Skip to content

test-builds.md

Latest commit

 

History

History
102 lines (84 loc) · 5.26 KB

test-builds.md

File metadata and controls

102 lines (84 loc) · 5.26 KB
Raw

Test builds and artifact lifetime

tools/rust-fast-gate.sh checks formatting, runs Clippy with all targets/features, then runs the complete workspace test suite and doctests. It does not select a smaller test subset. PostgreSQL checks remain a separate serial CI step using the same compiled test binaries. Cargo commands use --locked.

Normal runs keep Cargo's build cache. --ephemeral overrides CARGO_TARGET_DIR with a newly created directory and removes it on exit, including failure and handled SIGINT/SIGTERM. Both modes put test temporary files under a directory owned by that invocation and remove it on exit. Existing target directories are never cleaned. SIGKILL and host crashes cannot execute the cleanup handler; Cargo's shared download cache is intentionally retained.

Local builds use line-tables-only debug data. CI and ephemeral builds default to no debug data and no incremental state. Assertions and overflow checks remain active. CI panic backtraces lose source lines; assertion diagnostics retain their call-site location. For local variable inspection, set CARGO_PROFILE_DEV_DEBUG=2. Explicit debug/incremental environment overrides are respected by the ephemeral gate. Release settings are unchanged.

These choices follow the Cargo profile documentation. The existing Rust cache action already disables incremental compilation and caches dependencies; the change does not add caches or expand their retention. The first CI run with the new profile must rebuild dependencies whose compiler settings changed.

Measurements, 2026-09-28

Source baseline: 51e9af13832132e70632166456e38e1964336a79. Host: Mac14,6, 12 logical CPUs, Rust/Cargo 1.95.0, aarch64-apple-darwin. Both configurations used CARGO_INCREMENTAL=0, the same lockfile and test set, and separate task-owned target directories. This is a local measurement, not a GitHub-hosted Linux CI benchmark.

With dependencies already built and only the 18 workspace packages cleaned:

Workspace rebuild and complete test run Previous CI settings New CI settings
Wall time 259.43 s 148.83 s
Aggregate child-process CPU time 386.44 s 299.40 s
Tests passed / ignored 2,407 / 9 2,407 / 9

This single sequential pair took 42.6% less wall time and 22.5% less CPU time. It models a restored dependency cache, not a no-change rerun. Background load on the shared host varied, so these figures are observations rather than a CI speed guarantee.

Cold build and complete test run Previous CI settings New CI settings
Debug data line-tables-only 0
Target disk allocation (du -sk) 2,814,276 KiB 1,585,476 KiB
Total wall time 158.58 s 282.52 s
Tests passed / ignored 2,407 / 9 2,407 / 9

The target allocation fell by 43.7%. The cold runs had varying background load and do not establish a build-speed improvement. Test names and outcomes were identical. Tests requiring optional service credentials return early without those credentials; these runs do not establish live PostgreSQL/Redis/provider coverage. The existing CI PostgreSQL service and serial checks remain enabled.

To reproduce a cold comparison with rustup on macOS:

export PATH="$HOME/.cargo/bin:$PATH"
bench_root="$(mktemp -d "${TMPDIR:-/tmp}/openplotva-build-comparison.XXXXXX")"
trap 'rm -rf -- "$bench_root"' EXIT
for debug in line-tables-only 0; do
  mkdir -p "$bench_root/$debug/tmp"
  CARGO_TARGET_DIR="$bench_root/$debug/target" \
    TMPDIR="$bench_root/$debug/tmp" CARGO_INCREMENTAL=0 \
    CARGO_PROFILE_DEV_DEBUG="$debug" \
    /usr/bin/time -p cargo +1.95.0 test --locked --workspace --timings
  du -sk "$bench_root/$debug/target"
done

For a restored-dependency comparison, clean only the 18 workspace packages with cargo clean -p <package> inside each disposable target directory, then repeat its test command. Do not clean a shared or user-owned target directory to run this measurement. Preserve Rust 1.95.0 at the front of PATH for both Cargo and rustc; selecting a Cargo version alone can still invoke a system rustc.

Verification

With Rust 1.95.0, CARGO_INCREMENTAL=0, CARGO_PROFILE_DEV_DEBUG=0, and a task-owned CARGO_TARGET_DIR:

  • tools/rust-fast-gate.sh: passed, including cargo fmt --all -- --check, cargo clippy --locked --workspace --all-targets --all-features -- -D warnings, and cargo test --locked --workspace (2,407 passed, 9 ignored). The test step reused compiled binaries after Clippy; it emitted no compilation steps. Its temporary directory was empty after the gate exited.
  • PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s tools/tests -v: 6 tests passed, covering retained caches, disposable builds, profile overrides, failure propagation, SIGINT/SIGTERM cleanup, and invalid arguments.
  • PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s tools/maintenance/tests -p test_required_checks.py -v: 2 tests passed; required CI jobs remain present on pull requests.
  • bash -n tools/rust-fast-gate.sh, shellcheck tools/rust-fast-gate.sh, actionlint .github/workflows/ci.yml, and git diff --check: passed.

GitHub Actions and live service integrations were not run for this local change.