A connected-vehicle telemetry pipeline's cost is growing faster than the fleet. Where does the cost go, and what reduces it?

Where the cost goes

• Volume at ingestion, which for fleet telemetry is dominated by messages carrying almost no information — a stationary vehicle reporting identical coordinates every few seconds.
• Retention in the broker, which is storage plus the replication factor.
• Downstream processing, proportional to volume.
• Storage of the raw stream, retained for reprocessing, which grows monotonically.
• The serving store, where high-cardinality per-device state is expensive.

What reduces it, in order of leverage

• Edge filtering. The device transmits on meaningful change — moved beyond a threshold, changed heading, changed state — plus a low-frequency heartbeat. A very large reduction achieved at the cheapest possible point, before anything is transmitted, stored or processed.
• Aggregation before the expensive stages. Most consumers need "where is this vehicle now" and "what route did it take", neither of which requires every point at full fidelity in real time.
• Separating paths by consequence. Live tracking wants the latest value and can drop the rest — it should conflate. Route reconstruction wants completeness and tolerates minutes. Forcing both through one pipeline pays the strictest requirement of each, which is the expensive combination.
• Retention tiering: short retention at full fidelity in the broker, longer retention downsampled, archival compressed in object storage.
• Binary encoding, since serialisation dominates CPU at this volume and text formats cost several times more — and it reduces bandwidth on a metered mobile connection, which is a device-side cost too.

The conflation insight

Where updates supersede one another, keeping only the latest value per device bounds the work by fleet size rather than by message rate. That converts an unbounded problem into a bounded one, and it is the correct overload behaviour for the live path.

It is categorically unavailable to anything where each event matters — a fault code, a trip completion, a firmware update acknowledgement — which is exactly why those belong on a different transport.

The unit metric

Cost per vehicle per day, trended. Total cost rises with the fleet and tells you nothing; unit cost falling as the fleet grows is the only evidence the efficiency work is compounding.