This bundle deploys KubeFlow to a Juju K8s model. The individual charms that
make up this bundle can be found under
Read the documentation.
Setting up this bundle requires a Ubuntu installation or VM.
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
You'll also need to install the
sudo snap install microk8s --classic
Next, you will need to add yourself to the
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
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 220.127.116.11. To do this, edit the coredns configmap with this command:
microk8s.kubectl edit configmap -n kube-system coredns
Edit the line with
18.104.22.168 22.214.171.124 to use your local DNS, e.g.
Setup Charmed Kubernetes
You'll also need to install the
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
Most interactions will go through the central dashboard, which is available via
/. The deploy scripts will print out the address you can point
your browser to when they are done deploying.
Pipelines are available either by the main dashboard, or from within notebooks via the fairing library.
Note that until https://github.com/kubeflow/pipelines/issues/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.
You can view pipelines from the Pipeline Dashboard available on the central
dashboard, or by going to
To submit a TensorFlow job to the dashboard, you can run this
kubectl create -n <NAMESPACE> -f path/to/job/definition.yaml
<NAMESPACE> matches the name of the Juju model that you're using,
path/to/job/definition.yaml should point to a
similar to the
mnist.yaml example found here.
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
You can destroy the controller itself with this command:
# For microk8s juju destroy-controller $(juju show-controller | head -n1 | sed 's/://g') --destroy-storage
To run the test suite included in this repository, start by installing the Python dependencies:
pip install --user -r requirements.txt -r test-requirements.txt
Next, ensure that you either have the
juju-helpers snap package installed, or you have
kubectl binary available with
~/.kube/config set up correctly.
Then, run the tests with this command:
pytest tests/ -m <bundle>
<bundle> is whichever bundle you have deployed, one of
If you have Charmed Kubeflow deployed to a remote machine with an SSH proxy available (for example, if you have MicroK8s running on an AWS VM), you can run the tests like this to run them against the remote machine:
pytest tests/ -m <bundle> --proxy=localhost:9999 --url=http://10.64.140.43.xip.io/ --password=password
Additionally, if you'd like to view the Selenium tests as they're in progress, you can
pass in the
--headful option like this:
pytest tests/ -m <bundle> --headful