Build speed, final: v0.27.8 against v0.27.9

The milestone measurement of the build-speed work, taken after every change in the v0.27.9 notes had landed. It supersedes the part-way run in ../build-speed-20261015/ (compiler 44770d5cd) for the release numbers.

The benchmark (before.json, after.json)

scripts/bench-build.sh --reps 3 --sizes "50 100 200" --cases "app skyshop" --targets "web:app web", run from the branch tree, with --sky / --tree pointing at the v0.27.8 worktree for before. The same xtask bench-gen app (Model fields = update arms = N, 8 modules, seed 1) is built by both.

Scenarios: prime is a build in a fresh directory with the normal Sky Go cache; warm the same directory again; newdir the same sources in a new, deeper directory; edit one string literal changed in one view module (a per-run token, so the edited program is new to the cache); cold a fresh directory with an empty temporary GOCACHE. Wall and CPU in seconds (CPU is user + sys of sky and all its children), peak RSS of the largest single process in MB. Each cell is the median of 3, except first, which is rep 1 of prime (see "Reading it").

--target web:app

Nscenariowall beforewall afterCPU beforeCPU afterpeak RSS beforepeak RSS after
50first35.6326.8157.03153.22610793
50warm3.131.193.932.96145127
50newdir33.742.8756.585.73664581
50edit28.9511.0238.3829.11604600
50cold53.4825.90161.12133.76756782
100first69.5611.64157.6031.421027616
100warm17.201.2914.073.08275138
100newdir50.593.0379.115.69985590
100edit46.3511.6360.8431.63979603
100cold70.3826.82178.73134.71987784
200first139.3513.98189.0744.092781623
200warm64.941.6364.873.18773171
200newdir129.703.85183.236.932976594
200edit123.4512.82163.8544.252989598
200cold153.5227.20277.86140.253015812

--target web (Sky.Live)

Nscenariowall beforewall afterCPU beforeCPU afterpeak RSS beforepeak RSS after
50first7.322.6720.323.67655610
50warm0.620.461.140.878072
50newdir6.311.3420.392.20666581
50edit2.501.756.173.60594580
50cold20.8017.96101.7293.65781788
100first7.302.9420.774.16666585
100warm0.720.551.171.1110781
100newdir6.541.4320.772.35698586
100edit2.662.056.514.27598586
100cold21.3118.61102.9898.14789762
200first7.792.5322.064.89691592
200warm1.070.641.541.00186106
200newdir6.981.5022.062.27694585
200edit3.202.227.774.96603591
200cold20.2219.01102.8994.93764813

examples/13-skyshop (Go FFI, 76k symbols; medians of 3)

scenariowall beforewall afterCPU beforeCPU afterpeak RSS beforepeak RSS after
prime31.937.6786.386.1440411269
warm4.143.203.293.33313310
newdir26.725.3480.755.8539731272
edit10.306.5818.7413.2412971258
cold52.5155.99227.54202.8539724015

The genwhole shape (genwhole200-*.log, genwhole200-*.time)

genwhole.py examples/62-app-notes <dir> 200 adds, per k in 1..200, two Model fields, three Msg arms (a pure edit, a server read, a server write that reads the whole Model) and an input row in the view. This is the real web:app shape where the type-table, CAF and wire-record changes matter most. Each run generates a fresh directory and changes the header text to a per-run token, so the app packages are new to the Go cache; the runtime and stdlib packages were already cached by the benchmark above. One sky build --target web:app src/Main.sky per compiler with SKY_TIMINGS=1 (the logs hold the per-phase tables; scratch paths are replaced by <scratch>). One run each: the planned three were stopped after the first pair, because one v0.27.8 run takes 26 minutes.

genwhole N = 200, first buildv0.27.8v0.27.9
wall1575.7 s29.3 s
CPU (user + sys)2171.7 s119.8 s
peak RSS (largest process)4,775 MB1,112 MB

Before, each leg type-checked for about 238 s and lowered for about 263 s, and the two legs ran one after the other (frontend 787 s, backend 781 s). After, the legs run together (27.7 s for both); the largest step is in-process brotli of the new wasm (20.4 s, beside the backend leg).

The front end once (front-end-once.md)

The stream's own measurement of "run the front end once per program in a split build" (b5be6ed09), copied as it was written: median of 3 warm runs, against the branch before that change (c4d17f467), with the byte-identity and diagnostics comparisons it ran.

Reading it