Table of Contents
A unified namespace (UNS) is an architecture pattern, not a product: plant systems publish their current state to one shared, hierarchically named set of topics, usually on an MQTT broker, and any authorised application subscribes to what it needs. MQTT supplies the publish/subscribe transport. Your naming convention, payload rules and access controls are what turn it into a namespace.
Key takeaways
- MQTT is an OASIS standard for publish/subscribe messaging; version 5.0 was approved in 2019, and version 3.1.1 is also published as ISO/IEC 20922 (OASIS, IEC).
- MQTT deliberately defines neither topic names nor payload format. A UNS convention or the Sparkplug B specification (ISO/IEC 20237) fills that gap (Eclipse Foundation).
- A broker keeps at most the last retained message per topic, not history. A UNS feeds the historian; it does not replace it.
- The MQTT standard leaves authentication and authorisation to the implementer. Plan TLS, per-topic access control and the broker's place in the Purdue model before the first device connects.
What is a unified namespace?
Plant data often moves point to point: the historian polls the DCS, the MES reads the historian, a dashboard queries both. With 10 systems, a full mesh needs up to 45 links.
A unified namespace replaces that mesh with a hub. HiveMQ, a broker vendor, defines it as "an architecture and design pattern for structuring data producers and consumers", giving a "single source of truth" about the structure and current state of the enterprise (HiveMQ, 2026). Each producer publishes once; each consumer subscribes to what it needs.
The term does not appear in the MQTT or Sparkplug specifications; vendors and integrators promote it. The building blocks are standardised: MQTT for transport, Sparkplug for an industrial topic and payload format, and ISA-95 for an equipment hierarchy that can structure topic names.
How does MQTT work?
MQTT is "a Client Server publish/subscribe messaging transport protocol" designed for constrained environments such as machine-to-machine communication (OASIS MQTT 5.0). The specification calls the broker the Server. Clients connect to it, publish messages to topics, and subscribe to topics; publishers and subscribers never connect to each other.
The features that matter in a plant:
| MQTT feature | What the standard says | What it means in a plant |
|---|---|---|
| Topic levels | The forward slash separates levels and gives topic names a hierarchy | site-1/area-utilities/boiler-2/steam-flow can mirror the plant layout |
| Wildcards | + matches one level, # matches any number of levels at the end | A dashboard subscribes to site-1/area-utilities/# and receives everything in that area |
| Retained message | The broker stores the last retained message per topic and delivers it to future subscribers | A new application gets the current value immediately, without waiting for a change |
| Will message | Published by the broker when a client disconnects without a normal DISCONNECT, for example after a network failure | Consumers learn that an edge gateway has gone offline |
| QoS 0, 1, 2 | At most once, at least once, exactly once | Trade-off between overhead and delivery guarantees |
| Shared subscriptions | Each message goes to only one of the sessions sharing the subscription | Several workers can split a heavy processing load |
| Ports | TCP 1883 and 8883 are registered with IANA for MQTT without and with TLS | Firewall rules and audits can expect these ports |
Source: OASIS MQTT 5.0.
The standard notes that retained messages "are useful where publishers send state messages on an irregular basis". A UNS relies on this: when a device publishes only on change, retention lets a new subscriber see the current state at once.
On QoS, the standard's example for "at most once" is ambient sensor data, where a lost reading is soon replaced by the next; for "exactly once", billing systems.
What does MQTT leave undefined?
Three things, on purpose:
- Topic names. MQTT sets syntax rules (topics are case-sensitive, may contain spaces, and topics starting with
$are set aside for the broker's own use) but no meaning. - Payload. MQTT is "agnostic to the content of the payload". A value can be a number, a JSON document or a binary structure.
- Security choices. The standard leaves the choice of network, privacy, authentication and authorisation technologies to implementers and calls including them "the implementer's responsibility" (OASIS MQTT 5.0, section 5).
The Sparkplug specification puts it plainly: MQTT was originally designed as a message transport for real-time SCADA, but for MQTT solutions to be interoperable, "the Topic Namespace, Payload representation, and session state must be defined" (Sparkplug 3.0). Otherwise two integrators on one broker build two incompatible namespaces.
What does Sparkplug B add?
