I spent years pushing everything to the cloud. Sensors, PLCs, vibration monitors — you name it, we streamed it up to some data center. It worked fine in demos. Then I stood on a factory floor watching a robotic arm miss a quality check by three seconds because the data had to round-trip through a cloud region two hundred miles away. That's when edge computing stopped being a buzzword and became something I actually needed.
The first mistake I made was thinking edge meant less infrastructure. It doesn't. You still need servers, storage, and networking — just closer to where the action happens. We started with a single industrial PC bolted to a DIN rail near the assembly line. It ran a lightweight inference model for detecting weld defects. The model was garbage at first. We'd trained it on pristine lab data, not the dusty, vibrating, temperature-swung reality of a real plant. Retraining it on edge-collected data took weeks of iterative failures.
Another thing nobody warns you about is the mess of device protocols. Modbus, Profinet, OPC-UA, MQTT — they all speak different dialects. We spent more time writing protocol translators than we did on the actual analytics logic. Eventually we settled on a thin middleware layer that normalized everything into a common schema before it ever touched the edge processor. That abstraction saved us later when the plant added a new conveyor system with its own proprietary interface.
Connectivity is the hidden tax. In an office, you assume WiFi works. On a factory floor, you've got metal racks, motor drives, and welding arcs that kill wireless signals like they're swatting flies. We ran CAT6 everywhere, shielded, in conduit. Even then, a forklift driver backed into a junction box and took out half the network for an afternoon. We learned to design for physical abuse — redundant paths, armored cables, and edge nodes that could buffer hours of data if the uplink died.
The real win came when we stopped treating edge as a cloud substitute and started treating it as a partner. The edge handles real-time decisions: shut down a press if the vibration exceeds threshold, adjust coolant flow when temperature spikes. The cloud handles the long view: trend analysis, fleet-wide model updates, capacity planning. That split cut our cloud bandwidth costs by sixty percent and dropped our reaction time from seconds to milliseconds.
I still see architects try to cram everything into the cloud because it's simpler to manage. But simplicity at the design phase often means complexity in operations. The best edge deployments I've seen are the ones where the plant floor engineers own the edge nodes and the central IT team owns the cloud backend. They meet in the middle, and they argue about latency budgets like it's a blood sport.
If you're starting an IIoT project today, start by walking the floor. Find the one machine where a five-second delay causes a scrap part. That's your edge candidate. Everything else can wait.