• Overview of the Kubeflow Platform
• Kubeflow components versions
• Installation
  • Prerequisites
  • Install with a single command
  • Install individual components
  • Connect to your Kubeflow Cluster
  • Change default user password
• Upgrading and extending
• Release process
• CVE Scanning
• Frequently Asked Questions

This repository is owned by the Manifests Working Group. If you are a contributor authoring or editing the packages please see Best Practices. You can join the CNCF Slack and access our meetings at the Kubeflow Community website. Our channel on the CNCF Slack is here #kubeflow-platform. You can also find there our biweekly meetings, including the commentable Agenda.

The Kubeflow Manifests repository is organized under three main directories, which include manifests for installing:

All components are deployable with kustomize. You can choose to deploy the whole Kubeflow platform or individual components.

This repo periodically syncs all official Kubeflow components from their respective upstream repos. The following matrix shows the git version that we include for each component:

The following is also a matrix with versions from common components that are used from the different projects of Kubeflow:

This is for the installation from scratch. For the in-place upgrade guide please jump to the upgrading and extending section.

The Manifests WG provides two options for installing Kubeflow official components and common services with kustomize. The aim is to help end users install easily and to help distribution owners build their opinionated distributions from a tested starting point:

1. Single-command installation of all components under apps and common
2. Multi-command, individual components installation for apps and common

Option 1 targets ease of deployment for end users.
Option 2 targets customization and ability to pick and choose individual components.

The example directory contains an example kustomization for the single command to be able to run.

⚠️ In both options, we use a default email ([email protected]) and password (12341234). For any production Kubeflow deployment, you should change the default password by following the relevant section.

• This is the master branch which targets Kubernetes 1.29+
• For the specific Kubernetes version per release consult the release notes
• Either our local Kind (installed below) or your own Kubernetes cluster with a default StorageClass
• Kustomize 5.2.1+
• Kubectl in a version that is compatible with your Kubernetes cluster

NOTE

kubectl apply commands may fail on the first try. This is inherent in how Kubernetes and kubectl work (e.g., CR must be created after CRD becomes ready). The solution is to simply re-run the command until it succeeds. For the single-line command, we have included a bash one-liner to retry the command.

• 32 GB of RAM recommended
• 16 CPU cores recommended
• kind
• docker
• Linux kernel subsystem changes to support many pods
  • sudo sysctl fs.inotify.max_user_instances=2280
  • sudo sysctl fs.inotify.max_user_watches=1255360

cat <<EOF | kind create cluster --name=kubeflow --config=-
kind: Cluster
apiVersion: kind.x-k8s.io/v1alpha4
nodes:
- role: control-plane
  image: kindest/node:v1.29.4
  kubeadmConfigPatches:
  - |
    kind: ClusterConfiguration
    apiServer:
      extraArgs:
        "service-account-issuer": "kubernetes.default.svc"
        "service-account-signing-key-file": "/etc/kubernetes/pki/sa.key"
EOF

kind get kubeconfig --name kubeflow > /tmp/kubeflow-config
export KUBECONFIG=/tmp/kubeflow-config

docker login

kubectl create secret generic regcred \
    --from-file=.dockerconfigjson=/home/to/.docker/config.json \
    --type=kubernetes.io/dockerconfigjson

You can install all Kubeflow official components (residing under apps) and all common services (residing under common) using the following command:

while ! kustomize build example | kubectl apply -f -; do echo "Retrying to apply resources"; sleep 20; done

Once, everything is installed successfully, you can access the Kubeflow Central Dashboard by logging in to your cluster.

Congratulations! You can now start experimenting and running your end-to-end ML workflows with Kubeflow.

In this section, we will install each Kubeflow official component (under apps) and each common service (under common) separately, using just kubectl and kustomize.

If all the following commands are executed, the result is the same as in the above section of the single command installation. The purpose of this section is to:

• Provide a description of each component and insight on how it gets installed.
• Enable the user or distribution owner to pick and choose only the components they need.

Troubleshooting note

We've seen errors like the following when applying the kustomizations of different components:

error: resource mapping not found for name: "<RESOURCE_NAME>" namespace: "<SOME_NAMESPACE>" from "STDIN": no matches for kind "<CRD_NAME>" in version "<CRD_FULL_NAME>"
ensure CRDs are installed first

This is because a kustomization applies both a CRD and a CR very quickly, and the CRD hasn't become Established yet. You can learn more about this in kubernetes/kubectl#1117 and helm/helm#4925.

