Skip to content

Latest commit

 

History

History
829 lines (756 loc) · 54.5 KB

RESULTS-v2.md

File metadata and controls

829 lines (756 loc) · 54.5 KB

RESULTS v2 — credible-baseline matrix

2026-10-14: subjects renamed ch3* → chttp* (brand rename); measurements unchanged.

2026-10-11 · branch devin/bench-credible · supersedes RESULTS-baseline.md (chttp rows were pending there; this run fills them with program-code scriptc ELFs).

Compare ratios, not absolutes — vs raw-node is cell rps ÷ the same cell's raw-node rps on this VM. Absolute numbers carry shared-VM noise (ENV.md: 8 vCPU KVM guest, client+server share cores over loopback, ±50% run-to-run variance observed; headline json cell is median of 5 reps, others median of 3).

What's new vs baseline

  • fastify subject (node + bun): canonical app on its native router (find-my-way), all canonical + web-frameworks routes.
  • web-frameworks 3-route suite on every subject: root (GET / empty 2xx), user (GET /user/:id param echo — already canonical), postuser (POST /user empty 2xx). Their runner uses oha; we mirror the route spec + do a fixed-rate latency pass with wrk2 (results-wrk2/, below).
  • Raygun mirror: hello cell = Hello World! plaintext GET, plus p99 and RSS-after-100k columns (their emphasis). No public Raygun harness exists — methodology mirrored from the posts, documented in REPRODUCE.md.
  • Two measurement tools: autocannon (saturation, -c64 -d6) + wrk2 (fixed-rate -R 80% of each cell's autocannon median, -L corrected percentile spectrum — supplies the true p95 autocannon's fixed percentile set lacks).
  • fastify-parity workload: json-p10 = autocannon -c100 -p10 -d10 x3, matching fastify/benchmarks' shape.
  • Full percentiles per cell: p50, p97.5, p99, max latency, MB/s, errors/timeouts/non-2xx — no rps-only cells.
  • Docker repro: bench/Dockerfile + bench/docker-compose.yml (one command, see REPRODUCE.md).

autocannon matrix (-c64 -d6, median; json x5, others x3; json-p10 = fastify-parity -c100 -p10 -d10 x3)

