Edge data management is how you collect, clean, model, store and govern data on machines, gateways and site servers before any of it reaches the cloud. In 2026 it usually means an MQTT-based data layer on site, analytics and AI models running next to the equipment, and only filtered, contextualised data sent upstream. Done well, it cuts bandwidth and latency and keeps working when the network does not.
Updated October 2026
What changed in 2026
- Edge data planes replaced point tools. Microsoft positions Azure IoT Operations as a unified edge data plane on Arc-enabled Kubernetes, with an MQTT broker, data flows and connectors, and up to 72 hours of offline operation.
- AWS IoT Greengrass keeps moving. Greengrass v2.17.0 (16 Apr 2026) added non-root installs and a much smaller secure tunnelling component (about 4 MB, down from about 36 MB).
- Cloud-only IoT analytics services are gone. AWS IoT Analytics reached end of support on 15 Dec 2025. Azure Time Series Insights retired on 31 Mar 2025, with Azure Data Explorer as the named path.
- The unified namespace is mainstream. Sparkplug 3.0 is an international standard, ISO/IEC 20237. Vendors such as HighByte with Siemens Industrial Edge sell UNS as a reference architecture.
- AI agents now read edge data directly. Litmus Edge added an API so agents can query asset models and history, plus locally hosted small language models for air-gapped sites.
What edge data management covers
Think of it as five jobs, done close to where data is made.
| Job | What it means on site | Typical tools in 2026 |
|---|---|---|
| Connect | Read PLCs, sensors, cameras and historians over OPC UA, Modbus, MQTT | Akri connectors, Litmus Edge, Kepware, Siemens Industrial Edge |
| Contextualise | Turn raw tags into named assets with units and hierarchy | HighByte Intelligence Hub, Azure IoT Operations data flows |
| Store and buffer | Keep recent data locally; buffer when the uplink drops | Local time-series DB, Greengrass Stream Manager, MQTT broker persistence |
| Analyse and act | Run rules, models and agents next to the machine | Edge AI runtimes, stream processors, PLC write-back |
| Govern and forward | Decide what leaves the site, who can see it, and log it | Schema registry, role-based access, audit logs, cloud data flows |
Edge analytics is the fourth job. Edge data management is all five. You cannot run useful analytics on data you have not connected, named and trusted.
Why process data at the edge
The reasons have not changed much since the first version of this post. What changed is that the tools now exist to act on them.
- Latency. A defect camera or safety interlock needs an answer in milliseconds. A round trip to a cloud region is too slow and too variable.
- Bandwidth and cost. High-rate vibration or video data is expensive to ship. Sending features, events and summaries is cheaper.
- Resilience. Sites lose connectivity. Local buffering and local decisions keep the line running.
- Data residency and privacy. Some data cannot leave the site or country. Processing it locally keeps it in place.
- Context. Operators and engineers know the assets. Modelling data at the site captures that knowledge before it is lost.
A reference architecture for 2026
Most new deployments follow the same layered pattern.
- Device layer. PLCs, sensors, robots and cameras.
- Connectivity layer. Protocol connectors normalise OPC UA, Modbus and vendor protocols.
- Site data layer (UNS). An MQTT broker holds the current state of every asset, organised by an ISA-95-style hierarchy (enterprise / site / area / line / cell). Sparkplug B adds a standard payload and state handling.
- Edge compute layer. Containers on Kubernetes or a lightweight runtime run analytics, AI models and agents. They subscribe to the UNS and publish results back to it.
- Cloud layer. Contextualised data flows to a lakehouse or real-time analytics service for fleet-wide reporting and model training. Fabric Real-Time Intelligence (GA Nov 2024), Databricks and Snowflake are common targets.
The key design rule: every consumer reads from the site data layer, not directly from devices. That stops the point-to-point spaghetti that made older IIoT projects hard to scale.
