Platform Architecture
How EmberNet is put together, from the equipment on the floor up to the console.
The layers
┌─────────────────────────────────────────────┐
│ Cloud Control Plane │
│ Fleet Management, Tenant Orchestration │
├─────────────────────────────────────────────┤
│ Fabric │
│ Flux (L4+ overlay, primary) · ArcNet (L3)│
├─────────────────────────────────────────────┤
│ Edge Runtime Layer │
│ EmberRTOS, Container Orchestration │
├─────────────────────────────────────────────┤
│ Data & Telemetry Layer │
│ Time-Series Storage, Protocol Adapters │
├─────────────────────────────────────────────┤
│ Industrial Device Layer │
│ OPC UA · MQTT · Modbus · SNMP · BACnet │
└─────────────────────────────────────────────┘
Edge runtime
Every edge node runs EmberRTOS, our own immutable Linux distribution, with a read-only root filesystem and atomic, snapshot-backed updates. Platform services run on top of it as containers.
The immutability is the point. An edge node sitting in a plant for three years with nobody logging into it should be in a known state, and a failed update should roll back rather than leave a half-configured host that nobody can diagnose remotely.
On top of that:
- Container orchestration schedules platform services and deployed apps, restarts what fails, and reschedules off unhealthy nodes
- Service discovery connects services within a cluster without static addressing
- Health monitoring drives both restarts and what you see in Ops
Clusters are single-node for small sites, multi-node where you need workloads to survive losing a machine, and distributed where nodes across locations act as one.
Data pipeline
Data moves through five stages:
- Protocol adapters collect from devices in whatever protocol the device already speaks
- Normalization puts readings into a common shape
- Edge processing applies filtering and aggregation locally, before anything leaves the site
- Time-series storage persists it
- Cloud sync selectively forwards upward
Stage 3 is what makes the platform work on constrained links. Deciding what is worth sending at the edge is cheaper than shipping everything and deciding centrally.
Storage tiers
| Tier | Location | Retention | What it's for |
|---|---|---|---|
| Hot | Edge node | 7 days | Live dashboards, alerting |
| Warm | Edge cluster | 30 days | Historical analysis, trending |
| Cold | Cloud storage | 365 days | Compliance, long-term analytics |
These are the defaults. Hot retention is the tier that scales with deployment size; see System Requirements. Warm and cold are downsampled on ingest. Changing them is covered in Configuration Reference.
Volume-level storage is handled by Cinder, which replicates across nodes and heals automatically.
Networking
The fabric uses two layers for two different jobs.
Flux is the primary path and the application overlay at layer 4 and above. Access is granted between cryptographic identities rather than between addresses, it rides outbound TCP/443 so it survives restrictive networks, and it carries the cluster API and remote app interfaces back to your browser.
ArcNet is the encrypted layer-3 tunnel underneath it, giving nodes plain IP reachability to each other for agent check-ins and storage replication. It is faster in principle, but it needs UDP the network has to permit, so it is ranked below Flux — and at most sites it is carried over Flux rather than running beside it.
The ordering is the point. Flux is the transport that survives an industrial network, so it is the one the platform relies on; ArcNet is the optimisation you get when the network cooperates.
Both share the same properties:
- No exposed ports. Nodes dial out; nothing dials in.
- Mutually authenticated. Both ends prove who they are.
- Identity-based. Routing follows identity, not IP, so there is no flat network to scan.
- Automatic. Nodes join on enrollment and reconnect through outages and address changes on their own.
Protocols
The platform speaks the common industrial and IT protocols directly:
OPC UA, MQTT, Modbus TCP/RTU, SNMP, BACnet/IP, and REST/HTTP.
Device Monitor extends that to sixteen industrial protocols for metric collection, including EtherNet/IP, PROFINET, DNP3, S7comm, and IEC 61850. EmberBurn republishes tag data across a further set, including Sparkplug B and Kafka.
Deployment models
| Model | Who runs what |
|---|---|
| On-premises | Everything on customer hardware; no data leaves the network |
| Hybrid | Edge processing local, fleet management and long-term storage cloud-hosted |
| Managed | Fireball Industries operates the platform end to end, including hardware and updates |
Enterprises running their own clusters under the platform's control plane are covered in External Tenants.
Availability
- Failover. Workloads reschedule onto healthy nodes.
- Replication. Storage is replicated across nodes, with a default of three replicas. This is why three nodes is the recommended minimum.
- Offline resilience. Edge nodes keep collecting, storing, and alerting locally through a cloud outage, and reconcile when the link returns.
- Rolling updates. EmberRTOS updates in a maintenance window and reboots into a snapshot it can roll back from.
Next steps
- Multi-Tenancy: how tenant isolation works
- System Requirements: hardware and OS prerequisites
- Zero-Trust Networking: the security model in depth