If you bump into this error we advise to re-apply the kustomization of the component.

Cert-manager is used by many Kubeflow components to provide certificates for admission webhooks.

Install cert-manager:

kustomize build common/cert-manager/cert-manager/base | kubectl apply -f -
echo "Waiting for cert-manager to be ready ..."
kubectl wait --for=condition=ready pod -l 'app in (cert-manager,webhook)' --timeout=180s -n cert-manager
kubectl wait --for=jsonpath='{.subsets[0].addresses[0].targetRef.kind}'=Pod endpoints -l 'app in (cert-manager,webhook)' --timeout=180s -n cert-manager

In case you get this error:

Error from server (InternalError): error when creating "STDIN": Internal error occurred: failed calling webhook "webhook.cert-manager.io": failed to call webhook: Post "https://cert-manager-webhook.cert-manager.svc:443/mutate?timeout=10s": dial tcp 10.96.202.64:443: connect: connection refused

This is because the webhook is not yet ready to receive request. Wait a couple seconds and retry applying the manfiests.

For more troubleshooting info also check out https://cert-manager.io/docs/troubleshooting/webhook/

Istio is used by most Kubeflow components to secure their traffic, enforce network authorization and implement routing policies.

Install Istio:

echo "Installing Istio configured with external authorization..."
cd common/istio-1-22
kustomize build common/istio-1-22/istio-crds/base | kubectl apply -f -
kustomize build common/istio-1-22/istio-namespace/base | kubectl apply -f -
kustomize build common/istio-1-22/istio-install/overlays/oauth2-proxy | kubectl apply -f -

echo "Waiting for all Istio Pods to become ready..."
kubectl wait --for=condition=Ready pods --all -n istio-system --timeout 300s

The oauth2-proxy extends your Istio Ingress-Gateway capabilities, to be able to function as an OIDC client:

echo "Installing oauth2-proxy..."
kustomize build common/oauth2-proxy/overlays/m2m-self-signed/ | kubectl apply -f -
kubectl wait --for=condition=ready pod -l 'app.kubernetes.io/name=oauth2-proxy' --timeout=180s -n oauth2-proxy

It supports user sessions as well as proper token-based machine to machine atuhhentication.

Dex is an OpenID Connect Identity (OIDC) with multiple authentication backends. In this default installation, it includes a static user with email [email protected]. By default, the user's password is 12341234. For any production Kubeflow deployment, you should change the default password by following the relevant section.

Install Dex:

kustomize build common/dex/overlays/oauth2-proxy | kubectl apply -f -

Knative is used by the KServe official Kubeflow component.

Install Knative Serving:

kustomize build common/knative/knative-serving/overlays/gateways | kubectl apply -f -
kustomize build common/istio-1-22/cluster-local-gateway/base | kubectl apply -f -

Optionally, you can install Knative Eventing which can be used for inference request logging:

kustomize build common/knative/knative-eventing/base | kubectl apply -f -

Create the namespace where the Kubeflow components will live in. This namespace is named kubeflow.

Install kubeflow namespace:

kustomize build common/kubeflow-namespace/base | kubectl apply -f -

Create the Kubeflow ClusterRoles, kubeflow-view, kubeflow-edit and kubeflow-admin. Kubeflow components aggregate permissions to these ClusterRoles.

Install kubeflow roles:

kustomize build common/kubeflow-roles/base | kubectl apply -f -

Install the Multi-User Kubeflow Pipelines official Kubeflow component:

kustomize build apps/pipeline/upstream/env/cert-manager/platform-agnostic-multi-user | kubectl apply -f -

This installs argo with the runasnonroot emissary executor. Please note that you are still responsible to analyze the security issues that arise when containers are run with root access and to decide if the kubeflow pipeline main containers are run as runasnonroot. It is in general strongly recommended that all user-accessible OCI containers run with Pod Security Standards [restricted] (https://kubernetes.io/docs/concepts/security/pod-security-standards/#restricted)

Multi-User Kubeflow Pipelines dependencies

• Istio
• Kubeflow Roles
• OIDC Auth Service (or cloud provider specific auth service)
• Profiles + KFAM

Alternative: Kubeflow Pipelines Standalone

You can install Kubeflow Pipelines Standalone which

• does not support multi user separation
• has no dependencies on the other services mentioned here

You can learn more about their differences in Installation Options for Kubeflow Pipelines .

