retrain.yml

name: Scheduled Retraining

on:
  schedule:
    # Every Monday at 02:00 UTC
    - cron: "0 2 * * 1"
  workflow_dispatch:
    inputs:
      lookback_days:
        description: "Days to look back for drift detection"
        default: "30"
        required: false
      retrain_data_path:
        description: "Path to labelled dataset for retraining"
        default: "data/synthetic_dataset.parquet"
        required: false

env:
  MODEL_STORE_TYPE: ${{ vars.MODEL_STORE_TYPE || 'local' }}

permissions:
  id-token: write
  contents: write
  issues: write
  pull-requests: read

jobs:
  detect-and-retrain:
    runs-on: ubuntu-latest

    steps:
      - name: Checkout repository
        uses: actions/checkout@v4

      - name: Set up Python
        uses: actions/setup-python@v5
        with:
          # Keep formatting checks on the same interpreter as the main CI lint
          # job; Black's safety parser is interpreter-dependent.
          python-version: "3.12"
          cache: pip

      - name: Install dependencies
        run: |
          pip install -r requirements.txt
          pip install ruff black

      - name: Lint check
        run: |
          ruff check .
          black --check .

      - name: Run drift detection and retraining
        id: retrain
        run: |
          python -m scripts.retrain_if_drifted \
            --lookback-days ${{ github.event.inputs.lookback_days || 30 }} \
            --retrain-data-path ${{ github.event.inputs.retrain_data_path || 'data/synthetic_dataset.parquet' }} \
            2>&1 | tee retrain_output.txt
          echo "exit_code=$?" >> $GITHUB_OUTPUT

      - name: Upload drift reports
        if: always()
        uses: actions/upload-artifact@v4
        with:
          name: retrain-reports
          path: reports/
          retention-days: 90

      - name: Commit model artifacts on promotion
        if: steps.retrain.outputs.exit_code == '2'
        run: |
          git config user.name "ledgerlens-bot"
          git config user.email "bot@ledgerlens.ai"
          git add models/ metrics.json 2>/dev/null || true
          git diff --cached --quiet || \
            git commit -m "auto: promote retrained models [skip ci]"
          git push

      - name: Create or update tracking issue with drift report
        if: always()
        env:
          GH_TOKEN: ${{ github.token }}
        run: |
          summary=$(python -c "
          import json, os, glob

          report_dir = 'reports'
          if not os.path.isdir(report_dir):
              print('No reports directory — skipping.')
              exit(0)

          drift_files = sorted(glob.glob(f'{report_dir}/drift_report_*.json'))
          retrain_files = sorted(glob.glob(f'{report_dir}/retrain_report_*.json'))

          lines = ['## Weekly Retraining Report\\n']

          if drift_files:
              with open(drift_files[-1]) as f:
                  d = json.load(f)
              lines.append(f'**Drift detected:** {d[\"any_drift_detected\"]}')
              lines.append(f'**Features drifted:** {d[\"n_features_drifted\"]}/{d[\"n_features_checked\"]}')
              for feat in d['features']:
                  flag = '⚠️' if feat['drift_flag'] else '✅'
                  lines.append(f'- {flag} {feat[\"feature\"]}: PSI={feat[\"psi\"]:.4f}')

          if retrain_files:
              with open(retrain_files[-1]) as f:
                  r = json.load(f)
              lines.append(f'\\n**Retraining triggered:** {r[\"drift_report\"][\"any_drift_detected\"]}')
              lines.append(f'**Promotion decision:** {r[\"promotion_decision\"]}')
              lines.append(f'**Reason:** {r[\"reason\"]}')

          body = '\\n'.join(lines)
          print(body)
          with open('/tmp/summary.txt', 'w') as f:
              f.write(body)
          ")

          if [ -f /tmp/summary.txt ]; then
            summary=$(cat /tmp/summary.txt)
            # Find most recent open PR to comment on
            pr_url=$(gh pr list --state open --limit 1 --json url --jq '.[0].url // empty')
            if [ -n "$pr_url" ]; then
              gh pr comment "$pr_url" --body "$summary"
            else
              # Create a tracking issue
              today=$(date +%Y-%m-%d)
              existing_issue=$(gh issue list --label retraining --state open --json number --jq '.[0].number // empty')
              if [ -n "$existing_issue" ]; then
                gh issue comment "$existing_issue" --body "$summary"
              else
                gh issue create \
                  --title "Retraining Report — $today" \
                  --label retraining \
                  --body "$summary"
              fi
            fi
          fi

  # Optional: upload to S3/GCS if MODEL_STORE_TYPE is remote
  upload-artifacts:
    if: ${{ vars.MODEL_STORE_TYPE == 's3' || vars.MODEL_STORE_TYPE == 'gcs' }}
    needs: detect-and-retrain
    runs-on: ubuntu-latest
    permissions:
      id-token: write
      contents: read
    steps:
      - name: Checkout repository
        uses: actions/checkout@v4

      - name: Configure AWS credentials (OIDC)
        if: ${{ vars.MODEL_STORE_TYPE == 's3' }}
        uses: aws-actions/configure-aws-credentials@v4
        with:
          role-to-assume: ${{ vars.AWS_ROLE_ARN }}
          aws-region: ${{ vars.AWS_REGION || 'us-east-1' }}

      - name: Upload models to S3
        if: ${{ vars.MODEL_STORE_TYPE == 's3' }}
        run: |
          aws s3 sync models/ s3://${{ vars.MODEL_BUCKET }}/${{ github.repository }}/
          aws s3 sync reports/ s3://${{ vars.MODEL_BUCKET }}/${{ github.repository }}/reports/