SEC-04 / Data Foundations

Moving OT data to a cloud historian: architecture, security and governance

Move plant data to the cloud through a one-way, outbound-only path from a DMZ, add context before it leaves the site, and treat every new data flow as a governed conduit with a named owner.

7 min read Updated OT engineersCIOsCISOsPlant managers

Key takeaways

  • There are three common patterns: replicate an on-prem historian, use an edge gateway, or publish through an MQTT broker. Most mid-market plants end up with a mix.
  • Nothing in the cloud should be able to open a connection into Level 2 or Level 3. Outbound-only from a DMZ is the baseline.
  • Context (asset, product, batch, units) is cheaper to add at the edge than to reconstruct in the cloud later.
  • Cloud costs are driven by tag count, sample rate and query patterns. Decide what you actually need before you replicate everything.
  • Every tag sent off-site needs an owner, a meaning and a retention rule, or your AI projects will inherit the confusion.

Almost every industrial AI project starts the same way: someone needs plant data somewhere other than the plant. A data science team wants three years of trends. A vendor wants live data for a predictive maintenance pilot. Corporate wants OEE across five sites on one screen. The answer is usually "put it in the cloud," and the question that matters is how to do that without opening a path back into the control network or creating a pile of unowned data nobody trusts.

This page covers the architecture options, the security baseline, where to add context, the cost traps, and a governance checklist you can use before the first tag leaves the site.

Does this apply to you?

Yes, if any of these are true:

  • You have an on-prem historian (OSIsoft/AVEVA PI, Ignition, a SCADA vendor's built-in historian, or similar) and someone has asked for its data off-site.
  • An OEM or AI vendor has proposed an edge box that "just needs internet."
  • You are planning analytics or AI across more than one plant.
  • You already send data to the cloud, but nobody can tell you exactly which tags, by which path, or who approved it.

Three architecture patterns

There is no single right answer. Most mid-market plants use one primary pattern and tolerate one or two exceptions.

1. On-prem historian replication

The site historian stays the system of record. A replication service or connector pushes a selected set of tags (or the whole archive) to a cloud historian or data platform.

  • Good for: sites that already have a mature historian with clean tag structures and want long-term storage and cross-site analytics.
  • Watch for: replication of compressed data. If the site historian uses swinging-door or deadband compression, the cloud copy inherits it. That may be fine for dashboards and bad for some models. See industrial data quality.

2. Edge gateway

A gateway device or software at Level 3 or in the DMZ reads from PLCs, OPC UA servers or the historian, buffers locally, and sends upstream.

  • Good for: sites without a historian, or where the historian is old, overloaded or OEM-locked.
  • Watch for: gateways that read directly from Level 1 and Level 2 devices are new conduits. They need hardening, patching and an owner. An unmanaged gateway is just another unmanaged computer on the control network.

3. MQTT broker

Data sources publish to an MQTT broker. Subscribers (cloud connectors, dashboards, MES) consume what they need. Often combined with Sparkplug B for a defined payload and state model. This is also the backbone of most Unified Namespace designs.

  • Good for: many producers and many consumers, event-driven architectures, sites building toward a UNS.
  • Watch for: brokers default to "anyone authenticated can publish anywhere" unless you configure access control lists. That matters more than people expect, because command messages travel the same way as data.

Comparison

Pattern Typical effort Strength Main risk
Historian replication Low to medium Reuses existing tag model and history Inherits compression and tag mess
Edge gateway Medium Works without a historian, local buffering New unmanaged device in OT
MQTT broker Medium to high Decoupled, scales to many consumers Weak ACLs allow writes into OT

Security baseline: outbound-only from a DMZ

Whatever the pattern, the baseline is the same. NIST SP 800-82 Rev. 3 and IEC 62443 both point you toward segmented zones with controlled conduits between them. In practice, for cloud data:

  Cloud data platform
          ^
          | outbound TLS only (site initiates)
  +-------+--------------------------+
  | IT/OT DMZ (Level 3.5)            |
  |  replica historian / broker /    |
  |  gateway, publish-only creds     |
  +-------+--------------------------+
          ^
          | pull from below, no inbound
  +-------+--------------------------+
  | Level 3: site historian, MES     |
  +-------+--------------------------+
          ^
  | Level 2/1: SCADA, HMIs, PLCs     |
  +----------------------------------+

The rules that make this work:

  1. The site initiates. Nothing in the cloud can open a session into the DMZ, let alone below it.
  2. No direct path from Level 2 to the internet. Data steps through the DMZ, ideally via a replica or broker that holds a copy, not a pass-through proxy.
  3. Publish-only credentials. The identity used to send data cannot read configuration, subscribe to command topics, or write anything back.
  4. Write-back is a separate decision. If a future AI use case wants to send setpoints or recommendations back, that is a new conduit with its own risk assessment and MOC. Do not let it ride on the read path.
  5. Buffer locally. Plan for internet outages. Operators should never lose local trends because the cloud is unreachable.

For high-consequence sites, a hardware data diode or unidirectional gateway is a proportionate option. For most mid-market plants, a well-configured firewall with outbound-only rules from the DMZ is a reasonable starting point. Document the path in your zones and conduits model either way.

Watch out: Vendor edge boxes that ship with a remote-access agent "for support" are the most common way an outbound-only design quietly becomes two-way. Ask the question directly in your vendor security questionnaire.

Add context before the data leaves the site

Raw tag values are close to useless off-site. FT_101.PV = 43.7 means nothing to a data scientist in another country. Context includes:

  • Asset hierarchy: site, area, line, cell, equipment. ISA-95 (IEC 62264) gives you a standard vocabulary for this.
  • Engineering units and ranges.
  • Production context: product or SKU, batch or order, recipe, shift, crew.
  • State: running, idle, changeover, fault, planned maintenance.

