NimbusCast.

NOTES

1. 1

   Decision

   I wrote one Express app and exported it as a Vercel serverless function, with a small dev.ts that runs the same app on a port locally. Production and local development share the same routes, and every /api route goes through a per-IP rate limiter first.

2. 2

   Trade-off

   I let Vercel reuse a weather answer for 5 minutes, since the data barely changes in that time. Repeat searches come back about 2.5 times faster and don't use my free API quota. The trade-off is that the weather can be up to 5 minutes old.

3. 3

   What I got wrong

   At first the whole page waited on a single loading flag, so the weather stayed hidden until the third-party poem API answered. Now the weather renders right away, the poem loads on its own, and a failed refetch keeps the last good data on screen.

4. 4

   Failing gracefully

   The ranker asks PoetryDB for 5 random poems and picks the one that uses the weather word most often, so 'clouds' also matches 'cloud'. If PoetryDB fails or has no match, it serves one of 10 local poems instead of an error.

5. 5

   Search that doesn't race

   Suggestions wait 300 ms after you stop typing. Each new keystroke cancels the previous request, so an old, slow response can't overwrite a newer one.

RESULTS

REPEAT SEARCH, FROM VERCEL'S EDGE CACHE

0.1s

MEDIAN FIRST SEARCH (3 UPSTREAM CALLS)

0.27s

UPSTREAM CALLS ON A REPEAT WITHIN 5 MIN

0