How RabbitMQ actually works in production: a mental model for operators

Most RabbitMQ incidents look mysterious only because the operator is reasoning about the wrong abstraction. RabbitMQ is not a daemon with threads and shared memory. It is an Erlang/OTP application running on the BEAM VM, and almost everything you interact with (a connection, a channel, a queue) is an independent Erlang process with its own mailbox, scheduled alongside thousands of others. Once you internalize that, the weird symptoms become predictable: why one hot queue can stall an entire connection, why a single node crossing a memory threshold freezes publishers cluster-wide, and why a “healthy” broker can appear to drop messages.

This article builds that mental model: flow control vs resource alarms, classic vs quorum vs stream queues, the memory watermark and paging ladder, the metadata store, and the single Erlang distribution port that carries all inter-node traffic. Nothing here is a fix. Everything here explains why the fixes work.

What it is and why it matters

RabbitMQ is an AMQP 0-9-1 broker (with STOMP, MQTT, and AMQP 1.0 via plugins) implemented as an Erlang/OTP application. The BEAM VM runs one scheduler thread per CPU core by default, and those schedulers execute lightweight Erlang processes. Each process gets a budget of reductions (roughly, function calls) before being preempted, so no single process can monopolize a scheduler. But each process also has a mailbox: a FIFO inbox of messages from other processes. A process that falls behind does not starve others, but its mailbox grows without bound, and that growth is memory.

RabbitMQ’s failure modes are process-model failure modes:

• A queue is a single Erlang process. It is single-threaded by construction. One queue cannot scale beyond roughly one core’s worth of work, no matter how many cores the node has.
• Backpressure is not TCP-level magic. It is an internal credit system between Erlang processes, and it propagates hop by hop from the disk back to the publisher’s socket.
• Resource alarms are process-level circuit breakers that flip a boolean and change connection behavior cluster-wide. There is no graceful degradation curve.

Operators who think in threads and locks misread these behaviors as bugs. Operators who think in mailboxes, credits, and watermarks can predict them.

How it works

Everything is an Erlang process

The core process inventory on a RabbitMQ node:

• Connection processes: one per TCP connection. Handles AMQP framing, heartbeats, and channel multiplexing. Each costs a file descriptor, a socket, and per-process memory.
• Channel processes: one per channel. Channels are where messaging work happens: publish, consume, ack, declare, bind. A client connection with 100 channels is 101 Erlang processes on the broker.
• Queue processes: one per queue. Maintains the message sequence, delivery state, per-consumer unacked tracking, and paging between memory and disk. For quorum queues, this process runs the Raft state machine.
• Message stores: shared processes (msg_store_persistent and msg_store_transient) that write message bodies to disk through a write-ahead journal and segment files, with queue indices tracking positions.
• Metadata store: on 3.x clusters this is Mnesia, Erlang’s distributed database holding cluster metadata (users, permissions, vhosts, queue/exchange/binding definitions) but not message data. Khepri, a Raft-based metadata store, is opt-in during the 3.13/4.0 era, became the default for new clusters in 4.2, and is the only supported backend since 4.3, when Mnesia was removed.
• Schedulers: typically one per core, running all of the above. The aggregate run queue length is your direct measure of CPU saturation.

flowchart LR
  PUB[Publisher] -->|TCP| CONN[Connection process]
  CONN --> CH[Channel process]
  CH -->|credit-limited send| QP[Queue process]
  QP -->|page out| MS[Message store / WAL]
  CONS[Consumer] <-->|deliver / ack| CH
  QP --> MN[(Metadata store: definitions only)]
  subgraph Other node
    QP2[Queue / Raft peer]
  end
  QP <-->|"Raft + metadata + stats (port 25672)"| QP2
  MS -.->|slow I/O spends credit back| CH

The arrow from the message store back to the channel is the important one. Credit is spent at each hop (connection to channel, channel to queue, queue to persistence). When a downstream process cannot keep up, credit runs out and the sender blocks. The queue slows because it is paging to disk, so the channel blocks, so the connection’s reader stops pulling from the socket, so the publisher’s TCP buffer fills. The publisher sees throughput collapse with no error. That is the system working as designed: it trades publisher latency for bounded broker memory.

Two throttling mechanisms, not one

Operators constantly confuse these, and the response to each is different:

1. Credit-based flow control is per-connection, surgical, and transient. It toggles many times per second. A connection shows state: flow in the management API. A few connections in flow during a burst is normal. Many connections sustained in flow means a systemic write-path bottleneck: slow disk, a hot queue process, or a congested cluster link.
2. Resource alarms are cluster-wide emergency stops. When memory crosses the watermark or disk free drops below disk_free_limit on any node, all publishers on all nodes are blocked. Connections show state: blocking (alarm active, has not tried to publish) or state: blocked (tried to publish, frozen). Consumers keep draining.

One practical consequence of the credit model: a single slow queue can throttle every channel on a connection, including channels publishing to healthy queues. This is why separating publishing and consuming onto different connections is a standard recommendation, and why “one bad queue froze my whole app” is a topology problem, not a broker bug.

The memory ladder: paging, then the wall

Memory behavior has two thresholds, and the distance between them is your only warning:

• Paging threshold: vm_memory_high_watermark_paging_ratio (default 0.5) multiplied by the watermark. Past this point, queues aggressively page message contents to disk. Throughput degrades and disk I/O climbs, but publishers are not blocked. This is the leading indicator almost nobody watches.
• High watermark: vm_memory_high_watermark. Default was 0.4 (40% of RAM) in 3.x and changed to 0.6 (60%) in 4.x. Crossing it fires the memory alarm and blocks all publishers cluster-wide. The transition is a cliff edge: there is no gradual slowdown at the alarm itself, only at the paging stage before it.

