01 / Full-stack platform
LiveFeb 2026 – now · 982 commitsLiveCity.cam
A multilingual directory of live webcams worldwide — beaches, cities, airports, volcanoes, aurora. Started as a hobby, grew into a production platform with real search traffic.



- live cameras
- 1,500+
- locales incl. RTL
- 13
- localised pages
- 30k+
- Google clicks, 28-day period
- +78%
live cameras
locales incl. RTL
localised pages
Google clicks, 28-day period
The problem
Serve ~30,000 localised pages from a small Postgres instance and a Vercel budget — while streams go offline every few days and crawlers hammer the site.
Stack
- Next.js 16
- React 19
- TypeScript
- Supabase / Postgres
- Vercel ISR
- YouTube Data API
- Python batch pipeline
- Leaflet + CARTO maps
Engineering decisions
- 01ISR strategy: English prerendered at build, other locales on demand, 30-day revalidation and tag-based invalidation after content deploys.
- 02Catalogue cached in 200-row chunks after the list silently exceeded the 2 MB data-cache limit and production showed zero cameras.
- 03Homepage payload cut from 219 KB to 61 KB gzipped (3.6×) by moving the full camera list to a CDN-cached API; TTFB ~1 s → 0.3–0.5 s.
- 04Disabled unthrottled Link prefetch and extended revalidation — daily Vercel usage down ~35%.
- 05Blocked a scraper responsible for 71% of requests at the Vercel firewall after it exhausted the database IO budget.
- 06Python batch pipeline: YouTube Data API checks for duplicates, live status and embeddability, preflight before insert, translation-completeness verification across 13 locales.
- 07Found pages that had revalidate but no generateStaticParams — /country, /live-webcams, /blog, blog posts and multi-screen ran a function on every view instead of being cached. Fixed; production now returns x-vercel-cache: HIT.
- 08Automatic failover: when the YouTube player reports a stream as gone, the tile switches to a backup camera in the same time zone, so a dead stream needs no deploy. Tested locally against a real dead stream.
- 09Search driven by data, not guesses: hand-written titles for low-CTR pages chosen from real Search Console queries (one city page went from 8.3% to 23.5% CTR), and noindex on ~15k thin single-camera area pages to focus crawl budget on strong ones.
- 10Stream health check across 1,620 YouTube streams: 86 dead streams swapped to their new live IDs (37 cameras back online after weeks offline), 25 with no live source taken offline.
- 11Shipped in one week: reusable event hubs (New Year's Eve with 37 cities in midnight order, Christmas markets), a random-camera page with swipe and a full-screen Reels mode on phones, and hand-written city guides in 13 languages.
- 12Postmortem — a cache stampede we caused: purging the catalogue tag twice in one day made every page refetch 1,500+ heavy rows at once. ~169k requests stalled the database, and functions waiting up to 300 s billed ~6× a normal day in provisioned memory. Fixes shipped: a lean catalogue rebuilt with one fetch (heavy fields load per page), a ~10 s database timeout to fail fast, at most one purge a day, batched, tested deploys — and functions moved from Washington to Frankfurt, next to the database, so queries no longer cross the Atlantic. The fix passed its first real purge: sampled pages answered in 0.15–0.7 s with no errors in the logs.
Outcome — Grew from ~190 to 1,500+ cameras in two months through a semi-automated pipeline. Approved for Google AdSense; organic traffic growing month over month.












