e2e.yml

# The J1-J9 browser journeys, on a schedule rather than on every pull request.
#
# WHY NOT ON PRs. The suite takes ~36 minutes against a fourteen-container
# stack, needs exclusive use of host port 8988, and must run as a single
# Playwright process against one clean appliance. None of that belongs on a
# change-by-change gate — tests.yml stays fast and this runs nightly, with
# workflow_dispatch for when someone wants it now.
#
# THE CONSTRAINTS BELOW ARE THE TEST, NOT PREFERENCES. They are set in
# e2e/playwright.config.ts and e2e/e2e.env and are NOT overridden here:
#
#   workers: 1, retries: 0, fullyParallel: false
#       The journeys share one appliance. Parallel workers would fight over
#       cameras, relay paths and recordings; retries would hide the flake
#       instead of fixing it.
#   one clean appliance
#       up.sh pins the compose project name (vms-e2e) so the stack has its own
#       containers, network and volumes, and down.sh -v cannot reach anything
#       else. The concurrency group below is what stops two runs overlapping.
#   no arbitrary sleeps
#       up.sh waits on service state with a deadline; the specs use bounded
#       expect.poll. Nothing here adds a sleep.
#   synthetic cameras / ONVIF / WS-Discovery, real product services
#       e2e/docker-compose.e2e.yml. DeepStream and Caddy are absent, no real
#       customer cameras are involved, and nothing is published to the host
#       except MediaMTX's HLS port, which the SPA hardcodes.
#
# ─────────────────────────────────────────────────────────────────────────────
# KNOWN LIMITATION — READ BEFORE TRUSTING A FIRST RUN ON A HOSTED RUNNER
#
# Smart Search does not bake its model weights into the image. Its Dockerfile
# says so explicitly ("Model weights are NOT baked in ... They download on first
# run into this directory, which is a volume in compose"), and the same is true
# of the plate localiser and OCR. On a developer appliance the `models` volume
# is warm, so this is invisible. On a FRESH runner it is not: the first start
# pulls CLIP plus the plate models — hundreds of MB — from Hugging Face and
# download.pytorch.org before ingest is ready.
#
# That means this workflow, as written, needs outbound network on its first run
# and is slower and more fragile than the rest of the suite. It has NOT yet been
# executed on GitHub Actions; the suite has only run on a developer appliance.
# Before relying on it, pick one:
#
#   • run it on a self-hosted runner whose `vms-e2e_models` volume is warm
#     (fastest, and the closest match to how the suite is actually used);
#   • cache the models volume between runs;
#   • or publish a pre-warmed smartsearch image to GHCR and pull it here.
#
# Until one of those is in place, treat a first-run failure in Smart Search
# start-up as an environment problem rather than a product regression. J8 is the
# only journey that needs the index; J1-J7 and J9 do not.
# ─────────────────────────────────────────────────────────────────────────────

name: e2e

on:
  workflow_dispatch:
  schedule:
    # 02:30 UTC nightly — after the weekly full-deps job's usual window and
    # well clear of working hours in IST.
    - cron: "30 2 * * *"

permissions:
  contents: read

# ONE E2E APPLIANCE AT A TIME, GLOBALLY. Not keyed on ref: two runs on
# different branches would still collide on the compose project name, host port
# 8988 and the relay's path table. `cancel-in-progress: false` because killing a
# run mid-journey leaves containers behind; let it finish and always tear down.
concurrency:
  group: e2e-appliance
  cancel-in-progress: false

jobs:
  journeys:
    runs-on: ubuntu-latest
    # Generous because the first start builds the api image (which builds the
    # SPA) and may warm the model cache. A hang should fail, not run forever.
    timeout-minutes: 120

    steps:
      - uses: actions/checkout@v4

      - uses: actions/setup-node@v4
        with:
          node-version: "22"
          cache: npm
          cache-dependency-path: e2e/package-lock.json

      # `ci`, not `install`: the lockfile is the contract, and a browser
      # version that drifts from it is a different test run.
      - name: Install E2E dependencies
        working-directory: e2e
        run: npm ci

      - name: Install Playwright browser
        working-directory: e2e
        # Chromium only — playwright.config.ts defines exactly one project.
        run: npx playwright install --with-deps chromium

      - name: Free disk before building the stack
        # The stack carries torch, OpenCV and Keycloak. Hosted runners ship
        # several GB of toolchains this build never touches.
        run: |
          df -h /
          sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc || true
          df -h /

      - name: Start the E2E appliance
        run: ./e2e/scripts/up.sh

      - name: Run J1-J9
        working-directory: e2e
        # No -g filter and no worker override: the whole suite, exactly as the
        # config defines it.
        run: npx playwright test

      # ── Diagnostics. `always()` so a failure is explainable, and a success
      #    costs one small upload. ───────────────────────────────────────────
      - name: Capture service logs
        if: always()
        run: |
          mkdir -p e2e/artifacts
          docker compose \
            --project-name vms-e2e \
            --env-file e2e/e2e.env \
            -f docker-compose.yml \
            -f docker-compose.bridge.yml \
            -f e2e/docker-compose.e2e.yml \
            logs --no-color --timestamps > e2e/artifacts/compose-logs.txt 2>&1 || true
          docker ps -a > e2e/artifacts/containers.txt 2>&1 || true

      - name: Upload Playwright traces, video and screenshots
        if: always()
        uses: actions/upload-artifact@v4
        with:
          name: e2e-playwright-results
          # retain-on-failure means a green run uploads almost nothing.
          path: |
            e2e/test-results/
            e2e/playwright-report/
          if-no-files-found: ignore
          retention-days: 14

      - name: Upload service logs
        if: always()
        uses: actions/upload-artifact@v4
        with:
          name: e2e-service-logs
          path: e2e/artifacts/
          if-no-files-found: ignore
          retention-days: 14

      # ALWAYS, and with -v. A runner that keeps containers or volumes poisons
      # the next run on a self-hosted box, and the volumes are project-scoped
      # so this cannot reach anything else.
      - name: Tear down the E2E appliance
        if: always()
        run: ./e2e/scripts/down.sh -v