nobody 2ab1c31993

feat: transparently append to compressed archives

When appending to a compressed archive (gzip, brotli, zstd), the tool
now handles compression automatically. Since some compression formats don't
support appending to compressed files in place, we write a new
compressed file with all the data and atomically rename it to replace
the original (assuming there is enough space on that filesystem).

This means you can work with compressed archives the same way as
uncompressed ones. Point the tool at your .json.gz file and append
values. No manual decompression/recompression needed.

2025-11-30 17:09:49 -08:00

1.8 KiB

Raw Permalink Blame History

Fuzz Testing

Fuzz testing throws random inputs at your code until something breaks.

Commands

List available fuzz targets:

cargo fuzz list

Run a fuzz target:

cargo fuzz run fuzz_apply_move

Runs until you kill it or it finds a crash.

Reading the Output

#787958    REDUCE cov: 1281 ft: 6423 corp: 1112/621Kb lim: 4096 exec/s: 13823 rss: 584Mb L: 19/3954 MS: 1 EraseBytes-
#788755    REDUCE cov: 1281 ft: 6424 corp: 1113/621Kb lim: 4096 exec/s: 13837 rss: 584Mb L: 767/3954 MS: 2 CMP-CrossOver- DE: "6\000\000\000"-
#789383    REDUCE cov: 1281 ft: 6424 corp: 1113/621Kb lim: 4096 exec/s: 13848 rss: 584Mb L: 59/3954 MS: 3 InsertByte-ShuffleBytes-EraseBytes-

The fields:

#787958 — test case number. How many inputs have been tried.
REDUCE — what happened. NEW means new code was reached. REDUCE means an input was shrunk while keeping the same coverage. pulse is just a heartbeat.
cov: 1281 — coverage. Number of code edges hit. This is what you care about.
ft: 6423 — features. Finer-grained coverage metric. Ignore it.
corp: 1112/621Kb — corpus. 1112 interesting inputs saved, 621KB total.
exec/s: 13823 — speed. Test cases per second.
rss: 584Mb — memory use.
L: 19/3954 — input length. This one was 19 bytes. Largest in corpus is 3954.
MS: 1 EraseBytes- — mutation. How the input was generated. Doesn't matter.

Is It Working?

Watch cov. If it goes up, the fuzzer is finding new code paths. If it stops going up, either you have good coverage or the fuzzer is stuck.

exec/s in the thousands is fine. If it drops to double digits, something is wrong.

Seeing NEW events means progress. Long stretches without NEW means diminishing returns.

When to Stop

When cov stops increasing and you're bored. Hours for a quick check, days for thoroughness.

1.8 KiB Raw Permalink Blame History