Edge telemetry at the forecourt: moving fuel stations from polling to event streams

Most forecourt supervision systems we inherit still look like something designed around a serial cable: a central poller asks each pump, each tank probe, each price sign for its current state, once a second or slower. It works when a single operator has six sites. It starts to misbehave the day a chain buys its sixtieth, and it goes from inefficient to unsafe the day the first tanker overfill alarm loses a race with the next polling cycle.

The real cost of polling

Polling has three hidden costs that are usually not visible to the people buying the system. The first is detection latency: if a pump reports a leak between polls, the alarm waits. The second is cellular data waste most polled values are unchanged, and on 4G APNs with per megabyte pricing this shows up in the budget. The third is the operator's mental model: dashboards built on polled data are averages of snapshots, not a trace of what actually happened.

What an event driven forecourt looks like

We replace the poller with a small on site broker typically MQTT over TLS 1.3, running on an industrial mini PC in the office cupboard. Every pump controller, tank gauge, ATG probe, and POS terminal publishes state changes, not state. Alarms are their own topic with QoS 2 and retained messages, so a dashboard connecting cold still sees the current alarm state. Business events transaction closed, delivery started, price updated go out the same broker to a regional cloud endpoint.

Getting there from where you are

Very few operators can stop their sites for a rip and replace, so we stage it. Step one: put the broker alongside the existing poller and mirror events. Step two: migrate alarms first they have the clearest safety case. Step three: migrate transactions and deliveries. Step four: retire the polling path. A realistic schedule is six to nine months per chain, depending on how homogenous the pump population is.

What surprised us

Two things. Operators reacted better to fewer but more meaningful dashboard updates than to constant refreshes the cognitive load dropped. And the cellular bill fell more than we modelled, because most "changes" at a quiet site overnight are noise, not information.