Deploying Kubeflow everywhere: desktop, edge, and IoT devices

by Rui Vasconcelos on 4 November 2020

Kubeflow, the ML toolkit on K8s, now fits on your desktop and edge devices! 🚀

Data science workflows on Kubernetes

Kubeflow provides the cloud-native interface between Kubernetes and data science tools: libraries, frameworks, pipelines, and notebooks.

> Read more about what is Kubeflow

Cloud-native MLOps toolkit gets heavy

To make Kubeflow the standard cloud-native tool for MLOps within the AI landscape, the open-source community has accomplished the aggregation and integration of many projects on top of Kubernetes.

Unfortunately, this notable accomplishment also has a downside. Deploying Kubeflow on your laptop or edge device has become impractical.

The very minimum memory necessary to deploy the full Kubeflow bundle is 12Gb of RAM.

On top of that, it is Linux-based. This means that on Windows and macOS you need to allocate 12+ Gb of memory to a Linux VM.

Last time I tried, my 16Gb of RAM MacBook Pro did not like the idea.

Kubeflow lite to experiment on your desktop: Windows, macOS or Linux

To allow users to conveniently try out Kubeflow directly on their laptops or workstations, Canonical has conveniently pre-selected and packaged a subset of the Kubeflow applications to run on 8Gb of RAM.

Kubeflow runs on top of Kubernetes. Hence, in order to provide an out-of-the-box Kubeflow experience, the underlying K8s had to be also provided in a streamlined way.

The simple way to get K8s with built-in Kubeflow on Windows, macOs or Linux is MicroK8s

Now, besides the full Kubeflow bundle, MicroK8s also includes a Kubeflow lite bundle. To install Kubelow lite, deploy MicroK8s and then run:

$ KUBEFLOW_BUNDLE=lite microk8s enable kubeflow

> Check out what’s inside Kubeflow-lite

Kubeflow edge for inference and distributed training

Going even smaller in terms of the memory footprint, the Kubeflow edge bundle was born.

Kubeflow edge uses the inference and distributed training pieces inside Kubeflow – including TF-job-operator, PyTorch-operator, Seldon Core and Kubeflow Pipelines – and packages them for a 4Gb of RAM device to run.

So far, we have seen this option generate the most impact in industries that leverage an IoT mesh, such as manufacturing, mobility, retail, or ag-tech.

To install Kubeflow-edge, deploy MicroK8s and then run:

$ KUBEFLOW_BUNDLE=edge microk8s enable kubeflow

> Check out what’s inside Kubeflow edge

DIY: Build your own Kubeflow deployment

With the vision to empower AI innovators leveraging Kubeflow, not constrain them, Canonical has created Kubeflow lite and Kubeflow edge to get you started quickly wherever you are.

Once you have familiarized yourself with all that Kubeflow can offer, you can quickly add any application inside Kubeflow to your current bundle.

You could, for example, start with Kubeflow lite and add Katib the hyperparameter tuning piece of Kubeflow later on. To do this, run:

$ microk8s.juju deploy <app name> 

In addition, you can integrate applications that have been deployed with the command:

$ microk8s.juju relate <app A> <app B>

> Check out the list of 60 Kubeflow related applications

Kubeflow operators: the magic behind Kubeflow everywhere

This easy to deploy and composable Kubeflow is only possible due to Charmed Kubeflow, the set of charm operators that wrap the 20+ apps inside upstream Kubeflow.

> Read Kubeflow operators: lifecycle management for data science

Get started with Kubeflow

If you haven’t yet got a taste of Kubeflow, you can follow the upstream Kubeflow workstation docs, or watch the video below:

> Read the docs on MicroK8s Kubeflow add-on for more details.

Further reading

Related posts

The software operator design pattern: disadvantages – part 5

The software operator is a design pattern – it is a proven design that has been applied in many situations and implemented by several frameworks. Its widespread adoption in the industry has allowed us to study its consequences, both good and bad. This is important because software developers and IT architects need to know when […]

The software operator design pattern: advantages – part 4

The software operator is a design pattern. Its design is based on successful applications where this approach was found useful. In other words, it’s a proven approach that can be recommended to others. But like all approaches, it’s important to understand their positive and negative impact. Software developers need to understand when the […]

Operator Day at Kubecon EU 2022 – recordings available!

The Operator Day at Kubecon EU 2022, hosted by Canonical, took place on Monday 16 May  2022. We thank everyone for attending the event. Our thanks go out especially to those who engaged with us via chat during the online event. We enjoyed answering questions and having conversations during the presentations.  If you missed Operator […]