This is an update to previous tutorials, focused on deploying Kubernetes 1.17.6 (released in May 2020, based on 1.17.0 released in December 2019).

We will use Kubespray 2.13.1, which first runs terraform to create the Openstack resources, then ansible to configure the servers to run all the Kubernetes services.

Create Jetstream Virtual machines with Terraform

Terraform allows to execute recipes that describe a set of OpenStack resources and their relationship. In the context of this tutorial, we do not need to learn much about Terraform, we will configure and execute the recipe provided by kubespray.


On a Ubuntu 18.04 install python3-openstackclient with APT. Any other platform works as well, also install terraform by copying the correct binary to /usr/local/bin/, see Install the newest 0.12.x release, I tested with 0.12.26.

Request API access

In order to make sure your XSEDE account can access the Jetstream API, you need to contact the Helpdesk, see the instructions on the Jetstream Wiki. You will also receive your TACC password, which could be different than your XSEDE one (username is generally the same).

Login to the TACC Horizon panel at, this is basically the low level web interface to OpenStack, a lot more complex and powerful than Atmosphere available at Use tacc as domain, your TACC username (generally the same as your XSEDE username) and your TACC password.

First choose the right project you would like to charge to in the top dropdown menu (see the XSEDE website if you don’t recognize the grant code).

Click on Compute / API Access and download the OpenRC V3 authentication file to your machine. Source it typing:


it should ask for your TACC password. This configures all the environment variables needed by the openstack command line tool to interface with the Openstack API.

Test with:

openstack flavor list

This should return the list of available “sizes” of the Virtual Machines.

Clone kubespray

I needed to make a few modifications to kubespray to adapt it to Jetstream:

git clone
git checkout -b branch_v2.13.1 origin/branch_v2.13.1

See an overview of my changes compared to the standard kubespray release 2.13.1.

Run Terraform

Inside jetstream_kubespray, copy from my template:

cp -LRp inventory/zonca inventory/$CLUSTER
cd inventory/$CLUSTER

Open and modify cluster.tfvars, choose your image and number of nodes. Make sure to change the network name to something unique, like the expanded form of $CLUSTER_network.

You can find suitable images (they need to be JS-API-Featured, you cannot use the same instances used in Atmosphere):

openstack image list | grep "JS-API"

The default is JS-API-Featured-Ubuntu18-Latest.

I already preconfigured the network UUID both for IU and TACC, but you can crosscheck looking for the public network in:

openstack network list

Initialize Terraform:


Create the resources:


The last output log of Terraform should contain the IP of the master node k8s_master_fips, wait for it to boot then SSH in with:

ssh ubuntu@$IP

or centos@$IP for CentOS images.

Inspect with Openstack the resources created:

openstack server list
openstack network list

You can cleanup the virtual machines and all other Openstack resources (all data is lost) with bash

Install Kubernetes with kubespray

Change folder back to the root of the jetstream_kubespray repository,

First make sure you have a recent version of ansible installed, you also need additional modules, so first run:

pip install -r requirements.txt

It is useful to create a virtualenv and install packages inside that. This will also install ansible, it is important to install ansible with pip so that the path to access the modules is correct. So remove any pre-installed ansible.

Or install ansible with conda, the minimum required version of ansible is 2.9.

Then following the kubespray documentation, we setup ssh-agent so that ansible can SSH from the machine with public IP to the others:

eval $(ssh-agent -s)
ssh-add ~/.ssh/id_rsa

Test the connection through ansible:

ansible -i inventory/$CLUSTER/hosts -m ping all

If a server is not answering to ping, first try to reboot it:

openstack server reboot $CLUSTER-k8s-node-nf-1

Or delete it and run to create it again.

check inventory/$CLUSTER/group_vars/all.yml, in particular bootstrap_os, I setup ubuntu, change it to centos if you used the Centos 7 base image.

Finally run the full playbook, it is going to take a good 10 minutes:


If the playbook fails with “cannot lock the administrative directory”, it is due to the fact that the Virtual Machine is automatically updating so it has locked the APT directory. Just wait a minute and launch it again. It is always safe to run ansible multiple times.

If the playbook gives any error, try to retry the above command, sometimes there are temporary failed tasks, Ansible is designed to be executed multiple times with consistent results.

