emulator can bind to the hard-coded address and port 80, and the traffic from the client Pods to this address will be routed to the emulator.

This mode does not clash with other eBPF-based tools running in the cluster, but it has the disadvantage of not being able to bind to any port, only port 80. And for this to work properly the emulator Pods need to run on the host network, so they occupy port 80 in the network namespace of the Node.

`None`

In this routing mode the emulator does not perform any network routing or traffic interception. Instead, it relies on the environment variables defined inside the Google libraries to redirect metadata requests to the emulator. Note that support for the environment variables below may vary between different Google libraries. The configuration of your Pods should include:

spec:
  containers:
  - name: your-app
    env:
    - name: HOST_IP
      valueFrom:
        fieldRef:
          fieldPath: status.hostIP
    - name: GKE_METADATA_SERVER_PORT
      value: "16321"
    - name: GCE_METADATA_HOST
      value: "$(HOST_IP):$(GKE_METADATA_SERVER_PORT)"
    - name: GCE_METADATA_ROOT
      value: "$(HOST_IP):$(GKE_METADATA_SERVER_PORT)"
    - name: GCE_METADATA_IP
      value: "$(HOST_IP):$(GKE_METADATA_SERVER_PORT)"

The port value must match the port configured on the emulator DaemonSet (default: 16321). This approach avoids conflicts with other tools that could be attaching eBPF programs to the connect() syscall or using the well-known IP address and port hard-coded in the Google libraries, but requires a bit of per-Pod configuration.

The `metadata.google.internal` DNS record

If you observe DNS lookup errors for metadata.google.internal in your Pods, you can add to them the following DNS configuration:

spec:
  hostAliases:
  - hostnames: [metadata.google.internal]
    ip: 169.254.169.254

Not all Google libraries hit metadata.google.internal, some just hit 169.254.169.254 directly. The gcloud CLI, for example, needs this DNS configuration to work properly. See this example.

If your cluster supports CoreDNS’s coredns-custom ConfigMap, like k3s or AKS, you can alternatively set dns.provider to CoreDNSCustom in the Helm Chart or Timoni Module. When set, this option applies a ConfigMap called coredns-custom in the kube-system namespace that instructs CoreDNS to resolve metadata.google.internal to 169.254.169.254 cluster-wide, eliminating the need for per-Pod hostAliases configuration.

Warning: Enabling CoreDNSCustom will overwrite any existing coredns-custom ConfigMap in the kube-system namespace. If you already have custom CoreDNS entries in that ConfigMap, they will be lost. In that case, manually add the metadata.override entry to your existing ConfigMap instead of using this option.

Node initialization

The emulator Pods have toleration for any taints, so they will get scheduled earlier on the Nodes. This may be enough for ensuring the emulator is running before client Pods attempt to get credentials from them. But if you observe issues on your Pods due to this race condition, you can configure your Nodes to start with a taint that will prevent your Pods from being scheduled on them until the emulator has fully initialized. You can choose any key you want for the taint, as long as it has the prefix node.gke-metadata-server.matheuscscp.io. The emulator will remove any such taints from the Node after successfully initializing. See the taints and tolerations docs here.

Limitations and Security Risks

Pod identification

The strategy is picked per request by (routing mode, pod kind). There is no fallback between strategies — each combination has exactly one path.

routing mode	non-hostNetwork pod	hostNetwork pod
`eBPF`	kernel attestation (sockops 4-tuple → cgroup ID → pod UID)	same
`Loopback`	source IP → `GetByIP` (node-scoped)	netlink SOCK_DIAG → `/proc/<pid>/cgroup` → pod UID
`None`	source IP → `GetByIP` (node-scoped)	netlink SOCK_DIAG → `/proc/<pid>/cgroup` → pod UID

In eBPF mode every pod is identified by kernel attestation — the source IP is not consulted at all. In Loopback and None modes, regular pods are identified by their source IP (pod IPs are unique within a Node and GetByIP is node-scoped, so this is sound), while hostNetwork pods — which share the Node’s IP — fall back to the netlink-based attestation chain. See the host-network section below for the kernel-attestation details.

The trust root for source IP is the kernel’s report on the TCP connection — never any HTTP-level header (e.g. X-Forwarded-For).

If an attacker can perform source-IP impersonation in your cluster (e.g. ARP spoofing), the source-IP strategy used by Loopback/None for non-hostNetwork pods can be exploited to steal credentials. Please evaluate the risk of such attacks in your cluster before choosing this tool. The attack referenced typically requires Pods running with very high privileges, which should normally be blocked in production clusters via PSA, Gatekeeper or equivalent policy. The kernel-attestation paths (eBPF for any pod, netlink for hostNetwork pods) do not depend on source IP and are not affected by ARP spoofing.

