This article is based on an interview with Stephan Fabel, director of product management at Canonical Ltd, conducted by John Furrier of theCUBE at KubeCon + CloudNativeCon EU 2018 in Copenhagen, Denmark. Original source: SiliconAngle.com.
The Core Problem: Developers Shouldn’t Need to Be DevOps Engineers
Kubernetes has become one of the most important pieces of infrastructure in modern software delivery, but that power has historically come with a steep operational learning curve. For data scientists and application developers, understanding how to build container images, configure Kubernetes clusters, and manage deployments is often a distraction from their primary work: writing and iterating on applications.
Canonical Ltd — the company behind Ubuntu and a certified Kubernetes distributor — has been working to close this gap. According to Stephan Fabel, Canonical’s director of product management, the goal is to remove the DevOps burden from developers so they can concentrate on what they do best.
“The people who are actually writing those applications are not DevOps people, they’re data scientists, right? They shouldn’t have to learn how to deploy Kubernetes, how to create a container and all those things.”
— Stephan Fabel, Director of Product Management, Canonical Ltd
Canonical’s Approach: Pure Upstream Kubernetes with Cloud Portability
Canonical distributes a version of Kubernetes that stays conformant with upstream releases, meaning workloads are not locked into Canonical-specific extensions. This conformance is deliberate: it allows teams to migrate workloads freely between on-premises infrastructure and major public cloud providers without redesigning applications.
“Pure upstream, remain conformant to the main public clouds, such as to enable that workload migration and mobility from on-prem up to any of the other providers to accommodate all kinds of use cases.”
— Stephan Fabel, KubeCon EU 2018
This philosophy aligns with one of Kubernetes’ core promises: write your application once, deploy it anywhere. By staying close to the upstream project and maintaining CNCF conformance certification, Canonical avoids the fragmentation that can occur when distributions diverge significantly from the mainline.
Kubeflow: Machine Learning Workflows Without the Infrastructure Headache
A significant part of Canonical’s effort at KubeCon EU 2018 centered on Kubeflow, an open-source machine learning platform designed to run on Kubernetes. Kubeflow provides a composable, portable, and scalable stack for ML workflows, enabling data scientists to build, train, and deploy models without needing deep Kubernetes expertise.
Before tools like Kubeflow, deploying a machine learning application on Kubernetes and later migrating it to a different environment often required significant re-engineering. That rework introduced new bugs and slowed down iteration cycles. Kubeflow, in partnership with Canonical’s distribution, shifts the focus back to the application logic itself.
What Kubeflow Provides
- Portability: Run the same ML pipeline locally, on-premises, or on any major cloud provider.
- Composability: Assemble ML workflows from modular components rather than building monolithic pipelines.
- Scalability: Leverage Kubernetes’ native scheduling and resource management to scale training jobs on demand.
- Developer focus: Abstract away container orchestration so data scientists interact with familiar ML tooling instead of raw Kubernetes manifests.
Kubernetes as a Building Block for Next-Generation Applications
Fabel framed Kubernetes not as an end goal but as a foundational building block for composing more complex applications, including those that incorporate machine learning algorithms as a core component of a web application.
“Actually using things like Kubernetes as an effective building block to then build out web applications that use things like machine learning algorithms underneath — that’s a perfect use case for a next-gen workload and also something that we might use ourselves internally.”
— Stephan Fabel, KubeCon EU 2018
This composable, building-block model has practical benefits beyond machine learning. When individual components of an application are encapsulated in containers and orchestrated by Kubernetes, teams can update, scale, or replace those components independently. The result is an application architecture that is more resilient and significantly easier to evolve over time.
What This Means for Developer and DevOps Team Collaboration
The abstraction that Kubernetes and Kubeflow provide does not eliminate DevOps work — it repositions it. Platform and DevOps teams still design, secure, and maintain the underlying Kubernetes infrastructure. What changes is that developers and data scientists gain self-service deployment capabilities without needing to understand the full operational stack beneath them.
This separation of concerns reflects a maturing pattern in the industry:
- Platform teams build and maintain the Kubernetes environment, enforce policies, and manage upgrades.
- Application developers use standardized deployment interfaces (Helm charts, Kubeflow pipelines, operators) to ship code without writing raw infrastructure configuration.
- Data scientists interact with ML-focused tooling that sits on top of Kubernetes, submitting training jobs and serving models through purpose-built UIs and APIs.
This model reduces friction across all three groups and shortens the feedback loop between writing code and running it in a production-like environment.
Key Takeaways
- Canonical distributes a conformant, pure-upstream version of Kubernetes designed to support workload portability across on-premises and public cloud environments.
- Kubeflow abstracts Kubernetes complexity for data scientists, enabling portable and scalable ML workflows without deep infrastructure knowledge.
- The combination of Kubernetes and tools like Kubeflow positions Kubernetes as a composable building block for next-generation applications, not just a deployment target.
- The DevOps burden is shifted from individual developers to platform teams, improving overall team efficiency without eliminating operational rigor.
Disclosure: This article is based on editorial coverage of a theCUBE interview conducted at KubeCon + CloudNativeCon EU 2018. SiliconAngle Media sponsors theCUBE interviews at industry events.