You should have now a Kubernetes cluster running, test it:

$ ssh ubuntu@$IP
$ sudo su
$ kubectl get pods --all-namespaces
NAMESPACE       NAME                                                   READY     STATUS    RESTARTS   AGE
cert-manager    cert-manager-597bc67bf9-cf7xv                 1/1     Running   0          2m9s
ingress-nginx   ingress-nginx-controller-tfqbn                1/1     Running   0          2m31s
kube-system     coredns-76798d84dd-hhjs9                      1/1     Running   0          98s
kube-system     coredns-76798d84dd-l6fh2                      1/1     Running   0          83s
kube-system     dns-autoscaler-85f898cd5c-kld9w               1/1     Running   0          90s
kube-system     kube-apiserver-dummy2-k8s-master-1            1/1     Running   1          5m11s
kube-system     kube-controller-manager-dummy2-k8s-master-1   1/1     Running   0          5m11s
kube-system     kube-flannel-tzj4h                            1/1     Running   0          3m13s
kube-system     kube-flannel-xqq4j                            1/1     Running   0          3m13s
kube-system     kube-proxy-tfpzd                              1/1     Running   0          3m39s
kube-system     kube-proxy-z2djx                              1/1     Running   0          3m39s
kube-system     kube-scheduler-dummy2-k8s-master-1            1/1     Running   1          5m11s
kube-system     kubernetes-dashboard-77475cf576-ht5th         1/1     Running   0          82s
kube-system     kubernetes-metrics-scraper-747b4fd5cd-sdxsc   1/1     Running   0          82s
kube-system     nginx-proxy-dummy2-k8s-node-nf-1              1/1     Running   0          3m21s
kube-system     nodelocaldns-sg8x4                            1/1     Running   0          86s
kube-system     nodelocaldns-vljgx                            1/1     Running   0          86s

Compare that you have all those services running also in your cluster. We have also configured NGINX to proxy any service that we will later deploy on Kubernetes, test it with:

$ wget localhost
--2018-09-24 03:01:14--  http://localhost/
Resolving localhost (localhost)...
Connecting to localhost (localhost)||:80... connected.
HTTP request sent, awaiting response... 404 Not Found
2018-09-24 03:01:14 ERROR 404: Not Found.

Error 404 is a good sign, the service is up and serving requests, currently there is nothing to deliver. Finally test that the routing through the Jetstream instance is working correctly by opening your browser and test that if you access you also get a default backend - 404 message. If any of the tests hangs or cannot connect, there is probably a networking issue.

(Optional) Setup kubectl locally

We also set kubectl_localhost: true and kubeconfig_localhost: true. so that kubectl is installed on your local machine

Or install it yourself, I tested with 1.18.3.

it also copies admin.conf to:


I have a script to copy that to .config/kube and to replace the IP with localhost, because we cannot replace it with the public floating ip because the certificate is not valid for that.


Then make a SSH tunnel (lasts 3 hours):


(Optional) Setup helm locally

Install helm 3 from the release page on Github

I tested with v3.2.4.

Install Jupyterhub

It is preferable to run the Hub and the Proxy on the master node, just in case we want to downsize the cluster to only one node to save resources.

We need to remove the taint from the master node because currently the JupyterHub recipe doesn’t allow to add tolerations to the master node.

kubectl edit node $CLUSTER-k8s-master-1

Remove the 3 lines with the taint:

  - effect: NoSchedule

Now checkout the JupyterHub configuration files repository on the local machine (if you have setup kubectl and helm locally, otherwise on the master node).

git clone

Inside that, first run


to create the secret strings needed by JupyterHub then edit its output secrets.yaml to make sure it is consistent, edit the hosts lines if needed. For example, supply the Jetstream DNS name of the master node (XXX and YYY are the last 2 groups of the floating IP of the instance AAA.BBB.XXX.YYY).

kubectl create namespace jhub

Check some preliminary pods running with:

kubectl get pods -n jhub

Once the proxy is running, even if hub is still in preparation, you can check in browser, you should get “Service Unavailable” which is a good sign that the proxy is working.

Customize JupyterHub

After JupyterHub is deployed and integrated with Cinder for persistent volumes, for any other customizations, first authentication, you are in good hands as the Zero-to-Jupyterhub documentation is great.

Setup HTTPS with letsencrypt

Kubespray has the option of deploying also cert-manager, but I had trouble deploying an issuer, it was easier to just deploy it afterwards following my previous tutorial


Feedback on this is very welcome, please open an issue on the Github repository or email me at zonca on the domain of the San Diego Supercomputer Center (