faztest-fast-file-forge

Findings: the single authoritative record

Written 2026-11-19, when faztest joined the CHCE fleet. This document is the authority for any number quoted about faztest anywhere else — in the README, in a PR body, in the audit log. Where this document and another disagree, this one is right.

The reason it exists is the same one ziptastic’s docs/FINDINGS.md gives: a README line said faztest “beats xz and 7z in compression and speed”, nothing measured it, and a claim that is not tied to a table drifts from the product it describes. It is now tied to a table.


0. The machine every number here was measured on

A 2 GiB sandbox reporting nproc 2, Node v20.20.2, Brotli 1.1.0 (bundled in Node), Python 3.12. Ratio does not depend on how many cores ran and reproduces exactly; the seconds column is wall clock on a shared single core and should be read as “this order of magnitude”, not to two decimals. Every archive was decompressed and compared byte for byte before its size was recorded — 71 of 72 cells verified, and the one that is not is described in §5.

1. What faztest actually is

public/faztest.mjs is a Brotli front end: brotliCompressSync / brotliDecompressSync from node:zlib, with a quality knob (0–11, default 5), a bench mode, and a refusal to overwrite an existing output without --force. The browser path compresses in WebAssembly workers (src/lib/faztest-worker.ts), so the files never reach a server.

It ships no codec of its own. Every comparison below is therefore Brotli against the rest, not a new engine against the rest. That is worth stating plainly, because a Brotli front end competing with xz is competing with LZMA2 on xz’s home ground, and the numbers behave exactly as that framing predicts.

2. The claim, and what the measurement says

The README’s opening line was “Create faztest compressor that beats xz and 7z in compression and speed”. On the Silesia corpus:

claim measured verdict
beats xz -9e on size +16.98% larger at -q9, +3.87% larger at -q11 false
beats 7z LZMA2 on size +16.97% larger at -q9 false
beats xz -9e on speed 81 s against 241 s over the corpus (3.0× faster) true
beats ziptastic -9 on size +30.60% larger false
beats ziptastic -9 on speed 81 s against 622 s over the eleven whole-file rows (7.7× faster) true

So the honest one-line verdict is: faztest is a fast Brotli front end that loses on size to every baseline here and wins on time against all of them. At its maximum setting it still loses to xz on the corpus average — it wins on exactly two of the twelve files (osdb −0.68%, xml −1.04%) and pays 3.3× xz’s time to do it.

This is not a defect in the product; it is a defect in the sentence. Brotli at q9 is a deliberate speed/ratio trade, and it is a good one: 3× xz’s speed for 17% more bytes is a real product. The README now says that instead.

3. Method

4. The numbers

4a. Archive size, bytes

file raw faztest -q5 faztest -q9 ziptastic -9 xz -9e 7z LZMA2 9e
dickens 10,192,446 3,381,416 3,112,490 2,355,685 2,831,212 2,831,289
mozilla 51,220,480 16,473,409 16,048,786 12,353,529 13,376,240 13,376,320
mr 9,970,564 3,256,758 3,180,691 2,249,704 2,751,892 2,751,961
nci 33,553,445 2,234,365 2,005,974 1,434,131 1,449,272 1,449,336
ooffice 6,152,192 2,850,628 2,764,897 2,131,573 2,427,224 2,427,304
osdb 10,085,684 3,298,268 3,105,752 2,368,404 2,844,556 2,844,627
reymont 6,627,202 1,711,720 1,521,982 1,024,711 1,315,592 1,315,666
samba 21,606,400 4,461,158 4,197,043 3,737,179 3,739,524 3,739,600
sao 7,251,944 5,261,213 5,207,470 4,415,125 4,425,664 4,425,736
webster 41,458,703 10,600,222 9,710,099 7,158,216 8,368,672 8,368,749
x-ray 8,474,240 5,503,290 5,347,533 3,798,464 4,491,264 4,491,339
xml 5,345,280 520,795 479,494 374,814 434,892 434,957
total 211,938,580 59,553,242 56,682,211 43,401,535 48,456,004 48,456,884
ratio of raw 100% 28.10% 26.74% 20.48% 22.86% 22.86%

4b. faztest -q9 against both baselines, per file

file -q9 vs xz -9e -q9 vs ziptastic -9 ziptastic -9 vs xz -9e
dickens +9.93% +32.13% -16.80%
mozilla +19.98% +29.91% -7.65%
mr +15.58% +41.38% -18.25%
nci +38.41% +39.87% -1.04%
ooffice +13.91% +29.71% -12.18%
osdb +9.18% +31.13% -16.74%
reymont +15.69% +48.53% -22.11%
samba +12.23% +12.31% -0.06%
sao +17.67% +17.95% -0.24%
webster +16.03% +35.65% -14.46%
x-ray +19.07% +40.78% -15.43%
xml +10.26% +27.93% -13.81%
total +16.98% +30.60% -10.43%

