Benchmarks¶
What a command costs, measured end to end: process start to exit, against
the published fixture in testdata/fixture/ (16 nodes, 15 edges). At this
size, process start dominates, so these numbers say how quickly a CI step
or an editor integration gets an answer, not how the query engine scales.
Method¶
make build
store=$(mktemp -d)
build/passmcp-graph --store "$store" import testdata/fixture/graph.json
hyperfine -N --runs 50 --warmup 3 \
"build/passmcp-graph --store $store query 'agents -> tools where destructive'" \
"build/passmcp-graph --store $store analyze"
The machine was running other builds during the 2026-09-29 run (load average 46), so the spread is wide; the minimum is the better guide to what the command itself costs.
Results¶
| Scenario | Median | Min | Mean ± σ | Environment |
|---|---|---|---|---|
query "agents -> tools where destructive" |
13.9 ms | 9.6 ms | 15.9 ms ± 4.9 ms | Apple A18 Pro, Go 1.27.1, 2026-09-29, load average 46 |
analyze |
20.7 ms | 10.3 ms | 26.5 ms ± 18.8 ms | same |
An earlier measurement, recorded in the 0.0.1 README as the median of 50
runs of the whole process, gave 50.6 ms for the query and 37.4 ms for
analyze (Apple A18 Pro, Go 1.27.1, 2026-09-27); its load was not
recorded.
No benchmark is run in CI; these figures are re-measured by hand when a change could move them.