Book a consultation

HiveMQ Manufacturing: Plant Placement, Customers and Failure Modes

HiveMQ Multi-Site Architecture: Central Cluster, Broker per Plant or Cloud

HiveMQ multi-site architecture: one central cluster, a broker per plant or cloud, why bridging is the documented tool, and what happens when the WAN drops.

Vladimir Romanov


HiveMQ multi-site architecture is the choice of where brokers run when you have more than one plant: one central cluster that every plant connects to, a broker in each plant bridged to a central one, or a managed cloud broker. The answer turns on one question, what each plant must do when the WAN is down. HiveMQ’s documented tool for joining sites is bridging, and its documentation does not say whether a cluster may span a WAN. This page compares the layouts and gives our view. It sits under HiveMQ manufacturing and is current as of November 2026. For the bridge mechanics themselves, see edge-to-cloud bridging.

The three HiveMQ multi-site layouts

One central clusterBroker per plant, bridgedManaged cloud broker
Plant connects toThe central cluster over the WANIts own broker on the plant networkA HiveMQ Cloud broker over the internet
When the WAN dropsThe plant’s publishers lose their brokerPlant clients keep working; bridged traffic must bufferSame as central
What you runOne cluster and its operationsA broker per plant plus a cluster and bridgesLittle, but you depend on the service
Fits whenPlants tolerate losing data for a while and one team runs everythingPlants must keep logging and each has local SCADAYou do not want to run brokers, and the WAN risk is acceptable
Main costWAN as a production dependencyMore brokers and bridge configuration to patchPlan and retirement risk: Serverless ends 31 December 2026

HiveMQ’s Siemens reference page draws the middle column: a HiveMQ instance per factory, with selected topics bridged or tunnelled to a central HiveMQ cluster. The page is marketing and says local brokers “continue operating during WAN interruptions”, a design statement, not a measured result.

Choosing a multi-site HiveMQ layout A decision tree. First question: must a plant keep running and logging when the WAN is down. If yes, ask whether a central team needs data from every plant: yes leads to a broker per plant bridged to a central cluster, no leads to a broker per plant with no bridge yet. If no, ask whether you will run brokers yourselves: yes leads to one central cluster that plants connect to over the WAN, no leads to a managed cloud broker with the same WAN dependency. Must a plant keep running and logging when the WAN link is down? Does a central team need data from every plant? Will you run brokers yourselves, with a team to patch them? Yes No A broker per plant, bridged to a central cluster A broker per plant, no bridge yet; add one when a consumer asks One central cluster; plants connect by WAN A managed cloud broker, with the same WAN limit Yes No Yes No Both right-hand outcomes make the WAN a production dependency. In our view, accept that only for data a plant can lose for a while.
Plant autonomy decides the first branch. Everything that depends on the WAN being up is a central design, and the bridge is the documented way to keep plants independent.

Bridging is the documented tool, WAN clustering is not

HiveMQ documents bridging as the way to connect sites. The Enterprise Bridge extension connects a broker or cluster to one or more remote MQTT brokers, with identically configured extensions needed on every cluster node. HiveMQ Edge has its own MQTT bridges. Commercial offline buffering on an Edge bridge holds QoS 1 and 2 traffic during an outage, in memory by default or in a file-native store that reloads after a restart, and forwards it once the connection returns.

WAN clustering is undocumented either way. HiveMQ’s cluster page says nothing we found about spanning sites. We found no HiveMQ page that forbids a stretched cluster and none that supports one over a plant WAN. The page describes a “cluster merge” when an unreachable node returns, without explaining how conflicting state is resolved. In our view, a WAN partition inside a cluster is a far worse event than a bridge that reconnects, because replicated state is what the cluster protects. We would not stretch one across plants without written guidance from HiveMQ.

Plant autonomy when the WAN drops

Aerial view of long white-roofed industrial buildings beside a highway and railway
Each building is a plant that must keep producing; what the central site loses during a network outage is visibility, not control.

A plant broker keeps local publishers, SCADA and historian running. The remaining question is the bridge. MQTT quality of service applies on each hop between a client and a broker, so a chain of brokers is not one end-to-end guarantee. The buffer is finite: illustrative: a line that publishes 500 messages a second of 200 bytes each produces about 100 KB a second, roughly 360 MB an hour, so a four-hour outage needs about 1.4 GB of buffer for that bridge alone. Size from your own rates, and learn what happens at the limit, because Edge’s default queue limits and eviction behaviour are not documented in what we read.

Also decide what the central site does with stale or backfilled data. A historian that timestamps on arrival will put four hours of buffered data at one moment. Publish the source timestamp and make consumers use it.

Questions to settle before choosing

  • What must each plant do offline? List the functions: local alarms, batch records, OEE counters. Each needs its data path to be inside the plant.
  • Who consumes data centrally, and at what delay? A dashboard that is five minutes late during an outage is a different requirement from a billing record.
  • How many sites, and how different? Three similar plants can share a template. Thirty plants across vendors need a namespace team.
  • Who can patch a broker at a plant? If nobody on site can, a central or managed broker may be the honest answer, with the WAN risk accepted in writing.
  • What does a reconnect storm cost? After a long outage every bridge replays at once, and a central cluster must absorb it.