Edge analytics: types and use cases
Edge analytics runs four kinds of analysis on the edge layer:
| Type | Question it answers | Edge example |
|---|---|---|
| Descriptive | What is happening now? | Live OEE per line |
| Diagnostic | Why did it happen? | Correlating a pressure drop with a valve state |
| Predictive | What will happen next? | Bearing failure risk from vibration features |
| Prescriptive | What should we do? | An agent proposing a speed change, approved by an operator |
The common use cases from the original post still hold: remote monitoring and maintenance, retail shopper analytics, and smart video surveillance (see real-time video analytics). In 2026 add machine vision quality inspection and AI agents that draft work orders from edge events.
When edge analytics is the wrong choice: if decisions can wait minutes, the data is small, and the network is reliable, cloud analytics is simpler. Edge adds devices to patch, secure and update.
Governance at the edge
This is where most projects stall. Plan for it from day one.
- Schema and naming. Agree an asset model and topic structure before you connect the second site.
- Data contracts. Each consumer gets a versioned schema. Breaking changes need sign-off.
- Access control. Who can read which topics, and who can write back to equipment, must be explicit.
- Audit. Log every model version and every automated action taken at the edge.
- Retention. Decide what stays local, for how long, and what is summarised upstream.
Retired services: what to migrate from
If your architecture diagram still shows any of these, it needs updating.
| Service | Status | Common path |
|---|---|---|
| Google Cloud IoT Core | Retired Aug 2023 | Partner platforms or self-managed MQTT |
| Azure Time Series Insights | Retired 31 Mar 2025 | Azure Data Explorer or Fabric Real-Time Intelligence |
| AWS IoT Analytics | End of support 15 Dec 2025 | Other AWS data services (see the AWS migration page) |
How to get started
- Pick one line or site with a clear problem, such as unplanned downtime.
- Inventory the data sources and protocols there.
- Stand up a site broker and model the first 20 to 50 assets.
- Run one analytic or model at the edge and measure the result against a baseline.
- Only then design the cloud flow and the second site.
Where it's heading: 2026 → 2030
| Period | Edge data management |
|---|---|
| Around 2020 | Point-to-point gateways; cloud IoT hubs and cloud-only analytics services; few shared data models |
| 2026 (now) | Site-level UNS on MQTT/Sparkplug; edge data planes on Kubernetes; IDC put 2025 edge spending at about $265B (IDC, Feb 2026) |
| By 2028 (expected) | IDC forecasts edge spending of about $450B by 2029, growing roughly 15% a year (IDC); shared asset models across sites expected to be the norm |
| By 2030 (expected) | AI agents expected to be routine consumers of edge data, the foundation of physical AI, with governed write-back to equipment |
How NexaStack helps
NexaStack Edge AI deploys, updates and monitors AI models and agents across edge sites from one control plane. It connects to your site data layer, runs inference next to the equipment, and logs every model version and automated decision so actions stay governed and auditable.
Frequently asked questions
What is edge data?
Edge data is data created and first processed outside a central data centre, on devices, gateways or site servers. Examples are sensor readings, PLC tags, machine logs and camera frames.
What is edge analytics?
Edge analytics is running analysis, rules or AI models on edge data where it is created, so you get answers without waiting for the cloud.
What are some examples of edge devices?
Industrial PCs, IoT gateways, PLCs with compute modules, smart cameras, robots, vehicles and ruggedised servers on a factory floor or cell site.
Who uses edge data centres?
Manufacturers, telecom operators, retailers, energy companies and hospitals use them for low-latency or data-heavy workloads that cannot depend on a distant cloud region.
Is edge data management the same as a unified namespace?
No. A UNS is one part of it: the site data layer. Edge data management also covers connectivity, analytics, governance and what you forward to the cloud.
Get your edge data ready for AI
Bring one site and its data sources, and we will map the path from raw tags to governed edge AI with NexaStack. Book a 30-minute edge data readiness call →
