Ada Audits the Data-Rate and Latency Budgets
Ada re-derives this chapter’s own numbers step by step, at full precision
ADA · CALCULATION AUDIT
Ada Audits the Data-Rate and Latency Budgets
One thousand vibration sensors streaming 10,000 samples a second at 2 bytes each would push 1.728 TB/day of raw waveform, while a 20-feature summary collapses that to 6.912 GB/day. On the control side, an edge decision totals 25 ms against a cloud path of 235 ms. This audit re-derives both budgets and asks whether the 250× reduction and that 235 ms-versus-50 ms deadline really force the decision to the edge rather than the cloud.
Companion to the chapter Edge Processing for Big Data — every number here comes from that chapter.
Edge processing is justified by two calculations: how many bytes the raw stream would cost, and whether the decision path can meet its deadline. Both worked examples in this chapter supply the inputs; I only carry the arithmetic through.
1. Raw rate is sensors × sample rate × bytes.
One thousand vibration sensors at 10,000 samples/s and 2 bytes each, then multiplied out over a full day:
2. Features collapse each sensor-second to a handful of bytes.
Twenty features × 4 bytes = 80 bytes/s per sensor, so the fleet and the reduction factor follow:
3. Latency is additive along a path.
Sum each stage, then compare against the 50 ms machine-protection deadline:
| Design question | Arithmetic shown | Audit result |
| Raw data rate (sensors × rate × bytes) | 1,000 × 10,000 × 2 | 20 MB/s |
| Raw daily volume | 20 × 86,400 | 1.728 TB/day |
| Feature stream rate (1,000 × 80 B/s) | 1,000 × 80 | 80 KB/s |
| Feature daily volume | 80 × 86,400 | 6.912 GB/day |
| Reduction factor | 1,728 / 6.912 | 250× smaller |
| Edge decision latency | 12 + 8 + 5 | 25 ms — meets 50 ms |
| Cloud decision latency | 80 + 45 + 20 + 90 | 235 ms |
| Cloud vs 50 ms deadline | 235 / 50 | 4.7× too slow |
What this means for your design: the 250× reduction is why you forward RMS, peak, and spectral-band features instead of raw waveforms — 6.9 GB/day is a manageable uplink while 1.7 TB/day is not — but keep a short raw window on the gateway so the feature evidence stays auditable. And the latency sum settles the control question outright: a 50 ms machine-protection action cannot wait on a 235 ms cloud round-trip that is 4.7× over budget, so the edge acts locally and ships the decision, model version, and raw-window pointer upstream for audit and learning.
Every number above is taken from the chapter’s own examples and re-derived step by step.