14.3 Workflow, rules, events

Workflow engine

• MVP: Camunda 7 embeddable in JVM monolith. BPMN models for sanction-to-disbursement saga, application lifecycle, collection cycle, settlement.
• Scale: Temporal for code-first durable workflows.

See 5.5 Workflow and rules engines for selection rationale.

Rule engine

• MVP: decision tables in DB + JSON Logic for rule expressions. Sandbox testing via clone DB. Versioned.
• Scale: Drools when complexity demands.

Event bus

• MVP: RabbitMQ.
• Scale: Kafka with schema registry.

Queue / async workers

• Spring Cloud Stream or Spring AMQP for RabbitMQ.
• Spring Kafka for Kafka.
• Idempotency patterns: outbox + dedupe.
• DLQ for permanent failures with inspection UI.

Saga orchestration

• Camunda / Temporal for multi-step sagas.
• Local Java state machines (Spring Statemachine) for in-service flows.

ML platform

• Year 1: defer.
• Year 2+:
  • Feast (or homegrown on warehouse) as feature store.
  • MLflow for experiment + model registry.
  • Training: Python notebooks → MLflow.
  • Serving: Java-side via ONNX runtime or via a Python serving service called from Java.
  • Monitoring: Evidently / Arize / homegrown dashboards.

Notification service

• Cross-cutting, low-level shared service.
• Backends: SMS, WhatsApp, email, push, IVR.
• Templates centrally managed; multi-language.

Scheduler

• Spring @Scheduled for in-process.
• Quartz or K8s CronJob for distributed scheduled.
• Daily classification, daily NACH, monthly bureau, etc.