A medium production web:app, before and after
The build-speed work started from one real application: a medium production
Sky.Spa app built with --target web:app. It is a private project, so its
name, its modules, its types and its fields are withheld here. Only totals and
per-stage timings are given.
Field data before the work (v0.27.8 line)
These are the user's own timings of a full --target web:app build, reported
when the work started (.claude/AUTONOMOUS_GOAL.md). They are wall-clock
observations, not runs of a script in this repository.
| Machine | Build time |
|---|---|
| Spot VM, 8 CPUs, 32 GB | 29 min |
| the user's Mac | about 25 min |
| GitHub-hosted runner | 30 to 38 min |
On this branch
The user ran SKY_TIMINGS=1 sky build --target web:app on the 16 GB Apple
silicon Mac (8 CPUs) used for the build-speed work, go1.26.1, with the
compiler built from two commits of perf/build-speed. The figures are the
totals from the SKY_TIMINGS report.
Commit 58fec7490 (an intermediate commit)
| Stage | Time |
|---|---|
| total | 142.5 s |
| lowering, parent process | 49.4 s |
| backend leg | 44.9 s |
| frontend leg | 13.0 s |
| brotli | 27.6 s |
| peak memory | 3.5 GB |
Commit 5d30f9127 (after the call-site harvest fix)
| Stage | Time |
|---|---|
| total | 21.8 s |
| type check, parent process | 1.99 s |
| lowering, parent process | 1.80 s |
| split | 3.33 s |
| frontend leg | 5.79 s |
| backend leg | 8.15 s |
| peak memory | 1.25 GB |
In this run go build and brotli were cache hits: the app's source had not
changed since the previous build, so the Go packages and the .br / .gz
files came from the caches. The 21.8 s is therefore a rebuild of unchanged
source, not a cold build. A cold build of the same commit, and a build on the
VM or the runner from the field data above, were not measured.
Reading it
- The two runs are on a different machine from the field data, so the step from 25 to 38 min to these figures is not one controlled comparison.
- Between the two commits, lowering in the parent fell from 49.4 s to 1.80 s.
Commit
5d30f9127merges the call-site harvest fix (the#166close demanded a harvest whose answer it never read, and ran one inference per callee instead of per caller). - No other numbers for this app are claimed.