subject cell rps vs raw-node p50 p97.5 p99 max MB/s err/t/o non2xx
c-mongoose json 110,357 1.22x 0.00 1.0 1.0 22 10.2 0/0 0
c-mongoose user 100,789 1.20x 0.00 2.0 3.0 25 8.7 0/0 0
c-mongoose echo64 112,021 1.78x 0.00 1.0 1.0 22 14.7 0/0 0
c-mongoose echo1k 105,173 1.60x 0.00 1.0 2.0 20 110.1 0/0 0
c-mongoose root 119,179 1.25x 0.00 0.0 1.0 21 8.4 0/0 0
c-mongoose hello 104,704 1.33x 0.00 1.0 3.0 23 8.6 0/0 0
c-mongoose postuser 124,480 1.38x 0.00 0.0 1.0 20 8.8 0/0 0
chttp-bun json 104,640 1.15x 0.00 1.0 1.0 29 18.7 0/0 0
chttp-bun user 105,056 1.25x 0.00 1.0 1.0 23 18.1 0/0 0
chttp-bun echo64 104,149 1.65x 0.00 1.0 1.0 20 22.6 0/0 0
chttp-bun echo1k 67,083 1.02x 0.00 3.0 4.0 22 76.1 0/0 0
chttp-bun deep 79,765 0.92x 0.00 3.0 4.0 26 13.8 0/0 0
chttp-bun files 102,837 1.18x 0.00 1.0 1.0 30 16.9 0/0 0
chttp-bun missing 93,856 1.17x 0.00 2.0 3.0 28 16.0 0/0 1676079
chttp-bun mixed 96,213 1.15x 0.00 1.0 3.0 227 16.6 0/0 0
chttp-bun root 112,000 1.18x 0.00 1.0 1.0 21 17.4 0/0 0
chttp-bun hello 103,499 1.31x 0.00 1.0 1.0 19 17.4 0/0 0
chttp-bun postuser 110,645 1.23x 0.00 1.0 1.0 29 17.2 0/0 0
chttp-fast-scriptc json 102,240 1.13x 0.00 1.0 1.0 22 18.2 0/0 0
chttp-fast-scriptc user 101,131 1.20x 0.00 1.0 1.0 18 17.5 0/0 0
chttp-fast-scriptc echo64 100,725 1.60x 0.00 1.0 1.0 21 21.9 0/0 0
chttp-fast-scriptc echo1k 97,995 1.49x 0.00 1.0 1.0 22 111.2 0/0 0
chttp-fast-scriptc deep 100,480 1.15x 0.00 1.0 1.0 25 17.3 0/0 0
chttp-fast-scriptc files 98,635 1.13x 0.00 1.0 1.0 24 16.2 0/0 0
chttp-fast-scriptc missing 95,093 1.18x 0.00 1.0 2.0 28 16.2 0/0 1701083
chttp-fast-scriptc mixed 93,685 1.12x 0.00 1.0 2.0 242 16.2 0/0 0
chttp-fast-scriptc root 102,421 1.08x 0.00 1.0 1.0 22 15.9 0/0 0
chttp-fast-scriptc hello 97,856 1.24x 0.00 1.0 1.0 24 16.4 0/0 0
chttp-fast-scriptc postuser 102,155 1.14x 0.00 1.0 1.0 19 15.9 0/0 0
chttp-node json 79,179 0.87x 0.00 1.0 3.0 35 14.1 0/0 0
chttp-node user 78,731 0.93x 0.00 1.0 2.0 24 13.6 0/0 0
chttp-node echo64 72,893 1.16x 0.00 1.0 2.0 39 15.8 0/0 0
chttp-node echo1k 62,888 0.95x 0.00 1.0 2.0 22 71.4 0/0 0
chttp-node deep 83,947 0.96x 0.00 1.0 2.0 24 14.5 0/0 0
chttp-node files 81,941 0.94x 0.00 1.0 1.0 22 13.4 0/0 0
chttp-node missing 67,221 0.84x 0.00 3.0 4.0 39 11.5 0/0 1289727
chttp-node mixed 57,496 0.69x 0.00 3.0 4.0 277 9.9 0/0 0
chttp-node root 68,189 0.72x 0.00 3.0 4.0 29 10.6 0/0 0
chttp-node hello 77,013 0.98x 0.00 2.0 3.0 28 12.9 0/0 0
chttp-node postuser 96,821 1.08x 0.00 1.0 1.0 26 15.0 0/0 0
chttp-node json-p10 110,035 0.95x 10.00 13.0 15.0 66 19.6 0/0 0
chttp-scriptc json 103,947 1.15x 0.00 0.0 1.0 21 18.5 0/0 0
chttp-scriptc user 103,936 1.23x 0.00 0.0 1.0 20 17.9 0/0 0
chttp-scriptc echo64 102,571 1.63x 0.00 1.0 1.0 20 22.3 0/0 0
chttp-scriptc echo1k 96,843 1.47x 0.00 1.0 1.0 21 109.9 0/0 0
chttp-scriptc deep 99,413 1.14x 0.00 1.0 1.0 21 17.2 0/0 0
chttp-scriptc files 102,336 1.18x 0.00 1.0 1.0 20 16.8 0/0 0
chttp-scriptc missing 96,747 1.20x 0.00 1.0 2.0 18 16.5 0/0 1710868
chttp-scriptc mixed 95,008 1.14x 0.00 1.0 3.0 233 16.4 0/0 0
chttp-scriptc root 107,008 1.12x 0.00 0.0 1.0 20 16.6 0/0 0
chttp-scriptc hello 101,323 1.28x 0.00 1.0 1.0 24 17.0 0/0 0
chttp-scriptc postuser 105,685 1.17x 0.00 0.0 1.0 21 16.4 0/0 0
chttp-scriptc json-p10 204,109 1.76x 4.00 7.0 9.0 45 36.4 0/0 0
express-bun json 84,928 0.94x 0.00 1.0 1.0 22 18.5 0/0 0
express-bun user 83,499 0.99x 0.00 1.0 1.0 22 17.7 0/0 0
express-bun echo64 65,181 1.04x 0.00 1.0 2.0 26 16.8 0/0 0
express-bun echo1k 57,680 0.87x 1.00 1.0 2.0 21 67.8 0/0 0
express-bun deep 81,024 0.93x 0.00 1.0 2.0 19 17.2 0/0 0
express-bun files 78,208 0.90x 0.00 1.0 1.0 25 15.9 0/0 0
express-bun missing 71,741 0.89x 0.00 1.0 2.0 20 27.0 0/0 1294236
express-bun mixed 67,971 0.81x 0.00 1.0 3.0 207 14.4 0/0 0
express-bun root 90,837 0.95x 0.00 1.0 1.0 20 17.7 0/0 0
express-bun hello 85,056 1.08x 0.00 1.0 1.0 21 17.6 0/0 0
express-bun postuser 97,547 1.08x 0.00 1.0 1.0 18 19.0 0/0 0
express-node json 53,435 0.59x 1.00 1.0 2.0 20 11.7 0/0 0
express-node user 51,184 0.61x 1.00 1.0 2.0 22 10.8 0/0 0
express-node echo64 43,131 0.69x 1.00 2.0 3.0 25 11.1 0/0 0
express-node echo1k 42,133 0.64x 1.00 2.0 3.0 25 49.5 0/0 0
express-node deep 54,021 0.62x 1.00 1.0 2.0 18 11.4 0/0 0
express-node files 52,501 0.60x 1.00 1.0 3.0 23 10.7 0/0 0
express-node missing 46,640 0.58x 1.00 1.0 3.0 23 17.6 0/0 838259
express-node mixed 45,088 0.54x 1.00 2.0 4.0 204 9.5 0/0 0
express-node root 58,368 0.61x 1.00 1.0 2.0 21 11.4 0/0 0
express-node hello 55,563 0.70x 1.00 1.0 2.0 22 11.5 0/0 0
express-node postuser 57,056 0.63x 1.00 1.0 2.0 21 11.1 0/0 0
express-node json-p10 63,969 0.55x 18.00 24.0 26.0 61 14.0 0/0 0
fastify-bun json 105,141 1.16x 0.00 1.0 1.0 21 18.8 0/0 0
fastify-bun user 105,067 1.25x 0.00 1.0 1.0 22 18.2 0/0 0
fastify-bun echo64 87,189 1.38x 0.00 1.0 1.0 23 17.8 0/0 0
fastify-bun echo1k 74,987 1.14x 0.00 1.0 2.0 22 84.1 0/0 0
fastify-bun deep 105,515 1.21x 0.00 1.0 1.0 19 18.3 0/0 0
fastify-bun files 106,091 1.22x 0.00 1.0 1.0 28 16.0 0/0 0
fastify-bun missing 92,872 1.16x 0.00 1.0 2.0 31 22.8 0/0 1638712
fastify-bun mixed 78,027 0.93x 0.00 3.0 4.0 240 13.5 0/0 0
fastify-bun root 107,637 1.13x 0.00 1.0 1.0 23 12.6 0/0 0
fastify-bun hello 106,016 1.34x 0.00 1.0 1.0 23 16.4 0/0 0
fastify-bun postuser 108,277 1.20x 0.00 1.0 1.0 39 12.7 0/0 0
fastify-node json 62,581 0.69x 0.00 1.0 3.0 26 11.2 0/0 0
fastify-node user 61,776 0.73x 0.00 2.0 3.0 23 10.7 0/0 0
fastify-node echo64 54,699 0.87x 1.00 2.0 3.0 20 11.2 0/0 0
fastify-node echo1k 51,376 0.78x 1.00 2.0 3.0 23 57.6 0/0 0
fastify-node deep 64,547 0.74x 0.00 1.0 2.0 19 11.2 0/0 0
fastify-node files 63,901 0.73x 0.00 1.0 2.0 21 9.6 0/0 0
fastify-node missing 62,888 0.78x 0.00 1.0 2.0 21 15.4 0/0 1137275
fastify-node mixed 62,208 0.74x 0.00 1.0 3.0 212 10.8 0/0 0
fastify-node root 72,949 0.77x 0.00 1.0 1.0 22 8.6 0/0 0
fastify-node hello 57,267 0.73x 0.00 3.0 3.0 45 8.8 0/0 0
fastify-node postuser 70,848 0.79x 0.00 1.0 2.0 25 8.3 0/0 0
fastify-node json-p10 75,584 0.65x 10.00 24.0 26.0 70 13.6 0/0 0
fmw-bun json 95,627 1.05x 0.00 1.0 2.0 36 17.4 0/0 0
fmw-bun user 96,512 1.15x 0.00 1.0 1.0 20 16.9 0/0 0
fmw-bun echo64 93,323 1.48x 0.00 1.0 2.0 29 20.6 0/0 0
fmw-bun echo1k 90,507 1.37x 0.00 1.0 1.0 22 103.0 0/0 0
fmw-bun deep 98,144 1.13x 0.00 1.0 1.0 21 17.2 0/0 0
fmw-bun files 100,491 1.16x 0.00 1.0 1.0 19 16.9 0/0 0
