รูปแบบเอาต์บ็อกซ์สำหรับการเผยแพร่เหตุการณ์ที่เชื่อถือได้

A service often needs to update its database and publish a Kafka event. Doing both as separate steps risks one succeeding and the other failing.

This dual-write problem leaves the database and Kafka inconsistent.

If you commit the DB then publish, a crash in between loses the event.

If you publish then commit, a failed commit emits a phantom event. There is no atomic transaction spanning a relational DB and Kafka by default.

The outbox pattern records the event in the same database transaction as the business change, in a dedicated outbox table.

One atomic commit now contains both the state change and the intent to publish.

A typical outbox schema captures everything needed to build the Kafka record.

Within one transaction, write the business row and the outbox row together.

A separate relay reads new outbox rows and publishes them to Kafka.

Two common approaches: poll the table, or use Change Data Capture to stream the inserts.

Debezium has a dedicated Outbox Event Router SMT. It tails the DB transaction log, picks up outbox inserts, and routes each to the right Kafka topic — no polling, low latency.

The relay guarantees at-least-once delivery: an event is never lost, but it may be published more than once after a crash and retry.

Therefore consumers must be idempotent.

Use the event id (or aggregate id + version) to deduplicate downstream.

To preserve per-aggregate order, use the aggregate_id as the Kafka message key.

All events for one account then land on the same partition and are consumed in order.

Pros: no dual-write loss, works with any DB, decouples publishing.

Cons: extra table and relay, at-least-once duplicates, slight latency. Still the standard solution for reliable event publishing.

You learned the outbox pattern.

• Solves the dual-write problem by committing state + event atomically.
• A relay (polling or CDC/Debezium) publishes outbox rows to Kafka.
• Delivery is at-least-once, so consumers must be idempotent.
• Key by aggregate id to preserve ordering.