ci: single build recipe (reusable) + ordered tag release pipeline; two-tier opt cache

Fixes two coupled bugs: (1) a tag deployed the demo by snapshotting the STALE registry with no build/publish, so v0.1.1 shipped v0.1.0-era wasm; (2) publish-wasm built -O1/3D-off while ci-ubicloud built -O1/3D-on — two diverging recipes.

wasm-build.yml (reusable workflow_call): THE single build+test recipe, parameterized by opt_level / build_3d_viewer / run_tests / no_cache / upload_output. Two-tier output cache around build.sh's new --compile-only/--postprocess-only split: a BASE cache keyed opt-INDEPENDENTLY (compile output + sysroot headers, shared across -O1/-O2) and a FINAL cache keyed opt-SPECIFICALLY. final-hit ⇒ skip all; final-miss+base-hit ⇒ restore base, run only the asyncify/-O tail; full-miss ⇒ compile then postprocess. Both keys now include the 3D flag, closing the cache-poisoning divergence.

ci-ubicloud.yml: thin caller (main/PR/dispatch) → wasm-build at -O1. release.yml (tag v*): meta → build (wasm-build at -O2 + e2e gate, uploads output, reuses main's base cache so only the -O2 tail rebuilds) → publish-wasm (content-addressed push + manifest-<tag>.json) → deploy-demo (build-demo pinned to that manifest + libs/kicad/$LIB_TAG → Pages). So the shipped build is exactly the tested build, and the demo can never point at stale wasm.

deploy-demo.yml: demoted to manual dispatch-only re-deploy (keeps the --from-registry snapshot for re-shipping an already-published tag). publish-wasm.yml: deleted (folded into release.yml + wasm-build.yml).

Validated: actionlint clean (bar the known ubicloud custom-label note), YAML + needs-graph + reusable-path checks pass. NOT runtime-tested — the build can't run locally; needs a CI smoke (a PR exercises ci-ubicloud→wasm-build; a throwaway tag exercises release.yml) before relying on it.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Gergő Törcsvári 2026-06-22 14:15:00 +02:00
commit 913bc90fa0
No known key found for this signature in database
GPG key ID: 8E75F2CDE64E5322
5 changed files with 487 additions and 381 deletions

View file

@ -1,299 +1,49 @@
# The build + e2e CI: full KiCad WASM build (all 6 tools, Binaryen 130, host-side
# asyncify + a lighter wasm-opt -O1 shrink pass via BINARYEN_OPT_LEVEL) + wxWidgets e2e
# (291 tests) + KiCad e2e (38 tests) on a Ubicloud runner. Runs on pushes to
# main/ubicloud-ci and on PRs against main.
# Replaced the ephemeral-Hetzner-VM workflows (ci-full-build.yml / ci.yml,
# removed 2026-06-12): Ubicloud runners are plain runs-on labels — no
# create/delete-runner jobs, no shared one-slot lock — and benchmarked
# faster end-to-end. Proven green: run 27375556034 (329 tests, 1h52m).
# ubicloud-standard-30 (x86, 30 vcpu) ≈ the old Hetzner ccx53.
name: CI full build + e2e (Ubicloud)