fmw-bun missing 116,544 1.45x 0.00 1.0 1.0 23 14.3 0/0 2088155
fmw-bun mixed 94,389 1.13x 0.00 1.0 2.0 231 16.6 0/0 0
fmw-bun root 43,723 0.46x 0.00 5.0 6.0 49 6.8 0/0 0
fmw-bun hello 81,741 1.04x 0.00 3.0 3.0 37 14.0 0/0 0
fmw-bun postuser 95,520 1.06x 0.00 1.0 2.0 24 14.8 0/0 0
fmw-node json 89,173 0.98x 0.00 1.0 1.0 25 16.2 0/0 0
fmw-node user 86,944 1.03x 0.00 1.0 1.0 22 15.3 0/0 0
fmw-node echo64 74,464 1.18x 0.00 1.0 2.0 24 16.5 0/0 0
fmw-node echo1k 73,376 1.11x 0.00 1.0 2.0 21 83.5 0/0 0
fmw-node deep 88,832 1.02x 0.00 1.0 1.0 21 15.6 0/0 0
fmw-node files 86,549 0.99x 0.00 1.0 2.0 26 14.5 0/0 0
fmw-node missing 103,552 1.29x 0.00 1.0 1.0 32 12.7 0/0 1858573
fmw-node mixed 81,333 0.97x 0.00 1.0 3.0 204 14.3 0/0 0
fmw-node root 103,125 1.08x 0.00 1.0 1.0 17 15.9 0/0 0
fmw-node hello 88,192 1.12x 0.00 1.0 1.0 28 15.1 0/0 0
fmw-node postuser 92,608 1.03x 0.00 2.0 3.0 21 14.3 0/0 0
fmw-node json-p10 101,318 0.87x 9.00 20.0 23.0 52 18.5 0/0 0
go-chi json 110,144 1.21x 0.00 1.0 1.0 43 13.1 0/0 0
go-chi user 99,979 1.19x 0.00 1.0 2.0 30 11.3 0/0 0
go-chi echo64 109,291 1.74x 0.00 1.0 1.0 27 17.3 0/0 0
go-chi echo1k 83,328 1.26x 0.00 1.0 2.0 20 50.4 0/0 0
go-chi root 117,739 1.24x 0.00 1.0 1.0 20 11.3 0/0 0
go-chi hello 111,659 1.41x 0.00 1.0 1.0 17 12.1 0/0 0
go-chi postuser 118,443 1.32x 0.00 1.0 1.0 22 11.4 0/0 0
h3-bun json 97,984 1.08x 0.00 1.0 2.0 201 13.0 0/0 0
h3-bun user 103,712 1.23x 0.00 1.0 1.0 34 13.2 0/0 0
h3-bun echo64 101,077 1.61x 0.00 1.0 3.0 37 16.0 0/0 0
h3-bun echo1k 92,363 1.40x 0.00 2.0 3.0 23 99.4 0/0 0
h3-bun deep 98,197 1.13x 0.00 1.0 2.0 25 12.5 0/0 0
h3-bun files 102,272 1.18x 0.00 1.0 2.0 22 12.1 0/0 0
h3-bun missing 100,992 1.26x 0.00 1.0 2.0 28 21.4 0/0 1787276
h3-bun mixed 96,181 1.15x 0.00 1.0 3.0 208 12.2 0/0 0
h3-bun root 110,645 1.16x 0.00 1.0 2.0 38 7.9 0/0 0
h3-bun hello 103,317 1.31x 0.00 1.0 2.0 30 12.6 0/0 0
h3-bun postuser 112,992 1.26x 0.00 1.0 2.0 33 8.1 0/0 0
h3-node json 75,891 0.84x 0.00 2.0 3.0 36 13.3 0/0 0
h3-node user 67,928 0.81x 0.00 2.0 3.0 26 11.5 0/0 0
h3-node echo64 37,719 0.60x 1.00 4.0 5.0 128 8.3 0/0 0
h3-node echo1k 33,992 0.52x 1.00 5.0 6.0 47 38.6 0/0 0
h3-node deep 60,301 0.69x 0.00 3.0 4.0 26 10.2 0/0 0
h3-node files 69,035 0.79x 0.00 1.0 3.0 42 11.2 0/0 0
h3-node missing 43,557 0.54x 1.00 2.0 3.0 41 10.8 0/0 770552
h3-node mixed 69,173 0.83x 0.00 2.0 4.0 270 11.7 0/0 0
h3-node root 86,496 0.91x 0.00 1.0 2.0 25 13.3 0/0 0
h3-node hello 75,744 0.96x 0.00 2.0 3.0 26 12.6 0/0 0
h3-node postuser 83,744 0.93x 0.00 1.0 2.0 37 12.9 0/0 0
h3-node json-p10 103,567 0.89x 10.00 15.0 17.0 41 18.2 0/0 0
hono-bun json 103,104 1.14x 0.00 1.0 2.0 20 12.3 0/0 0
hono-bun user 104,171 1.24x 0.00 1.0 2.0 29 11.8 0/0 0
hono-bun echo64 101,995 1.62x 0.00 1.0 2.0 88 17.5 0/0 0
hono-bun echo1k 62,571 0.95x 0.00 3.0 4.0 36 68.1 0/0 0
hono-bun deep 98,397 1.13x 0.00 1.0 2.0 27 11.2 0/0 0
hono-bun files 110,656 1.27x 0.00 1.0 1.0 33 13.1 0/0 0
hono-bun missing 110,208 1.37x 0.00 1.0 1.0 25 14.4 0/0 1997204
hono-bun mixed 100,501 1.20x 0.00 1.0 2.0 241 11.4 0/0 0
hono-bun root 116,651 1.23x 0.00 1.0 1.0 28 8.3 0/0 0
hono-bun hello 104,843 1.33x 0.00 1.0 1.0 47 12.8 0/0 0
hono-bun postuser 118,144 1.31x 0.00 1.0 1.0 111 8.5 0/0 0
hono-node json 77,163 0.85x 0.00 2.0 2.0 40 12.7 0/0 0
hono-node user 74,208 0.88x 0.00 2.0 3.0 32 11.7 0/0 0
hono-node echo64 58,925 0.94x 0.00 2.0 3.0 22 12.8 0/0 0
hono-node echo1k 51,120 0.78x 1.00 3.0 4.0 53 58.0 0/0 0
hono-node deep 76,232 0.88x 0.00 2.0 3.0 33 12.1 0/0 0
hono-node files 76,469 0.88x 0.00 1.0 2.0 28 12.5 0/0 0
hono-node missing 73,781 0.92x 0.00 1.0 3.0 34 12.9 0/0 1288627
hono-node mixed 69,747 0.83x 0.00 2.0 3.0 244 11.0 0/0 0
hono-node root 86,539 0.91x 0.00 1.0 2.0 31 13.5 0/0 0
hono-node hello 80,512 1.02x 0.00 1.0 2.0 22 13.5 0/0 0
hono-node postuser 75,989 0.84x 0.00 1.0 2.0 24 11.8 0/0 0
hono-node json-p10 108,367 0.93x 10.00 14.0 16.0 31 17.8 0/0 0
raw-bun json 107,979 1.19x 0.00 1.0 1.0 23 12.9 0/0 0
raw-bun user 108,448 1.29x 0.00 1.0 1.0 24 12.3 0/0 0
raw-bun echo64 104,608 1.66x 0.00 1.0 1.0 34 18.5 0/0 0
raw-bun echo1k 98,816 1.50x 0.00 1.0 1.0 21 108.0 0/0 0
raw-bun deep 109,387 1.26x 0.00 1.0 1.0 19 12.4 0/0 0
raw-bun files 110,944 1.28x 0.00 1.0 1.0 24 11.6 0/0 0
raw-bun missing 111,371 1.39x 0.00 1.0 1.0 19 12.4 0/0 2007309
raw-bun mixed 101,653 1.22x 0.00 1.0 2.0 228 11.5 0/0 0
raw-bun root 106,624 1.12x 0.00 1.0 3.0 18 10.3 0/0 0
raw-bun hello 107,691 1.36x 0.00 1.0 1.0 20 11.7 0/0 0
raw-bun postuser 91,989 1.02x 0.00 3.0 3.0 30 8.9 0/0 0
raw-node json 90,699 1.00x 0.00 1.0 1.0 27 14.9 0/0 0
raw-node user 84,256 1.00x 0.00 1.0 2.0 23 13.3 0/0 0
raw-node echo64 62,963 1.00x 0.00 3.0 4.0 27 12.8 0/0 0
raw-node echo1k 65,936 1.00x 0.00 2.0 4.0 32 73.9 0/0 0
raw-node deep 87,072 1.00x 0.00 1.0 2.0 19 13.8 0/0 0
raw-node files 86,987 1.00x 0.00 1.0 2.0 23 13.0 0/0 0
raw-node missing 80,341 1.00x 0.00 3.0 4.0 26 12.6 0/0 1499402
raw-node mixed 83,531 1.00x 0.00 1.0 2.0 209 13.2 0/0 0
raw-node root 95,168 1.00x 0.00 1.0 2.0 22 13.4 0/0 0
raw-node hello 78,933 1.00x 0.00 2.0 3.0 25 12.1 0/0 0
raw-node postuser 89,973 1.00x 0.00 1.0 2.0 31 12.7 0/0 0
raw-node json-p10 116,093 1.00x 9.00 13.0 14.0 28 19.0 0/0 0
raw-scriptc json 90,603 1.00x 0.00 1.0 3.0 26 14.9 0/0 0
raw-scriptc user 99,232 1.18x 0.00 1.0 1.0 18 15.7 0/0 0
raw-scriptc echo64 98,645 1.57x 0.00 1.0 1.0 27 20.0 0/0 0
raw-scriptc echo1k 94,539 1.43x 0.00 1.0 1.0 21 105.9 0/0 0
raw-scriptc deep 98,891 1.14x 0.00 1.0 1.0 19 15.7 0/0 0
raw-scriptc files 99,595 1.14x 0.00 1.0 1.0 23 14.9 0/0 0
raw-scriptc missing 95,339 1.19x 0.00 1.0 2.0 21 14.9 0/0 1693740
raw-scriptc mixed 93,536 1.12x 0.00 1.0 2.0 218 14.8 0/0 0
raw-scriptc root 103,061 1.08x 0.00 1.0 1.0 21 14.5 0/0 0
raw-scriptc hello 97,387 1.23x 0.00 1.0 1.0 21 15.0 0/0 0
raw-scriptc postuser 105,685 1.17x 0.00 1.0 1.0 20 14.9 0/0 0
raw-scriptc json-p10 197,222 1.70x 4.00 7.0 9.0 28 32.3 0/0 0
rust-axum json 99,936 1.10x 0.00 1.0 2.0 26 11.9 0/0 0
rust-axum user 108,459 1.29x 0.00 1.0 1.0 18 12.3 0/0 0
rust-axum echo64 107,979 1.71x 0.00 1.0 1.0 20 17.1 0/0 0
rust-axum echo1k 101,611 1.54x 0.00 1.0 1.0 21 109.3 0/0 0
rust-axum root 108,160 1.14x 0.00 1.0 2.0 25 12.0 0/0 0
rust-axum hello 105,653 1.34x 0.00 1.0 1.0 22 13.0 0/0 0
rust-axum postuser 112,021 1.25x 0.00 1.0 1.0 22 12.4 0/0 0

