Publication transactionnelle d’événements et boîte de sortie

In an event-driven Spring Modulith application, modules talk to each other by publishing application events. A module raises an event inside a transaction, and listeners in other modules react to it.

The danger: by default, when you publish an event and a listener processes it asynchronously (or in a separate transaction), the listener's work can fail after the publisher already committed. The result is a lost event — the publisher's state changed, but the downstream side effect never happened.

• Publisher commits an Order as PAID.
• The async listener that sends a confirmation email crashes.
• Nobody retries — the customer never gets the email.

This lesson shows how Spring Modulith's Event Publication Registry implements the transactional outbox pattern to guarantee delivery.

The transactional outbox pattern solves the dual-write problem: you must atomically (1) change your business state and (2) record that an event needs to be delivered.

Instead of trying to write to the database and a message broker in one transaction (impossible without distributed transactions), you write both into the same database in the same local transaction:

• The business change (e.g. the order row).
• A row in an outbox table describing the event.

A separate process then reads unprocessed outbox rows and delivers them, marking each as completed. Because the outbox write shares the business transaction, an event is recorded if and only if the business change committed.

Spring Modulith ships a ready-made outbox: the Event Publication Registry. When an event has a transactional listener (annotated with @ApplicationModuleListener), Modulith automatically:

• Writes an event publication row before the listener runs.
• Marks it completed when the listener finishes successfully.
• Leaves it incomplete if the listener throws — so it can be retried.

You enable it by adding the starter and a persistence module. The registry persists publications in a table such as event_publication.

The key annotation is @ApplicationModuleListener. It is a composed annotation that combines three behaviors:

• @Async — the listener runs on a separate thread, decoupling modules.
• @Transactional(propagation = REQUIRES_NEW) — the listener runs in its own transaction.
• @TransactionalEventListener(phase = AFTER_COMMIT) — it fires only after the publisher's transaction commits.

Combined with the registry on the classpath, every event handled by such a listener gets an outbox entry. If the listener fails, the publication stays incomplete and survives restarts.

The publishing side stays simple. Inside a normal @Transactional service method you call ApplicationEventPublisher.publishEvent(...). Spring Modulith intercepts the publication and, because there is a transactional module listener, writes the outbox row in the same transaction as your business change.

Use immutable Java records for events — they are concise, serializable, and clearly value-typed.

Here is the crucial sequence that guarantees delivery:

• Your complete() method opens a transaction and changes the order.
• publishEvent causes Modulith to INSERT an incomplete event_publication row in that same transaction.
• The transaction commits — order change and outbox row land together, atomically.
• After commit, the @ApplicationModuleListener runs in a new transaction.
• On success, the publication is marked completed.

If the JVM crashes between commit and listener success, the row is still incomplete in the database, ready to be republished. No event is ever silently dropped.

The JPA persistence module stores publications in a table. Each row identifies a serialized event and its target listener, plus timestamps. Knowing the schema helps you reason about retries and monitoring.

• id — UUID primary key.
• listener_id — fully-qualified listener method that must handle it.
• event_type + serialized_event — the event payload (JSON by default via Jackson).
• publication_date — when it was created.
• completion_date — NULL while incomplete; set when the listener succeeds.

A row with completion_date IS NULL is an outstanding event awaiting (re)delivery.

Incomplete publications are useless unless something retries them. Spring Modulith can republish outstanding events on application startup, which recovers from crashes that happened mid-delivery.

Enable it in application.properties:

• spring.modulith.republish-outstanding-events-on-restart=true

On boot, Modulith reads all incomplete event_publication rows and re-invokes their listeners. Because listeners should be idempotent, replaying a partially-processed event is safe.

Because an event may be delivered more than once (retry after a crash, or scheduled resubmission), the at-least-once guarantee forces your listeners to be idempotent. Processing the same event twice must produce the same end state as processing it once.

Common techniques:

• Use a natural business key (the order id) and check whether the side effect already happened.
• Track processed event ids in a dedup table with a unique constraint.
• Make the operation naturally idempotent (UPSERT, or set-to-state instead of increment).

Startup republishing only helps when you restart. For long-running services you also want periodic recovery of stuck publications (e.g. a listener that threw a transient error). Modulith offers a completion / resubmission scheduler.

• spring.modulith.events.completion-mode — controls whether completed rows are deleted, archived, or updated.
• Enable a recurring resubmission so incomplete events older than a threshold are retried automatically.

Combined with idempotency, this turns the outbox into a self-healing delivery channel without a separate message broker.

The same registry bridges to external messaging. Spring Modulith provides externalization modules (Kafka, RabbitMQ, AMQP, SQS, etc.). You annotate an event with @Externalized, and a transactional listener publishes it to the broker — backed by the very same outbox.

This means the at-least-once guarantee extends across process boundaries: the broker send is itself an event publication that is only marked complete once the message is accepted by the broker.

You learned how to guarantee event delivery in Spring Modulith using the transactional outbox pattern:

• The Event Publication Registry persists an event_publication row in the same transaction as your business change — solving the dual-write problem.
• @ApplicationModuleListener = async + REQUIRES_NEW transaction + AFTER_COMMIT, so listeners run after the publisher commits and get their own outbox entry.
• A publication is incomplete until its listener succeeds (completion_date IS NULL); failures leave it for retry.
• Republish on restart and scheduled resubmission recover stuck events — making delivery at-least-once.
• Because delivery is at-least-once, listeners must be idempotent.
• @Externalized bridges the same outbox to Kafka/RabbitMQ/SQS for cross-process guarantees.