Skip to content

01 - Argo Workflows

Domain 1: Argo Workflows (36%) - the largest domain.

A container-native workflow engine. Each step runs as a pod; the controller orchestrates them.

The object model

  • Workflow - one execution. Contains templates and an entrypoint.
  • WorkflowTemplate - a reusable, namespaced definition invoked by other workflows.
  • ClusterWorkflowTemplate - the cluster-scoped equivalent.
  • CronWorkflow - a Workflow on a schedule.

Templates

Work templates do something:

Type Runs
container A container with image, command, args
script A container plus an inline script; its stdout becomes an output parameter
resource A Kubernetes resource operation (create, apply, delete, patch), with success conditions
suspend Pauses the workflow until resumed manually or after a duration
http An HTTP request without a pod

Orchestration templates call other templates:

Type Semantics
steps A list of lists. Items within an inner list run in parallel; outer lists run in sequence, forming a barrier between them
dag Tasks with explicit dependencies. A task starts as soon as its dependencies finish 
# steps: b and c run in parallel, but both wait for a, and d waits for both
- - name: a
- - name: b
  - name: c
- - name: d

# dag: d waits only for b and c; a fast chain is not blocked by a slow sibling
tasks:
  - name: a
  - name: b
    dependencies: [a]
  - name: c
    dependencies: [a]
  - name: d
    dependencies: [b, c]

Prefer dag when task durations vary or the graph is non-trivial.

Parameters and artifacts

Parameters carry small string values.

  • Inputs declared under inputs.parameters, referenced as {{inputs.parameters.name}}
  • Outputs declared with valueFrom.path (a file), valueFrom.parameter, or the script template's stdout
  • Consumed downstream as {{steps.<step>.outputs.parameters.<name>}} or {{tasks.<task>.outputs.parameters.<name>}}

Artifacts carry files, through an artifact repository: S3, GCS, Azure Blob, MinIO, HTTP, Git, or raw.

  • Output artifacts are uploaded when the step ends; input artifacts are downloaded before the next step starts
  • Artifact garbage collection (artifactGC) deletes them on workflow completion or deletion, which matters because artifact storage otherwise grows without bound

A volumeClaimTemplate is the alternative for sharing data: a PVC created for the workflow and mounted by multiple steps, avoiding the object storage round trip at the cost of scheduling constraints.

Control flow

  • Conditionals: when: "{{steps.check.outputs.result}} == success"
  • Loops: withItems (a static list), withParam (a JSON array from a previous step's output), withSequence (a numeric range)
  • Recursion: a template invoking itself, bounded by a when condition
  • Parallelism: limits at workflow, template, and controller level
  • Suspend and resume for manual approval gates

Reliability and lifecycle

Feature Purpose
retryStrategy Retry limits, backoff, and retryPolicy (Always, OnFailure, OnError, OnTransientError)
activeDeadlineSeconds A hard time limit at workflow or template scope
timeout on a template Per-step limit
onExit handler Runs after the workflow finishes, whatever the outcome; used for notification and cleanup
Lifecycle hooks Run templates on specified expressions, such as a step entering Failed
ttlStrategy Deletes the Workflow object after completion, per outcome
podGC Deletes pods on completion or success, freeing cluster resources sooner
Workflow archive Persists completed workflows to a database so the UI retains history after the object is deleted

Synchronization limits concurrency with a mutex (one at a time) or semaphore (N at a time), backed by a ConfigMap. This is how you stop twenty workflows hammering the same database.

Security

  • Workflows run pods under a service account; the workflow's own RBAC determines what resource templates may do
  • The executor (emissary is the current default) is how the controller supervises the main container and collects outputs
  • podSpecPatch allows targeted pod spec changes such as resource limits
  • Restrict who may create Workflows in a namespace, since a Workflow is effectively arbitrary pod creation

Key terms

  • Workflow - the Argo custom resource representing one execution of a set of templates
  • WorkflowTemplate - a reusable namespaced workflow definition invoked by other workflows
  • ClusterWorkflowTemplate - the cluster-scoped reusable workflow definition
  • CronWorkflow - a Workflow executed on a schedule with a concurrency policy
  • Entrypoint - the template a Workflow starts from
  • Container template - a template running a single container as a pod
  • Script template - a template running an inline script whose stdout becomes an output parameter
  • Resource template - a template performing a Kubernetes resource operation with success conditions
  • Suspend template - a template pausing the workflow until resumed or until a duration elapses
  • Steps template - an orchestration template of nested lists where inner lists run in parallel and outer lists in sequence
  • DAG template - an orchestration template where tasks declare explicit dependencies
  • Output parameter - a small value produced by a step and referenced by later steps
  • Artifact - a file passed between steps through an artifact repository
  • Artifact repository - the object storage backing artifact transfer, such as S3, GCS, or MinIO
  • artifactGC - the configuration deleting artifacts when a workflow completes or is deleted
  • withParam - the loop construct expanding a JSON array from a previous step into parallel iterations
  • retryStrategy - per-template retry configuration including limit, backoff, and retry policy
  • onExit handler - a template that runs after a workflow completes regardless of outcome
  • ttlStrategy - configuration deleting completed Workflow objects after a specified time
  • podGC - configuration deleting workflow pods on completion to free cluster resources
  • Synchronization - mutex or semaphore configuration limiting concurrent workflow or template execution
  • Workflow archive - database persistence of completed workflows so history survives object deletion