A video platform burning bandwidth on thumbnails nobody re-watched

A video platform burning bandwidth on thumbnails nobody re-watched

A short-video app generated six thumbnail sizes per clip and served them straight from edge to origin on every miss. With 1.2M clips and a long tail, most thumbnails were cold, so misses constantly punched back to a single bucket in Oregon. Origin egress ran $5,900/month and Asian users waited 900ms for grids to fill.

The issue: a flat edge with no middle layer for the long tail.

What we changed: added a regional tier-1 cache between edge POPs and origin. Tokyo and Singapore POPs now checked a shared Asian mid-tier before crossing the Pacific, so one cold thumbnail warmed the whole region, not each POP.

Origin egress dropped to $1,300/month. Grid load time in Jakarta fell from 900ms to 210ms. Mid-tier hit ratio settled at 88%.

The number that mattered: origin egress down 78%.