# Main/PR gate: build all 6 KiCad WASM tools + run the wxWidgets + KiCad e2e
# suites on a Ubicloud runner. The actual recipe lives in the reusable
# wasm-build.yml (the SINGLE build definition) — this is just the main/PR caller,
# pinned to -O1 for fast feedback. The tag release (release.yml) calls the same
# recipe at -O2 (the shipped demo opt level) and then publishes, so the build can
# never diverge from what ships.
#
# Routine edits under scripts/ do NOT bust the WASM output cache — only the subset
# in scripts/deploy/wasm-cache-hash.mjs does. For inputs the sc-hash can't see
# (base-image / apt drift, a bad cache) bump .ci-cache-epoch; for a one-off
# rebuild put [no-cache] or [rebuild-wasm] in the commit message / PR title.
on:
push:
branches: ["main", "ubicloud-ci"]
# docs-only commits must not burn a paid runner slot
paths-ignore: ["docs/**", "**.md"]
pull_request:
branches: ["main"]
paths-ignore: ["docs/**", "**.md"]
workflow_dispatch:
inputs:
binaryen_version: { description: "Binaryen version for the host asyncify step", required: false, default: "130" }
no_cache: { description: "Bypass the KiCad WASM output cache (force a full rebuild this run)", type: boolean, required: false, default: false }
binaryen_version:
description: "Binaryen version for the host asyncify step"
required: false
default: "130"
no_cache:
description: "Bypass the KiCad WASM output cache (force a full rebuild this run)"
type: boolean
required: false
default: false
concurrency:
# Per-ref: a new push to a PR (or to main) queues behind its own in-flight run
# instead of cancelling it (cancel-in-progress: false). Note GitHub keeps at most
# one pending run per group — rapid successive pushes collapse to "running + latest
# pending", superseding any earlier still-pending run.
# Per-ref: a new push queues behind its own in-flight run (don't cancel).
group: ci-ubicloud-${{ github.ref }}
cancel-in-progress: false
jobs:
build-and-test:
name: Build all tools + KiCad e2e (Ubicloud)
# Don't run untrusted fork PRs on the paid runner (push/dispatch always run;
# same-repo PRs run). Remove this `if:` to enable fork-PR CI.
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubicloud-standard-30
timeout-minutes: 300
env:
KICAD_LOG_NESTED: "1"
BINARYEN_VERSION: ${{ inputs.binaryen_version || '130' }}
# Use the lighter wasm-opt -O1 shrink pass instead of -O2
# (scripts/common/apply-asyncify.sh). -O1 still runs CoalesceLocals, so the
# asyncify-instrumented coroutine functions stay under V8's per-function
# locals limit: it produces the same ~187 MB pcbnew.wasm as -O2 and is green
# on the load-pcb chromium-ci spec, but is a faster pass. (Skipping the
# optimization entirely via ASYNCIFY_ONLY=1 shipped a 336 MB wasm that
# crashed headless Chromium — that flag is benchmark-only, not for e2e.)
BINARYEN_OPT_LEVEL: "-O1"
# Stable docker-compose project → deterministic build-cache volume name
# (kicad-wasm-ci_kicad-build-cache) for the deps cache below.
COMPOSE_PROJECT_NAME: kicad-wasm-ci
steps:
# Docker comes preinstalled on Ubicloud images (validated by the arm64
# repro run) — only the Binaryen from-source toolchain is needed.
- name: Install build toolchain (Binaryen from-source)
run: |
export DEBIAN_FRONTEND=noninteractive
sudo apt-get update
# xvfb: kicad e2e runs headed Firefox under a virtual display — headless
# Firefox cannot create any GL context on the GPU-less VM (see
# tests/playwright-kicad.config.ts CI prefs).
# autoconf/automake/make: the host-side wxWidgets + wx test-app builds
# (build-wx-wasm.sh regenerates ./configure on fresh checkouts).
sudo apt-get install -y cmake ninja-build g++ libjemalloc2 xvfb autoconf automake make