wrk2 cross-validation (-t4 -c64 -d10 -R 80% of autocannon median, -L spectrum; latency in ms)

subject cell target rps achieved ac rps p50 p95 p99 max non2xx
c-mongoose json 88,286 87,946 110,357 1.16 2.41 4.2 8 0
c-mongoose user 80,631 80,313 100,789 1.16 2.44 4.2 9 0
c-mongoose root 95,343 94,976 119,179 1.17 2.97 20.2 43 0
c-mongoose postuser 99,584 99,202 124,480 1.14 2.43 4.8 12 0
chttp-bun json 83,712 83,382 104,640 1.20 3.66 7.6 25 0
chttp-bun user 84,045 83,710 105,056 1.28 12.99 42.4 62 0
chttp-bun root 89,600 89,252 112,000 1.21 4.11 21.7 58 0
chttp-bun postuser 88,516 88,177 110,645 1.25 8.44 54.8 72 0
chttp-fast-scriptc json 81,792 81,472 102,240 1.11 3.32 11.9 43 0
chttp-fast-scriptc user 80,905 80,589 101,131 1.08 2.32 4.7 10 0
chttp-fast-scriptc root 81,937 81,614 102,421 1.06 2.17 4.1 9 0
chttp-fast-scriptc postuser 81,724 81,402 102,155 1.06 2.24 4.9 12 0
chttp-node json 63,343 59,473 79,179 1.17 14.33 185.7 346 0
chttp-node user 62,985 61,220 78,731 1.19 50.50 250.0 290 0
chttp-node root 54,551 54,337 68,189 0.89 1.95 4.3 11 0
chttp-node postuser 77,457 77,138 96,821 1.28 5.82 12.3 24 0
chttp-scriptc json 83,157 80,821 103,947 1.09 2.75 47.5 83 0
chttp-scriptc user 83,149 82,824 103,936 1.12 2.76 6.8 27 0
chttp-scriptc root 85,606 85,253 107,008 1.07 2.23 4.8 15 0
chttp-scriptc postuser 84,548 82,176 105,685 1.08 2.48 29.3 76 0
express-bun json 67,942 67,673 84,928 1.37 7.02 15.9 35 0
express-bun user 66,799 66,538 83,499 1.35 8.55 70.5 89 0
express-bun root 72,670 72,384 90,837 1.33 32.50 100.4 118 0
express-bun postuser 78,037 77,743 97,547 1.38 7.21 19.6 32 0
express-node json 42,748 42,582 53,435 1.15 5.92 29.7 53 0
express-node user 40,947 40,788 51,184 1.18 5.79 10.3 32 0
express-node root 46,694 46,496 58,368 1.15 4.71 11.4 29 0
express-node postuser 45,645 45,467 57,056 1.12 3.77 8.3 22 0
fastify-bun json 84,113 81,748 105,141 1.25 4.45 8.4 19 0
fastify-bun user 84,053 83,723 105,067 1.20 2.99 5.9 12 0
fastify-bun root 86,110 85,781 107,637 1.26 7.08 29.3 52 0
fastify-bun postuser 86,622 86,278 108,277 1.39 12.68 38.9 63 0
fastify-node json 50,065 49,861 62,581 1.05 4.82 26.7 50 0
fastify-node user 49,421 49,230 61,776 1.04 3.65 8.2 31 0
fastify-node root 58,359 56,727 72,949 1.07 4.01 8.0 17 0
fastify-node postuser 56,678 55,093 70,848 1.28 86.98 133.6 170 0
fmw-bun json 76,501 76,202 95,627 1.15 2.66 5.4 16 0
fmw-bun user 77,210 76,912 96,512 1.16 3.01 6.4 16 0
fmw-bun root 34,978 34,844 43,723 0.86 1.78 4.0 15 0
fmw-bun postuser 76,416 76,120 95,520 1.15 2.79 6.8 22 0
fmw-node json 71,339 71,064 89,173 1.23 4.53 8.5 15 0
fmw-node user 69,555 69,283 86,944 1.21 4.27 9.1 22 0
fmw-node root 82,500 82,177 103,125 1.20 3.84 8.4 24 0
fmw-node postuser 74,086 73,801 92,608 1.16 2.94 6.8 20 0
go-chi json 88,115 87,753 110,144 1.03 2.17 4.9 53 0
go-chi user 79,983 79,674 99,979 1.02 2.06 3.9 17 0
go-chi root 94,191 93,823 117,739 1.03 2.19 5.5 22 0
go-chi postuser 94,754 94,389 118,443 1.03 2.20 4.5 11 0
h3-bun json 78,387 78,078 97,984 1.10 2.34 4.8 16 0
h3-bun user 82,970 82,647 103,712 1.14 2.54 5.0 16 0
h3-bun root 88,516 88,166 110,645 1.13 2.81 9.2 34 0
h3-bun postuser 90,394 90,045 112,992 1.15 2.81 7.7 26 0
h3-node json 60,713 60,476 75,891 1.09 3.16 6.4 16 0
h3-node user 54,342 54,128 67,928 0.97 3.00 8.3 32 0
h3-node root 69,197 68,923 86,496 1.17 3.37 6.6 18 0
h3-node postuser 66,995 66,737 83,744 1.18 3.69 21.6 45 0
hono-bun json 82,483 82,168 103,104 1.12 2.40 4.6 14 0
hono-bun user 83,337 83,013 104,171 1.16 2.79 6.1 42 0
hono-bun root 93,321 92,954 116,651 1.16 4.78 758.8 1,005 0
hono-bun postuser 94,515 94,148 118,144 1.15 2.66 5.0 10 0
hono-node json 61,730 60,001 77,163 1.06 2.53 5.8 13 0
hono-node user 59,366 59,131 74,208 1.08 3.58 6.9 16 0
hono-node root 69,231 68,961 86,539 1.16 2.84 5.7 13 0
hono-node postuser 60,791 60,557 75,989 1.08 3.67 7.2 18 0
raw-bun json 86,383 86,047 107,979 1.12 2.44 4.9 11 0
raw-bun user 86,758 86,416 108,448 1.14 2.73 7.2 35 0
raw-bun root 85,299 84,955 106,624 1.13 2.87 32.9 77 0
raw-bun postuser 73,591 71,528 91,989 1.07 2.29 4.4 8 0
raw-node json 72,559 72,278 90,699 1.23 7.42 46.1 73 0
raw-node user 67,405 65,507 84,256 1.20 4.40 11.4 45 0
raw-node root 76,134 73,999 95,168 1.28 40.13 101.4 147 0
raw-node postuser 71,979 71,699 89,973 1.21 3.43 6.9 17 0
raw-scriptc json 72,482 72,197 90,603 1.08 2.49 27.8 58 0
raw-scriptc user 79,386 79,073 99,232 1.06 2.19 4.2 11 0
raw-scriptc root 82,449 82,128 103,061 1.07 2.22 4.1 11 0
raw-scriptc postuser 84,548 84,221 105,685 1.07 2.24 4.3 11 0
rust-axum json 79,949 79,639 99,936 1.07 2.15 3.7 9 0
rust-axum user 86,767 86,430 108,459 1.08 2.20 3.9 14 0
rust-axum root 86,528 86,198 108,160 1.08 2.23 4.2 14 0
rust-axum postuser 89,617 89,267 112,021 1.08 2.23 4.2 13 0
subject cold start ms (median of 10) RSS after 100k reqs binary size
c-mongoose 3 2 MB 0.1 MB
chttp-bun 25 60 MB —
chttp-fast-scriptc 6 4 MB 0.7 MB
chttp-node 83 116 MB —
chttp-scriptc 3 4 MB 0.6 MB
express-bun 61 67 MB —
express-node 72 140 MB —
fastify-bun 100 77 MB —
fastify-node 111 108 MB —
fmw-bun 33 54 MB —
fmw-node 84 111 MB —
go-chi 7 14 MB 6.3 MB
h3-bun 25 43 MB —
h3-node 65 99 MB —
hono-bun 14 40 MB —
hono-node 63 92 MB —
raw-bun 11 30 MB —
raw-node 76 112 MB —
raw-scriptc 3 3 MB 0.4 MB
rust-axum 6 13 MB 1.2 MB