4c. faztest at its maximum setting, -q11

The default is -q5 and the CLI’s own bench uses -q5; -q9 is the setting anyone reaching for “best” would pick. -q11 is the ceiling, so it is measured too — otherwise “faztest loses to xz” could be answered with “you did not use the best setting”.

file faztest -q11 -q11 vs xz -9e -q11 vs ziptastic -9
dickens 2,853,158 +0.78% +21.12%
mozilla 14,077,215 +5.24% +13.95%
mr 2,825,713 +2.68% +25.60%
nci 1,616,326 +11.53% +12.70%
ooffice 2,479,758 +2.16% +16.33%
osdb 2,825,123 -0.68% +19.28%
reymont 1,335,805 +1.54% +30.36%
samba 3,823,126 +2.24% +2.30%
sao 4,589,967 +3.71% +3.96%
webster 8,788,708 +5.02% +22.78%
x-ray 4,687,298 +4.36% +23.40%
xml 430,390 -1.04% +14.83%
total 50,332,587 +3.87% +15.97%

Total 50,332,587 bytes — 23.75% of raw — in 796 s, against xz’s 241 s. It is both larger and slower than xz -9e on the corpus average, which is the whole of the “use the best setting” answer.

4d. Wall clock

codec seconds, whole corpus slowest file
faztest -q5 23.0 mozilla 5.3 s
faztest -q9 81.3 mozilla 21.6 s
ziptastic -9 622.4 nci 166.4 s
xz -9e 241.2 webster 59.9 s
7z LZMA2 9e 243.6 webster 63.6 s
faztest -q11 796.4 mozilla 243.5 s

4e. The corpus

file bytes class
dickens 10,192,446 English text
mozilla 51,220,480 executable
mr 9,970,564 medical image
nci 33,553,445 chemical text
ooffice 6,152,192 office binary
osdb 10,085,684 database
reymont 6,627,202 Polish text
samba 21,606,400 source code
sao 7,251,944 binary
webster 41,458,703 English text
x-ray 8,474,240 medical image
xml 5,345,280 XML
total 211,938,580 twelve files

5. What these numbers do not say

  1. mozilla is measured as two halves for ziptastic only. A whole-file ziptastic -9 on 51,220,480 bytes did not finish inside a 300-second sandbox step in four attempts (its own scripts/silesia.sh warns that mozilla alone does not fit). It is therefore measured as two 25,610,240-byte halves, compressed and verified separately and summed: 12,353,529. To keep the comparison like-for-like, xz -9e on the same two halves is 13,507,344, so the split is applied to both sides before any delta is computed. Every other cell is a whole file.
  2. The 7z column is LZMA2, not 7-Zip. The 7-Zip binary is not installable in this sandbox (the vendor host is refused by the network policy), so the column is LZMA2 at preset 9 | extreme in a 7z container written by py7zr. It lands within 0.01% of the xz column on every file, which is what the same LZMA2 core should do; treat it as “7z-class ratio”, not as a measurement of 7-Zip’s own encoder.
  3. The xz column is liblzma, not the xz CLI — lzma.compress(preset=9 | PRESET_EXTREME), the same library the CLI calls. There is no xz binary in this sandbox.
  4. Times include process startup (~40 ms of Node per invocation). On a 5 MB file that is noise; on a 10 KB file it is the whole measurement, which is why no small file is in the table.
  5. Brotli’s window is bounded at 16 MiB by the format. Files above that cannot see the whole input, which is part of why the two largest files (mozilla, webster) show the widest gaps. That is a property of the format, not a tuning mistake.
  6. This is one corpus. Silesia is text- and binary-heavy and contains no images or video. On already-compressed content every codec here declines; the fleet’s own PNG/BMP work (ziptastic docs/FINDINGS.md §6i) is where that is measured.

6. Reproduce it

git clone --depth 1 https://github.com/MiloszKrajewski/SilesiaCorpus.git /scratch/work/silesia
cd /scratch/work/silesia && mkdir -p corpus && for z in *.zip; do unzip -qo "$z" -d corpus; done
python3 scripts/bench.py            # writes bench.csv, resumes where it stopped

7. Open questions