Compute where the latency lives, store where the scale lives

On the edge side we deploy industrial PCs or SBC-class hardware running containerized workloads, so the software stack is reproducible and updatable rather than a hand-configured box nobody dares reboot. Typical edge duties include ML inference near the sensor, protocol translation, local buffering, and downsampling high-rate data before it crosses the WAN. The link upward is usually MQTT or gRPC over TLS with backpressure handling, and fleet management — remote deployment, health monitoring, staged rollouts — is designed in from the start, because version fifteen of your edge software is a certainty.

In production, the test is what happens during a WAN outage: the edge keeps operating, buffers what matters, and reconciles when the link returns, and nobody near the machine notices. We measure success in inference latency at the edge, data completeness in the cloud, and how long it takes to roll an update to fifty sites — which should be an afternoon, not a road trip.

Design starts with a workload inventory rather than a hardware catalog. For each function — inference, protocol translation, buffering, aggregation, dashboards — we record its latency requirement, its data volume, and what must keep working when the link is down, then place it at the edge or in the cloud on that evidence. The resulting integration map shows every data path, the interface at each boundary, and who owns the components on either side. Where equipment or software vendors keep their side closed, we integrate at the interfaces they document and treat their component as a black box with a defined contract.

Each boundary gets its own test before the system does: the edge stack is proven on the bench with the cloud unreachable, the transport is exercised with the link cut, throttled, and restored, and the cloud side is loaded with replayed edge data at realistic rates. Rollout is phased — one site first, watched for long enough to trust it, then the rest in waves through the fleet tooling, with rollback to the previous image at any point. Documentation records every connection, its transport, its certificate, and the named owner for the edge hardware, the network, and the cloud account.