Fleet telemetry that still makes sense at 10,000 devices
Fleet telemetry breaks quietly. Nothing fails; the dashboards simply stop answering questions, because the schema was designed for a hundred devices and nobody revisited it at ten thousand.
Schema is a contract with your future self
Every payload carries a version. Without it, the day you add a field you have two incompatible populations in the same table and no way to tell them apart. With it, migration is a query. This is thirty minutes of work on the first device and weeks of work on the ten-thousandth.
Thirty minutes of work on the first device; weeks of work on the ten-thousandth.
Sample on change, not on a timer
Fixed-interval sampling produces mostly redundant data and still misses the transient that mattered. Reporting on change, with a heartbeat to prove liveness, cuts ingest volume dramatically and improves the resolution of the events you actually care about.
Retention is a per-field decision
Raw high-rate signals rarely need to live past a fortnight; aggregates and state transitions often need years. Treating the whole payload as one retention class means paying long-term storage rates for short-term data, or throwing away the record you needed for a warranty claim.
