Skip to main content
A workload is a single unit of work running on the platform. Every piece of code the platform executes, whether it’s a step in a workflow, an interactive workstation session, or a deployment answering requests, runs as a workload. The platform schedules each workload onto the compute available to its perimeter and tracks it until it finishes. On the cluster, a workload is a Kubernetes pod, sometimes managed under a job. You never interact with these objects directly; the platform creates, schedules, and cleans them up for you.

​
Where workloads come from

Workloads come from the three kinds of work you can run on the platform:

​
Workloads vs. workflows

A workflow is the definition of the work: Python code you author that declares a graph of steps. A workload is the work itself, physically running on the cluster. When you trigger a run, the platform decomposes it into one workload per step task. A flow with five steps runs as five workloads, and a foreach step over 100 items runs as 100. If a workflow is the recipe, workloads are what’s actually cooking on the burners. The distinction matters because not every workload comes from a workflow. Workstations and deployments are workloads too, so the platform uses workloads as the common unit for everything consuming compute, regardless of what launched it.

​
What the platform tracks

For every workload, the platform records:
  • Its status as it moves through its lifecycle: pending, running, and then finished, failed, or cancelled.
  • The queue and compute pool it runs on, and the CPU and memory it requests.
  • The container image it runs, the node it’s scheduled on, and when it starts and stops.
  • The perimeter it belongs to and the user or machine user that owns it.
For task workloads, the platform also records which run, step, and task the workload belongs to, so you can move between the workflow view and the infrastructure view.

​
Viewing workloads

Because every kind of work on the platform runs as a workload, workloads give administrators a single place to see everything consuming cluster resources, regardless of what launched it. The platform’s workloads view groups workloads by kind: tasks grouped by their queue, workstations, and deployments. For the compute workloads run on, see What is compute?