This repository showcases an example of a simple outbox pattern implementation using Spring Boot and Kafka. For more information on the outbox pattern, see: microservices.io - Transactional Outbox.

• Problem Statement
• Setup
  • Prerequisites
  • Project Structure
  • Running the Application
• Anomaly Simulation
  • Order Service Code Example
  • Failed Commit Anomaly
  • Failed Broker Delivery Anomaly
• Outbox Pattern
  • Pros
  • Cons
• Other Approaches
  • Integrated Transactional Messaging
  • Database Triggers
• Out of Scope

In event-driven architecture, applications often need to persist data and send messages to a message broker like Kafka or RabbitMQ. However, naive implementations may encounter the following issues:

• Sending event while transaction is open: Data might be lost if the transaction fails.
• Sending event outside of transaction: Event might be lost if sending fails.

This repository demonstrates these consistency anomalies and provides a solution using the outbox pattern.

• Java 21
• Docker
• Kafka
• Postgres

This application consists of a multi-module Gradle project with two services:

• Responsibilities: Taking orders and sending events to Kafka.
• Port: 8078
• Endpoint: POST /orders
  • Optional argument: simulationStrategy
    • OUTBOX (default)
    • FAILED_COMMIT_ANOMALY
    • FAILED_BROKER_DELIVERY
• Swagger UI: https://localhost:8078/swagger-ui/index.html
• Database: Postgres (stores order data)
• Kafka: Produces orders to Kafka

• Responsibilities: Consumes events and processes orders.
• Kafka: Consumes orders and logs the result.

1. Start Docker Services: Ensure Docker is running and execute docker-compose up to start Kafka and Postgres services.
2. Run Order Service:

   ./gradlew :order-service:bootRun

3. Run Logistics Service:

   ./gradlew :logistics-service:bootRun

To create an order, use the following example:

curl -X 'POST' \
  'https://localhost:8078/orders?simulationStrategy=OUTBOX' \
  -H 'accept: */*' \
  -H 'Content-Type: application/json' \
  -d '{
  "userId": "3fa85f64-5717-4562-b3fc-2c963f66afa6",
  "items": [
    {
      "productId": "3fa85f64-5717-4562-b3fc-2c963f66afa6",
      "quantity": 1
    }
  ]
}'

@RequiredArgsConstructor
@Slf4j
public abstract class AbstractOrderService implements SimulatedOrderService {
    private final OrderMapper orderMapper;
    private final CrudRepository<Order, UUID> orderRepository;
    private final KafkaTemplate<String, String> kafkaTemplate;
    private final String topic;

    @Override
    @Transactional
    public OrderCreatedDto create(final OrderDto order) {
        var orderEntity = orderRepository.save(orderMapper.toEntity(order));
        log.info("Saved order from user {} with id {}", order.getUserId(), orderEntity.getId());
        var dto = orderMapper.toDto(orderEntity);
        log.info("Preparing to send order with id to kafka {}", orderEntity.getId());
        kafkaTemplate.send(topic, JsonUtil.toJson(dto))
                .whenComplete(
                        (result, ex) -> {
                            if (result != null && ex == null) {
                                log.info("Sent order with id {} to kafka", dto.getId());
                            } else {
                                log.error("Failed to send order with id {} to kafka", dto.getId(), ex);
                            }
                        });
        return dto;
    }
}

To simulate the failed commit anomaly, the code is modified to throw an exception after the message is sent to Kafka - FailedDatabaseOrderService. This results in the message being sent to Kafka, but the transaction is rolled back, so the data is not persisted in the database. The logistics-service will consume the message, but the entry will not be present in the order-service database.

To simulate the failed broker delivery anomaly a CorruptedKafkaTemplate is created, that will throw an exception in the KafkaTemplate call. The data is persisted in the database, but no event is sent. The order-service database will have the entry, but the message will not be consumed by the logistics-service.

This example also demonstrates an often misunderstood concept of the KafkaTemplate - since send() is asynchronous and returns a ListenableFuture, the exception is not thrown immediately, but rather when the future is completed. This means that the exception is not caught by the @Transactional method and the transaction is committed.

Naive attempt to resolve by awaiting future completion:

Before implementing the outbox pattern, a naive attempt can be made to resolve the consistency issues by awaiting the completion of the future returned by the KafkaTemplate.send method. This approach involves blocking the transaction until the Kafka broker confirms the message delivery. While this ensures that the transaction would only commit if the message was successfully sent, it introduces significant latency and blocking behavior into the system. Additionally, this approach does not fully address the issue of failed commits.

The outbox pattern resolves the anomalies mentioned above by persisting the event in the database within the same transaction as the data. A separate scheduler then processes the events, ensuring at-least-once delivery semantics.

Why does the outbox pattern only guarantee the at-least-once delivery?

The outbox pattern does not provide exactly-once delivery semantics because the message delivery is not transactional. If the message is sent but the scheduler fails before marking the message as processed, the message will be sent again.

• Non-blocking message sending
• Retry mechanism in case of failure

• Additional implementation overhead
• Increased message delivery latency

Some messaging systems, like Kafka, support transactions natively.

Pros:

• Simplifies architecture by using Kafka's transactional capabilities

Cons:

• Tightly couples application logic with Kafka's transactional API
• Longer response times due to Kafka's producer characteristics

Using database triggers to publish events after the transaction commits.

Pros:

• Ensures consistency using database features

Cons:

• Depends on database-specific features
• Can be complex to manage

• Provision of partition keys to ensure message ordering within partitions
• Retry of "stuck" messages, which may cause deadlocks in the outbox table