Skip to content


Pachyderm has support for Kubernetes Role-Based Access Controls (RBAC), which is a default part of all Pachyderm deployments. In most use cases, Pachyderm sets all the RBAC permissions automatically. However, if you are deploying Pachyderm on a cluster that your company owns, security policies might not allow certain RBAC permissions by default. Therefore, you need to contact your Kubernetes administrator and provide the following list of required permissions:

Rules: []rbacv1.PolicyRule{{
    APIGroups: []string{""},
    Verbs:     []string{"get", "list", "watch"},
    Resources: []string{"nodes", "pods", "pods/log", "endpoints"},
}, {
    APIGroups: []string{""},
    Verbs:     []string{"get", "list", "watch", "create", "update", "delete"},
    Resources: []string{"replicationcontrollers", "services"},
}, {
    APIGroups:     []string{""},
    Verbs:         []string{"get", "list", "watch", "create", "update", "delete"},
    Resources:     []string{"secrets"},
    ResourceNames: []string{client.StorageSecretName},

The following table explains how Pachyderm uses those permissions:

Permission Description
Access to nodes Required for the coefficient option in the parallelism parameter of the pipeline spec. coefficient determines the number of worker nodes to run for your pipeline. If this permission cannot be granted, constant can be used instead.
Access to pods, replica controllers, and services Pachyderm uses this permission to monitor the created pipelines. The permissions related to replicationcontrollers and services are used in the setup and deletion of pipelines. Each pipeline has its own RC and service in addition to the pods.
Access to secrets Required to give various kinds of credentials to pipelines, including storage credentials to access S3 or other object storage backends, Docker credentials to pull from a private registry, and others.


In older Kubernetes versions, kube-dns did not work properly with RBAC. To check if your cluster is affected by this issue, run:

kubectl get all --namespace=kube-system

System response:

deploy/kube-dns   1         1         1            0           3m

NAME                     DESIRED   CURRENT   READY     AGE
rs/kube-dns-86f6f55dd5   1         1         0         3m

NAME                            READY     STATUS    RESTARTS   AGE
po/kube-addon-manager-oryx      1/1       Running   0          3m
po/kube-dns-86f6f55dd5-xksnb    2/3       Running   4          3m
po/kubernetes-dashboard-bzjjh   1/1       Running   0          3m
po/storage-provisioner          1/1       Running   0          3m

NAME                      DESIRED   CURRENT   READY     AGE
rc/kubernetes-dashboard   1         1         1         3m

NAME                       TYPE        CLUSTER-IP     EXTERNAL-IP   PORT(S)         AGE
svc/kube-dns               ClusterIP     <none>        53/UDP,53/TCP   3m
svc/kubernetes-dashboard   NodePort   <none>        80:30000/TCP    3m

In the output above, po/kubernetes-dashboard-bzjjh has only two out of three pods ready and has restarted four times. To fix this issue, run:

kubectl -n kube-system create sa kube-dns
kubectl -n kube-system patch deploy/kube-dns -p '{"spec": {"template": {"spec": {"serviceAccountName": "kube-dns"}}}}'

These commands enforce kube-dns to use the appropriate ServiceAccount. Kubernetes has created the ServiceAccount, but does not use it until you run the above commands.

Resolving RBAC Permissions on GKE

When you deploy Pachyderm on GKE, you might see the following error:

Error from server (Forbidden): error when creating "STDIN": "pachyderm" is forbidden: attempt to grant extra privileges:

To fix this issue, run the following command and redeploy Pachyderm:

kubectl create clusterrolebinding cluster-admin-binding --clusterrole=cluster-admin --user=$(gcloud config get-value account)

Last update: July 16, 2020