Besides installation instructions in Kubeflow Pipelines Standalone documentation, you need to apply two virtual services to expose Kubeflow Pipelines UI and Metadata API in kubeflow-gateway.

KFServing was rebranded to KServe.

Install the KServe component:

kustomize build contrib/kserve/kserve | kubectl apply -f -

Install the Models web application:

kustomize build contrib/kserve/models-web-app/overlays/kubeflow | kubectl apply -f -

Install the Katib official Kubeflow component:

kustomize build apps/katib/upstream/installs/katib-with-kubeflow | kubectl apply -f -

Install the Central Dashboard official Kubeflow component:

kustomize build apps/centraldashboard/upstream/overlays/kserve | kubectl apply -f -

Install the Admission Webhook for PodDefaults:

kustomize build apps/admission-webhook/upstream/overlays/cert-manager | kubectl apply -f -

Install the Notebook Controller official Kubeflow component:

kustomize build apps/jupyter/notebook-controller/upstream/overlays/kubeflow | kubectl apply -f -

Install the Jupyter Web App official Kubeflow component:

kustomize build apps/jupyter/jupyter-web-app/upstream/overlays/istio | kubectl apply -f -

It is still in development.

Install the PVC Viewer Controller official Kubeflow component:

kustomize build apps/pvcviewer-controller/upstream/default | kubectl apply -f -

Install the Profile Controller and the Kubeflow Access-Management (KFAM) official Kubeflow components:

kustomize build apps/profiles/upstream/overlays/kubeflow | kubectl apply -f -

Install the Volumes Web App official Kubeflow component:

kustomize build apps/volumes-web-app/upstream/overlays/istio | kubectl apply -f -

Install the Tensorboards Web App official Kubeflow component:

kustomize build apps/tensorboard/tensorboards-web-app/upstream/overlays/istio | kubectl apply -f -

Install the Tensorboard Controller official Kubeflow component:

kustomize build apps/tensorboard/tensorboard-controller/upstream/overlays/kubeflow | kubectl apply -f -

Install the Training Operator official Kubeflow component:

kustomize build apps/training-operator/upstream/overlays/kubeflow | kubectl apply -f -

Finally, create a new namespace for the default user (named kubeflow-user-example-com).

kustomize build common/user-namespace/base | kubectl apply -f -

After installation, it will take some time for all Pods to become ready. Make sure all Pods are ready before trying to connect, otherwise you might get unexpected errors. To check that all Kubeflow-related Pods are ready, use the following commands:

kubectl get pods -n cert-manager
kubectl get pods -n istio-system
kubectl get pods -n auth
kubectl get pods -n knative-eventing
kubectl get pods -n knative-serving
kubectl get pods -n kubeflow
kubectl get pods -n kubeflow-user-example-com

The default way of accessing Kubeflow is via port-forward. This enables you to get started quickly without imposing any requirements on your environment. Run the following to port-forward Istio's Ingress-Gateway to local port 8080:

kubectl port-forward svc/istio-ingressgateway -n istio-system 8080:80

After running the command, you can access the Kubeflow Central Dashboard by doing the following:

1. Open your browser and visit http://localhost:8080. You should get the Dex login screen.
2. Login with the default user's credentials. The default email address is [email protected] and the default password is 12341234.

In order to connect to Kubeflow using NodePort / LoadBalancer / Ingress, you need to setup HTTPS. The reason is that many of our web applications (e.g., Tensorboard Web Application, Jupyter Web Application, Katib UI) use Secure Cookies, so accessing Kubeflow with HTTP over a non-localhost domain does not work.

Exposing your Kubeflow cluster with proper HTTPS is a simple proces, but dependent on your environment. There are also third-party commercial distributions available.

NOTE

If you absolutely need to expose Kubeflow over HTTP, you can disable the Secure Cookies feature by setting the APP_SECURE_COOKIES environment variable to false in every relevant web app. This is not recommended, as it poses security risks.

