Camunda 8 Vs Argo

Camunda 8 (BPMN + Zeebe engine) y Argo Workflows (Kubernetes controller) NO son competidores reales. Camunda apunta a business workflows con human tasks y BPMN compliance. Argo apunta a containerized pipelines (CI/CD, ML, ETL) en Kubernetes. Son complementarios — pueden coexistir. Para MVP "drop-in Camunda": NO replicar Argo, pero asegurar deployment K8s-native.

Diferente target use case¶

Dimensión	Camunda 8	Argo Workflows
Primary workflow type	Business processes	Containerized pipelines
Sweet spot	Long-running + human tasks	Short batch + parallel containers
Audience	Business + devs	DevOps + data engineers
Stakeholder visibility	Business analysts ven BPMN	DevOps ven YAML
Example workflow	Loan approval (días)	ML training pipeline (horas)

Estos engines optimizan diferentes ejes — no son alternativas directas.

Architecture comparison¶

Aspecto	Camunda 8	Argo Workflows
Engine	Custom (Zeebe, Java)	Kubernetes controller (Go)
State store	RocksDB embedded	etcd (via K8s API)
Scheduling	Job worker pattern	K8s scheduler
Execution unit	External worker	Container pod
Scaling	Brokers + partitions	K8s HPA + nodes
Multi-tenancy	Native (8.5+)	K8s namespaces
Monitoring	Operate + Optimize	Argo UI + Prometheus

Argo es un addon de Kubernetes. Camunda es un engine independiente que puede correr en K8s.

Workflow definition¶

Camunda BPMN¶

<bpmn:process id="loan-approval">
  <bpmn:userTask id="review-application">
    <zeebe:assignmentDefinition assignee="manager" />
  </bpmn:userTask>
  <bpmn:serviceTask id="check-credit">
    <zeebe:taskDefinition type="credit-check" />
  </bpmn:serviceTask>
  <bpmn:exclusiveGateway id="decision" />
</bpmn:process>

Visual, hierarchical, BPMN-compliant.

Argo YAML¶

apiVersion: argoproj.io/v1alpha1
kind: Workflow
metadata:
  generateName: ml-pipeline-
spec:
  entrypoint: main
  templates:
  - name: main
    dag:
      tasks:
      - name: preprocess
        template: preprocess
      - name: train
        template: train
        dependencies: [preprocess]
      - name: evaluate
        template: evaluate
        dependencies: [train]
  - name: preprocess
    container:
      image: my-preprocessor:v1.2
      resources:
        requests: { memory: "1Gi", cpu: "1" }

YAML-as-code, container-based, K8s-native.

Capabilities¶

Camunda exclusivo¶

Human tasks (Tasklist UI)
BPMN visual modeling
DMN decision tables
Process mining (Optimize)
Message correlation
Long-running (days/weeks)
Multi-stakeholder workflows

Argo exclusivo¶

Container-native (every step is an image)
K8s-native (Pods, ConfigMaps, Secrets, RBAC)
GPU scheduling
Resource limits per step
Image-based isolation
Parameter passing between containers
Loop with withItems / withParam

Operational comparison¶

Aspecto	Camunda 8	Argo Workflows
Install	Helm chart o standalone	Helm chart en K8s
State location	RocksDB en brokers	etcd compartido con K8s
Logs	Worker logs + Operate	`kubectl logs` pods
Scaling	Manual broker count	K8s HPA automático
Costs	Brokers always running	Pods on-demand

Argo tiene menor cost en idle (no brokers corriendo). Camunda tiene lower latency per step (no pod scheduling overhead).

Performance characteristics¶

Métrica	Camunda 8	Argo Workflows
Workflow startup	ms (acquire job)	seconds (pod schedule + start)
Per-step overhead	ms (worker invocation)	seconds (pod startup)
Best for	Many short steps	Fewer long steps
Throughput	375 TPS (Intuit benchmark)	Limited by K8s scheduler

Cuando importa: - Camunda gana si tienes 1000s de workflows/sec con steps quick - Argo gana si tienes 10s de workflows con steps that take minutes

Complementariedad¶

Estos engines pueden coexistir en una organización:

flowchart TD
    BW[Business workflows<br/>human + system]
    CM[Camunda 8 / MVP]
    TA[Trigger Argo Workflow desde Camunda]
    AW[Argo Workflows<br/>containers, parallel]
    RR[Return result a Camunda]
    BW --> CM
    CM -->|cuando necesitas containerized pipeline| TA
    TA --> AW --> RR

Patrón típico: - Camunda: loan approval workflow - Step: "evaluate creditworthiness via ML model" - Implementation: launch Argo workflow with ML pipeline - Camunda waits for Argo workflow to complete - Continue process

Para el MVP¶

Decisión: NO replicar Argo¶

Razones: 1. Target del MVP es BPMN business workflows, no containerized pipelines 2. Replicar Argo sería un engine completamente diferente 3. Argo ya existe y es excelente para su use case 4. K8s-native está fuera del scope core del MVP

Patterns aplicables¶

Limitados — pero hay algunos:

K8s-native deployment: el MVP DEBE deployarse fácil en K8s. Tomar prestado de Argo:
Helm chart oficial
CRDs opcionales para definir processes via K8s API
Integration con K8s logs, metrics, RBAC

Container step option: para casos avanzados, permitir que un service task se ejecute como K8s pod en vez de worker:

<bpmn:serviceTask id="train-model">
  <zeebe:taskDefinition type="builtin:k8s-pod" />
  <zeebe:taskHeaders>
    <zeebe:header key="image" value="ml-trainer:v2" />
    <zeebe:header key="cpu" value="4" />
    <zeebe:header key="memory" value="16Gi" />
  </zeebe:taskHeaders>
</bpmn:serviceTask>

El engine spawn un Pod, espera completion, recolecta logs/output.

Trigger external Argo workflow: feature de integration con Argo si el usuario ya lo tiene.

Lo que NO copiar¶

Engine architecture (Kubernetes controller pattern)
YAML-only workflows (mantener BPMN como primary)
DAG-only model (BPMN tiene más expressiveness)
State en etcd (Postgres es mejor para business state)

Conclusión¶

Argo Workflows y Camunda 8 son complementarios, no competidores. El MVP puede coexistir con Argo: - MVP: business workflows (BPMN, human tasks, DMN) - Argo: technical pipelines (CI/CD, ML, ETL)

Patterns prestables son limitados — principalmente K8s-native deployment. NO replicar Argo architecture.

Esta comparación cierra el "competitive landscape" del MVP: - vs Camunda original: simplified clone - vs Temporal: different audience (BPMN required) - vs Conductor: different language (BPMN vs JSON) - vs Argo: different use case (business vs technical)

Cada uno tiene su nicho. El MVP apunta firmemente al nicho de Camunda.