Concepts
Seven mental models that show up across every Workbench tab.
Workspace
Your team's slice of the cluster. Workspaces hold quota (CPUs, memory, GPUs, storage), an RBAC scope, and a cost envelope. Resources you create in Workbench live inside one workspace at a time. The current workspace is shown in the top-right of the screen — switch with the workspace picker.
Preset
A reusable template (WorkspaceKind) that defines what a session looks like before you create it. A preset bundles an IDE type (JupyterLab, VS Code, RStudio), a catalog of container images, a set of compute sizes (pod sizes), and default storage volumes. When you create a session, you pick a preset first and the wizard fills in the options around it. Presets also exist for training jobs and endpoints, but those are read-only or admin-managed.
Service
A lightweight, browser-accessible tool that runs on cluster compute. Services include Cloud Shell (terminal access), Remote Desktop (VNC GUI), and PVC Viewer (file browser). They share a common lifecycle — create, open, delete — and are cheaper and faster to spin up than a full session.
Compute pool
A named bundle of where a workload runs: GPU vendor and count per node, instance type, autoscaling min/max, and a default cost rate. A single pool is reused by sessions, training jobs, sweeps, and endpoints — pick one once and the rest of the form auto-tunes around it.
Lifecycle
Most resources move through the same set of phases:
| Phase | Meaning |
|---|---|
| Pending | Provisioning compute, pulling images, mounting storage. |
| Running | Healthy and accepting work. |
| Succeeded | Finite job completed cleanly. Logs and artifacts persist. |
| Paused | Compute released. Storage retained. Resume to restart. |
| Failed | Last reconcile errored. Inspect Activity on the detail page. |
Cost
Every active resource accumulates spend. Workbench shows three numbers: $/hr (the live burn rate), accumulated (since the resource started), and idle % (compute reserved but not running work). Idle spend is the most common preventable cost — sessions left running overnight are the usual culprit.
Observability
Detail pages link out to Grafana dashboards scoped to that resource. Cluster-wide metrics live on the Observability tab.