- uses: actions/checkout@v4
with: { submodules: recursive }
- uses: actions/setup-node@v4
with: { node-version: 20 }
# --- KiCad WASM output cache -------------------------------------------
# The expensive work is `./docker/build.sh all` below: the in-Docker compile
# of all tools + the 1-2h host-side asyncify/wasm-opt chain. Its output bytes
# are fixed by the kicad + wxwidgets submodule SHAs (sources), the Binaryen
# version + opt-level (job env), and the build-logic files hashed by
# scripts/deploy/wasm-cache-hash.mjs (the "sc" hash). When all match, the
# artifacts are bit-identical — so we cache output/ (+ the sysroot headers the
# host GAL build needs) and skip Docker entirely on a hit: build.sh, the deps
# restore/seed, and the sysroot export are all gated on a miss below.
#
# Routine edits under scripts/ do NOT auto-bust this — only the subset listed
# in wasm-cache-hash.mjs does. For inputs the sc-hash can't see (base-image /
# apt drift, a bad cache), bump .ci-cache-epoch in a commit; for a one-off
# rebuild, put [no-cache] or [rebuild-wasm] in the commit message / PR title.
#
# Backend: Ubicloud's transparent cache (NOT GitHub's — gh's cache API reads
# 0), 30 GB/repo/week, LRU + 7-day eviction. The *.wasm.debug.wasm files
# (5+ GB, unused by the tests) are excluded to keep the entry ~0.5 GB.
- name: Compute build inputs
id: keys
run: |
echo "kicad=$(git -C kicad rev-parse HEAD)" >> "$GITHUB_OUTPUT"
echo "wx=$(git -C wxwidgets rev-parse HEAD)" >> "$GITHUB_OUTPUT"
echo "sc=$(node scripts/deploy/wasm-cache-hash.mjs)" >> "$GITHUB_OUTPUT"
echo "epoch=$(cat .ci-cache-epoch 2>/dev/null || echo 0)" >> "$GITHUB_OUTPUT"
- name: Cache control (commit message / dispatch)
id: cachectl
env:
HEAD_MSG: ${{ github.event.head_commit.message }}
PR_TITLE: ${{ github.event.pull_request.title }}
DISPATCH_NOCACHE: ${{ github.event.inputs.no_cache }}
run: |
SKIP=false
if printf '%s\n%s' "$HEAD_MSG" "$PR_TITLE" | grep -qiE '\[(no-cache|rebuild-wasm)\]'; then SKIP=true; fi
[ "$DISPATCH_NOCACHE" = "true" ] && SKIP=true
echo "skip=$SKIP" >> "$GITHUB_OUTPUT"
echo "KiCad WASM output-cache restore skip=$SKIP"
- name: Restore KiCad WASM output cache
id: output-cache
if: steps.cachectl.outputs.skip != 'true'
uses: actions/cache/restore@v4
with:
path: |
output/*.js
output/*.wasm
output/*.wasm.map
output/*.worker.js
output/images.tar.gz
build-wasm/sysroot/include
!output/*.wasm.debug.wasm
key: kwasm-${{ runner.os }}-bin${{ env.BINARYEN_VERSION }}${{ env.BINARYEN_OPT_LEVEL }}-k${{ steps.keys.outputs.kicad }}-wx${{ steps.keys.outputs.wx }}-sc${{ steps.keys.outputs.sc }}-e${{ steps.keys.outputs.epoch }}
# Cache the ~10-min --build-deps output (boost/cairo/occ/... sysroot + stamps,
# which live in the kicad-build-cache docker volume). Key on the deps inputs
# only — deps don't depend on the kicad/wx submodule SHAs, so it stays warm
# across most PRs. On a hit we seed the volume so build.sh's per-dep
# check_stamp short-circuits; on a miss the deps build and we tar them out.
- name: Restore deps cache
id: deps-cache
if: steps.output-cache.outputs.cache-hit != 'true'
uses: actions/cache@v4
with:
path: deps-cache
key: deps-${{ runner.os }}-${{ hashFiles('scripts/deps/**','scripts/common/versions.sh','scripts/common/functions.sh','scripts/common/env.sh','docker/Dockerfile','docker/docker-compose.yml') }}
- name: Seed deps volume from cache
if: steps.output-cache.outputs.cache-hit != 'true' && steps.deps-cache.outputs.cache-hit == 'true'
run: |
docker volume create kicad-wasm-ci_kicad-build-cache
docker run --rm -v kicad-wasm-ci_kicad-build-cache:/bw -v "$PWD/deps-cache":/cache \
alpine sh -c 'tar xzf /cache/deps.tar.gz -C /bw'
