Delivery Guarantees

---

Delivery guarantee defines how reliably a message is delivered from producer → Kafka → consumer

Types

• At Most Once
• At Least Once
• Exactly Once (practical understanding)

---

End-to-End Flow

Producer → Kafka → Consumer → Offset Commit

👉 Delivery guarantee depends on:

• Producer config
• Kafka behavior
• Consumer offset handling

---

At Most Once

Message is delivered at most once (may be lost)

Producer sends → Offset committed → THEN processing

If crash happens:

• ❌ Message lost

Producer Config

acks: 0

Consumer Config

enable-auto-commit: true

Problem

• No retries
• No guarantee
• Data loss possible

Money debited → event lost → audit system not updated

Not acceptable

---

At Least Once

Message is delivered at least once

👉 No data loss
👉 But duplicates possible

Flow

Producer → Kafka → Consumer → Process → Commit offset

If crash before commit:

Message reprocessed → duplicate

Producer Config

acks: all retries: 3

Consumer Config

enable-auto-commit: false ack-mode: manual

Consumer Code

@KafkaListener(topics = "account.transaction.completed.v1")
public void consume(Event event, Acknowledgment ack) {

    process(event);

    ack.acknowledge();
}

Debit processed twice → duplicate transaction

👉 Must handle using idempotency

---

Exactly Once (Realistic Understanding)

Kafka cannot guarantee full exactly-once across systems easily

But we can achieve it practically

Message is processed only once logically

How to Achieve

1. Idempotent Producer

enable-idempotence: true

1. Idempotent Consumer

Check before Processing

if (alreadyProcessed(event.getTransactionId())) {
    return;
}

1. Database Constraint

UNIQUE(transaction_id)

---

Guarantee	Data Loss	Duplicates	Use Case
At Most Once	❌ Possible	❌ No	logs, metrics
At Least Once	✅ No	⚠️ Possible	most systems
Exactly Once	✅ No	✅ No	banking (with idempotency)

---

Where Problems Happen

Case 1 — Producer Retry

Send → timeout → retry → duplicate

Case 2 — Consumer Crash

Processed → crash before commit → reprocess

Case 3 — Network Issues

• duplicates
• delayed delivery

---

Producer (Transaction Service)
      ↓
Kafka
      ↓
Consumer (Ledger Service)
      ↓
Check transactionId (idempotency)
      ↓
Insert into DB (unique constraint)
      ↓
Commit offset