BENCHMARKS · PERFORMANCE

Performance measurement

Fast anchor scanning does not imply that complete morphological searches run as fast as ripgrep. kfind also pays for query compilation, resource initialization, and candidate verification. Read execution time together with search quality in the search-strategy comparison.

Workloads

Morphology measures fresh-process initialization, cases per second, p95, and RSS. Query compile, matcher scan, 1 GiB file scan, npm startup, and TUI indexing have separate entrypoints.

The regex baseline runs each of seven queries against the same single file in a fresh process. It reports batch wall time separately for kfind and each rg or grep combination.

Lock wait and build time are excluded from workload time. They remain part of the recorded environment and revision.

Statistics

Morphology backends use one warm-up and five measurements; the regex baseline uses two warm-ups and ten measurements. Reports include the median, minimum, maximum, and p95. A single run or successful smoke test is not performance evidence.

Throughput, latency, seconds, and bytes are not collapsed into one percentage score.

Regression decision

Baseline and candidate use identical toolchains, resources, input checksums, and options. A quality projection prevents adoption of a faster but incorrect candidate.

Unfavorable changes and noise ranges remain visible, and unknown causes are labeled as inference.