# Same build recipe as ci-full-build.yml: full KiCad WASM build of ALL 6
# tools, host-side asyncify on Binaryen v130 (BINARYEN_VERSION in env →
# get-wasm-opt.sh) + a lighter -O1 shrink pass (BINARYEN_OPT_LEVEL in job env).
# 'all' = pcbnew eeschema calculator pl_editor symbol_editor gerbview,
# built in sequence. --build-deps short-circuits on a cache hit.
- name: Build all KiCad tools WASM with Binaryen ${{ env.BINARYEN_VERSION }}
if: steps.output-cache.outputs.cache-hit != 'true'
run: |
# Lift the docker-compose dev-Mac caps and pipeline the host-side
# asyncify pass with the next tool's container compile — same tuning
# as the Hetzner workflow (rationale in ci-full-build.yml and
# docs/ci-build-slowness-findings.md). standard-30: 30 vcpu / 120 GB.
# nproc-1: Ubicloud's preinstalled docker daemon rejects a cpus limit
# equal to the host CPU count ("range of CPUs is from 0.01 to 30.00",
# run 27330725978); the Hetzner box accepted 32-of-32.
export KICAD_DOCKER_CPUS="$(( $(nproc) - 1 ))" KICAD_DOCKER_MEM=110G
export KICAD_PIPELINE=1 BINARYEN_CORES=16
# Self-built wasm-opt: official x86_64-linux tarballs run asyncify
# ~4x slower than a stock gcc -O3+LTO build (A/B run 27276830256).
export BINARYEN_BUILD_FROM_SOURCE=1
# Build the experimental WASM 3D viewer (default OFF) so tests/kicad/3d-viewer.spec.ts
# has a viewer-enabled pcbnew to open. Propagates into the container via docker/build.sh.
export BUILD_3D_VIEWER=ON
echo "Building ALL tools with BINARYEN_VERSION=${BINARYEN_VERSION}, compile -j $(nproc) on ${KICAD_DOCKER_CPUS} CPUs, pipelined wasm-opt"
./docker/build.sh all --build-deps -j "$(nproc)"
echo "wasm-opt used:"; ./scripts/common/get-wasm-opt.sh --version 2>/dev/null || true
ls -lh output/*.wasm
# On a cache miss, tar the freshly-built deps (sysroot + stamps) out of the
# volume so actions/cache saves them (post-job) under the key above.
- name: Package deps for cache
if: steps.output-cache.outputs.cache-hit != 'true' && steps.deps-cache.outputs.cache-hit != 'true'
run: |
mkdir -p deps-cache
docker run --rm -v kicad-wasm-ci_kicad-build-cache:/bw -v "$PWD/deps-cache":/cache \
alpine sh -c 'cd /bw && tar czf /cache/deps.tar.gz sysroot stamps'
# Cache the host wxWidgets build (build-wasm/wxwidgets, ~11 min, incremental
# but uncached) keyed on the wx submodule SHA + the build script/versions.
# On a hit, touch the tree so make treats the restored objects as current
# (same SHA ⇒ identical sources), so build-wx-wasm.sh just relinks fast.
- name: Restore wx build cache
id: wx-cache
uses: actions/cache@v4
with:
path: build-wasm/wxwidgets
key: wx-${{ runner.os }}-${{ steps.keys.outputs.wx }}-${{ hashFiles('scripts/build-wx-wasm.sh','scripts/common/versions.sh') }}
- name: Mark restored wx objects current
if: steps.wx-cache.outputs.cache-hit == 'true'
run: find build-wasm/wxwidgets -exec touch {} +
# Host-side wxWidgets build producing the wx test apps for `npm run test`
# (separate from the in-docker wxWidgets that KiCad links against).
# Auto-installs emsdk into tools/ on first run.
- name: Build wxWidgets (wxUniversal WASM)
