# broker-comparison

Companion project for **Kafka vs RabbitMQ vs Pulsar for Java Teams: A Decision Framework with
Benchmarks** on [ankurm.com](https://ankurm.com).

Three brokers, three questions, one set of measurements. Ordering, replay and consumer scaling are
asked identically of all three by driving the client libraries directly, so that what is compared
is the brokers' delivery models rather than three sets of Spring defaults. What each Spring
integration adds on top is [chapter 6](docs/06-what-spring-adds.md).

## Verified stack

| Component | Version | Source of the number |
|---|---|---|
| JDK (tests) | 25.0.4.1+1 (Temurin) | `java -version` |
| JDK (brokers) | 21.0.12.1+1 (Temurin) | `java -version` |
| Spring Boot | 4.1.1 | `maven-metadata.xml` on Maven Central |
| kafka-clients | 4.2.1 | `spring-boot-dependencies-4.1.1.pom` |
| amqp-client | 5.30.0 | `spring-boot-dependencies-4.1.1.pom` |
| pulsar-client | 4.2.4 | `spring-boot-dependencies-4.1.1.pom` |
| Kafka broker | 4.2.1 (KRaft) | `kafka_2.13-4.2.1.tgz` |
| RabbitMQ broker | 3.10.25 | generic-unix tarball |
| Pulsar broker | 4.2.4 standalone | `apache-pulsar-4.2.4-bin.tar.gz` |

RabbitMQ is pinned at 3.10.25 because the box that produced these transcripts has Erlang 24;
3.11 and later need Erlang 25. The AMQP 0-9-1 semantics measured here are unchanged in 4.x.

## Quickstart

The Kafka measurements need no external broker — `spring-kafka-test` starts a real KRaft broker
in-process. The other two need a broker each, and the scripts start both without Docker and
without root.

```bash
mvn -Dgroups=kafka test

ERL_ROOT=... RABBITMQ_HOME=... scripts/rabbit-broker.sh
mvn -Dgroups=rabbit test

PULSAR_HOME=... scripts/pulsar-broker.sh
mvn -Dgroups=pulsar test

scripts/footprint.sh kafka|rabbit|pulsar     # the operational numbers
scripts/run-all.sh                           # regenerates every docs/output/ file
```

## Documentation

| Chapter | What it settles |
|---|---|
| [01 Three models](docs/01-three-models.md) | Where a message lives, and why everything else follows |
| [02 Ordering](docs/02-ordering.md) | What survives the second consumer, in all three |
| [03 Replay](docs/03-replay.md) | Reading it twice, and the one broker that cannot |
| [04 Consumer scaling](docs/04-consumer-scaling.md) | The partition ceiling, and the client buffer that defeats fan-out in the other two |
| [05 Operational footprint](docs/05-operational-footprint.md) | Startup, memory, ports, processes, configuration surface |
| [06 What Spring adds](docs/06-what-spring-adds.md) | The three integrations, and the Boot 4 starter trap |
| [07 The decision table](docs/07-the-decision-table.md) | Every row backed by a measurement, and how to choose |

## Captured output

| File | What it shows |
|---|---|
| [`kafka-ordering.txt`](docs/output/kafka-ordering.txt) | Per-key order kept, global order not — and why A, B and C were the wrong keys to demonstrate it |
| [`kafka-replay.txt`](docs/output/kafka-replay.txt) | The same 12 records read three times |
| [`kafka-consumer-scaling.txt`](docs/output/kafka-consumer-scaling.txt) | 5 consumers, 3 partitions, 2 idle |
| [`rabbit-ordering.txt`](docs/output/rabbit-ordering.txt) | Message 1 finishing tenth |
| [`rabbit-replay.txt`](docs/output/rabbit-replay.txt) | 12 messages, then 0, then an empty queue |
| [`rabbit-consumer-scaling.txt`](docs/output/rabbit-consumer-scaling.txt) | 40/0/0/0/0 without `basicQos`, 8 each with it |
| [`pulsar-ordering.txt`](docs/output/pulsar-ordering.txt) | `Shared` spreading a key, `Key_Shared` pinning it |
| [`pulsar-replay.txt`](docs/output/pulsar-replay.txt) | `seek(earliest)` and a new subscription |
| [`pulsar-consumer-scaling.txt`](docs/output/pulsar-consumer-scaling.txt) | The 1000-message receiver queue defeating fan-out |
| [`footprint.txt`](docs/output/footprint.txt) | Startup, memory, ports and config surface for all three |
| [`tests.txt`](docs/output/tests.txt) | The nine tests behind all of the above |

## What is not measured here

Throughput. A messages-per-second number from one 2-core container says nothing about any of
these brokers, and publishing one would be worse than publishing nothing. The measurements above
are structural: they hold on any hardware, because they are properties of the delivery models
rather than of the machine.