Sparkplug is an Eclipse Foundation specification; version 3.0 was released in 2022 and published in November 2023 as the international standard ISO/IEC 20237 (Eclipse Foundation, 2023). It defines:
- A fixed topic structure:
spBv1.0/group_id/message_type/edge_node_id/[device_id]. - Message types: birth and death certificates for edge nodes and devices (NBIRTH, NDEATH, DBIRTH, DDEATH), data (NDATA, DDATA), commands (NCMD, DCMD) and host application state (STATE).
- A binary payload based on Google Protocol Buffers, with data types, timestamps and metadata.
- Session state. A birth message carries all current values; afterwards, values should be published only when they change (report by exception). If an edge node's connection drops, its death certificate tells every host application to mark that node's data as STALE. A host can request a rebirth to get the full state again.
Two details surprise newcomers. Sparkplug data messages must use QoS 0 and must not be retained, so the current state comes from birth messages, not from retained topics. And the topic is organised by edge node, not by plant equipment, so the plant context has to go into group_id, whose examples in the specification include a pipeline segment or a manufacturing line, or into metric names (Sparkplug 3.0).
| Plain MQTT with JSON | Sparkplug B | |
|---|---|---|
| Topic structure | Your own convention, often following ISA-95 | Fixed: group, message type, edge node, device |
| Payload | Readable text; schema is your choice | Compact binary (Protocol Buffers) with data types |
| Current state | Retained messages | Birth certificates and rebirth requests |
| Offline detection | Will messages, if you design them | Built in: death certificates and STALE quality |
| Easiest for | IT tools, scripts, data teams | SCADA hosts and edge devices that support it |
The choice is not exclusive. One option is to run Sparkplug between edge devices and the broker, and republish values into readable ISA-95 topics for consumers.
How should you structure the topic hierarchy?
The ISA-95 equipment hierarchy is a practical basis. The OPC UA companion specification for ISA-95 lists its levels: enterprise, site and area, then process cells and units for batch and continuous processes, production lines and work cells for discrete manufacturing, and storage zones and units for material movement (OPC 10030).
A topic for a refinery could read company/site-1/crude-unit/pump-p101/discharge-pressure. Rules that save rework:
- Mirror the equipment hierarchy, not the network. Names should survive a change of PLC or gateway.
- Agree on one spelling convention. Topics are case-sensitive, so
Pump-P101andpump-p101are two topics. Use lowercase and no spaces. - Never start a topic with
$. The standard sets those topics aside, and subscriptions that start with a wildcard do not match them. - Put the value, unit, timestamp and quality in the payload. A bare number loses its meaning once it leaves the plant.
- Separate raw, derived and command data into different branches, so access rules can treat them differently.
- Decide what is retained. Retain state, such as current values and equipment status; do not retain events, such as alarms already acknowledged.
- Give the namespace an owner. A convention without an owner drifts. Our article on industrial data governance covers naming ownership.
Does a unified namespace replace the historian or OPC UA?
Not the historian. For each topic, the broker replaces the previous retained message with the new one (OASIS MQTT 5.0). It holds the present, not the past. Trends, reports and model training still need a historian subscribed to the namespace. Our guide to extracting process historian data for machine learning covers that side.
Not OPC UA either. OPC UA Part 14 defines how OPC UA PubSub messages travel over an MQTT broker, with JSON or binary UADP encoding and a default topic pattern starting with opcua/json/data (OPC 10000-14). A practical split is to let OPC UA read from controllers and servers, and an MQTT broker distribute the result. See OPC UA explained for data scientists.
Some vendors present the UNS as a replacement for the ISA-95 "pyramid" (HiveMQ, 2026). For data flow, that is a fair description. For security, it is not: network levels and the boundary between OT and IT still apply, and the broker has to sit somewhere on that map.
How do you secure an MQTT broker in a plant?
The broker becomes one place holding the state of the plant. Protect it accordingly:
- Require authentication. In Eclipse Mosquitto, an open-source broker, versions from 2.0 require you to choose authentication options explicitly before clients can connect (Mosquitto). Check the equivalent default in whichever broker you use.
- Use TLS and client certificates. Mosquitto's
require_certificateoption refuses clients without a valid certificate (mosquitto.conf). - Restrict topics per client. Mosquitto's access control file grants read, write, readwrite or deny per topic, with patterns based on the client ID or username. The Sparkplug specification suggests letting each edge node publish only on its own group and edge node IDs, which makes spoofing another node difficult.
- Treat command topics as writes into OT. Sparkplug NCMD and DCMD messages, or any equivalent, can change the process. Deny them by default and route any write through the existing control and change process.