Pods running on the host network

Pods with spec.hostNetwork: true share the Node’s IP address, so the source IP the emulator sees on the connection is not unique to the pod. Both routing-mode-specific kernel-attestation paths are described below; the eBPF chain is the same path used for every pod in eBPF mode (non-hostNetwork pods are also resolved through it), and the netlink chain is only used for hostNetwork pods in Loopback and None modes.

In eBPF routing mode, a cgroup/sock_ops BPF program records the connecting task’s cgroup ID into a 4-tuple-keyed map at TCP connect time. The emulator looks up the 4-tuple from the accepted connection, walks /sys/fs/cgroup to find the matching directory, and extracts the pod UID from the kubelet’s pod<UID> cgroup naming convention.
In Loopback and None routing modes, the emulator queries the host’s socket table via NETLINK_SOCK_DIAG to resolve the 4-tuple to its socket inode, walks /proc/<pid>/fd to find the owning PID, then reads /proc/<pid>/cgroup for the same pod<UID> extraction.

Both paths produce a kubernetes pod UID without trusting source IP or any HTTP-level information. The pod’s regular ServiceAccount (configured via spec.serviceAccountName) is then used; there is no node-level shared ServiceAccount.

Hard requirements: a Linux kernel with cgroup v2 unified hierarchy and BPF cgroup-hook support (in practice any kernel from the last few years — bpf_get_current_cgroup_id has been available since 4.18). No specific CNI is required. CI exercises three configurations on every PR: cilium-default (Cilium 1.19.x with default values), cilium-kpr (Cilium with kubeProxyReplacement=true, which makes Cilium attach its own cgroup/connect4 program alongside ours), and kindnet (kind’s default CNI, no Cilium).

Sandboxed container runtimes (gVisor, Kata) are unsupported because their userspace kernels are not visible to host BPF programs and not represented in the host’s /proc.

Token Cache

When the emulator is configured to cache tokens, the issued Google OAuth 2.0 Access and OpenID Connect Identity Tokens are cached and returned to client Pods on every request until their expiration.

This means that even if you revoke the required permissions for the Kubernetes ServiceAccount to issue those tokens, the client Pods will still get those tokens from the emulator until they expire. This is a limitation of how the permissions are evaluated: they are evaluated only when the tokens are issued, which is what caching tries to avoid. If your use case requires immediate revocation of permissions, one alternative is to disable token caching. The Helm Chart and Timoni Module distributed here have options to disable token caching, look at their respective values APIs. Another alternative is manually deleting all the Pods of the emulator DaemonSet:

kubectl delete pods -l app=gke-metadata-server -n kube-system

New Pods will be created by the DaemonSet and they will not have the cached tokens.

The emulator implements an 80% refresh rule (similar to kubelet for ServiceAccount token rotation) and supports configurable maximum token duration limits. The 80% rule means tokens are refreshed when they have consumed 80% of their original lifetime. The maximum token duration setting allows capping the effective duration of cached tokens (default: 1 hour). This can be configured through the Helm Chart and Timoni Module values APIs.

With the default configuration, tokens expire in at most one hour.

Disclaimer

This project was not created by Google. Enterprise support from Google is not available. Use this tool at your own risk. (But please do feel free to report bugs and CVEs, request help, new features and contribute.)

Furthermore, this tool is not necessary for using GCP Workload Identity Federation inside non-GKE Kubernetes clusters. This is just a facilitator. Kubernetes and GCP Workload Identity Federation work together by themselves. This tool just makes your Pods need much less configuration to use GCP Workload Identity Federation for Kubernetes, by making the configuration as close as possible to how Workload Identity Federation for GKE is configured.

This site is open source. Improve this page.

gke-metadata-server

GA is here!

gke-metadata-server

Usage

Configure GCP Workload Identity Federation for Kubernetes

Create a Workload Identity Pool and Provider pair for the cluster

Grant Kubernetes ServiceAccounts permission to impersonate Google Service Accounts

Alternatively, grant direct resource access to the Kubernetes ServiceAccount

Deploy gke-metadata-server in your cluster

Using the Helm Chart

Using the Timoni Module

Using only the container image (with your own Kubernetes manifests)

Verify the OCI artifacts

Verify the container image

Automatic verification

Verify the Helm Chart

Automatic verification

Verify the Timoni Module

Automatic verification

Further documentation

Routing Modes

eBPF

Loopback

None

The metadata.google.internal DNS record

Node initialization

Limitations and Security Risks

Pod identification

Pods running on the host network

Token Cache

Disclaimer

Deploy `gke-metadata-server` in your cluster

`eBPF`

`Loopback`

`None`

The `metadata.google.internal` DNS record