Namespace design across sites

Make the site a topic level and fix the names once. HiveMQ’s Unified Namespace guidance (22 February 2026) says to define context centrally, let local teams add detail, use standard templates per asset type, fix units and naming conventions and validate payloads against schemas. Edge can prefix topics with the ISA-95 hierarchy. In practice:

  1. Agree the top levels centrally: enterprise, site, area, line, asset.
  2. Keep local detail below that, owned by the plant.
  3. Use the same unit and naming rules everywhere, or the central view is a translation project.
  4. For Sparkplug, Group ID plus node ID must be unique, so make the site part of the Group ID.
  5. Decide which topics bridge. Bridge the modelled, contextualised ones, not every raw tag.

The design questions behind this are in Unified Namespace and the HiveMQ Unified Namespace page.

The Siemens reference pattern, step by step

The Siemens Architecture Hub page for HiveMQ describes four steps, which are HiveMQ’s and Siemens’ design, not ours. PLCs, drives and sensors connect to Siemens Industrial Edge devices through OT connectors. A local HiveMQ per factory enforces topic structure, metadata and access control for HMI, SCADA and analytics. Local brokers bridge or tunnel selected topics to a central HiveMQ cluster that serves ERP, MES, history and cloud analytics. Industrial Edge Management handles deployment and lifecycle. The page contains no partnership text, so read it as an architecture sketch.

Where it gets hard

  • Two sources of truth. A topic changed at a plant and not centrally creates silent mismatches. Version the namespace.
  • Operating many brokers. Every plant broker is something to patch, back up and monitor, on a monthly release cadence.
  • Sparkplug across a bridge. Primary host and STATE behaviour per broker is not covered in the sources we read. Test it before design.
  • Ownership. Decide who owns the bridge when it breaks at 02:00: the plant, IT or the vendor.
  • Cluster basics. Each central or plant cluster needs a replica count of at least 2. See clustering and high availability and Cloud versus self-hosted.

For the broker-neutral version of this decision, see UNS broker. Back to HiveMQ manufacturing and the HiveMQ hub. Sibling: OT network security.

Frequently asked questions

Should I run one HiveMQ cluster for all plants or one broker per plant?

In our view, a broker per plant bridged to a central cluster whenever a plant must keep working through a WAN outage. One central cluster is simpler but makes the WAN a production dependency. HiveMQ's documented multi-site tool is bridging, and its Siemens reference page shows a local broker per factory.

Can a HiveMQ cluster span multiple sites over a WAN?

We found no HiveMQ page that forbids or supports a stretched cluster over a plant WAN. Treat it as undocumented and ask HiveMQ for written guidance. In our view, a WAN partition would then split the cluster.

What happens to plant data when the link to the central broker drops?

It depends on the layout. A plant broker keeps serving local clients. Bridged traffic needs buffering: HiveMQ Edge documents commercial offline buffering, in memory or on file, and the broker queues messages for persistent sessions up to a configurable maximum. Check your sizing, because both are finite.

How should topic namespaces be designed across several sites?

Define context centrally and let each plant add detail. HiveMQ's Unified Namespace guidance says to use standard templates per asset type, fixed units and naming conventions, and to validate payloads against schemas. Make the site a topic level, and keep Sparkplug Group ID plus node ID unique across the estate.

Is HiveMQ Cloud a good central broker for several plants?

It can be, if the plants can tolerate the WAN dependency and you accept a managed service. HiveMQ's documentation states that HiveMQ Cloud Serverless retires on 31 December 2026, so check which Cloud plan you would be buying. Bridging from a plant broker works the same way as to a self-hosted cluster.

References and further reading

  1. HiveMQ cluster documentation · HiveMQReplica count, discovery, transport, network split and merge.
  2. HiveMQ Enterprise Bridge Extension · HiveMQThe documented tool for connecting a broker or cluster to remote brokers.
  3. HiveMQ architecture page · Siemens Architecture HubA local HiveMQ per factory, bridged to a central cluster.
  4. Foundations of Unified Namespace architecture · HiveMQPublished 22 February 2026: central context, local detail, templates and schemas.
  5. HiveMQ Edge configuration · HiveMQBridges and offline buffering modes.
  6. HiveMQ Cloud documentation · HiveMQStates the Serverless retirement date of 31 December 2026.

← Manufacturing

Have a project that needs technical ownership?

Send us the situation. If it is not something we should take on, we will say so.

The Plant-Side Brief

The weekly brief for the people who own a plant’s control systems.

One email every Wednesday for engineering and operations leaders responsible for controls, SCADA, MES and OT networks. Skim the top in two minutes; read the Deep Dive when you have fifteen.

  • One modernization problem, worked through to cost, risk and a schedule you can defend. The Deep Dive.
  • End-of-support dates before they reach your plant Windows, servers, controllers and HMI software. Lifecycle Watch.
  • A checklist your engineers can run that week not a framework to admire. Field Kit.
  • Every number sourced, no vendor agenda we are paid by plants, never by vendors.