Same shape, but plan for the cold CDN
Everyone's diagram assumes the CDN is warm. Mine assumes it just restarted.
Day one, config push that flushes the edge, new PoP coming online: for a while your "92% absorbed" is 0% absorbed, and whatever's behind the edge eats the difference. You can't size the origin for the full 50k (that defeats the whole design), but you can decide how gracefully it degrades. So: one extra app.large (four total, 52% at normal peak instead of 69%) and a third store replica (15k/s of origin read capacity instead of 10k). Under a cold edge the store saturates at 3x oversubscription instead of 5x, which is the difference between "slow for everyone" and "down for everyone" while the edge refills. The hot set refills fast, that's what makes it hot.
Costs $2.49/hr against kenji's $2.07. I'm paying 42 cents an hour for a softer worst Tuesday. Cheap for what it buys, and I'll take the cost-axis haircut.
3 Comments
Sign in to join the discussion.
- David Cho@david_cho
you buy resilience before the incident or you buy it during. before is cheaper
- Elena Petrova@elena_petrova
the honest framing is peak-vs-average for the ORIGIN, not the edge. normal peak your extra capacity idles at 52%. cold-edge peak it's the only thing keeping the store under 3x. sizing for the second distribution is a legit school of thought
- Sticky Session@sticky_session
lived this in 2019. cdn config push invalidated everything, origin was sized for misses only, site was down 40 minutes while the edge rewarmed. nobody models the cold edge until it happens to them. 42 cents is nothing