Ada Audits the Gorilla Compression Ratio

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

foundations
math-foundations
calculation-audit
analytics-ml
beginner
Ada ADA · CALCULATION AUDIT

Ada Audits the Gorilla Compression Ratio

The chapter compresses a time series two ways: Gorilla squeezes a 16-byte point to about 1.37 bytes — a lossless 12x — while a lossy FFT turns an 8,000 B/s vibration stream into peaks for a 50x-plus cut that cannot be rebuilt. A ratio is only trustworthy after you divide the real numbers and note what survives. This audit works the Gorilla compression ratio against the FFT to show why the ratio must be read alongside reconstruction.

Companion to the chapter Time-Series Compression Audit Limits — every number here comes from that chapter.

— 16 bytes to 1.37, and where lossy buys more, ~4 minutes

A compression ratio is only trustworthy after you divide the real numbers. Gorilla's 12× is lossless; the FFT's 50× is not — and that gap is the whole audit-fidelity story.

1. One raw point is 16 bytes.

A 64-bit timestamp plus a 64-bit double. Gorilla averages 1.37 bytes per point on production data, so:

ratio = 16 ÷ 1.37 = 11.68× ≈ 12× (lossless)

2. Scale it to one hour at 1 Hz.

3600 points × 16 B = 57,600 B raw; 3600 × 1.37 = 4,932 B compressed. That saves 57,600 − 4,932 = 52,668 B, or 91.4% — and every value still reconstructs exactly.

3. Lossy FFT trades reconstruction for a bigger cut.

A 4 kHz, 2-byte vibration stream is 4000 × 2 = 8,000 B/s. Ten spectral bins at roughly 12–16 bytes each ≈ 120–160 B/s, so 8000 ÷ 160 = 50× up to 8000 ÷ 120 = 67× — but the waveform cannot be rebuilt from peaks.

Method Arithmetic Result Reconstruction
Gorilla (per point) 16 ÷ 1.37 11.68× ≈ 12× Exact
Gorilla (per hour) 57,600 → 4,932 B 91.4% saved Exact
FFT peaks (per second) 8000 → 120–160 B 50–67× Lossy

What the audit buys you: the 12× Gorilla number and the 50× FFT number are not competitors — one keeps every bit for the compliance log, the other discards the waveform for a trend chart, so the ratio must always be read alongside the reconstruction column or you will size the wrong pipe for an audit-grade record.

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