Transactional Patterns

Establish the boundary

• Inspect the project's Boot/Framework version, transaction manager, persistence technology, rollback configuration and caller before changing propagation.
• Place multi-step business transactions at the application/service boundary. Preserve Spring Data repository transactions; repository annotations are not inherently wrong.
• Default propagation is REQUIRED — joins existing transaction or creates one
• Use readOnly = true for appropriate read units of work. It is an optimization hint, not write prevention, authorization or automatic routing to a read replica.
• Keep existing architecture and transaction-manager choices. A local database transaction does not make writes to another database or an HTTP provider atomic.

The snippets below are illustrative; domain classes and imports come from the application. Adapt the service template and compare the good and bad examples when editing a service.

@Service
@RequiredArgsConstructor
@Transactional(readOnly = true) // default for all methods in this service
public class OrderService {

    @Transactional // overrides readOnly for writes
    public Order createOrder(CreateOrderRequest request) {
        inventoryService.reserve(request.items()); // participates in same TX
        return orderRepository.save(Order.from(request));
    }

    public Optional<Order> findById(UUID id) {
        return orderRepository.findById(id); // readOnly = true inherited
    }
}

Propagation

Participating REQUIRED calls use the outer transaction's isolation, timeout and read-only settings. Catching an inner failure does not clear its rollback-only marker; outer commit may still throw UnexpectedRollbackException.

REQUIRES_NEW needs another connection while the outer transaction retains its resources. Account for pool capacity and lock contention. Reserve independent commits for records that must survive a failed operation (for example, an attempted action), not a success audit row that claims a rolled-back order was created. Do not reference an uncommitted parent row from an independent audit transaction.

// REQUIRES_NEW — for audit logging that must survive rollback
@Transactional(propagation = Propagation.REQUIRES_NEW)
public void logAuditEvent(AuditEvent event) {
    auditRepository.save(event); // commits independently of parent TX
}

// Illustrative: record an attempt for an existing order; failures are unchecked.
@Transactional
public void processOrder(Order order) {
    auditService.logAuditEvent(new AuditEvent("ORDER_START", order.getId()));
    try {
        // ... process, may throw
    } catch (RuntimeException e) {
        auditService.logAuditEvent(new AuditEvent("ORDER_FAILED", order.getId()));
        throw e; // parent TX rolls back, audit TX already committed
    }
}

Self-Invocation Pitfall

// ❌ BROKEN — self-invocation bypasses Spring proxy, @Transactional ignored
@Service
public class OrderService {
    @Transactional
    public void processAll(List<UUID> ids) {
        ids.forEach(id -> this.processSingle(id)); // bypasses proxy!
    }

    @Transactional(propagation = Propagation.REQUIRES_NEW)
    public void processSingle(UUID id) { ... } // never creates new TX
}

// ✅ FIX — extract the independently transactional operation to another bean
@Service
@RequiredArgsConstructor
public class OrderService {
    private final OrderProcessor orderProcessor; // separate bean

    @Transactional
    public void processAll(List<UUID> ids) {
        ids.forEach(id -> orderProcessor.processSingle(id)); // goes through proxy
    }
}

Handling Exceptions

// Default: RuntimeException and Error roll back; checked exceptions do not.
// Check project-wide rollback configuration before adding per-method rules.

@Transactional(rollbackFor = InsufficientInventoryException.class) // checked exception
public Order createOrder(CreateOrderRequest request) throws InsufficientInventoryException {
    ...
}

Never use noRollbackFor to recover from OptimisticLockException: the persistence provider marks the transaction rollback-only. Roll it back and retry the whole unit of work in a fresh transaction, with a bounded policy and a fresh entity read.

Optimistic Locking

@Entity
public class Order {
    @Version
    private Long version; // Hibernate handles conflicts automatically
}

// Handles concurrent updates
@Transactional
public Order updateStatus(UUID id, OrderStatus newStatus) {
    Order order = orderRepository.findById(id).orElseThrow();
    order.updateStatus(newStatus); // conflict can surface at flush or commit
    return orderRepository.save(order);
}

Use the version-specific retry guidance for dependencies and annotation imports. A retrying facade must not already hold a transaction that each attempt's REQUIRED method would join.

Distributed Transactions (Saga Pattern)

A saga coordinates durable local steps; wrapping remote calls in @Transactional does not implement one.

1. Commit the pending order, saga state and outgoing command in one local transaction.
2. Dispatch the command outside that transaction, using a stable provider idempotency key.
3. Persist each confirmed outcome and the next command atomically in another local transaction.
4. On rejection, persist failure state and a compensation command. Retry compensation independently, and reconcile timeouts where the remote outcome is unknown.

Do not save failure state and then throw a rollback-triggering exception from the same transaction: that erases the state. Recovery must resume from durable state after a crash. Use the project's existing workflow mechanism; see messaging/outbox and idempotency for delivery and duplicate effects.

Side Effects After Commit

For best-effort local notifications, a commit-bound listener prevents delivery on rollback. For required delivery, persist an outbox entry with the business change or use the project's durable publication mechanism. AFTER_COMMIT alone cannot recover a process crash or retry a failed delivery.

// Publisher — inside the TX
@Transactional
public Order place(UUID id) {
    Order order = orderRepository.findById(id).orElseThrow();
    order.place();
    eventPublisher.publishEvent(new OrderPlaced(order.getId())); // published now; listener defers handling
    return orderRepository.save(order);
}

// Listener — runs ONLY if the TX commits successfully
@TransactionalEventListener(phase = TransactionPhase.AFTER_COMMIT)
public void onOrderPlaced(OrderPlaced event) {
    emailService.sendConfirmation(event.orderId()); // best effort; no durable retry
}

AFTER_COMMIT runs after the DB commits, but resources may still be bound to the completed transaction. Delegate database writes to a separate REQUIRES_NEW bean. Listener failure cannot undo the committed business write; make delivery failures observable and recoverable. Without an active transaction the listener is skipped by default. @Async does not add durable delivery. See domain events.

Verification

Test through Spring proxies and verify committed data from a new transaction. Cover checked versus unchecked failures, inner rollback-only propagation, optimistic conflicts at commit, audit rollback, and listener behavior after commit and rollback. Use the production database engine for isolation and contention tests. A test-wide rollback can hide after-commit behavior.

Gotchas

• Agent removes repository transaction annotations - keep the unit of work and existing Spring Data behavior.
• Agent treats AFTER_COMMIT as durable messaging - use an outbox or persistent publication registry for required delivery.
• Agent uses readOnly for replica routing or write protection - neither follows from the flag alone.
• Agent commits success audit entries with REQUIRES_NEW - commit them with the business mutation.
• Agent suppresses optimistic-lock rollback - retry in a fresh transaction.
• Agent calls @Transactional methods on this — self-invocation bypasses proxy
• Agent expects checked exceptions to rollback — must add rollbackFor
• Agent uses @Transactional on private methods — Spring proxy can't intercept

Official sources

• Spring transaction propagation
• TransactionalEventListener lifecycle
• Jakarta Persistence optimistic-lock rollback semantics