World of Warcraft market data for 184 realms, connected across a web app, a desktop companion, and an in-game addon.
14,000+ active users in 30 daysAug 17–Sep 15, 2026, Google Analytics
I build and operate the path from upstream auction data to the page a player uses to decide what to buy or sell: collection, storage, APIs, caching, the web app, a Tauri desktop companion for Windows, and the in-game addon it installs. The platform runs as 17 services on Railway.
An upstream schema break must not look like an empty auction house. The collector distinguishes missing or malformed auction data from a valid empty result, preserving the difference between “nothing is listed” and “we could not read the market.”
A popular query can arrive many times while the cache is cold. The Redis layer coalesces misses, validates selected cached shapes with Zod, and uses bounded failure paths. Cache behavior is part of the API contract, not just a speed optimization.
Blizzard’s API covers retail across 184 connected realms in four regions, the TSM API covers the classic editions, and for WoW Forever the desktop companion uploads players’ own auction scans. Each edition has its own TimescaleDB instance, and collection follows each source’s cadence rather than pretending every market updates continuously.
Time-boxing server-side prefetch cut cold item pages from 9.8 s to 0.64 s. In September 2026 I traced a spike of 503 errors to an index that blocked PostgreSQL’s in-place updates and a rollup that stalled Node’s event loop for up to 16 seconds. After the fix, a 600-request sample showed p99 latency down from 36.9 s to 2.6 s and 503s from 22 to 0.
More than 66,000 tests run in CI, but most changes touch a small part of the code. In a replay of 110 failed CI runs, predictive test selection ran a median 13.7% of the test files and still caught 94 of the 110 failing test files. How it works.
One API serves the web app, the desktop companion, and the in-game addon.
Freshness, cost, and responsiveness pull in different directions. I use source-aware collection and layered caches so the user can get a useful answer without every request repeating the same upstream work.
More engineering work