Skip to main content
Sportz tests both halves of the system. The backend has 95 Vitest tests across three layers; the frontend has a Playwright end-to-end suite that drives the real UI in a browser. Each layer answers a different question and runs against a different level of realness.

​
The test pyramid, as built

​
Unit tests: validation & status logic

The pure functions: getMatchStatus (derives scheduled/live/finished from timestamps, including boundary cases like “now exactly equals start”) and the Zod schemas (valid + invalid inputs, coercion, the endTime > startTime rule). No DB, no network: fast and deterministic.

​
Integration tests: routes against real Postgres

These drive the real Express app with Supertest, against a real database (not a mock). Mocking the DB would mean testing the mock, not the SQL.
  • createApp() is built without server.listen() so Supertest can drive it in-memory (see Backend).
  • resetDb() runs TRUNCATE ... RESTART IDENTITY CASCADE in a beforeEach, so every test gets a clean, deterministic database with IDs starting at 1.
  • The WS broadcast functions are replaced with vi.fn() spies to assert the route calls them correctly, without needing a live socket.
The one thing mocked: Arcjet. Its protect() makes a real network call; a setup file (tests/setup/mock-arcjet.ts) mocks it to always allow (and sets test env vars before any import, because arcjet.ts throws at import time if ARCJET_KEY is missing).

​
WebSocket tests: real socket pairs

These start a real http.Server on port 0 (OS-assigned free port), attach the real WS server, and connect real ws clients. They verify: welcome on connect, subscribe/unsubscribe confirmations, malformed-JSON handling, room-scoped broadcast (subscribers of match 1 get it, subscribers of match 2 and unsubscribed clients don’t), and the Arcjet deny path (overriding the global mock for one call). Two real bugs were found by writing these tests: a message race and a missing close() causing hung timers. Both are post-mortemed in Issues.

​
End-to-end tests: Playwright against the real UI

These drive the actual sportz-ui in a real Chromium browser: load the page, click “Watch Live,” assert the commentary panel fills, toggle dark mode, page through matches, scan for accessibility violations. They verify the thing the backend tests can’t: that a user sees and does the right thing.
  • Mocked at the network boundary. REST is stubbed with page.route, the WebSocket with page.routeWebSocket. List/scale data is generated from a template (Array.from({ length: 8 }, ...)). So the suite is deterministic and offline: it never depends on the cold free-tier backend, whose latency would otherwise silently change what renders.
  • Runs against a production build, on its own port. webServer runs npm run build && npm run start -- -p 3100 with reuseExistingServer: false. next dev compiles routes on first request, so under fullyParallel several cold compiles serialize and blow past the 30s test timeout, producing flaky failures that shift between runs and vanish when you run files individually. A prod build serves pre-compiled routes instantly (reliable in parallel) and is the exact build CI and Vercel ship. The dedicated port (separate from next dev’s 3000) lets the dev server and test server coexist with no collision.
  • Accessibility is part of the suite. @axe-core/playwright scans the settled page (after the entry animation reaches opacity: 1, to avoid mid-animation false positives) for wcag2a/wcag2aa violations. Writing this caught two real contrast bugs: a theme-token that wasn’t theme-aware and a status pill that broke the established -700-on--50 pattern (see Issues).
  • First-run UI is kept out of the way. The onboarding tour auto-opens on a first visit, and Playwright starts each spec with fresh localStorage, so it would pop up in every test, cover the page with an overlay, and break the a11y scan + Watch-Live clicks. Two guards: the app skips the auto-tour under automation (navigator.webdriver), and the mock setup pre-sets the “seen” flag via page.addInitScript (deterministic belt-and-suspenders). The lesson generalizes: any first-run popup/overlay must be suppressible in automated runs.
The diagnostic lesson worth keeping: failures that move between runs, pass individually, and are all timeouts point at an environment/concurrency problem, not broken test logic.

​
Deployed smoke: the live stack (post-deploy)

One spec (e2e/smoke.deployed.spec.ts, tagged @deployed) runs against the real deployed stack: the Vercel frontend talking to the Render backend, no mocks. It answers the one question mocking can’t: is the deployed system actually wired together? Four read-only checks: the app loads → a real match renders (REST + CORS + baked API URL) → the WebSocket connects (the wss:// upgrade through Render’s proxy) → Watch Live opens the commentary panel.
  • It’s the top rung of the realness ladder (mocked → … → deployed). Each rung answers its own question; we don’t make one tier carry two jobs.
  • Read-only and presence-based. It never mutates production and never asserts exact data (real data changes); it only checks that things are present and connected. Generous timeouts absorb Render’s free-tier cold start.
  • Excluded from CI and the normal suite. npm run test:e2e uses --grep-invert @deployed; the deployed spec runs only via npm run test:e2e:deployed (which sets PLAYWRIGHT_NO_SERVER=1 + PLAYWRIGHT_BASE_URL). It’s a post-deploy check, not a per-push gate: it hits live prod and depends on a cold server. Automating it after a Vercel deploy is a documented next step (DevOps).

​
Running the tests

CI runs the backend suite against a postgres:16-alpine service container: fast, free, deterministic, and identical to the local throwaway DB. See DevOps.

​
Not yet built

  • Contract testing between frontend and backend types.
  • Load testing (k6/Artillery) of the WS server under thousands of concurrent subscribers, relevant before any real scale.
  • Visual regression, relevant once the UI stabilizes.