For security reasons, we don't want to use the default username and email for the default Kubeflow user when installing in security-sensitive environments. Instead, you should define your own username and email before deploying. To define it for the default user:

1. Edit common/dex/overlays/oauth2-proxy/config-map.yaml and fill the relevant field with your email and preferred username:

   ...
     staticPasswords:
     - email: <REPLACE_WITH_YOUR_EMAIL>
       username: <REPLACE_WITH_PREFERRED_USERNAME>

For security reasons, we don't want to use the default password for the default Kubeflow user when installing in security-sensitive environments. Instead, you should define your own password and apply it either before creating the cluster or after creating the cluster.

Pick a password for the default user, with email [email protected], and hash it using bcrypt:

```sh
python3 -c 'from passlib.hash import bcrypt; import getpass; print(bcrypt.using(rounds=12, ident="2y").hash(getpass.getpass()))'
```

For example, running the above command locally with required packages like passlib would look as follows:

python3 -c 'from passlib.hash import bcrypt; import getpass; print(bcrypt.using(rounds=12, ident="2y").hash(getpass.getpass()))'
Password:       <--- Enter the password here
$2y$12$vIm8CANhuWui0J1p3jYeGeuM28Qcn76IFMaFWvZCG5ZkKZ4MjTF4u <--- GENERATED_HASH_FOR_ENTERED_PASSWORD

1. Edit common/dex/base/dex-passwords.yaml and fill the relevant field with the hash of the password you chose:

   ...
     stringData:
       DEX_USER_PASSWORD: <REPLACE_WITH_HASH>

1. Delete the existing secret dex-passwords in auth namespace using the following command:

   kubectl delete secret dex-passwords -n auth

2. Create secret dex-passwords with new hash using the following command:

   kubectl create secret generic dex-passwords --from-literal=DEX_USER_PASSWORD='REPLACE_WITH_HASH' -n auth

3. Recreate the dex pod in auth namespace using the following command:

   kubectl delete pods --all -n auth

4. Try to login using the new dex password.

For modifications and in place upgrades of the Kubeflow platform we provide a rough description for advanced users:

• Never ever edit the manifests directly, use Kustomize overlays and components on top of the example.yaml.
• This allows you to upgrade by just referencing the new manifests, building with kustomize and running kubectl apply again.
• You might have to adjust your over the top overlays and components if needed.
• You might have to prune old resources. For that you would add labels to all your resources from the start.
• With labels you can use kubectl apply with --prune and --dry-run to list prunable resources.
• Sometimes there are major changes, e.g. in the 1.9 release we switch to oauth2-proxy, which need additional attention.
• Nevertheless with a bit of Kubernetes knowledge one should be able to upgrade.

The Manifest Working Group releases Kubeflow based on the release timeline. The community and the release team work closely with the Manifest Working Group to define the specific dates at the start of the release cycle and follow the release versioning policy, as defined in the Kubeflow release handbook.

To view all past security scans, head to the Image Extracting and Security Scanning GitHub Action workflow. In the logs of the workflow you can expand the Run image extracting and security scanning script step to view the CVE logs. You will find a per-image CVE scan and a JSON dump of per-WorkingGroup aggregated metrics. You can run the Python script from the workflow file locally on your machine to obtain the detailed JSON files for any git commit.

The Kubeflow security working group follows a responsible disclosure policy for CVE results:

• Internal Review: All CVE findings are initially reviewed internally by the security working group.
• Severity Assessment: Each CVE is assessed for severity and potential impact on the Kubeflow project.
• Disclosure: For high and critical severity CVEs, the security working group will:
  • Notify the maintainers and contributors
  • Try to provide a fix or mitigation strategy
  • Publicly disclose the CVE details

• Q: What versions of Istio, Knative, Cert-Manager, Argo, ... are compatible with Kubeflow?
  A: Please refer to each individual component's documentation for a dependency compatibility range. For Istio, Knative, Dex, Cert-Manager and OAuth2 Proxy, the versions in common are the ones we have validated.
• Q: Can I use earlier version of Kustomize with Kubeflow manifests? A: No, it is not supported anymore, although it might be possible with manual effort.