Jupyter as a Service on FlashBlade

How to manage dev environments for data scientists at scale.

Photo by Planet Volumes on Unsplash

Jupyter notebooks are a great tool for data scientists to explore datasets and experiment with model development. They enable developers to easily supplement code with analysis and visualizations.

Rather than the historical practice of having users manage their own notebook servers, JupyterHub can be deployed by an organization to offer a centralized notebook platform. JupyterHub also enables infrastructure teams to give each user access to centralized storage for: shared datasets, scratch space, and a persistent IDE.

This blog post presents an example of deploying Jupyter-as-a-Service on Pure Storage FlashBlade. Users are able to create new notebook servers on the fly within a Kubernetes cluster with zero-touch provisioning. IT teams are able to manage efficient use of compute and storage resources across users.

JupyterHub

JupyterHub is used to manage and proxy multiple instances of the “single-user” Jupyter notebook server. It provides a public HTTP proxy on your network so users can login to a central landing page from their browser. Once a user logs in, JupyterHub spins up a server (pod) for that user. It reconnects to that user’s persistent storage. So, users can have stateful dev environments, but the compute nodes are only used as needed.

We’ll deploy JupyterHub as a Kubernetes service so it’s easily manageable as part of a cluster.

FlashBlade in a Kubernetes environment

FlashBlade is an excellent storage backend for JupyterHub for a few reasons.

First, it enables access to training datasets in-place, eliminating the need to copy datasets between nodes. Data scientists can perform training and testing of models using shared datasets with minimal data management.

Second, FlashBlade supports Pure’s CSI driver, Pure Service Orchestrator (PSO). PSO fully automates creation and management of PersistentVolumes (PV) for applications in a Kubernetes cluster, which brings self-service to a JupyterHub deployment by eliminating manual storage administration for new users.

JupyterHub is just one of the many applications that together, form a complete AI platform for data scientists. All of these applications should be backed by the same, centralized storage for management simplicity and efficient data management.

Remove storage silos.

Installation

Prep Steps

• Clone our AI Platform repo
• On your local machine, ensure that your cluster config is active for kubectl.
• On your local machine, install Helm and the Helm repos for PSO and for JupyterHub:

helm repo add pure https://purestorage.github.io/helm-charts
helm repo add jupyterhub https://jupyterhub.github.io/helm-chart/

• Each node in the cluster needs to have access to the datasets on FlashBlade. Mount the datasets folder directly to each cluster node at /datasets.

Deploy PSO

Customize: You’ll need a psovalues.yaml file that describes your FlashBlade array. The easiest thing to do is copy our default ./psovalues.yaml and adjust the “arrays” section.

Example customization:

arrays:
  FlashBlades:
    - MgmtEndPoint: "10.61.169.20"                        # CHANGE
      APIToken: "T-c4925090-c9bf-4033-8537-d24ee5669135"  # CHANGE
      NFSEndPoint: "10.61.169.30"                         # CHANGE

Install:

helm install pure-storage-driver pure/pure-csi — namespace jhub -f ./psovalues.yaml

Deploy a PV for shared datasets

Customize:

The ./datasetpv.yaml file is used create a Persistent Volume Claim named “shared-ai-datasets”. Adjust datasetpv.yaml to use your FlashBlade Data VIP and filesystem name.

nfs:
  server: 10.61.169.100      # CHANGE to your data vip
  path: /datasets            # CHANGE to your filesystem name

Install:

kubectl create -f datasetpv.yaml

Deploy JupyterHub

Customize:

The only change required for the jupvalues.yaml file is to add a security token. Generate a random hex string:

openssl rand -hex 32

Copy the output and, in your jupvalues.yaml file, replace the phrase SECRET_TOKEN with your generated string:

proxy:
  secretToken: 'SECRET_TOKEN'   # CHANGE to 32-digit secret token 

Install:

helm install jhub jupyterhub/jupyterhub — namespace jhub — version 0.8.2 -f jupyterhub/values.yaml

Use Jupyter notebooks!

JupyterHub is now ready for use.

Installing JupyterHub creates a proxy service that serves traffic for end users. The public address (proxy-public) can be found via:

> kubectl --namespace=jhub get svc proxy-public
NAME           TYPE           CLUSTER-IP       EXTERNAL-IP                
proxy-public   LoadBalancer   10.43.197.255.   10.61.169.60    

When a user navigates to proxy-public’s external-IP address, they’ll get the JupyterHub login screen:

When Victor logs in, he has access to shared datasets (like cifar10 and openimages) as well as his home directory of personal notebooks, plots, and files.

Conclusion

Running JupyterHub as a service within a Kubernetes cluster is easy to deploy and manage. Data scientists not only have persistent storage backing their personal environments, but they also have access to all shared datasets without time-consuming data copying or complex data management.

Our code is on github: https://github.com/PureStorage-OpenConnect/ai-platform

Try out these quick installation steps and let us know how it goes.