Ordering Guarantees

Global ordering doesn't scale; brokers order within a partition/queue/message group. Kafka orders per partition (same key → same partition), SQS FIFO orders per message group ID, RabbitMQ orders per queue with a single consumer. Parallel consumers, retries and redeliveries can still reorder processing. Design for it: key messages by entity, make handlers tolerate out-of-order events (version numbers, "last writer wins by version"), or carry full state instead of deltas.

Ordering by broker

async def apply_customer_update(event: dict):
    # Ignore stale events delivered out of order
    updated = await db.execute(
        '''UPDATE customers SET name = $2, email = $3, version = $4
           WHERE id = $1 AND version < $4''',
        event["customer_id"], event["name"], event["email"], event["version"])
    if updated == "UPDATE 0":
        metrics.increment("events.stale_ignored")

// Same orderId → same partition → ordered for that order
kafkaTemplate.send("order-events", order.getId(), OrderEvent.paid(order));

Common pitfalls

• Assuming global order across partitions.
• Increasing partition count (keys move partitions).
• Delta events applied out of order corrupt state.

Key points

• Order is per partition/group
• Key by entity
• Make consumers order-tolerant

Further reading: Kafka: Concepts