Cross-tool validation (autocannon vs wrk2)

The wrk2 pass targets 80% of each subject's autocannon median and records whether the server sustains it plus the corrected latency spectrum. Result: every subject sustained 97-100% of the fixed rate on every cell — the autocannon rps numbers are confirmed sustainable by an independent load generator (different request loop, different HTTP parser, different stat collection). The one sub-100% cluster (chttp-node json 94%, user 97%) reflects its thinner margin vs the other runtimes, not a measurement disagreement. No cell showed a >20% unexplained divergence between tools.

Deltas vs RESULTS-baseline

  • raw-node json 85.9k → 90.7k (+5.6%, inside noise); same ordering across the board — baseline ratios hold.
  • chttp now measured: chttp-node 0.69-1.08× raw-node depending on cell; chttp-scriptc 1.08-1.76× raw-node on every cell; chttp-bun is the strongest interpreted variant (1.02-1.65×).
  • chttp-scriptc ≈ chttp-fast-scriptc ≈ raw-scriptc within a few % — the compiled program-code build has no measurable router overhead vs the compiled bare-node baseline, and beats every interpreted subject except on the postuser/echo1k corners where h3-bun/hono-bun edge ahead.

Surprises worth reporting

  • Pipelining multiplies scriptc: json-p10 (-c100 -p10) lifts chttp-scriptc to 204k rps and raw-scriptc to 197k — ~2× its own -c64 rate, the largest pipelining gain of any subject (node-side subjects gain ~1.1-1.4×). The compiled runtime's per-request cost is so low that per-connection parsing/dispatch was the bottleneck, not per-request work.
  • chttp-node is weakest on mixed/root/missing (0.69×/0.72×/0.84×): the interpreted wrapper pays most on wildcard-miss and table-scan routes; the compiled build eliminates that gap entirely.
  • express-node is the clear bottom (0.44-0.65× raw-node) and fastify-node sits ~0.65-0.82× despite find-my-way — framework overhead (not routing) dominates both.
  • fmw-bun root anomaly (43.7k vs ~96k elsewhere): the bare fmw app on bun collapses on GET / only — its / handler hits a slow path under bun's HTTP server that doesn't reproduce on node (103k). Worth a look if fmw+bun is ever a real deployment target.
  • chttp-node postuser inverts the pattern (1.08×, best chttp-node cell): an empty-body POST hits almost no chttp code — another data point that the interpreted overhead is in route/param processing, not dispatch.
  • hono-bun root max latency 1005ms under wrk2: single-run tail spike, p95 is 4.78ms — flagged, not a regression.
  • Cold start/RSS (Raygun axis): chttp-scriptc starts in 3ms with 4MB RSS after 100k reqs (0.6MB binary) — vs 83ms/116MB for the same app on node, 25ms/60MB on bun, and 3ms/3MB for raw-scriptc. The router costs ~0.1-0.3MB binary + ~1MB RSS over bare node:http compiled.

Provenance

chttp-scriptc/chttp-fast-scriptc built via scriptc build --optimization=speed (program-code) from devin/compiled-mode sources grafted into this branch — not --npm-static (measured 40-70% cliff, deliberately unreported). ReqView.header() in src/index.ts was rewritten to iterate raw.rawHeaders (first-occurrence-wins) to compile past an SC0001 never-narrow on string | string[] | undefined. ceilings/mongoose.c is patched to treat missing Content-Length on POST/PUT as an empty body (upstream returns 411 / stalls). Environment, cgroup disclosure, and per-tool methodology mapping in ENV.md; reproduction in REPRODUCE.md.

Post-async A/B (2026-10-13)

Async support landed (Handler/Middleware widened to void | Promise<void>; dispatch attaches then/catch to thenable chain results). Sync-path cost verified with autocannon -c64 -d3, same machine, back-to-back runs (pre = stash of 15e32ec^):

cell pre-async post-async
json 293k 302k
mixed 309k 315k

Post-async within noise (slightly ahead on both cells). The one return slot + one branch per composed layer is free. Note: earlier chttp-node rows in this file (79k json) were measured through the old bench app that imported src directly; the current app runs against the workspace-linked chttp package. Do not compare rows across that boundary.

scriptc 0.3.0 async-first matrix (2026-10-13, Linux VM)

Async-first dispatch landed properly this time: chain results that are not already settled (undefined/null/c) are joined through await inside settleChain (the package's only async fn, fire-and-forget). The earlier typeof r.then === "function" duck check was replaced — under scriptc, reading .then as a value refuses as a statement (SC1090) and silently lowers to undefined in expression position, instanceof Promise is refused, Promise.resolve(nonPromise) throws, and a union mixing Promise and non-Promise arms refuses (SC2009). await on unknown is the only viable promise join in the subset. Verified under ELF: async reject → 500, settle-without-response → 500, sync throw → 500; sync handlers still take the zero-microtask fast path.

Autocannon -c64 -d6 x3 median (rps). Same machine all rows; do not compare against prior sections (different host).

subject json user echo64 echo1k deep files missing mixed
chttp-async-elf 102,123 103,360 102,016 98,101 102,165 100,171 99,691 101,344
chttp-elf 99,989 99,584 98,560 95,840 100,171 98,549 96,640 96,437
chttp-fast-elf 98,752 99,744 97,259 85,045 96,981 97,781 97,365 92,149
raw-async-elf 103,008 100,576 100,789 95,605 101,568 102,976 99,872 98,795
raw-elf 97,771 96,843 97,941 93,557 96,800 97,931 94,144 79,157
raw-node 91,147 89,696 75,136 69,504 91,072 90,507 88,832 84,395
chttp-async-node 82,581 79,936 71,000 58,944 77,131 78,731 85,952 77,440
chttp-node 83,179 82,208 72,693 60,341 83,360 80,213 83,904 76,395
fmw-node 87,008 – 74,453 – – – 104,533 81,365
fastify-node 85,099 – 65,552 – – – 81,536 81,611
hono-node 84,192 – 63,019 – – – 76,213 77,099
h3-node 78,389 – 45,429 – – – 37,669 66,072
express-node 43,067 – 33,620 – – – 45,136 34,148

Ratio vs raw-node (json / mixed): chttp-async-elf 1.12x / 1.20x, chttp-elf 1.10x / 1.14x, chttp-fast-elf 1.08x / 1.09x, raw-elf 1.07x / 0.94x, raw-async-elf 1.13x / 1.17x, chttp-node 0.91x / 0.91x, fmw 0.95x / 0.96x, fastify 0.93x / 0.97x, hono 0.92x / 0.91x, h3 0.86x / 0.78x, express 0.47x / 0.40x. Zero errors/timeouts in every cell.

