Skip to content

ADR-0004: Apache Kafka as the Event Bus

Status: Accepted Date: 2026-01-20

Context

Services need to publish and consume domain events asynchronously. We evaluated Apache Kafka (Amazon MSK), RabbitMQ, Amazon SNS/SQS, and AWS EventBridge.

Decision

Apache Kafka via Amazon MSK. Topics are named chakra.{context}.{event-type} (e.g., chakra.orders.order-placed). Retention is 7 days for operational topics; 30 days for audit topics. Schema registry is not adopted at launch (see tradeoffs).

Rationale

Replayability: The saga compensation pattern requires consumers to replay events during debugging and recovery. Kafka's log retention enables this; RabbitMQ's queue model does not.

Consumer group model: Multiple services can independently consume the same event stream (Orders publishes OrderPlaced; Inventory consumes it to reserve stock; Analytics consumes it for reporting). Each consumer group maintains independent offsets.

Consumer lag observability: Kafka consumer lag is a first-class metric. observability/alerts/kafka-lag.yaml alerts when lag exceeds thresholds. This is not possible with SNS/SQS without additional instrumentation.

Consequences

Positive: Log retention for replay; independent consumer groups; lag observability; MSK handles broker management. Negative: Operational complexity vs. managed queues; schema evolution without a registry requires consumer-side forward compatibility; MSK cost vs. SQS.

  • ADR-0006: Events from EventStoreDB are forwarded to Kafka via the outbox pattern