Improve CI timings #8553
Description
Goal
Reduce required CI wall-clock time for faster PR merges, while preserving current required coverage in the first pass.
Current Required Merge Flow
flowchart TD
A[Detect Changes] --> B[Rust Validation]
A --> C[Repo Validation]
A --> D[Spell Check]
A --> E[cargo-test matrix: ubuntu/macos/windows]
A --> F[build-rolldown-native matrix: ubuntu/macos/windows]
A --> I[Node Validation]
F --> J[node-test: windows/macos/ubuntu]
F --> K[node-dev-server-test: windows/macos/ubuntu]
F --> L[Type Check]
Workflow Files (Hotspots)
.github/workflows/ci.yml.github/workflows/reusable-node-test.yml.github/workflows/reusable-node-dev-server-test.yml.github/workflows/reusable-cargo-test.yml.github/workflows/reusable-native-build.yml
Timing Baseline (Pre-optimization)
Measured from runs:
2269951960222678505235
Total required CI wall-clock:
- ~11.23 min
- ~11.38 min
Job Family Averages (Pre-optimization)
| Job family | Avg duration | Max duration |
|---|---|---|
node-dev-server-test |
5.69m | 7.32m |
cargo-test |
4.98m | 6.23m |
node-test |
4.38m | 4.75m |
Node Validation |
4.35m | 4.43m |
Build @rolldown/browser |
4.23m | 4.60m |
build-rolldown-wasi |
4.01m | 4.22m |
build-rolldown-native |
2.83m | 3.88m |
Progress
| PR | Title | Status | Slow steps addressed | Estimated savings |
|---|---|---|---|---|
| #8554 | ci: dedupe type-check from dev server workflow | ✅ Merged | node-dev-server-test (avg 5.69m, worst 7.32m) — removed redundant pnpm type-check step |
Reduces the slowest CI lane; type-check already covered in node-test-* jobs |
| #8573 | ci: skip redundant native binding build for browser and remove standalone job | ✅ Merged | build-browser → Build @rolldown/browser (avg 184.5s) and Node Validation → Build @rolldown/browser (avg 114.0s) |
Removes standalone build-browser job entirely (~4.23m avg); skips ~60-90s native binding build in browser package scripts |
| #8570 | ci: parallelize Node tests on ubuntu, single Node 24 on macOS/windows | ✅ Merged | node-dev-server-test and node-test — sequential Node 20/22/24 runs parallelized via matrix; reduced matrix on macOS/windows to Node 24 only |
~4min wall-clock savings; 46% reduction in billable minutes (79 vs 146 for node jobs) |
| #8574 | ci: use Dev Drive for Windows CI jobs | ✅ Merged | cargo-test (windows) (avg 4.98m, worst 6.23m) and build-rolldown-native (windows) (avg 2.83m, worst 3.88m) — ReFS Dev Drive bypasses Windows Defender minifilter for CARGO_HOME/RUSTUP_HOME |
~33% faster cargo-test-windows, ~20% faster build-rolldown-windows |
| #8577 | ci: remove unnecessary submodule checkouts | ✅ Merged | Checkout step in reusable-native-build, reusable-node-dev-server-test, ci.yml (type-check), reusable-wasi-build, reusable-wasi-test, vite-tests — removed submodules: true from workflows that don't use rollup/ or test262/ |
Saves ~10-20s per job by skipping submodule clone |
| #8578 | ci: optimize cache keys to fix race conditions and reduce usage | ✅ Merged | Cache contention across all debug jobs sharing debug-build key — assign unique keys per job type (lint, cargo-test, native-build), remove cache-warmup.yml, disable release build caching |
Estimated cache usage ~6.4 GB (down from ~9.9 GB); eliminates cache race conditions |
| #8580 | ci: remove WASI build & test pipeline | ✅ Merged | build-rolldown-wasi (avg 4.01m) and wasi-test matrix (3 OS) — removed entire WASI CI pipeline |
Removes ~4min from critical path; saves 4 jobs worth of billable minutes |
| #8586 | ci: move Windows cargo target dir to Dev Drive | ✅ Merged | cargo-test (windows) Build step (252s cold) and build-rolldown-native (windows) — moves target/ to Dev Drive via CARGO_TARGET_DIR |
Cold build: 252s → ~150-170s; compilation I/O on fast ReFS instead of slow C: drive |
| #8600 | ci: skip Windows CI jobs on PRs | 🔄 Open | cargo-test (windows), build-rolldown-windows, node-test-windows, node-dev-server-test-windows — all skipped on PRs, still run on main |
Removes Windows from PR critical path (~9.3m cold / ~6m warm); saves ~4 jobs of billable minutes per PR |
| #8610 | perf: merge 4 integration test binaries into 1 | 🔄 Open | cargo-test Build step — merges 4 integration test binaries (~297MB) into 1 (~82MB), eliminating 3 redundant link steps |
Local: -48% CPU, -17% wall; CI: -3s (modest, cached); -215MB disk per build |
Latest timing (run 22813402815 on main, 2026-03-08)
Attempt 1 (cold cache): Total wall-clock 9m16s
Attempt 2 (warm cache on Windows rerun): cargo-test (windows) 358s
cargo-test step breakdown
| OS | Build (compile) | Run Test | Total | Notes |
|---|---|---|---|---|
| Windows (cold cache) | 252s | 131s | 556s | First run, cache miss |
| Windows (warm cache) | 96s | 135s | 358s | Rerun with cache hit — 36% faster |
| Ubuntu | 144s | 91s | 294s | |
| macOS | 151s | 115s | 376s |
All jobs (attempt 1)
| Job | Duration |
|---|---|
| cargo-test (windows) | 556s (9.3m) — critical path (cold cache) |
| cargo-test (macos) | 376s (6.3m) |
| cargo-test (ubuntu) | 294s (4.9m) |
| build-rolldown-windows | 266s |
| Node Validation | 252s |
| build-rolldown-macos | 225s |
| node-dev-server-test-windows | 209s |
| Rust Validation | 171s |
| build-rolldown-ubuntu | 149s |
| node-test-windows | 140s |
| node-dev-server-test-ubuntu (×3) | ~117s |
| node-dev-server-test-macos | 115s |
| node-test-ubuntu (×3) | ~97s |
| node-test-macos | 96s |
| Type Check | 45s |
Critical path (warm cache): cargo-test (windows) at 358s. The Run Test step (135s) is now the dominant cost on Windows with warm cache.
Critical path (cold cache): cargo-test (windows) at 556s. The Build step (252s) dominates.
Next bottleneck: cargo-test parallelization with cargo-nextest
cargo test runs ~2000 tests but parallelizes poorly (sequential binary execution). cargo-nextest runs each test as a separate process with full cross-binary parallelism.
Expected Run Test savings:
- Windows: 135s → ~45-65s (~70-90s saved)
- macOS: 115s → ~40-55s
- Ubuntu: 91s → ~30-45s
Metadata
Metadata
Assignees
Labels
No labels
Type
Fields
Give feedbackPriority
None yet