Atlas — Live Fleet Dashboard
A real-time operations dashboard streaming vehicle telemetry onto a live map, with geofence alerts and route replay.
320
Vehicles streamed
<2s
Update latency
60fps
Map render
The build
Fleet operators usually learn about a late delivery the next morning, from a report. Atlas is the alternative: a dashboard where a dispatcher sees a delay while it is still recoverable. We built the full path — a telemetry ingest gateway, a streaming layer, a live map, geofence evaluation and historical route replay — and load-tested it at 320 concurrent vehicles.
This project was delivered under our Web Development service. Modern, scalable and high-performance web solutions.
- Scope
- Full build — ingest, backend, dashboard
- Duration
- 6 weeks
- Stack
- React · Node.js · WebSockets · Redis · TimescaleDB · Mapbox GL
What made this hard
Dispatch teams run on batch reports. A vehicle idles for forty minutes, the SLA is already blown by the time anyone sees a number, and the customer calls before the operator knows. The hard part is not drawing dots on a map — it is keeping hundreds of positions current without melting the browser, and turning raw GPS into something a dispatcher can act on in the ten seconds of attention they actually have.
How we built it
- 1
Ingest before interface
We started at the wrong end on purpose. A gateway accepts position reports over HTTP and MQTT, normalises them, and writes to a time-series store. Getting the data model right first meant the dashboard never had to compensate for a messy feed.
- 2
Stream, do not poll
Positions fan out over a Redis pub/sub channel to a WebSocket layer. Each connected dashboard subscribes only to the vehicles in its current map viewport, so panning changes the subscription rather than the payload size.
- 3
Make the map cheap to draw
Vehicles render as a single Mapbox GL symbol layer fed by one GeoJSON source, updated in place. One source swap per tick instead of hundreds of marker objects is the difference between 60fps and a stuttering tab.
- 4
Turn position into meaning
Geofences are evaluated server-side on every report, not in the browser. Entering or leaving a zone raises an event, and dwell time past a threshold escalates it. The dispatcher sees "Vehicle 114 stationary 38 min inside Depot B", not a dot that has not moved.
How the pieces fit
Devices → ingest gateway (Node) → Redis pub/sub for live fan-out, TimescaleDB for history → WebSocket server with per-viewport subscriptions → React dashboard rendering a single Mapbox GL source. History and live share one query interface, so route replay is the same component as the live view with a different time window.
Decisions we had to defend
WebSockets over polling
Polling 320 vehicles at a useful interval is thousands of redundant requests a minute. A persistent socket per dashboard costs one connection and pushes only what changed.
Viewport-scoped subscriptions
The obvious build pushes every vehicle to every client. That works at 30 vehicles and collapses at 300. Scoping to the visible map area keeps payloads flat as the fleet grows.
Downsample history, keep raw hot
Full-resolution GPS for a year is expensive and nobody replays a route at one-second fidelity. Recent data stays raw; older data is downsampled on a schedule. Replay stays smooth, storage stays sane.
Where it ended up
The build sustains 320 simulated vehicles reporting every five seconds with sub-two-second end-to-end latency and no dropped frames on the map. Route replay scrubs a full shift smoothly. It is the reference we point to for any real-time operations interface — the same shape works for field service, plant telemetry or delivery tracking.