Шаблон транзакционного исходящего сообщения

In an event-driven FastAPI service, a single request often needs to do two things: persist state to your database and publish an event to Kafka or Pulsar.

The trap is that these are two separate systems with no shared transaction. If you commit the DB row then crash before publishing, downstream services never hear about it. If you publish first then the DB commit fails, you emit an event for state that does not exist.

• db.commit() succeeds, producer.send() fails → lost event
• producer.send() succeeds, db.commit() fails → phantom event

This is the dual-write problem, and no amount of try/except fully solves it.

A common first instinct is: "I'll just wrap the publish in a retry loop." But retries do not close the gap.

• The process can be killed (OOM, deploy, k8s eviction) between commit and publish — no code runs to retry.
• Retrying after a broker timeout can produce duplicates if the first send actually landed.
• You cannot roll back a Kafka write once partially acknowledged.

The core issue is that the commit and the intent to publish are not atomic. We need to make the publish-intent part of the same database transaction as the state change. That is exactly what the Transactional Outbox pattern does.

The Transactional Outbox pattern adds an outbox table in the same database as your business data. Instead of publishing to the broker inline, you INSERT a row describing the event into the outbox table — inside the same transaction that changes your state.

Because the order row and the outbox row are written in one transaction, they either both commit or both roll back. There is no window where one exists without the other.

A separate relay process later reads unpublished outbox rows and pushes them to Kafka/Pulsar, marking them sent. The broker write is decoupled from the request path.

• Atomicity comes from the DB transaction, not from a distributed transaction.
• Delivery becomes at-least-once — consumers must be idempotent.

The outbox table needs enough metadata for the relay to publish correctly and for you to debug. A typical schema:

• id — UUID primary key, also reused as the event id for idempotency
• aggregate_type / aggregate_id — what the event is about (e.g. order / order id), often used as the partition key
• event_type — e.g. OrderCreated
• payload — JSON body of the event
• created_at — ordering
• published_at — NULL until the relay sends it

Keeping aggregate_id lets the relay set the Kafka partition key so all events for one order stay ordered.

Here is the heart of the pattern. The order and the outbox row are added to the same session and committed together. Notice there is no broker call in the request handler at all.

If commit() fails, neither row exists. If it succeeds, both exist. The event will be delivered by the relay, guaranteed.

The simplest relay is a polling publisher: a background loop that repeatedly selects unpublished outbox rows, sends them to the broker, and marks them published.

• Select rows where published_at IS NULL, ordered by created_at.
• Publish each to Kafka/Pulsar using aggregate_id as the key.
• Set published_at = now() and commit.

If the process crashes after publishing but before marking, the row gets re-sent on the next poll — hence at-least-once and the need for idempotent consumers. The event id is the deduplication key.

If you run more than one relay replica for throughput, two workers might grab the same outbox rows and double-publish. Postgres gives you a clean fix: SELECT ... FOR UPDATE SKIP LOCKED.

• FOR UPDATE locks the selected rows for the transaction.
• SKIP LOCKED makes other workers skip already-locked rows instead of blocking.

Each worker claims a disjoint batch, publishes, marks them, and commits — releasing the locks. This turns the outbox into a safe concurrent work queue without an external queue system.

Because the outbox guarantees at-least-once delivery, every consumer must tolerate seeing the same event more than once. The standard technique is a processed-events table keyed by the event id.

Before applying an event, attempt to insert its id. If it already exists, you have seen it — skip. Do the dedup insert and the business effect in the same transaction so a crash cannot leave them out of sync.

Polling is simple but adds latency and DB load. A higher-performance alternative reads Postgres's write-ahead log (WAL) directly using Change Data Capture — typically Debezium.

• Debezium tails the WAL and emits a Kafka message for every INSERT into the outbox table.
• No published_at column or polling loop needed — the log is the queue.
• Its Outbox Event Router SMT maps outbox columns to topic, key, and payload.

Trade-off: CDC is operationally heavier (connector, replication slot) but gives near-real-time delivery and offloads the relay entirely from your FastAPI app.

Events for the same aggregate usually must arrive in order — OrderCreated before OrderShipped. Both the outbox and the broker preserve this if you are careful:

• Order the relay select by created_at (or a monotonic sequence).
• Use aggregate_id as the Kafka partition key so all of one order's events land on the same partition, which Kafka delivers in order.

Events for different aggregates can interleave freely — you only need per-aggregate ordering, which is what partition-by-key buys you. A pure global ordering across all orders is rarely required and kills parallelism.

A few practices keep the outbox healthy at scale:

• Prune published rows — periodically delete rows where published_at is older than a retention window, or move them to an archive, so the table stays small and the partial index stays fast.
• Partial index on unpublished rows: CREATE INDEX ... ON outbox (created_at) WHERE published_at IS NULL; keeps the relay's hot query cheap.
• Monitor lag — alert on the count or age of unpublished rows; a growing backlog means the relay is down or the broker is unreachable.
• Idempotent producer — enable Kafka enable.idempotence=true to suppress relay-side resend duplicates at the broker level.

Test your understanding of why the Transactional Outbox pattern actually solves the dual-write problem.

You learned how to publish events reliably from a FastAPI service:

• The dual-write problem: committing to the DB and publishing to Kafka/Pulsar are separate, non-atomic actions that can desync on a crash.
• The Transactional Outbox writes an event row into an outbox table in the same transaction as the state change, making them atomic.
• A relay (polling publisher or Debezium CDC) later sends outbox rows to the broker and marks them published.
• Use FOR UPDATE SKIP LOCKED for safe concurrent relays, aggregate_id as the partition key for ordering, and a processed-events table for idempotent, at-least-once consumers.
• Operate it with pruning, a partial index, and backlog monitoring.