Concepts
Kubeflow Workspaces uses an option-based model that separates the development environment a
data scientist wants from the Kubernetes configuration used to run it. Platform administrators
define WorkspaceKind resources containing reusable software, compute, and policy options,
while data scientists create and update Workspace resources through a guided interface using
those available options.
Workspace and WorkspaceKind
A Workspace represents an interactive development environment intended for data scientists.
It is a namespaced resource that records the WorkspaceKind and configuration options selected
by the data scientist.
A WorkspaceKind is a cluster-scoped resource managed by platform administrators. It defines
a type of development environment, such as JupyterLab, Visual Studio Code (code-server), or
RStudio, together with the options and policies available to users.
This separation allows users to work with meaningful environment choices without configuring the underlying Kubernetes resources directly.
How the concepts fit together
The administrator defines a WorkspaceKind containing the
environment options and policies available to users:
In the Workspaces frontend, a data scientist creates a
Workspace through a guided process:
- Choose an available
WorkspaceKind - Select from the Image Config and Pod Config options presented by the frontend
- Create the
Workspace
The frontend presents the administrator-defined options, so data scientists do not need to write Kubernetes resource definitions themselves. The Workspace controller translates the selected configuration into the Kubernetes resources needed to run the environment.
Image Configs and Pod Configs
A WorkspaceKind can define two primary groups of options:
- Image Configs define the software environment available to a Workspace, including the container image and ports exposed by that configuration.
- Pod Configs define reusable pod-level configuration, such as compute resources and scheduling requirements.
Example:
A data scientist might select a JupyterLab WorkspaceKind, a PyTorch Image Config,
and a GPU Pod Config. Workspaces translates those selections into the Kubernetes configuration
defined by the cluster administrator.
Filter Rules
Filter Rules control which WorkspaceKinds, Image Configs, or Pod Configs are visible or allowed for a given context.
Rules can use information such as namespace labels or previously selected options to present only the configurations that are relevant to the user.
Example:
An administrator can use Filter Rules to show GPU Pod Configs only when a compatible Image Config is selected, or restrict certain configurations based on namespace labels.
Option Redirects
Option Redirects allow administrators to point an existing Image Config or Pod Config to a replacement.
This provides a controlled way to evolve the available environment catalog. When a data scientist starts or edits a Workspace, the frontend can display redirect information to explain which replacement configuration will be used.
Example:
An older image option can redirect to a newer image option without changing the configuration of an already running Workspace unexpectedly.
Activity Rules
Activity Rules define lifecycle policies for Workspaces based on their activity.
Activity Probes allow administrators to define how activity is detected for a Workspace, rather than relying on a single fixed definition of activity.
Administrators can configure Activity Rules to determine when an inactive Workspace becomes eligible to be paused. These policies can be tailored to different teams or hardware requirements.
Example:
An administrator might configure a shorter inactivity period for GPU-backed Workspaces to conserve expensive compute resources, while allowing longer periods for other environments.
Feedback
Was this page helpful?
Thank you for your feedback!
We're sorry this page wasn't helpful. If you have a moment, please share your feedback so we can improve.