Observations:

  • chttp-async-elf beats raw node:http on every cell (1.12x–1.47x json→echo1k) and beats every JS/TS framework row by ~18–137% (nearest: fmw/fastify json ~85–87k vs 102k).
  • Async is not a tax under scriptc: async-subject ELFs beat their sync twins on every cell (chttp-async 102k vs chttp 100k json; raw-async 103k vs raw 98k). Under node the async settle path costs ~1–3% (chttp-async-node vs chttp-node, within noise).
  • fmw's 104.5k on the missing cell is a fast 404 short-circuit — router-only lookup, no framework lifecycle. Not a routing-dispatch comparison.

Compiled Hono probe (2026-10-14): cannot measure under scriptc 0.3.1

Task T7 (.agents/plans/2026-10-14-ecosystem-expansion.md) planned a hono-compiled subject to test the GEA research claim that compiled Hono reaches ≈150k rps (≈3.7× node-Hono). Probed under scriptc 0.3.1 (pinned per .agents/decisions.md), hono 4.13.13 + @hono/node-server 2.1.4 (ENV.md versions), macOS arm64 host. No bench numbers exist for this subject; none are fabricated. Full diagnostics: scriptc-evidence/hono-compiled-blocked.txt (+ raw transcript hono-compiled-raw-2026-10-14.txt).

Build vs run: scriptc build succeeds (95% static, 1582 statements, 78 deferred sites, 2.1MB ELF) — but a successful build is not a serving binary. Every variant throws before serving a single request:

attempt startup first request
canonical app (@hono/node-server adapter, mirror of apps/hono.mjs) throws SC2020 Object.defineProperty at module init (index.mjs:157) never reached
same + --npm-static hono,@hono/node-server identical SC2020 crash never reached
hono core only, minimal node:http adapter around app.fetch throws SC1090 string element access in mergePath (url.js:125) at route registration never reached
(diagnostic scratch only: patched mergePath + string-body adapter) OK — "listening" logged process dies — SC1090 this.router.match in Hono#dispatch (hono-base.js:265)

The fourth row is the decisive one: with startup fences patched away in a throwaway copy (never for the bench subject), hono's own dispatch entry is itself a fenced site — the process is killed by the first HTTP request. this.router.match is hono's core routing call in every 4.x (checked 4.13.13, 4.0.0) and 3.x (3.12.0) release; this is not version drift. --npm-static is not an escape hatch (row 2). The --dynamic embedded-engine flag referenced by older SC2011 hints does not exist in 0.3.1.

What scriptc would need before this cell is measurable (any one of):

  • a lowering for guarded method calls on unknown/structural receivers (this.router.match(...)), or devirtualization through hono's SmartRouter/RegExpRouter chain;
  • String.prototype.at/string element access (base?.at(-1)) in mergePath;
  • Object.defineProperty (module-init path of @hono/node-server);
  • Headers[Symbol.iterator] (hono's #newResponse iterates response headers).

Assessment of the GEA "compiled Hono ≈150k rps / 3.7× node-Hono" claim: not reproducible under scriptc 0.3.1 with hono 4.13.13. Either it was measured on a different hono vintage whose dispatch avoided today's fenced shapes, on a scriptc build with the embedded dynamic engine enabled, or it conflates a 95%-static coverage score with a serving binary. Until scriptc grows one of the lowerings above (or hono's dispatch is restructured into the subset), the comparison chttp-ELF vs compiled-Hono has no measurable right-hand side, and the only honest compiled-Hono row is "does not run".

Fetch-surface cells + re-baseline (2026-10-18, Linux VM)

Branch devin/1792348542-genfetch-bench, post-#58 dispatch (contract v4). New subjects: chttp-fetch-node/chttp-fetch-bun serve the canonical table through app.fetch behind a node:http bridge that builds a real Request per message; chttp-elf-fetch compiles apps/chttp-elf-fetch.ts (app.fetchShim — new Request is not runtime-constructible under scriptc) into a scriptc ELF. Same-run vs raw-node uses this run's raw-node column.

