in 𝕏
Glowing blue digital chain links with particle effects on a dark background, conveying connectivity and technology

Your Source for Edge Computing News, Events & Careers

The Edge Computing Association brings together industry professionals with curated news, career opportunities, and community resources.

Designing edge architectures for high-availability industrial systems

Industrial availability is often discussed as a percentage of uptime, but operators experience it as a process that keeps running when a link, controller, power feed or application fails. Designing edge architectures for high-availability industrial systems therefore begins with the consequences of disruption: a stopped conveyor, a lost production batch, an unsafe interlock or a control room working from stale information. The edge must preserve safe operation and useful production data even when the wider enterprise platform is unavailable.

A resilient design separates safety-critical control, supervisory control, edge analytics and cloud services into explicit layers. It defines which functions continue locally, how state moves between redundant nodes, and which recovery targets apply to each workflow. This avoids treating every application as equally urgent and makes failover repeatable enough for operators, engineers and automated orchestration to trust.

Availability starts at the plant floor

At a brownfield refinery, mining process plant or food factory, new edge infrastructure usually has to coexist with controllers, instruments and networks that have operated for years. Replacing every device can be slow and risky, so gateways should support established industrial protocols and isolate newer compute services from legacy segments. Resilient functions belong close to the process, while the control system remains independent of cloud connectivity.

Australian conditions place particular demands on these designs. Remote Pilbara mining operations combine autonomous haulage, ore processing and communications across vast areas, making local control and store-and-forward data essential. Coastal Queensland’s LNG and coal seam gas facilities must cope with heat, humidity, corrosion and cyclone exposure. Edge equipment needs suitably rated enclosures, thermal management and protected power supplies, while redundant communications paths need to span both terrestrial links and local radio coverage.

Melbourne’s advanced manufacturing sector adds another reality: production lines often combine newer digital systems with long-lived programmable logic controllers. Sydney water networks and metropolitan infrastructure similarly need remote assets to report alarms and preserve local commands during back-office outages. Availability targets should therefore be expressed around critical functions such as safe shutdown, dosage control, pressure regulation and alarm acknowledgement, each with its own recovery time objective and recovery point objective.

Redundancy must include state and recovery

Two edge servers provide little protection if they share a power supply, switch, storage system or faulty software path. Failure domains must be independent across power, cooling, networking, hardware and management planes. Active-passive pairs can provide a clear recovery model, while active-active nodes can reduce switching delays, but concurrent processing requires careful state synchronisation, fencing and conflict handling. Otherwise, a network partition can create split-brain behaviour in which two nodes believe they control the same asset.

Data consistency matters as much as server duplication. Industrial commands should carry identifiers, timestamps and sequence information so an edge node can reject stale or repeated messages. Checkpoints, local historians and replayable event logs help reconstruct the state of a process after a crash. For non-idempotent operations, the application must determine whether a partially completed transaction can safely resume, roll back or enter a controlled hold state.

Safety-critical decisions should remain within certified or appropriately engineered control layers. Edge analytics and artificial intelligence can improve condition monitoring, predict equipment faults or recommend process adjustments, yet they should not sit in the final decision path for a protective interlock unless the safety case explicitly permits it. A useful architecture degrades predictably: safety functions continue, supervisory tools operate with cached data, analytics pause, and enterprise services recover later.

Networks need local intelligence

Large distances between Australian control centres, mines, processing plants and support depots make the wide-area network an unreliable foundation for time-critical control. Traffic should prioritise established safety and process functions over bulk file transfers, telemetry uploads or software updates. Each site needs a local policy for bandwidth saturation, with queues and rate limits that protect control traffic when the backhaul becomes congested. Redundant paths alone are insufficient if they share a trench, mast, fibre route or upstream service.

Deterministic Ethernet can support predictable delivery for local motion and machine-control workloads where the selected equipment supports it. Time-sensitive networking features may improve scheduling, scheduling latency and bandwidth reservation, although operators must test them with real industrial traffic rather than idealised traffic models. Private LTE or 5G can support mobile equipment and remote sites, but a cellular bearer still needs resilient edge processing, authentication and fallback behaviour. Fibre, licensed radio and microwave links may be combined where their failure domains differ.

Edge nodes should keep a bounded local buffer and continue essential work during disconnection. Store-and-forward gateways can retain events, commands and historian records until connectivity returns, using backpressure to prevent unlimited memory growth. Accurate time is also essential for sequencing, event correlation and forensic analysis; a combination of precision time protocol and GNSS references may suit isolated sites. When time quality deteriorates, applications need to flag uncertainty rather than silently issue commands using unreliable timestamps.

