Deploy Dask Gateway with JupyterHub on Kubernetes


April 4, 2022

Tutorial obsolete, see the new version of the tutorial

Updated 28 April 2022: switched to Dask Gateway 2022.4.0

In this tutorial we will install Dask Gateway on Kubernetes and configure JupyterHub so Jupyter Notebook users can launch private Dask cluster and connect to them.

I assume to start from a Kubernetes cluster already running and JupyterHub deployed on top of it via Helm. And SSL encryption also activated (it isn’t probably necessary, but I haven’t tested that). I tested on Jetstream 2, but the recipe should be agnostic of that.


Clone on the machine you use to run helm and kubectl the repository with the configuration files and scripts:

git clone

Then you need to setup one API token, create it with:

openssl rand -hex 32

Then paste it both in dask_gateway/config_jupyterhub.yaml and dask_gateway/config_dask-gateway.yaml, look for the string TOKEN and replace it.

Launch dask gateway

We can install version 2022.4.0 with:

$ bash

You might want to check config_dask-gateway.yaml for extra configuration options, but for initial setup and testing it shouldn’t be necessary.

After this you should see the 3 dask gateway pods running, e.g.:

$ kubectl -n jhub get pods
NAME                                       READY   STATUS    RESTARTS   AGE
api-dask-gateway-64bf5db96c-4xfd6          1/1     Running   2          23m
controller-dask-gateway-7674bd545d-cwfnx   1/1     Running   0          23m
traefik-dask-gateway-5bbd68c5fd-5drm8      1/1     Running   0          23m

Modify the JupyterHub configuration

Only 2 options need to be changed in JupyterHub:

  • We need to run a image which has the same version of dask-gateway we installed on Kubernetes (currently 0.9.0)
  • We need to proxy dask-gateway through JupyterHub so the users can access the Dask dashboard

If you are using my script to deploy JupyterHub, you can modify it and add another values option at the end, --values dask_gateway/config_jupyterhub.yaml.

You can modify the image you are using for Jupyterhub in dask_gateway/config_jupyterhub.yaml.

To assure that there are not compatibility issues, the “Client” (JupyterHub session), the dask gateway server, the scheduler and the workers should all have the same version of Python and the same version of dask, distributed and dask_gateway. If this is not possible, you can test different combinations and they might work.

Then redeploy JupyterHub:

bash && cd dask_gateway && bash

Check that the service is working correctly, if open a browser tab and access, you should see:

{"status": "pass"}

If this is not working, you can open login to JupyterHub, get a terminal and first check if the service is working:

>  curl http://traefik-dask-gateway/services/dask-gateway/api/health

Should give:

{"status": "pass"}

Create a dask cluster

You can now login to JupyterHub and check you can connect properly to dask-gateway:

from dask_gateway import Gateway
gateway = Gateway(

Then create a cluster and use it:

cluster = gateway.new_cluster()
client = cluster.get_client()

Client is a standard distributed client and all subsequent calls to dask will go through the cluster.

Printing the cluster object gives the link to the Dask dashboard.

For a full example and screenshots of the widgets and of the dashboard see:

(Click on the Raw button to download notebook and upload it to your session).