# 📝 Worked decision walkthroughs: four use cases and a 2026 cost check

> This lesson is part of the paid section. Sign in and purchase the course to view the full content.

## Learning Objectives

By the end of this lesson you can:

- Run the same six-question checklist against any candidate workload.
- Defend a broker pick for payments, event sourcing, task distribution and IoT telemetry using delivery, ordering, operational and licensing arguments.
- Build a first-order monthly cost estimate for both brokers from message rate, message size and retention.
- Name the dominant cost driver in each architecture and the one number that most often surprises teams.

## Quick Review

Last lesson split ecosystem fit into three questions: which client library your code imports, whether you need a connector platform, and whether you need stateful stream processing.

We also established that operations and retention behave very differently: Kafka keeps bytes on disk for a configured window, RabbitMQ deletes a message the moment it is acknowledged. Hold those two facts close, because every walkthrough below leans on them.

## Theory: The Six-Question Checklist

Before you argue about products, answer these in order. Stop at the first question that decides the case.

| # | Question | Points to Kafka | Points to RabbitMQ |
| --- | --- | --- | --- |
| 1 | Must anyone replay old data later? | Yes, replay is the point | No, once processed it is gone |
| 2 | What ordering unit matters? | Per key, achieved by partition key | Per queue, single active consumer |
| 3 | What does failure recovery look like? | Offset rewind, reprocess a range | Per-message nack, retry, DLQ |
| 4 | What is the message rate? | Sustained high fan-in | Moderate, bursty, uneven task times |
| 5 | How complex is routing? | Simple topic fan-out | Exchanges, bindings, headers, priorities |
| 6 | What dominates cost? | Bytes retained × replication | In-flight backlog + node count |

> [!IMPORTANT]
> Question 1 is the one people skip. If nobody can name who replays the data and why, you probably do not need a log. If somebody can, you almost certainly do.