One-star reader
What people actually hate about the apps they use every day.
Apple publishes a public feed of one- and two-star App Store reviews. This site
reads them, groups them into the complaints that keep coming back, and prints a
real count and a real quote beside each one. Day 5 of a 26-day run.
Live at onestarreader.onedaybuilt.com.
The shape of it
| |
|---|
| Watched | ~11,800 apps, from every genre chart in four storefronts |
| Read | the apps whose review feed a collector has reached |
| Storefronts | United States, Germany, Britain, France |
| Model | claude-haiku-4-5, Batch API, for grouping and translation only |
A watched app has a name, a genre and a score. A read app also has its
complaints. The gap between those two numbers is the whole engineering problem
below.
Two Apple endpoints, two different rations
Both were measured, not assumed. Both measurements are in
src/lib/apple.ts and src/lib/schema.ts next to the code they explain.
The review feed hands out fifty reviews a page, ten pages deep — a hard
ceiling of 500 per app per storefront — and rations that to roughly one page per
address every ten minutes. A throttled request returns HTTP 200 with an empty
feed, not an error, so a site that trusts it publishes "no complaints" when it
means "we were blocked". Every attempt is recorded, answered or not.
The lookup endpoint gives a score, a rating count and a version. It answers
one request instantly from anywhere, including datacentre addresses. Asked
45,000 times from one address it refuses about three in four, and asking more
slowly does not help: five a second and forty a second were refused at the same
rate. The lever is the address, not the pace.
So neither feed is collected from Vercel. Both run as scheduled GitHub Actions,
four shards each, because a fresh runner is a fresh address:
.github/workflows/collect.yml — reviews, every 15 minutes, 14 pages a shard.github/workflows/ratings.yml — scores, hourly, 900 pairs a shard
watch holds the app/storefront pairs worth asking about — where the app
actually charted, plus all four storefronts for anything somebody requested.
That is 20,124 pairs instead of the 47,192 a blind cross join produces, and
27,000 questions a night that could never have had an answer.
Slugs are stored, never derived
apps.slug, unique index, filled by fillSlugs(). Five apps here are called
McDonald's and 62 have names written entirely in Chinese, Arabic or Japanese
that reduce to an empty string. Deriving a URL from a name in the browser — as
the search box did — sends a reader to another company's page.
The database has a cliff in it
Neon's free plan allows 5 GB of egress per project per month. Past it the
whole project answers HTTP 402 — reads, writes, psql, pg_dump — until the
first of the next month, and the data cannot be exported to escape.
This site walked into it on 12 September with 5.95 GB in ten days. The cause
was snapshot() reading the whole database on every cold render: 11,798 apps,
40,000 rating rows and every quoted review, about 14 MB, to draw two hundred
pages that needed a fiftieth of it.
Every query in that snapshot is now anchored to the apps that actually have
complaints, and review bodies are clipped in SQL rather than in the browser —
roughly 14 MB down to 2.5 MB. A watched app's page is one small query of its
own (watchedApp()), because eleven thousand pages nobody has asked for should
not be held in memory so that a handful can be visited.
Check the number with:
neon api "/projects/holy-bread-39202992"
data_transfer_bytes against 5e9, synthetic_storage_size against 0.5e9.
Costs
Free except the model. Grouping and translating everything read so far has cost
about $0.63 in total, at Batch API rates. No model is ever called while
somebody waits; the site only reads what the cron has already written.
Running it
pnpm install
cp .env.example .env.local # DATABASE_URL, ANTHROPIC_API_KEY, CRON_SECRET
pnpm db:push
pnpm dev
Scripts import server-only, so they run under
tsx --conditions=react-server:
| Script | What it does |
|---|
scripts/import-charts.mts | walks 224 genre charts, fills apps and watch |
scripts/ratings.mts | one turn of the score rotation |
scripts/collect.mts | one shard of the review collector |
scripts/cluster.mts | groups complaints via the Batch API (--stale for new reviews only) |
scripts/translate.mts | translates the quoted reviews |
scripts/slugs.mts | gives new apps their URL |
House rules that apply here
packageManager is pinned to [email protected] and pnpm.onlyBuiltDependencies
is set — without both, Vercel guesses a pnpm version and the install can exit 1
on a repo that builds locally every time. Neon is in Frankfurt and
vercel.json pins fra1; a mismatched region puts a transatlantic hop on
every query. Reads and writes go through a server-only module.