It is far cheaper to attach this at the edge or in the DMZ, where MES and the historian are close, than to reconstruct it later by joining timestamps across systems. Every reconstruction is a lineage gap.

Tag naming

You do not need to rename every PLC tag. You do need a consistent naming or modeling layer above them. Options:

  • An asset model in the historian (AF-style hierarchies or equivalent).
  • A UNS topic structure based on ISA-95 levels.
  • A mapping table maintained by a named owner, version-controlled.

Pick one, write down the convention, and enforce it for every new tag sent off-site. A convention that exists only in one engineer's head will not survive their next job change.

Cost traps

Cloud historian and data platform pricing varies by vendor, but the cost drivers are consistent:

  • Tag count. Many products price per tag or per stream.
  • Sample rate. One-second data for 20,000 tags is a very different bill from one-minute data for 2,000.
  • Query and egress. Dashboards that refresh every few seconds against raw data are expensive. So is pulling data back out for model training elsewhere.
  • Duplicate storage. Raw, cleaned and aggregated copies of the same data in three services.
  • Retention by default. "Keep everything forever" is a decision, and an expensive one.

Plant-floor reality: The tags you need for your first two AI use cases are usually a few hundred, not tens of thousands. Start there and grow on evidence.

Governance checklist before go-live

Use this before you switch on any new data flow off-site. It maps to the industrial data readiness scorecard.

Check Owner Evidence
Use case recorded in the AI use case register Business owner Register entry
Tag list scoped to the use case, with context tags OT engineer Version-controlled list
Data owner named for each tag group Plant / process owner RACI
Path drawn as a conduit in the zone model OT security lead Updated diagram
Outbound-only confirmed by firewall rule review IT/OT network Rule export
Credentials publish-only, rotated, not shared IT security Config screenshot
No write-back path exists, or one is separately approved OT security lead Risk assessment
Local buffering tested with link down OT engineer Test record
Retention set for raw and derived data Data owner Retention rule
Vendor access terms reviewed (who can see the data) Procurement / legal Contract clause
Compression and sample rate documented OT engineer Historian config

Common mistakes

  • Replicate everything first, decide later. You pay for it, and nobody can tell you which tags matter.
  • Letting the gateway vendor own the architecture. The gateway is theirs; the conduit is yours.
  • Skipping the DMZ because "it's only reading." Read-only paths become read-write paths when someone asks for a closed-loop pilot.
  • No owner for the cloud copy. IT thinks OT owns it, OT thinks IT owns it, and nobody fixes bad tags.
  • Ignoring time. Different clocks across PLCs, historian and gateway create offsets that ruin event correlation.

Monday-morning checklist

  1. List every existing data path from OT to anywhere off-site. Include vendor edge boxes and remote-access tools.
  2. For each, write down: source, path, destination, credential, owner. Gaps are your first actions.
  3. Pick the pattern (replication, gateway, broker) that fits your top AI use case and your existing historian.
  4. Draw the path through a DMZ and get it reviewed against the Purdue model.
  5. Scope the first tag list to that use case. Add context tags.
  6. Set a retention rule and a monthly cost review.

Effort: a first governed path for one site and one use case is typically weeks of part-time OT and IT effort, not months, if the historian is in reasonable shape. Cleaning up an existing undocumented setup takes longer.

What good looks like

Every tag off-site traces back to a use case and an owner. The path is drawn, reviewed and outbound-only. Context travels with the data. Costs are reviewed monthly. And when an AI vendor asks for "all your data," you can hand them a governed subset in a day instead of a firewall exception in a hurry. For the wider organizational picture, see IT/OT convergence governance.

Sources and further reading

  1. NIST SP 800-82 Rev. 3, Guide to Operational Technology (OT) Security
  2. ISA-95 Standard (ISA)
  3. OPC UA Online Reference: published specifications, OPC 10000 / IEC 62541 (OPC Foundation)
  4. MQTT Version 5.0, OASIS Standard

This guide is general information for manufacturers, not legal advice. Confirm regulatory obligations with qualified counsel.

FAQ

Frequently asked questions

Is it safe to send OT data to the cloud?

It can be, if the design is outbound-only from a DMZ, the cloud side cannot initiate connections into the control network, credentials are scoped to publish-only, and the path is documented as a conduit in your zone model. The bigger risk is usually scope creep: a read path that later gains write-back for a dashboard or AI recommendation without review.

Should we replicate our on-prem historian or replace it with a cloud historian?

For most mid-market plants, keep the on-prem historian as the system of record for operations and replicate a curated subset to the cloud. Operators need local trends during an internet outage. Replacing the local historian outright only makes sense if you have reliable connectivity, local buffering and a tested plan for when the link drops.

How many tags should we send to the cloud?

Start with the tags tied to a specific use case in your AI use case register, plus their context tags. Sending every tag 'just in case' raises cost, makes ownership impossible and buries the useful signals. Add tags when a named use case needs them, and record why.

What is the difference between an edge gateway and an MQTT broker?

An edge gateway collects data from PLCs or OPC UA servers, often normalizes and buffers it, and forwards it upstream. An MQTT broker is a message hub that routes published data to subscribers. Many deployments use both: gateways publish to a broker in the DMZ, and a bridge or connector forwards selected topics to the cloud.

Related guides

Know where your plants stand

Score your AI governance, OT security and data readiness in about 10 minutes each. Results stay in your browser.