bench.yml

# Criterion benchmarks + regression gate.
#
# Phase-B1b (PHASE-A-2 bench plan §3). Bench targets live under each
# crate's `benches/` directory (see `[[bench]]` stanzas in the affected
# Cargo.toml files); this workflow is the on-demand runner + baseline
# publisher.
#
# Current mode: BASELINE COLLECTION (v0.1.0).
#   `fail-on-alert: false`, alerts are surfaced as PR comments only.
#   This is a 1-week bake window while we accumulate a drift signature
#   on the GitHub-hosted runner. See `docs/RELEASE.md` for the flip
#   checklist that hard-enables the gate once the baseline has settled.
#
# Future mode (post-flip):
#   `fail-on-alert: true` blocks merges on >10% regressions.
#   Consider moving to a self-hosted Linux runner (CPU-pinned,
#   governor=performance, SMT off) per PHASE-A-2 §5, GitHub-hosted
#   runners exhibit 15-40% noise on the shared hypervisor and will
#   false-positive a 10% gate until that happens.
name: bench

on:
  push:
    branches: [main]
  pull_request:
    types: [labeled, synchronize, reopened]

permissions:
  # `benchmark-action/github-action-benchmark` commits to `gh-pages`
  # and posts PR comments. `contents: write` covers both.
  contents: write
  pull-requests: write
  deployments: write

jobs:
  bench:
    # On PRs, only run when the `benchmark` label is present, keeps
    # vanilla PRs fast and avoids burning runner minutes on doc-only
    # changes. Push-to-main always runs to keep the gh-pages baseline
    # current.
    if: >-
      github.event_name == 'push'
      || (github.event_name == 'pull_request' && contains(github.event.pull_request.labels.*.name, 'benchmark'))
    # TODO(post-v0.1.0): flip to `[self-hosted, linux, bench]` once
    # we stand up the CPU-pinned runner. Until then, GH-hosted is
    # acceptable because the gate is off (see `fail-on-alert` below).
    runs-on: ubuntu-22.04
    timeout-minutes: 60

    steps:
      - uses: actions/checkout@v4

      - name: Install Rust toolchain
        uses: dtolnay/rust-toolchain@stable
        with:
          toolchain: "1.95"

      - uses: Swatinem/rust-cache@v2
        with:
          shared-key: bench-cache

      - name: Run criterion benches
        # `--output-format bencher` is the libtest-compatible output
        # that `github-action-benchmark` expects when `tool: cargo`.
        # `tee` keeps the same stream visible in the job log so we
        # don't lose debug detail when the action re-parses.
        run: |
          cargo bench --workspace --benches -- --output-format bencher | tee output.txt

      - name: Upload bench output artifact
        uses: actions/upload-artifact@v7
        with:
          name: bench-output
          path: output.txt
          retention-days: 30

      - name: Check bench output non-empty
        id: check_bench
        run: |
          if [ -s output.txt ]; then
            echo "has_output=true" >> "$GITHUB_OUTPUT"
          else
            echo "has_output=false" >> "$GITHUB_OUTPUT"
            echo "::warning::No benchmark output produced; skipping baseline compare."
          fi

      - name: Store + compare against baseline
        if: steps.check_bench.outputs.has_output == 'true'
        uses: benchmark-action/github-action-benchmark@v1
        with:
          tool: cargo
          output-file-path: output.txt
          github-token: ${{ secrets.GITHUB_TOKEN }}
          # Baseline push happens only on push-to-main; PR runs are
          # compare-only and never rewrite gh-pages history.
          auto-push: ${{ github.event_name == 'push' }}
          gh-pages-branch: gh-pages
          benchmark-data-dir-path: bench
          # 1.10x = 10% slower than baseline triggers a warning.
          alert-threshold: "110%"
          # v0.1.0 posture: WARN only, don't block merges. Flip to
          # `true` after the 1-week baseline bake (see `docs/RELEASE.md`).
          fail-on-alert: false
          comment-on-alert: true
          summary-always: true