// the find
rabbitmq/rabbitmq-server
Open source RabbitMQ: core server and tier 1 (built-in) plugins
RabbitMQ is the reference implementation of a multi-protocol message broker, supporting AMQP 0-9-1, AMQP 1.0, MQTT 5.0, and STOMP out of the box. It's been running production workloads since 2007 and is what most developers mean when they say 'message queue'. Written in Erlang, which is not incidental — the fault-tolerance model is baked into the runtime.
Quorum queues (Raft-based, v3.8+) give you actual durability guarantees with per-message fsync and replication — classic mirrored queues were always a footgun and the team has been honest about deprecating them. The Streams feature adds append-only log semantics with non-destructive consumers, essentially Kafka-style offset replay inside RabbitMQ, without standing up a separate system. Prometheus/Grafana integration is first-class and ships metrics that actually matter (queue depth, publish/deliver rates, memory per queue, credit flow blocks). The Kubernetes Operator handles rolling upgrades across minor and patch versions without queue downtime, which is genuinely hard to get right.
Erlang is a dependency you will eventually need to debug — when something goes wrong in the internals (memory alarm triggers, credit flow, mnesia partition), you're reading BEAM crash dumps and Erlang stack traces whether you like it or not. The jump from single-node to cluster is operationally non-trivial: you need to understand partition handling strategies, quorum queue minimum sizes, and what happens to classic queues when a node goes down, and none of that is obvious from the docs. AMQP 1.0 was bolted on late and the management UI still treats it like a second-class citizen — filtering, tracing, and per-queue stats are weaker for 1.0 connections than for 0-9-1. Stream consumer performance degrades noticeably if you're not careful about subscription credits and flow control — the defaults are not tuned for high-throughput replay scenarios.