Security failures become availability failures

A compromised gateway, expired certificate or obstructed software update can stop an industrial process as effectively as a hardware failure. Operational technology should therefore use least-privilege access, strict segmentation between control and business networks, allow-listed communication and authenticated device identity. Security monitoring must be capable of distinguishing malicious traffic from a faulty controller, particularly where older equipment cannot support modern authentication methods. Passive inspection and carefully governed protocol gateways can bridge these gaps without exposing the entire plant network.

The software supply chain needs the same resilience discipline as the hardware design. Secure boot, signed firmware, protected configurations and recoverable update partitions can prevent a bad image from becoming a permanent outage. Remote nodes should support staged rollouts, health verification and rollback, with a documented procedure for sites that cannot receive an update immediately. A controller that is temporarily offline should remain in a safe, known state instead of being repeatedly restarted or remotely reconfigured.

Observability should combine machine health, network performance and process context. Operators need clear indications of which redundant path is active, when failover occurred, whether queues are approaching capacity and which data may have been lost. Logs should be protected from unauthorised alteration and kept close enough to the asset for diagnosis during a WAN outage. Before sharing identifiable operational or community information through association resources, readers should review the site privacy policy and apply relevant data-handling requirements.

Testing recovery under industrial reality

A failover demonstrated under ideal conditions may still fail during a plant event. Acceptance testing should include process interruption, switch failure, WAN loss, power loss, storage exhaustion, clock drift, certificate expiry and application crash. Each scenario needs a measurable expected result, including the maximum interruption, data loss, operator workload and restoration method. Tests that touch live equipment must use approved procedures, coordinated isolation and safe operating states.

Test environments should reproduce the constraints of the real installation, including legacy protocols, constrained bandwidth, environmental conditions and imperfect documentation. Engineers can capture normal traffic during commissioning and replay selected fault patterns in a representative testbed. Runbooks should then be exercised by the people who will use them during an incident, including control room staff, site technicians, network specialists and cybersecurity responders. Recovery procedures need clear triggers because hesitation during a partial outage can cause more damage than the original fault.

The Australian market strongly influences how this capability is delivered. Operators commonly coordinate original equipment manufacturers, local engineering firms, systems integrators and specialist technicians, with each party controlling part of the lifecycle. Imported components may have extended lead times, making spare parts, compatible alternatives and local repair arrangements important for remote operations. Vendors that support the full operating lifecycle are often preferred over suppliers offering hardware alone, particularly where on-site travel is costly and every intervention window must be planned.

Governance must evolve with the architecture. Edge inventories, ownership, software baselines, certificate expiry dates, recovery targets and test evidence should be maintained as operational data. Relevant functional-safety and industrial cybersecurity standards, including IEC 61511 and IEC 62443 where applicable, should be incorporated into design assurance rather than applied after commissioning. Australian teams may call a result a “fair dinkum result”, but the operational meaning should be exact: the system changed state as designed, recovered within the agreed window and produced an audit trail that supports the next decision.

Begin with a focused availability workshop involving operations, control engineering, networking, cybersecurity and maintenance. Select one critical workflow, map its failure modes and recovery dependencies, then define the recovery time and acceptable data loss for each step. Use the Edge Computing Association’s events and professional network to bring the relevant industrial, technology and supplier perspectives into the same room.

Turn that workshop into a small pilot at a representative site, whether that is a Pilbara processing facility, a Queensland energy installation, a Melbourne production line or a remote utility asset. Inject realistic faults, measure recovery, document operator actions and use the evidence to refine the platform. Make the first production rollout from measured behaviour rather than assumptions on a slide.

Industry Events & Highlights

Oct 2021
IDC FutureScape: IT Advances for 2022 and Beyond
Industry Report
Oct 2021
IBM Announces AI, Cloud & Edge Collaboration Deals at MWC LA
Los Angeles
Sep 2021
Edge AI Summit 2021
Industry Conference
Jul 2021
Edgetech Podcast: Cloudflare COO Michelle Zatlyn
Podcast Episode
May 2021
Victor Ai's Blueprint for Smart Cities
Featured Content
Apr 2021
Edgetech Podcast: Qnext Corp CEO Anthony Decristofaro
Podcast Episode

Stay Informed

Subscribe to our bi-weekly newsletter for the latest edge computing news, events, and career opportunities.