A reader opens the file and asks one of the demo's questions. Bytes and requests are
what this page's reader fetched from R2 on 2026-11-14, averaged over each graph's questions.
A GB here is 230 bytes.
Prices, checked 2026-11-14:
- Cloudflare R2: $0.015 per GB a
month after 10 GB free, reads $0.36 per million after 10 million free, no charge for data
out. The free allowance is per account; here nothing else is assumed to use it.
- Amazon S3 Standard, us-east-1: $0.023 per
GB a month, $0.0004 per 1,000 GET requests, data out $0.09 per GB after the first 100 GB a
month, less past 10 TB.
- Hugging Face buckets, checked 2026-11-18: public
storage is free on a best-effort basis. Past "the first few gigabytes" Hugging Face asks you
to upgrade to PRO, $9 a month, which includes up to 10 TB of public storage
(storage limits). It names no
number, so files up to 5 GB are priced on a free account. Data out is included up to 8 times
what you store; it does not say over what period, or what happens past that. A bucket's
download link redirects to Hugging Face's CDN and this page's reader follows it: the hermetic
books' questions read the same bytes as from R2 and took 1.5 to 2.4 times as long.
- GitHub Pages:
free for a public repository, a 1 GB site and a soft limit of 100 GB of traffic a month. Git
refuses files over 100 MiB.
- Zenodo: free, up to 50 GB
per record.
- AWS Graviton machines, on demand,
us-east-1, 730 hours a month: $0.0042 an hour for 0.5 GiB of memory up to $0.1071 for 16 GiB,
each with 2 vCPUs. Disk (gp3) is $0.08 per GB a month. Data out is charged as for S3.
The servers are QLever (commit
18c097c, 2026-11-14) and Oxigraph 0.5.13. On
2026-11-15 we loaded each graph into both, from the same statements, and asked them the demo's
questions. Each ran in Docker with 2 CPUs and 4 GB of memory, and each question returned as many
rows as it does from the file. A server's price comes from what it used:
- the smallest machine whose memory holds the server's peak while answering, its index and
512 MiB for the system, because the answers were timed with the index in memory;
- a disk of 8 GB plus the index the server built, in whole GB;
- each reader's answer sent out once.
At a million readers a month a server gets a question about every 2.6 seconds. The slowest
answer we measured took 0.95 s, so one machine keeps up. We did not measure busier traffic,
so the page stops there.
The OpenCitations Index has no server price: 15.8 billion statements is far past what fits in
those runs, and we show only what we measured. The scripts and every number are in
experiments/server-bench.
Not counted: the time it takes someone to run a server, backups, and building its index.