Your submission was sent successfully! Close

Thank you for signing up for our newsletter!
In these regular emails you will find the latest updates from Canonical and upcoming events where you can meet our team.Close

Thank you for contacting our team. We will be in touch shortly.Close

  1. Blog
  2. Article

Andreea Munteanu
on 16 September 2022

FAQ: MLOps with Charmed Kubeflow

Charmed Kubeflow is Canonical’s Kubeflow distribution and MLOps platform. The latest release shipped on 8 September. Our engineering team hosted a couple of livestreams to answer the questions from the community: a beta-release webcast and a technical deep-dive. In case you missed them, you can read the most frequently asked questions (FAQ) about MLOps and access helpful resources in this blog post.

Note that you can also watch the videos on Youtube: Beta-release & a technical deep-dive.

Upstream Kubeflow and Charmed Kubeflow: the differences explained

What’s the main feature of the new Kubeflow 1.6?

One of the themes of this version of Kubeflow was improved user experience and pipelines 2.0 in particular. The new release comes with improved input-output rules, faster meta-data support, and simpler authoring components. Learn more about the new Kubeflow pipelines in our livestream.
Kubeflow 1.6 also supports Kubernetes 1.22, and many bug fixes related to Notebooks. Learn more about this here.

What’s the difference between Kubeflow upstream and Charmed Kubeflow?

Charmed Kubeflow is an official distribution of the upstream project. It includes the same features and follows the same release cycle and roadmap development. The main difference is that Charmed Kubeflow uses charms as operators that manage its lifecycle.

Charms are Kubernetes operators that automate maintenance and security operations. They accelerate workload deployment,  allowing data scientists to take models to market more efficiently.

Will Charmed Kubeflow always release at the same time as the upstream?

Canonical made a considerable effort to align the release cycle with the upstream release. Part of our team also actively contributes to the upstream project.

Are Charmed Kubeflow’s pipelines aligned with the upstream ones?

Yes. Charmed Kubeflow’s pipelines have the same features as the upstream project. It applies to all components that Canonical’s product has.

Charmed Kubeflow: understanding its features 

As a Charmed Kubeflow user, why would you choose the latest stable release?

Charmed Kubeflow has more versions that can be found on CharmHub. Latest stable is what most of the users aim for because it has been extensively tested and verified by the engineering team. However, edge is dedicated to those who are interested in testing bleeding edge features the engineering team has been working on.

What’s the difference between manifests and charms?

Kubeflow Manifests provide a static reference deployment for Kubeflow.  They can be what you need if you’re interested in being your own expert by managing your own configuration changes, catching typical errors, and general maintenance.  Charmed Kubeflow wraps the same applications of Kubeflow in operators that handle a lot of the configuration and maintenance hassles.  From deployment to day-2 operations, these operators make managing Kubeflow easier by automating and handling common situations.  These operators also provide easier integrations with other tools, such as observability, using Grafana or Prometheus, through interoperability with the rest of the charming ecosystem found on CharmHub.

Why is Charmed Kubeflow integrated with an observability stack?

Whenever there is a big deployment,  system administrators are interested in understanding what’s happening with the product. The integration with Grafana and Prometheus gives them further information about the status of the deployment through the collection of metrics and logs. The administrator can see operational details, like how many resources are being used or how many deployments are live.

What’s the difference between the beta and general availability release?

The main difference between Charmed Kubeflow 1.6 Beta and the general availability release consists of small bug fixes, related to various components such as Tensorboard or Notebooks. All known issues in Charmed Kubeflow 1.6 Beta are available here.  They have been addressed and fixed for the general availability version.

Integrations with Charmed Kubeflow

How can you host the dashboard publicly?

Charmed Kubeflow can host and access the dashboard from a public domain. In order to install it, please follow the quick start guide. To access the dashboard, follow our docs. For having your dashboard accessible through the public domain you would have to expose the istio-ingressgateway service through a public domain, like you would with any other Kubernetes Service. Be aware of the additional security risks that come with exposing dashboards to the public internet.

Will the old pipelines work after the update?

Yes. You can learn more about kpf v2 from the upstream documentation.
If you would like to upgrade from Charmed Kubeflow 1.4 to Charmed Kubeflow 1.6 please follow our guide.

Any suggestions for storage class used in multiple nodes under Microk8s 1.22?

The Microk8s team provides an OpenEBS addon with some more advanced storage features than default Microk8s storage.

Charmed Kubeflow: what’s next

What will be the focus for the next release?

In the future, Charmed Kubeflow is going to keep evolving. From a CI/CD angle, there will be more scheduled testing, for both charms and bundles. Canonical will improve the out-of-the-box experience, by having more detailed documentation and projects to help users get started. Moreover, the observability integration is going to grow.  There will be a set in the Grafana dashboard, such that once deployed, system administrators have all the information handy.

At the moment, the upstream project is building the new release team. Shortly after, the team is going to start working on the roadmap. Join the community meetings if you want to stay up to date on this.

What other apps are you thinking of integrating with?

Charmed Kubeflow is a product that Canonical is looking forward to integrating with other applications such as MLFlow, Spark or Mindspore. Improving model management is one of the challenges that Canonical wants to address by integrating with other applications.

Do you consider ML model monitoring as well?

On our next roadmap, we are looking into this feature and how to enable it for our users. Our engineering team is looking into model monitoring and model drift monitoring, analysing the various options and apps that could be used for these tasks.

Join the community

 How can we get involved in the upstream community?

Anyone who is interested in Kubeflow can contribute to the community in different ways. Follow the upstream guidelines and start contributing right away. The release team is a great way to get started because you get an overview of the whole project. You can sign up until 28 September!
If you are a Charmed Kubeflow user, you can share your feedback on Discourse or raise issues on our Git.

If you have other questions about MLOps or Charmed Kubeflow, please contact us on Discourse.

Related posts

Andreea Munteanu
19 February 2024

Edge AI: what, why and how with open source

Ubuntu Article

Edge AI is transforming the way that devices interact with data centres, challenging organisations to stay up to speed with the latest innovations. From AI-powered healthcare instruments to autonomous vehicles, there are plenty of use cases that benefit from artificial intelligence on edge devices. This blog will dive into the topic, capt ...

Andreea Munteanu
5 January 2024

AI in 2024 – What does the future hold?

AI Article

AI in 2024 ...

Andreea Munteanu
24 November 2023

Building a comprehensive toolkit for machine learning

AI Article

In the last couple of years, the AI landscape has evolved from a researched-focused practice to a discipline delivering production-grade projects that are transforming operations across industries. Enterprises are growing their AI budgets, and are open to investing both in infrastructure and talent to accelerate their initiatives – so it’ ...