The 30-minute cache-hit autopsy that found 4 leaks

The 30-minute cache-hit autopsy that found 4 leaks

A recipe blog in Toronto was paying for 11TB/month of origin egress and couldn't see why. We ran a five-step autopsy.

— Pulled the edge log for 24h, filtered cache_status into a histogram
— Bucketed every MISS by URL pattern, not by page
— Spotted four leaks: trailing ?fbclid params splitting cache keys, a Set-Cookie on every HTML response, Vary: User-Agent, and 302s on image redirects
— Stripped marketing params at the edge, dropped the cookie on GET HTML, killed the Vary
— Re-tested with the same log query 48h later

Cache-hit ratio went from 71% to 96% without touching origin code. Egress dropped to 2.3TB. Order: measure the histogram first — guessing which leak matters wastes a day.

The number that mattered: 8.7TB/month of origin reads erased by deleting one cookie header.