Skip to content

Unsupported workflows

Dagweave Local checks every workflow for settings that need a real cluster, and stops if it can’t honour one. To run the check on its own:

Terminal window
dagweave lint my-workflow.yaml

A clean workflow prints:

my-workflow.yaml: nothing here needs a cluster feature this machine lacks.

Otherwise it prints a refuse or warn line per finding, with the field, its template and what to do. It exits non-zero if anything is refused.

The check covers the pod spec each step would get, including spec.templateDefaults and podSpecPatch, plus the workflow, each template and each lifecycle hook. A refused field stops the run. A warned one runs, with a note on what differs. A field Dagweave Local has no rule for is refused.

  • nodeSelector, affinity, nodeName, runtimeClassName, schedulingGates and resourceClaims. The step would never start.
  • serviceAccountName. No such account exists locally.
  • resource and plugin templates.
  • templateRef and workflowTemplateRef. There are no stored WorkflowTemplates locally.
  • A workflow or template that suspends, since nothing can resume it.
  • A {{workflow.*}} variable Dagweave Local can’t supply, or a variable that points at an input, output or step that doesn’t exist.
  • A when, depends or {{=...}} that Argo can’t parse, or that reads an input, output or step that doesn’t exist or hasn’t run yet. Test an expression with dagweave expr.
  • An expression that reads a workflow parameter spec.arguments doesn’t declare. It only has a value if you pass it with -p.
  • imagePullSecrets, hostPath volumes and persistent volume claims.
  • CPU and memory limits. On the host, nothing enforces them.
  • hostNetwork, hostPID and hostIPC. They point at the local cluster’s container, not your laptop.
  • synchronization and memoize. No locks are held and nothing is cached.
  • metrics. Nothing collects them.
  • On the host only: activeDeadlineSeconds on the workflow or a template, and parallelism limits.

A warning about a Secret goes away once you store it with dagweave secret.

Scheduling settings that mean nothing on one node, such as tolerations, priorityClassName, topologySpreadConstraints, podDisruptionBudget and resource requests.

The check can’t see what a kubectl or helm command will do. Five catalog steps are refused, with their reason: helm-upgrade, kubectl-apply, kubectl-delete, scale-deployment and wait-for-rollout.

Your own steps run as the default service account in the local cluster’s argo namespace. It can create, get and patch pods, and create and get workflows, there. Anything else, such as a Deployment or a Secret, fails on permissions.

Watch out for kubectl apply. A Pod or Workflow applied to argo lands in the local cluster and the step succeeds, but nothing reaches the cluster you meant.