The whole thing costs less than a coffee per hour
Start with the bill and work backwards: $1.98/hr, and I'll defend every cent.
The data is 60M pastes at ~9KB, call it 540 GB. That fits ONE db.large with a terabyte and a half to spare, so it's a single shard, sharding none, and the only db money I'm spending is RF 3 quorum because the durable-link promise is the one non-negotiable in the contract.
Reads: the hot 1% of pastes is about 5 and a half GB, so a pair of cache.smalls (32 GB between them) holds the entire working set with room for the TTL churn. 300s TTL, which is exactly the staleness budget, not a second under it. Why give away hit ratio to be fresher than anyone asked.
The app tier is four app.smalls at $0.04 each. They front the full 1,515/s evening peak at 76% utilization, which is honest sizing, not tight sizing. Losing one at steady load leaves 3x more capacity than needed.
Unit economics check: call it 40M views a day for ~$48 a day all in. That's about a tenth of a cent per thousand views. If your pastebin design costs more than your coffee habit you built the wrong thing.
4 Comments
Sign in to join the discussion.
- Omar Farouk@omar_farouk
@tom_schneider i rounded up to be safe with money. old habit
- Tom Schneider@tom_schneider
checked it: 60M x 9KB is 540GB logical, x3 copies is 1.6TB physical spread over 3 instances with 2TB disks each. 3.7x of disk headroom on the shard, and the 30-day TTL means it only grows with adoption, not with time. pedantic note: it's $47.52 a day, not 48
- Zoé Laurent@zoe_laurent
the 300s TTL detail is the part i like. everyone reflexively sets 60 and pays for freshness the contract never asked for
- Marcus Lee@marcus_lee
under a cent per thousand views. that's the whole review honestly