🧠 Delivery Semantics

When messages/events are sent between systems (like Kafka, RabbitMQ, or a microservice), there are three common delivery guarantees:

1. At-most-once – message may be delivered once or not at all
2. At-least-once – message is delivered one or more times
3. Exactly-once – message is delivered exactly once, no duplicates

We’ll focus on At-least-once vs Exactly-once.

---

1️⃣ At-Least-Once Delivery

• Ensures every message is processed at least once
• Messages can be duplicated due to retries
• Simpler to implement
• Often used in streaming and messaging systems

🔹 Example

Suppose a payment system processes an event:

1. Event sent to service
2. Service crashes before acknowledging
3. Event is retried → message processed twice

Event: "Charge $100 to user"
Processing: 2 times → user may get charged twice

✅ Pros: No message lost ❌ Cons: Duplicate messages, need idempotent processing

---

2️⃣ Exactly-Once Delivery

• Ensures each message is processed exactly once
• No duplicates, no data loss
• Harder to implement, usually requires transactional processing
• Common in streaming systems like Kafka with transactional writes

🔹 Example

Same payment system:

1. Event sent to service
2. Service crashes before acknowledging
3. Event retried
4. System ensures charge happens only once

✅ Pros: No duplicates, consistent state ❌ Cons: More complex, may impact performance

---

3️⃣ Comparison Table

Feature	At-Least-Once	Exactly-Once
Guarantee	Process ≥1 times	Process =1 time
Risk	Duplicates	No duplicates
Complexity	Low	High
Latency	Low	Higher (transactional overhead)
Implementation	Retry logic, idempotent operations	Transactions, deduplication, atomic writes
Use Case	Logging, analytics, notifications	Financial transactions, inventory updates

---

🔧 How Exactly-Once is Implemented

1. Idempotent Consumers

   • Consumer ignores duplicate messages based on a unique message ID

2. Atomic Writes

   • Write to output topic/DB only if processing is successful

3. Transactional Messaging

   • Systems like Kafka transactions guarantee atomic commit across multiple partitions

---

🔑 Analogy

Think of mailing letters:

• At-least-once: You mail a letter. If the postman loses it, you resend → recipient may get 2 letters.
• Exactly-once: You mail a letter, and the system ensures the recipient gets exactly 1 letter, no duplicates, even if retries occur.

---

⚡ Summary

• At-least-once: No data loss, duplicates possible, simpler
• Exactly-once: No duplicates, guaranteed processing once, harder & heavier
• Trade-off: Performance vs reliability

---