KubeFlow Bundle

Overview

This bundle deploys KubeFlow to a Juju K8s model. The individual charms that make up this bundle can be found under charms/.

Deploying

Setup

If you are on macOS or Windows, you will need to use an Ubuntu VM. You can install multipass and access an Ubuntu VM with these commands:

multipass launch --name kubeflow --mem 16G
multipass shell kubeflow

16G is the recommended amount of memory to apportion the VM. Less than that may run into issues with pods not coming up properly.

Once you have an Ubuntu environment, you'll need to install these snaps to get started:

sudo snap install juju --classic
sudo snap install juju-wait --classic
sudo snap install juju-helpers --classic

Next, check out this repository locally:

git clone https://github.com/juju-solutions/bundle-kubeflow.git
cd bundle-kubeflow

The below commands will assume you are running them from the bundle-kubeflow directory.

Then, follow the instructions from one of the subsections below to deploy Kubeflow to either microk8s or Charmed Kubernetes.

Setup microk8s

You'll also need to install the microk8s snap:

sudo snap install microk8s --classic

Next, you will need to add yourself to the microk8s group:

sudo usermod -aG microk8s $USER
newgrp microk8s

Finally, you can run these commands to set up microk8s:

python3 scripts/cli.py microk8s setup --controller uk8s
python3 scripts/cli.py deploy-to uk8s

The deploy-to command allows manually setting a public address that is used for accessing Kubeflow on MicroK8s. In some deployment scenarios, you may need to configure MicroK8s to use LAN DNS instead of the default of 8.8.8.8. To do this, edit the coredns configmap with this command:

microk8s.kubectl edit configmap -n kube-system coredns

Edit the line with 8.8.8.8 8.8.4.4 to use your local DNS, e.g. 192.168.1.1.

Setup Charmed Kubernetes

You'll also need to install the kubectl snap:

sudo snap install kubectl --classic

You will then need to create an AWS account for juju to use, and then add the credentials to juju:

$ juju add-credential aws
Enter credential name: kubeflow-test

Using auth-type "access-key".

Enter access-key: <YOUR ACCESS KEY>

Enter secret-key: <YOUR SECRET KEY>

Credential "kubeflow-test" added locally for cloud "aws".

Next, you can run these commands to set up Charmed Kubernetes:

python3 scripts/cli.py ck setup --controller ckkf
python3 scripts/cli.py deploy-to ckkf

Authentication

By default, this bundle deploys Dex configured with basic login credentials. Dex is able to user other services for authentication, such as GitHub's OAuth service. To configure Kubeflow to use one of these other services for authentication, check out the Dex documentation. The YAML shown in those examples should be configured with:

juju config dex-auth connectors="$CONNECTOR_YAML"

As an example, this is what you might use for $CONNECTOR_YAML if you are deploying with support for LDAP:

[{
    "id": "ldap",
    "name": "OpenLDAP",
    "type": "ldap",
    "config": {
        "bindDN": "cn=admin,dc=example,dc=org",
        "bindPW": "admin",
        "groupSearch": {
            "baseDN": "cn=admin,dc=example,dc=org",
            "filter": "",
            "groupAttr": "DN",
            "nameAttr": "cn",
            "userAttr": "DN"
        },
        "host": "ldap-service.auth.svc.cluster.local:389",
        "insecureNoSSL": true,
        "userSearch": {
            "baseDN": "cn=admin,dc=example,dc=org",
            "emailAttr": "DN",
            "filter": "",
            "idAttr": "DN",
            "nameAttr": "cn",
            "username": "cn"
        },
        "usernamePrompt": "Email Address"
    }
}]
'

For more thorough examples, see https://github.com/dexidp/dex/tree/master/examples If you'd like to disable the default basic credentials, run:

juju config dex-auth static-username='' static-password=''

Using

Main Dashboard

Most interactions will go through the central dashboard, which is available via Ambassador at /. The deploy scripts will print out the address you can point your browser to when they are done deploying.

Pipelines

Pipelines are available either by the main dashboard, or from within notebooks via the fairing library.

Note that until kubeflow/pipelines#1654 is resolved, you will have to attach volumes to any locations that output artifacts are written to, see the attach_output_volume function in pipline-samples/sequential.py for an example.

Argo UI

You can view pipelines from the Pipeline Dashboard available on the central dashboard, or by going to /argo/.

TensorFlow Jobs

To submit a TensorFlow job to the dashboard, you can run this kubectl command:

kubectl create -n <NAMESPACE> -f path/to/job/definition.yaml

Where <NAMESPACE> matches the name of the Juju model that you're using, and path/to/job/definition.yaml should point to a TFJob definition similar to the mnist.yaml example found here.

TensorFlow Serving

You can submit a model to be served with TensorFlow Serving:

# For a single model
juju deploy cs:~kubeflow-charmers/kubeflow-tf-serving --storage models=storage-class,, --config model=/path/to/base/dir/model-name

# For a model.conf:
juju deploy cs:~kubeflow-charmers/kubeflow-tf-serving --storage models=storage-class,, --config model-conf=/path/to/model.conf

Removing

Kubeflow model

To remove Kubeflow from your Kubernetes cluster, first run this command to remove Kubeflow itself:

juju destroy-model kubeflow --destroy-storage

If you encounter errors while destroying the model, you can run this command to force deletion:

juju destroy-model kubeflow --yes --destroy-storage --force

Alternatively, to simply release storage instead of deleting it, run with this flag:

juju destroy-model kubeflow --release-storage

Kubeflow controller

You can destroy the controller itself with this command:

# For microk8s
juju destroy-controller $(juju show-controller | head -n1 | sed 's/://g') --destroy-storage

lqdev / bundle-kubeflow