benchmarks.md

Benchmarks

Numbers come from node bench/bench.ts. CI runs it on every push and writes the result to the job summary, so Linux numbers can be found in each CI run.

Synthetic graph: a layered DAG with libs, services and apps, where each project depends on up to 3 lower projects (dense on purpose). "plan" covers file ownership, the config diff, graph closure and sorting.

Windows 11 (local dev machine)

In-process (best of 5, ms)

projects	changed files	parse+validate	graph build	plan	affected
10	10	0.25	0.02	0.07	10
100	100	0.91	0.16	0.48	83
500	1000	2.43	0.31	2.23	485
1000	10000	4.66	0.54	12.7	1000
5000	100000	21.2	3.38	121.3	4951
10000	100000	35.1	6.80	141.7	9529

End to end (bundled action, real git repo, push event; best of 5)

Bundle: dist/index.js = 71255 bytes. Node startup baseline (node -e 0): 74.5 ms

projects	changed files	total wall time (ms)
10	5	516.2
100	50	538.0
1000	1000	1048.5
5000	10000	672.4

Platform: win32/x64, Node v24.21.0

Note: on Windows, each git process spawn costs about 50 ms, which dominates the end-to-end times. Linux runners spawn processes far faster; see the Linux table below.

Linux: GitHub-hosted ubuntu-latest runner (CI run 37253383631)

Bundle: dist/index.js = 71,255 bytes. Node startup baseline (node -e 0): 20.5 ms. Node v24.21.0, linux/x64.

In-process (best of 5, ms)

projects	changed files	parse+validate	graph build	plan	affected
10	10	0.29	0.01	0.07	10
100	100	1.08	0.20	0.63	83
500	1000	3.48	0.39	2.35	485
1000	10000	4.04	0.55	13.9	1000
5000	100000	19.6	4.14	120.8	4951
10000	100000	32.5	9.44	143.7	9529

End to end (bundled action, real git repo, push event; best of 5)

projects	changed files	total wall time (ms)
10	5	50.2
100	50	57.2
1000	1000	105.2
5000	10000	232.9

Conclusion: at 1,000 projects the whole action, including Node startup and git, finishes in about 0.1 s on a hosted runner. A compiled (Rust) binary could save at most the ~20 ms Node startup plus a few ms of CPU, which isn't worth shipping and maintaining five-plus platform binaries.