DevOps for Manufacturing & Industry 4.0 platforms

What regulates the infrastructure

What actually goes wrong here

• Precautionary OT shutdown following an IT-side compromise. The costly decision is usually the containment action, not the malware: when IT and OT are not demonstrably segmented, the safe response is to stop the line, and in continuous processes the restart of a furnace, paint shop or reactor is measured in days.
• Unpatchable installed base. PLCs, HMIs and drives commonly run for fifteen to thirty years on operating systems long out of support, and cannot be patched inside a production window, so compensating controls (segmentation, monitoring, strict conduit control) carry the entire security burden — a constraint that has no analogue in an all-IT estate.
• MES or ERP dependency stopping healthy machines. Where the line cannot issue a work order, print a compliant label or serialise a part without the system, the equipment sits idle even though nothing mechanical has failed — a failure mode invisible in equipment-availability metrics.
• Tier-n supplier IT outage halting the customer under just-in-time and just-in-sequence supply. Buffers are deliberately small, so a supplier that cannot transmit or receive schedules stops the customer's line within hours; the 2022 compromise of Toyota supplier Kojima Industries halted output across the customer's domestic plants.
• Historian and traceability record loss creating an evidential gap. Where genealogy data for a period cannot be reconstructed, the affected material may have to be scrapped or a recall scoped conservatively wide, because the record — not the physical product — is what proves conformity.
• Flat, unsegmented plant networks allowing lateral movement from a single engineering workstation or remote-maintenance connection into shared control networks across multiple lines or sites.

How demand behaves

Baseline load follows shift patterns — commonly two or three shifts, with weekend and annual shutdown troughs — but machine and sensor telemetry is continuous and does not stop when the office does. Data volume steps up sharply at model or product changeover, during commissioning of a new line, and at period-end when quality and traceability records are consolidated for release. Unlike consumer-facing sectors, the peak is not a demand spike but a production event.

Data you will be holding

The crown jewels are trade secrets rather than personal data: process recipes and setpoints, CAD/PLM designs, tooling parameters and yield data. Some product lines carry export-control (dual-use) obligations on technical data. Quality and traceability records are separately sensitive because they are the evidentiary basis for batch release and recall scoping.

Architecture this pushes you toward

The prevailing reference model is the Purdue/ISA-95 hierarchy, with control systems at levels 0-2, MES at level 3 and enterprise systems at level 4, separated by a DMZ. Design consequences follow directly: control traffic stays on-premises and deterministic, historians and MES sit at the boundary, and only aggregated or replicated data crosses into cloud analytics — typically one-way. Remote vendor access to machinery is a recurring architectural weak point because it is operationally necessary and often implemented per-vendor rather than through a single brokered path.

Availability expectation

The sector generally does not express this as a headline availability percentage. Tolerance is measured instead as unplanned downtime cost per minute of a given line and as OEE impact, which varies by orders of magnitude between a discrete assembly cell and a continuous process (a furnace, paint shop or reactor) where an unplanned stop imposes hours or days of restart regardless of how quickly IT is restored.