Two details bite people. First, unacknowledged messages in classic queues cannot be paged to disk; they are pinned in RAM so the broker can redeliver them if a consumer dies. Unacked accumulation is therefore more dangerous per message than ready-queue depth. Second, the alarm has no hysteresis: it clears at the same threshold where it fired, so a node hovering at the watermark can flap.

Disk has its own alarm at disk_free_limit, default 50MB. On any modern server that default is dangerously small; a burst of persistent writes or a log spike can trip it in seconds. Consumers continue working during a disk alarm, which is the one recovery lever you get for free.

Queue types are different machines

The word “queue” hides three distinct implementations:

• Classic queues (v1, and v2 since 3.10): a single Erlang process backed by a per-queue index and the shared message store. Pages to disk under pressure. Classic mirrored queues (the old HA mechanism) were deprecated in 3.9 and removed in 4.0; if you still run them, migrating to quorum queues is an upgrade prerequisite.
• Quorum queues: the queue process runs Raft. Each queue has a leader plus followers, its own write-ahead log, and a majority requirement for commits. They are the recommended type for durability and HA, but they cost more: more memory per message, more file descriptors for WAL segments, more disk I/O, and a new failure mode (minority state, where the queue refuses writes). Leader placement matters because the leader does all the work; followers only replicate.
• Streams (3.9+): append-only log segments, Kafka-style. Consumer position is an offset, so monitoring shifts from queue depth to consumer lag. Excellent for high-throughput replayable event streams, wrong for task-queue semantics.

The distribution port is one TCP connection

In a cluster, every pair of nodes talks over the Erlang distribution protocol on a single TCP connection (port 25672, with EPMD on 4369 for node discovery). That one pipe carries metadata replication, quorum queue Raft traffic, management stats aggregation, and internal RPC. It is flow-controlled, invisible to most monitoring stacks, and a well-known bottleneck.

When it saturates, the symptoms arrive scattered: metadata transaction timeouts, Raft commit lag, and heartbeat delays. Heartbeats matter because partition detection depends on net_ticktime (default 60s). A sustained GC pause or scheduler saturation that delays ticks past that window produces a false partition detection. So “network partition” alerts during CPU overload are often the distribution path, not the network.

Where this model shows up in production

The characteristic failure archetypes map directly onto the machinery above:

• Memory wall: unacked or ready messages grow, paging starts, disk I/O climbs, credit flow throttles publishers, then the watermark fires and publishing stops cluster-wide. Cliff edge at the end of a visible ramp.
• Hot queue: one queue process saturates its single core. Publish latency to that queue climbs while the rest of the node idles. Adding nodes does not help; only sharding the queue does.
• Disk halt: disk_free_limit breached, publishers blocked, consumers still draining. Often a 50MB default on a terabyte disk.
• FD and process exhaustion: connections, channels, queues, and WAL files all consume FDs; each also consumes an Erlang process from a finite table (default limit 1,048,576). Default OS ulimits around 1024 are insufficient for any real deployment.
• False partitions: run queue saturation delays net_ticktime heartbeats and the cluster reports a split the network never had.
• Distribution port congestion: inter-node traffic saturates the single 25672 pipe and every cluster subsystem degrades at once.

Tradeoffs and when they bite

Quorum vs classic vs stream. Quorum queues buy durability and safe failover at the cost of memory, FDs, disk I/O, and write latency. Streams buy throughput and replay at the cost of AMQP queue semantics. Classic queues are cheap and fast but offer no HA on their own. Pick per workload, not per cluster.

High watermark headroom. A higher watermark (the 4.x default of 0.6) gives more breathing room but shrinks the margin between the alarm and an actual OOM kill, especially in containers. RabbitMQ does not always detect cgroup memory limits reliably; an absolute watermark is the safer choice in containerized deployments.

Alarm tuning. Raising disk_free_limit from the 50MB default (1-2x RAM as a relative value, or a multi-GB absolute floor) converts a guaranteed false alarm into a useful one. The tradeoff is less usable disk, which is almost always the right exchange.

Prefetch. High prefetch improves throughput but lets a single stuck consumer pin thousands of unacked messages in RAM, unreachable by paging and by other consumers. Low prefetch protects memory but starves fast consumers. There is no correct default; it is per-workload.

Signals to watch in production

These are the model’s load-bearing signals. Everything else in a RabbitMQ monitoring setup hangs off these.

How Netdata helps

• Netdata’s RabbitMQ collector pulls the management API signals that map to this model: node memory ratio vs mem_limit, alarm booleans, FD and socket ratios, Erlang process usage, and per-connection states, so the two throttling mechanisms stay visibly distinct on a dashboard instead of being grepped out of /api/connections by hand.
• Ready vs unacknowledged depth are charted separately per queue, which is exactly the split that distinguishes memory-alarm risk (unacked) from consumer absence (ready).
• Run queue, memory ratio, and paging indicators can be correlated on one timeline, so you see the paging ramp that precedes the cliff-edge alarm rather than discovering the alarm itself.
• Cluster link traffic, partition status, and queue state can be viewed together, which is the corroboration you need before treating zero inter-node traffic as a partition.
• Per-second collection catches the transient flow toggling that 5-second management API windows smooth over, while sustained alerting conditions keep normal surgical backpressure from paging you.

Related guides

This article is the foundation piece for the RabbitMQ section. Troubleshooting runbooks that build on this mental model (memory alarms, flow control, partitions, quorum queue issues) will be listed at the RabbitMQ guides hub as they are published.