pdl: a phone-first scoreboard for a weekly padel group

Problem

I play padel every week with the same group: me, Jói, Bjössi and Gísli. Padel is always two against two, so after a few weeks the obvious questions are who is actually winning, who is on a run and who should pair with whom tonight.

Nobody keeps that in their head. I wanted something I could open on my phone at the court, tap four names, mark the winners, and have the table update.

Constraints

• It is used on a phone, at the court. Logging a match has to take a few taps. It installs as a PWA with a standalone manifest and a service worker.
• Private, but no accounts. Four people don't need sign-up. The API sits behind one shared basic-auth password.
• The real scores must be hard to break. I develop against the same Cloudflare account the live app runs on, so local work can't be allowed to write to the group's data.
• Close to free to run. One Worker and one D1 database, on Cloudflare.
• A fair table. A player who has played twice and won twice shouldn't top someone who is 8 and 4. Ranking has to account for that.

Architecture

One Cloudflare Worker serves both the React app and the /api/* routes. Static files are served first. Anything that matches no file falls through to the Worker, which answers API calls and hands every other path back to the assets binding so deep links load the app.

The schema is two tables. Because a match is always 2v2, a match row holds four player columns, winner_1, winner_2, loser_1 and loser_2, instead of a join table. Standings aren't stored. The Worker reads every player and match and computes the table on each request in scoring.ts, a pure module with no I/O.

Scoring is one point per win. The current streak is counted backwards from the newest match. Players with at least five matches are ranked by win rate. Everyone else, and any comparison between a qualified and an unqualified player, falls back to points.

Types run end to end. Request schemas live in shared/types.ts and both sides import them. The Worker's route chain exports its own type, and the client is hc<AppType>(), so a renamed field breaks the build instead of the app. Worker bindings are inferred from the infrastructure file in alchemy.run.ts, so they can't drift from what is deployed.

Only the prod stage points at the live pdl Worker and database. Every other stage, including alchemy dev, gets its own stage-scoped copies.

What shipped

• Scores tab. Standings with played, wins, losses, points and streak, a match logger and match history. In the logger, each tap on a name cycles it through winner, loser and unselected. It only submits with exactly two of each.
• Tools tab. Pick four players, roll, and it shuffles them into two teams. Adding a player is a name field, capped at 40 characters.
• API. Six routes: health, list players, list matches, standings, add player, add match. The server rejects a match without four distinct players.
• Tests. 22 in total. 13 cover scoring and sort order. 9 hit the HTTP routes against a real local D1 (Miniflare) with the migrations replayed from disk, so the tests always run on the real schema.
• One gate. bun run check runs format, lint, typecheck, test and build in that order. CI runs it on every push to master and on pull requests.
• Structured logs. Each request emits one wide event through evlog. The allowed fields and event names are a closed type, so logging a new field means adding it to the schema first.

The project has 13 commits. Two v1 commits in April were the app. The August work moved it to Vite 8, React 19, StyleX and Base UI, and replaced wrangler.toml with Alchemy. I adopted the live Worker and database into Alchemy's state without recreating them.

What I'd do differently

• Make a missing secret fail the deploy from day one. The Worker treats an empty password as "no auth" so local dev works without one. That meant a missing secret would have shipped the app wide open. Prod now refuses to plan or deploy without PDL_PASSWORD.
• Test the routing I declared, not the routing I assumed. The config said single-page-application fallback, but asset-first routing sent unknown paths to the Worker, so the setting never ran and deep links returned 404 in production. Two tests now cover it.
• Read what a config option actually routes. I added runWorkerFirst: ['/api/*'] to be explicit. It sent everything else, including Vite's dev module runner, through the asset handler, and local dev died with Expected 101 status code. The default already did the right thing.
• Ship a way to undo a match. There is no edit or delete route. A mis-tapped result stays in the standings until someone fixes it in D1 by hand.
• Commit the icons the manifest promises. The manifest lists icon-192.png, icon-512.png, a maskable icon and apple-touch-icon.png, but only favicon.svg is in public/.
• Finish the custom domain. The app still lives on pdl.einargudni.workers.dev. The delegation runbook is written, but the old Vercel records for pdl block it.