/.github/workflows/partition-benchmark.yaml

name: Partition Benchmark

# PRs are gated against a rolling median baseline; pushes to main record a new
# baseline sample. A manual run (workflow_dispatch) on main is a read-only
# inspection; a manual run on any OTHER branch is a self-test that records +
# uploads to a throwaway `-selftest` prefix -- so the write path can be validated
# on a branch, before merge, without touching the real baseline. (Keying on the
# ref rather than an input is deliberate: dispatch inputs are validated against
# the default-branch workflow, so a branch-only input can't be passed pre-merge.)
on:
  pull_request:
    branches: [main]
  push:
    branches: [main]
  workflow_dispatch:

permissions:
  contents: read

env:
  NLTK_DATA: ${{ github.workspace }}/nltk_data
  PYTHON_VERSION: "3.12"
  # Number of times to run the full benchmark suite.
  NUM_ITERATIONS: "3"
  # Allowance over the rolling median baseline. Sized to absorb runner-speed
  # variance on shared ubuntu-latest runners (observed ~1.6x), not just a code delta.
  REGRESSION_THRESHOLD: "0.30"
  # How many recent main-run samples the baseline median is taken over.
  BASELINE_WINDOW: "20"
  # Below this many samples, the gate is in warm-up (records, does not fail).
  BASELINE_MIN_SAMPLES: "5"
  # Increment to change cache key when benchmark-affecting dependencies are updated, to ensure clean slate runs.
  CACHE_VERSION: "v2"
  # S3 location for metrics – matches core-product convention.
  S3_METRICS_BUCKET_KEY: utic-metrics/ci-metrics
  # Prefix holding one immutable JSON record per main run (the rolling history).
  S3_HISTORY_PREFIX: open-source/partition-benchmark/history

jobs:
  setup:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: ./.github/actions/base-cache
        with:
          python-version: ${{ env.PYTHON_VERSION }}

  benchmark:
    name: Measure and compare partition() runtime
    runs-on: ubuntu-latest
    needs: [setup]

    steps:

      - uses: actions/checkout@v4

      - uses: ./.github/actions/base-cache
        with:
          python-version: ${{ env.PYTHON_VERSION }}

      - name: Install system dependencies
        run: |
          sudo apt-get update
          sudo apt-get install -y libmagic-dev poppler-utils libreoffice
          sudo add-apt-repository -y ppa:alex-p/tesseract-ocr5
          sudo apt-get update
          sudo apt-get install -y tesseract-ocr tesseract-ocr-kor


      - name: Restore HuggingFace model cache
        uses: actions/cache/restore@v4
        with:
          path: ~/.cache/huggingface
          key: hf-models-${{ runner.os }}-${{ env.CACHE_VERSION }}-${{ github.sha }}
          restore-keys: |
            hf-models-${{ runner.os }}-${{ env.CACHE_VERSION }}-
            hf-models-${{ runner.os }}-


      - name: Run partition benchmark
        env:
          NUM_ITERATIONS: ${{ env.NUM_ITERATIONS }}
        run: |
          uv run --no-sync python scripts/performance/benchmark_partition.py \
            benchmark_results.json

      - name: Save HuggingFace model cache
        uses: actions/cache/save@v4
        with:
          path: ~/.cache/huggingface
          key: hf-models-${{ runner.os }}-${{ env.CACHE_VERSION }}-${{ github.sha }}


      # Decide, in one place, which baseline mode this run is in:
      #   push:main                          -> record + upload to the real prefix
      #   workflow_dispatch on a non-main ref -> record + upload to a throwaway -selftest prefix
      #   pull_request / dispatch on main     -> read-only, real prefix (gate only, never write)
      - name: Resolve baseline mode
        id: mode
        run: |
          prefix="${{ env.S3_HISTORY_PREFIX }}"
          record=""
          upload="false"
          if [ "${{ github.event_name }}" = "push" ]; then
            record="--record"; upload="true"
          elif [ "${{ github.event_name }}" = "workflow_dispatch" ] && \
               [ "${{ github.ref_name }}" != "main" ]; then
            prefix="${prefix}-selftest"; record="--record"; upload="true"
          fi
          {
            echo "prefix=$prefix"
            echo "record=$record"
            echo "upload=$upload"
          } >> "$GITHUB_OUTPUT"
          echo "mode: prefix=$prefix record='$record' upload=$upload"


      # Pull the whole history prefix (one object per past run). On fork PRs the S3
      # secrets are absent, so this no-ops -> empty dir -> warm-up/pass.
      - name: Download benchmark history from S3
        continue-on-error: true
        env:
          AWS_ACCESS_KEY_ID: ${{ secrets.S3_EVAL_ACCESS_KEY }}
          AWS_SECRET_ACCESS_KEY: ${{ secrets.S3_EVAL_SECRET_KEY }}
        run: |
          mkdir -p bench_history
          aws s3 cp --recursive \
            "s3://${{ env.S3_METRICS_BUCKET_KEY }}/${{ steps.mode.outputs.prefix }}/" \
            bench_history/


      # Gate against the rolling median. When recording (push, or dispatch self-test),
      # --record also writes this run's record into bench_history/ for the upload below.
      - name: Compare against rolling baseline
        id: compare
        run: |
          uv run --no-sync python scripts/performance/compare_benchmark.py \
            benchmark_results.json \
            bench_history \
            --threshold ${{ env.REGRESSION_THRESHOLD }} \
            --window ${{ env.BASELINE_WINDOW }} \
            --min-samples ${{ env.BASELINE_MIN_SAMPLES }} \
            ${{ steps.mode.outputs.record }}


      # Extend the baseline. `sync` (no --delete) uploads the one new record written
      # above; existing objects are unchanged. PR/read-only runs skip this entirely.
      # `!cancelled()` (instead of the implicit success() &&) so the record still
      # persists when the compare step fails the gate on a regression -- otherwise a
      # red run never reaches S3 and the baseline can't track the new reality. The
      # record is written before the gate check, so the local file exists either way.
      - name: Upload new baseline record to S3
        if: ${{ !cancelled() && steps.mode.outputs.upload == 'true' }}
        continue-on-error: true
        env:
          AWS_ACCESS_KEY_ID: ${{ secrets.S3_EVAL_ACCESS_KEY }}
          AWS_SECRET_ACCESS_KEY: ${{ secrets.S3_EVAL_SECRET_KEY }}
        run: |
          aws s3 sync \
            bench_history/ \
            "s3://${{ env.S3_METRICS_BUCKET_KEY }}/${{ steps.mode.outputs.prefix }}/"


      - name: Upload benchmark artifacts
        if: always()
        uses: actions/upload-artifact@v4
        with:
          name: benchmark-results-${{ github.sha }}
          path: |
            benchmark_results.json
            bench_history/
          retention-days: 30