System Design for Industrial IoT
A tour through the architecture patterns I reach for when the sensor count jumps from a hundred to a hundred thousand.
Starting from the sensor
Every IIoT system I have designed begins with the same question: how noisy is the source? A pressure transducer sampling at 1 kHz is a very different animal from a temperature probe that wakes up every 30 seconds. The downstream architecture follows from that one number.
The pattern I keep reaching for
- Edge aggregation — a lightweight gateway buffers and batches.
- Broker — Kafka or RabbitMQ depending on replay needs.
- Stream processor — Flink or a Node worker for enrichment.
- Hot store — time-series database (TimescaleDB or Influx).
- Cold store — object storage for raw replay.
Where it breaks
It breaks whenever someone skips step 1. An enrichment service that consumes raw sensor traffic will flatline the broker the first time a line goes into fault and everything starts screaming at once. Push the debouncing to the edge and the rest of the stack stays calm.
Design for the worst minute of the year, not the average one. That is where plant trust is won or lost.