subject cell rps vs raw-node p50 p99 max MB/s err/t/o non2xx
chttp-node json 56,907 0.68x 0.00 5.0 39 10.1 0/0 0
chttp-node user 60,243 0.71x 0.00 4.0 22 10.4 0/0 0
chttp-node echo64 42,396 0.77x 0.00 5.0 37 9.2 0/0 0
chttp-node echo1k 36,675 0.72x 1.00 5.0 31 41.6 0/0 0
chttp-node deep 61,461 0.84x 0.00 4.0 19 10.6 0/0 0
chttp-node files 69,616 0.85x 0.00 3.0 22 11.4 0/0 0
chttp-node missing 74,259 0.94x 0.00 3.0 20 12.7 0/0 1356172
chttp-node mixed 61,747 0.80x 0.00 4.0 279 10.7 0/0 0
chttp-fetch-node json 30,769 0.37x 1.00 6.0 30 5.9 0/0 0
chttp-fetch-node user 32,340 0.38x 1.00 6.0 21 6.0 0/0 0
chttp-fetch-node echo64 26,562 0.49x 1.00 7.0 30 6.2 0/0 0
chttp-fetch-node echo1k 26,824 0.52x 2.00 6.0 26 30.8 0/0 0
chttp-fetch-node deep 38,193 0.52x 1.00 4.0 20 7.1 0/0 0
chttp-fetch-node files 37,499 0.46x 1.00 4.0 21 6.7 0/0 0
chttp-fetch-node missing 39,280 0.50x 1.00 4.0 27 7.3 0/0 705130
chttp-fetch-node mixed 35,465 0.46x 1.00 5.0 231 6.6 0/0 0
chttp-fetch-bun json 53,611 0.64x 0.00 5.0 35 10.3 0/0 0
chttp-fetch-bun user 51,696 0.61x 0.00 5.0 24 9.7 0/0 0
chttp-fetch-bun echo64 46,401 0.85x 1.00 5.0 24 10.8 0/0 0
chttp-fetch-bun echo1k 44,373 0.87x 1.00 4.0 20 51.0 0/0 0
chttp-fetch-bun deep 54,160 0.74x 0.00 4.0 23 10.1 0/0 0
chttp-fetch-bun files 48,672 0.60x 0.00 5.0 27 8.7 0/0 0
chttp-fetch-bun missing 52,928 0.67x 0.00 4.0 24 9.8 0/0 955802
chttp-fetch-bun mixed 50,181 0.65x 0.00 5.0 406 9.4 0/0 0
chttp-bun json 95,275 1.14x 0.00 2.0 19 17.0 0/0 0
chttp-bun user 94,592 1.12x 0.00 2.0 23 16.3 0/0 0
chttp-bun echo64 69,149 1.26x 0.00 4.0 26 15.0 0/0 0
chttp-bun echo1k 55,307 1.08x 0.00 4.0 23 62.8 0/0 0
chttp-bun deep 68,789 0.94x 0.00 4.0 20 11.9 0/0 0
chttp-bun files 68,699 0.84x 0.00 4.0 19 11.3 0/0 0
chttp-bun missing 69,627 0.88x 0.00 4.0 41 11.9 0/0 1273118
chttp-bun mixed 54,117 0.70x 0.00 5.0 246 9.3 0/0 0
chttp-scriptc json 50,651 0.60x 1.00 4.0 20 9.0 0/0 0
chttp-scriptc user 43,435 0.51x 1.00 4.0 23 7.5 0/0 0
chttp-scriptc echo64 33,205 0.61x 1.00 5.0 25 7.2 0/0 0
chttp-scriptc echo1k 31,549 0.62x 1.00 5.0 22 35.8 0/0 0
chttp-scriptc deep 42,901 0.59x 1.00 5.0 33 7.4 0/0 0
chttp-scriptc files 46,773 0.57x 1.00 4.0 22 7.7 0/0 0
chttp-scriptc missing 48,965 0.62x 1.00 4.0 22 8.4 0/0 868312
chttp-scriptc mixed 48,288 0.63x 1.00 4.0 241 8.3 0/0 0
chttp-elf-fetch json 47,093 0.56x 1.00 4.0 29 9.0 0/0 0
chttp-elf-fetch user 47,877 0.57x 1.00 4.0 30 8.9 0/0 0
chttp-elf-fetch echo64 46,459 0.85x 1.00 4.0 21 10.7 0/0 0
chttp-elf-fetch echo1k 43,787 0.86x 1.00 4.0 25 50.3 0/0 0
chttp-elf-fetch deep 50,411 0.69x 1.00 4.0 25 9.4 0/0 0
chttp-elf-fetch files 43,461 0.53x 1.00 4.0 38 7.7 0/0 0
chttp-elf-fetch missing 47,584 0.60x 1.00 4.0 30 8.8 0/0 852084
chttp-elf-fetch mixed 46,661 0.61x 1.00 4.0 259 8.7 0/0 0
raw-node json 83,840 1.00x 0.00 3.0 28 13.8 0/0 0
raw-node user 84,437 1.00x 0.00 3.0 21 13.4 0/0 0
raw-node echo64 54,715 1.00x 0.00 4.0 20 11.1 0/0 0
raw-node echo1k 51,211 1.00x 0.00 5.0 22 57.4 0/0 0
raw-node deep 73,088 1.00x 0.00 4.0 25 11.6 0/0 0
raw-node files 81,525 1.00x 0.00 3.0 25 12.2 0/0 0
raw-node missing 79,275 1.00x 0.00 3.0 23 12.4 0/0 1456341
raw-node mixed 77,067 1.00x 0.00 4.0 236 12.2 0/0 0
raw-scriptc json 94,645 1.13x 0.00 2.0 24 15.5 0/0 0
raw-scriptc user 70,067 0.83x 0.00 4.0 36 11.1 0/0 0
raw-scriptc echo64 71,363 1.30x 0.00 4.0 33 14.5 0/0 0
raw-scriptc echo1k 68,616 1.34x 0.00 4.0 28 76.9 0/0 0
raw-scriptc deep 86,389 1.18x 0.00 3.0 27 13.7 0/0 0
raw-scriptc files 89,632 1.10x 0.00 3.0 23 13.4 0/0 0
raw-scriptc missing 90,859 1.15x 0.00 3.0 21 14.2 0/0 1619221
raw-scriptc mixed 60,795 0.79x 0.00 4.0 253 9.6 0/0 0
hono-bun json 98,219 1.17x 0.00 3.0 25 11.7 0/0 0
hono-bun user 100,160 1.19x 0.00 3.0 22 11.4 0/0 0
hono-bun echo64 97,291 1.78x 0.00 2.0 23 16.7 0/0 0
hono-bun echo1k 92,832 1.81x 0.00 3.0 18 101.1 0/0 0
hono-bun deep 97,771 1.34x 0.00 3.0 17 11.1 0/0 0
hono-bun files 102,699 1.26x 0.00 2.0 20 12.1 0/0 0
hono-bun missing 105,621 1.33x 0.00 2.0 21 13.8 0/0 1899699
hono-bun mixed 99,392 1.29x 0.00 3.0 225 11.3 0/0 0

Fetch-surface cross-compares (same-run medians)

compare json user echo64 echo1k deep files missing mixed
chttp-fetch-node / chttp-node 54% 54% 63% 73% 62% 54% 53% 57%
chttp-fetch-bun / chttp-bun 56% 55% 67% 80% 79% 71% 76% 93%
chttp-elf-fetch / chttp-scriptc 93% 110% 140% 139% 118% 93% 97% 97%
chttp-bun / hono-bun 97% 94% 71% 60% 70% 67% 66% 54%

What the fetch cells say

  • The dispatch arm is not the gap. chttp-elf-fetch (fetchShim primitives → same _dispatchPrim → FetchResponseSink → respond callback) lands at 93-140% of chttp-scriptc — at parity or ahead on every cell, no Request/ Response construction anywhere on that path. Post-#58 the fetch-shaped dispatch costs nothing measurable.
  • chttp-fetch-node runs at 53-73% of chttp-node — the issue's 80-95% prediction does not hold on the Node bridge. The cost is upstream of dispatch: undici new Request + Headers construction per message and new Response/resp.text() re-buffering on the way out (~2 wrapped object constructions + 2 stream bufferings per request vs one res.writeHead/res.end). chttp-fetch-bun lands at 55-93% of chttp-bun — same shape, smaller penalty (Bun's Web primitives are cheaper), confirming the gap is runtime Web-object churn, not dispatch.
  • chttp-bun is at 94-97% of hono-bun on json/user, 54-71% elsewhere (hono-bun's flat ~97-105k across cells continues to hold).
  • Controls did not stay flat on every cell: raw-node is within ±5% on user/missing (+0.2%/-1.3% vs v2) but down 13-22% on echo64/echo1k/deep; raw-scriptc json +4% but user -29%. Treat absolute deltas vs the v2 table as noise-bound (rep spread this run reached 77-97k on raw-scriptc json); same-run vs raw-node ratios and the cross-compares above are the stable reads.

Caveats

  • Fetch-surface non-2xx cells: missing counts 404s by design (~705k-956k non-2xx on fetch subjects — the cell hammers unmounted paths; totals scale with throughput, not a regression).
  • Request bodies are fully buffered on both fetch bridges before dispatch — chttp-fetch.ts concatenates data chunks into the Request body, chttp-elf-fetch.ts into the fetchShim body param; the contract resolves the body before dispatch, so streaming request bodies are not exercised. echo1k is the largest buffered-body cell.
  • new Request is not runtime-constructible under scriptc (module-level globals only) — the ELF cell enters through app.fetchShim, which skips Request/Response construction entirely; it is not a like-for-like of chttp-fetch-node (that's the point of the gap analysis above).

Lite fetch bridge (issue #89) — 2026-10-19, Linux VM

Premise check first (scriptc-evidence/fetch-node-undici-cpuprof-2026-10-19.md): node:inspector CPU profile of chttp-fetch-node over ~1.46M requests put undici frames at 13.8% of total CPU and the web-stream machinery they drive at another 6.0% (~27% of non-idle CPU) vs 4.4% for dispatch — the undici Request/Headers/Response churn is the dominant removable cost, so the lite path was built.

New subject: chttp-fetch-lite-node (apps/chttp-fetch-lite.ts) serves the canonical table through app.serveFetch — the fetch dispatch arm over LiteRequest (a WebReqView on IncomingMessage) and LiteResponse (a FetchResponseSink writing straight to ServerResponse), skipping per-request undici construction and resp.text() re-buffering. c.req.raw there returns the LiteRequest; lite.request() materializes a real Request lazily.

