An 8x Backhaul Gap No Schedule Can Close

Ada re-derives this chapter’s own numbers step by step, at full precision

foundations
math-foundations
calculation-audit
fundamentals
Ada ADA · CALCULATION AUDIT

An 8x Backhaul Gap No Schedule Can Close

A factory streams 500 vibration sensors at 1 kHz, 4 bytes a sample — 16 Mbps of raw data — over a cellular backhaul of only 2 Mbps, an 8x oversubscription. One day piles up 172.8 GB against a link that can carry just 21.6 GB, so a single day’s batch upload takes about 8 days. This audit asks whether any smarter schedule or bigger buffer can close that gap, or whether the architecture must discard roughly 87.5% of the data at the edge.

Companion to the chapter IoT Architecture Planner — every number here comes from that chapter.

Ada: The factory-sensor question above rests on a few numbers — 16 Mbps of raw data against a 2 Mbps link, and a batch upload that “takes nearly 8 days.” Let me verify them and then show why edge processing is not optional here.

The raw rate:

  • 500 sensors x 1000 samples/s x 4 bytes = 2,000,000 bytes/s = 2 MB/s
  • In bits: 2,000,000 x 8 = 16,000,000 bit/s = 16 Mbps

Against a 2 Mbps backhaul, the oversubscription is 16 / 2 = 8x. That factor is the whole problem, and batching cannot dodge it. One day generates 2 MB/s x 86,400 s = 172.8 GB, while the link can carry only (2 / 8) MB/s x 86,400 = 21,600 MB = 21.6 GB/day. The ratio is the same 172.8 / 21.6 = 8, so uploading a single day’s data takes 172.8 / 21.6 = 8 days — and every real day adds another 172.8 - 21.6 = 151.2 GB to a backlog that can only grow.

So the only way through is to shrink the data at the source. To fit 16 Mbps into 2 Mbps the edge must discard (16 - 2) / 16 = 87.5% of the volume before it reaches the link — which is exactly why the answer cites a 90-99% reduction, not a modest trim.

The design meaning is that an 8x capacity gap is a structural fact, not a tuning knob: no smarter schedule, off-peak window, or larger buffer can move 172.8 GB/day through a 21.6 GB/day pipe, so the architecture either reduces data at the edge by roughly seven-eighths or it fails.

Every number above is taken from the chapter’s own material and re-derived step by step.