- Place the broker on the Purdue map. Joint guidance led by CISA prefers "push-based or brokered architectures that move required features or summaries out of OT without granting persistent inbound access" (CISA et al., 2025). Business users should subscribe to a broker on their side of the DMZ, fed from OT, not to the broker inside the control network.
Our articles on the Purdue model and where AI fits, data diodes and OT cybersecurity when you add AI cover these boundaries in more detail.
Where does AI fit on a unified namespace?
On a UNS, a model is one more subscriber.
- Inference on live data. An anomaly or quality model subscribes to the topics it needs, without polling the DCS.
- Results as new topics. The model publishes to its own branch, such as
.../pump-p101/predicted-failure-risk, for operators and dashboards. It advises; it does not write setpoints. - Training from history. Train on historian data, since the broker holds only current values, and compute features the same way in both.
- Context for assistants. An on-premise LLM at the business level can read current values, read-only, from the business-side broker.
See integrating AI with SCADA and OT systems for the wider architecture.
What we have seen at KABAS AI
Our information systems team integrates existing systems, and our data science team builds predictive models and the MLOps pipelines that run them. The advice below is practical guidance from that kind of work, not a published benchmark.
Write the naming convention before installing the broker. A one-page document listing the hierarchy levels, spelling rules, payload fields and retained topics avoids renaming topics once consumers depend on them.
Start with one area and read-only data. Publish one unit's values, connect one consumer such as the historian or a dashboard, and add command topics later, if ever, under the control team's change process.
Decide who owns each branch. The control team owns raw equipment data, the data team owns derived values and model outputs. Each owner approves who can publish in their branch.
Sources
- OASIS, MQTT Version 5.0, OASIS Standard, 7 March 2019
- IEC Webstore, ISO/IEC 20922:2016, Message Queuing Telemetry Transport (MQTT) v3.1.1
- Eclipse Foundation, Sparkplug Specification Version 3.0
- Eclipse Foundation, The Eclipse Foundation Announces Sparkplug as an International Standard, 7 November 2023
- Kudzai Manditereza, What is Unified Namespace (UNS) and Why Does it Matter?, HiveMQ, February 2026
- OPC 10030, OPC UA Common Object Model: ISA-95, ISA95EquipmentElementLevelEnum
- OPC 10000-14, OPC UA Part 14: PubSub, MQTT transport
- Eclipse Mosquitto, Authentication methods
- Eclipse Mosquitto, mosquitto.conf manual
- CISA, ASD's ACSC and partners, Principles for the Secure Integration of Artificial Intelligence in Operational Technology, December 2025
Frequently asked questions
Is a unified namespace the same thing as an MQTT broker?
No. The broker is the software that receives and forwards messages. The unified namespace is the agreed structure on top of it: the topic names, the payload format, what is retained and who may publish where. Installing a broker is quick; agreeing on the namespace takes work between control, IT and data teams.
Do we need Sparkplug B to build a unified namespace?
No. Plain MQTT with a documented topic convention and JSON payloads also works and is easier for IT tools to read. Sparkplug B adds a standard topic structure, a compact binary payload, data types and birth and death messages that tell consumers when data is stale. The two can be combined: Sparkplug between edge devices and the broker, and readable topics for consumers.
Can a unified namespace replace the process historian?
No. An MQTT broker keeps at most the last retained message on each topic, so it shows current state, not history. You still need a historian or time-series database to store values over time, for trends, reports and model training. The historian can be one of the subscribers to the namespace.
Should we choose OPC UA or MQTT?
It is rarely one or the other. OPC UA gives a rich information model and secure client-server access to controllers and servers, while MQTT distributes data through a broker to many consumers. The OPC UA specification itself defines how to publish OPC UA data over MQTT, with JSON or binary encoding.
Which MQTT QoS level should plant data use?
Frequent process values can usually use QoS 0, because the next value follows soon; the MQTT standard gives ambient sensor data as an example of this case, and Sparkplug requires QoS 0 for its data messages. Use QoS 1 for messages that must arrive, such as events or state changes, and make consumers tolerate duplicates. QoS 2 adds a longer handshake and is seldom needed for process values.
Where should the MQTT broker sit in the Purdue model?
It depends on who publishes and who consumes. One pattern is an edge or site broker at Level 3 that collects data from the control network, and a copy pushed through the DMZ to business systems. Current CISA-led guidance prefers push-based or brokered architectures that move data out of OT without persistent inbound access.



