loader-benchmark

Overview ¶

Command loader-benchmark measures one codescan scan and reports what it cost.

It exists because the in-repo benchmarks cannot answer "how much did we improve": they live inside the module, so they do not exist in any earlier release. This one talks only to the public codescan.Run API, which every version has, so the same source compiles against a released tag and against the working tree — and run.sh builds one probe per version to put them side by side.

One scan per process, deliberately. Peak RSS is a process-wide high-water mark, so a second scan in the same process would report the first one's peak plus its own garbage.

Usage is through run.sh; see ../README.md. Running this directly measures a single cell:

loader-benchmark -dir /path/to/corpus -corpus mycorpus -label current
loader-benchmark -summarize results/warm.jsonl -baseline-label v0.33.3