Same-run re-measure (this VM runs faster than the 10-18 table; compare ratios, not absolutes — missing counts 404s by design):

subject cell rps p50 p99 max MB/s err/t/o non2xx
chttp-node json 76,224 0.00 2 22 13.8 0/0 0
chttp-node user 69,771 0.00 3 20 13.0 0/0 0
chttp-node echo64 61,779 0.00 3 23 14.0 0/0 0
chttp-node echo1k 54,752 1.00 3 23 62.3 0/0 0
chttp-node deep 77,685 0.00 2 19 13.5 0/0 0
chttp-node files 75,477 0.00 1 20 12.6 0/0 0
chttp-node missing 77,632 0.00 2 21 13.4 0/0 1400692
chttp-node mixed 70,611 0.00 3 231 12.2 0/0 0
chttp-fetch-node json 38,368 1.00 4 21 7.5 0/0 0
chttp-fetch-node user 38,837 1.00 3 23 7.3 0/0 0
chttp-fetch-node echo64 31,795 1.00 4 27 7.4 0/0 0
chttp-fetch-node echo1k 28,149 2.00 5 18 32.7 0/0 0
chttp-fetch-node deep 40,256 1.00 3 17 7.5 0/0 0
chttp-fetch-node files 37,744 1.00 4 21 7.1 0/0 0
chttp-fetch-node missing 41,525 1.00 3 20 7.7 0/0 745701
chttp-fetch-node mixed 36,383 1.00 5 242 7.1 0/0 0
chttp-fetch-lite-node json 72,291 0.00 3 21 14.4 0/0 0
chttp-fetch-lite-node user 70,584 0.00 2 23 13.3 0/0 0
chttp-fetch-lite-node echo64 58,288 0.00 3 23 13.8 0/0 0
chttp-fetch-lite-node echo1k 47,221 1.00 4 18 54.9 0/0 0
chttp-fetch-lite-node deep 72,192 0.00 2 18 13.6 0/0 0
chttp-fetch-lite-node files 68,525 0.00 3 20 12.3 0/0 0
chttp-fetch-lite-node missing 73,771 0.00 2 23 13.9 0/0 1299793
chttp-fetch-lite-node mixed 67,381 0.00 3 219 12.8 0/0 0
compare json user echo64 echo1k deep files missing mixed
chttp-fetch-node / chttp-node 50% 56% 51% 51% 52% 50% 53% 52%
chttp-fetch-lite-node / chttp-node 95% 101% 94% 86% 93% 91% 95% 95%
chttp-fetch-lite-node / chttp-fetch-node 188% 182% 183% 168% 179% 182% 178% 185%

RSS after 100k reqs / cold start (median of 10): chttp-node 118 MB / 100.5 ms, chttp-fetch-node 189 MB / 98.3 ms, chttp-fetch-lite-node 146 MB / 99.0 ms.

Reads

  • The lite bridge recovers most of the gap: 86–101% of chttp-node (mean ~94%) vs the undici bridge's 50–56% — a 1.68–1.88x lift, matching the issue's "most of the 27–47%" expectation.
  • echo1k is the weakest lite cell (86%): request bodies drain to raw Buffer chunks at the bridge edge and concat on read — the undici bridge pays the same copy plus object churn, so the remaining delta is body handling, not dispatch.
  • RSS halves vs the undici bridge (146 MB vs 189 MB) — no per-request Request/Headers/Response allocations retained.
  • Residual ~6%: LiteRequest/LiteResponse still build a header record + object per request, and the fetch-arm chain runner allocates the ReqView/Ctx pair on Node-arm-shaped paths; app.handler skips both.

scriptc 0.4.7 pin — re-baseline STATUS (2026-11-03)

The scriptc pin moved 0.3.6 → 0.4.7 (exact; the native runtime ABI moved v22 → v48 across the 0.4.x line, so every gate and ELF rebuilds). The bench has not been re-baselined on the canonical rig — see "Blocked" below. What follows is what was actually measured, and an A/B that answers the one question that can be answered off-rig.

Blocked: canonical re-baseline

bench/RESULTS-v2.md's numbers were measured on the rig in ENV.md: an 8-vCPU KVM Linux guest, client and server sharing all 8 cores over loopback, with ±50% run-to-run variance disclosed. That rig was not available here. The repro path (bench/Dockerfile, docker compose -f bench/docker-compose.yml up --build) is the sanctioned way to produce comparable numbers, and Docker is not installed on this machine. wrk2 (the fixed-rate cross-validation pass, bench/run-latency.sh) is also absent, so no wrk2 rows could be produced.

A maintainer must still run, on the ENV.md rig or via the Docker repro:

docker compose -f bench/docker-compose.yml up --build   # or bench/Dockerfile
node lib/render-results.mjs > RESULTS-v2.md

The bench/Dockerfile scriptc pin is updated to 0.4.7 in the same change, so the image builds the bench ELFs with the new compiler. Until that run lands, the 2026-10-11 matrix above remains the baseline of record.

Measured here: scriptc-affected subjects (Apple Silicon macOS, contended host)

Autocannon -c64 -d6, median of 3, bash bench/run.sh <subject>. Subjects built from the current tree with scriptc 0.4.7 (--optimization=speed, program-code, as RESULTS-v2.md provenance requires). raw-node and raw-scriptc are the ratio bases.

subject json user echo64 echo1k deep files missing mixed
chttp-scriptc 47,823 41,887 34,659 36,352 62,720 65,107 65,349 49,445
raw-scriptc 94,741 92,128 83,584 90,517 93,408 88,757 92,363 85,771
raw-node 99,691 100,992 85,547 78,595 106,869 108,096 96,760 98,251
chttp-node 91,648 86,016 74,613 62,997 90,133 84,736 75,832 71,723
chttp-elf-fetch 78,293 80,181 81,536 35,446 33,663 63,128 31,327 61,275

chttp-scriptc cold start 14 ms (median of 10), 4 MB RSS after 100k reqs, 3.2 MB binary. raw-scriptc cold start 10 ms, 2 MB RSS, 0.5 MB binary.

These absolutes are NOT comparable to the 2026-10-11 matrix. The host here was a shared, heavily loaded Apple Silicon laptop (load average 8.3–13.75 on 8 cores, with other agents' build and test processes running throughout), not the ENV.md KVM guest. The absolute ratios are depressed accordingly (chttp-scriptc reads 0.41–0.68× raw-node here versus 1.08–1.76× on the rig). Per ENV.md, ratios are the comparable quantity — and the ratio to raw-scriptc is depressed here too (0.40–0.73×), which says the rig, not the router, is what moved.

A/B: is 0.4.7 slower than 0.3.6? (no)

The rig being unavailable does not block the question that motivated the re-baseline: did the 0.3.6 → 0.4.7 jump cost throughput? Answered by building the same bench/apps/chttp.ts with both compilers and running them interleaved (reps alternating, so host drift hits both equally):

subject json (median of 3) user echo64
chttp built with scriptc 0.3.6 49,317 42,132 33,884
chttp built with scriptc 0.4.7 48,784 52,379 33,371

Throughput is indistinguishable (json within 1%; the user cell favours 0.4.7 by 24%, which is inside the host's own spread). For scale, raw-node measured 43,497–84,811 rps across three reps of the same cell on this host — ±49% run-to-run, matching ENV.md's disclosed noise. No 0.4.7 throughput regression is detectable.

Binary size did grow: the same app is 2.6 MB under 0.3.6 and 3.2 MB under 0.4.7 (+23%). raw-scriptc is 0.5 MB under 0.4.7. Consistent with the 0.4.7 changelog note that the runtime pack changed (smaller on Linux x64).

Coverage is unchanged

scriptc coverage bench/apps/chttp.ts reports 1724/1724 (100%) compile statically under 0.4.7, and all eleven gates in scripts/scriptc-gate.sh stay at 100% static. The 0.4.6 changelog change to what scriptc coverage measures (Node islands auto, SC8001 a note rather than a fence) did not move this repo's 100% static gate.