run: ./scripts/build-wx-wasm.sh
- name: Build wxWidgets test apps
run: ./scripts/build-wasm-test.sh
# The GAL test compiles against kicad headers that include boost from
# build-wasm/sysroot — a real host dir on dev machines, but inside the
# docker build-cache volume on CI (run 27370316200 failed with
# "boost/ptr_container/ptr_vector.hpp not found"). Its Makefile only
# needs $(SYSROOT)/include, so copy just the headers out of the volume.
- name: Expose docker sysroot headers to host builds
if: steps.output-cache.outputs.cache-hit != 'true'
run: |
VOL=kicad-wasm-ci_kicad-build-cache
echo "sysroot volume: $VOL"
mkdir -p build-wasm/sysroot
docker run --rm -v "$VOL":/bw -v "$PWD/build-wasm/sysroot":/host alpine \
sh -c 'cp -r /bw/sysroot/include /host/'
sudo chown -R "$(id -u):$(id -g)" build-wasm/sysroot
# Persist the freshly built KiCad WASM artifacts (+ the sysroot headers the
# GAL build needs) for future runs. Split save (not a combined cache action)
# so the [no-cache]/dispatch bypass can skip the restore yet still refresh
# here. Runs on a miss OR a bypass (cache-hit != 'true'); placed before the
# tests so artifacts are cached regardless of test outcome. No-op if the key
# already exists.
- name: Save KiCad WASM output cache
if: steps.output-cache.outputs.cache-hit != 'true'
uses: actions/cache/save@v4
with:
path: |
output/*.js
output/*.wasm
output/*.wasm.map
output/*.worker.js
output/images.tar.gz
build-wasm/sysroot/include
!output/*.wasm.debug.wasm
key: kwasm-${{ runner.os }}-bin${{ env.BINARYEN_VERSION }}${{ env.BINARYEN_OPT_LEVEL }}-k${{ steps.keys.outputs.kicad }}-wx${{ steps.keys.outputs.wx }}-sc${{ steps.keys.outputs.sc }}-e${{ steps.keys.outputs.epoch }}
# gal_webgl_test.{js,wasm} are gitignored build artifacts with their own
# build script — without this, all 29 gal-webgl scenarios time out on
# galTest.isReady() because the page 404s the wasm (run 27359020746).
- name: Build GAL WebGL test app
run: ./scripts/build-gal-webgl-test.sh
- name: Install test deps
working-directory: tests
run: npm ci
# The collab e2e specs rebuild apps/kicad/collab-bundle.js with esbuild
# (collab/build.mjs), which bundles web/standalone/src/wasm/collab/* and
# its `@pcbjam/shared` import. That package is the web/pcbjam-shared
# submodule, resolved through the web/ pnpm workspace — so install it
# (and the bundle's runtime deps: zod, @ts-rest/core, yjs) here. Without
# this the bundle build fails: [ERROR] Could not resolve "@pcbjam/shared".
- name: Install web workspace deps (collab bundle)
working-directory: web
run: |
corepack enable
pnpm install --frozen-lockfile
- name: Install Playwright browsers
working-directory: tests
run: npx playwright install --with-deps firefox chromium
- name: Stage KiCad WASM for tests
working-directory: tests
run: npm run setup:kicad
# wxWidgets e2e: bundled headless Chromium + SwiftShader (CI-gated launch
# arg in tests/playwright.config.ts).
- name: wxWidgets e2e (npm run test)
working-directory: tests
run: npm run test
- name: KiCad e2e (npm run test:kicad:ci)
working-directory: tests
# xvfb-run: headed Firefox on a virtual display — WebGL needs GLX + Mesa
# llvmpipe, which headless Firefox can't reach on a GPU-less VM.
# test:kicad:ci adds the chromium-ci project: pcbnew's wasm exceeds
# SpiderMonkey's x86-64 code budget, so those specs run on Chromium.
run: xvfb-run -a npm run test:kicad:ci
- name: Upload test logs & screenshots
if: always()
uses: actions/upload-artifact@v4
with:
name: ubicloud-e2e-${{ github.run_id }}
path: |
tests/logs/**
tests/test-results/**
tests/playwright-report/**
if-no-files-found: ignore
build:
uses: ./.github/workflows/wasm-build.yml
with:
# -O1: faster asyncify shrink, sufficient for the test gate (the demo ships
# -O2 via release.yml). 3D viewer ON so 3d-viewer.spec.ts has a viewer.
opt_level: "-O1"
build_3d_viewer: "ON"
binaryen_version: ${{ inputs.binaryen_version || '130' }}
run_tests: true
no_cache: ${{ inputs.no_cache || false }}