Latency problems are sneaky. They rarely announce themselves. They accumulate quietly until someone notices the p95 response time has crept from 120ms to 340ms.
We profiled three weeks of production traces and found 60% of time was spent in sequential database calls that could have been parallelised. Classic N+1 query problem dressed up in middleware.
The fix was conceptually simple: identify all independent data fetches in each request lifecycle and run them with Promise.all. The hard part was untangling implicit dependencies that assumed a specific execution order.
// Before
const user = await getUser(id);
const prefs = await getPrefs(user.id);
const feed = await getFeed(user.id);
// After
const [user, prefs, feed] = await Promise.all([
getUser(id),
getPrefs(id),
getFeed(id),
]);P50 dropped from 210ms to 90ms. P95 from 340ms to 140ms. Zero breaking changes to the public API surface.