Requirements and sizing
| Profile | For | Servers | Each server |
|---|---|---|---|
| Pilot | Proof of concept | 1 | 4 vCPU, 16 GB RAM, 200 GB SSD |
| Standard | A fintech or lender in production | 1 app + 1 database replica | 8 vCPU, 32 GB RAM, 500 GB SSD |
| High availability | Banks | 3 app nodes, PostgreSQL primary and replica, HSM | 8 vCPU, 32 GB RAM each |
These are starting points per install. The file redactor (speech and image recognition) is the heaviest service; leave it out if you don’t redact files.
Each install needs
- Linux with Docker Engine and Compose, in the country its data must stay in.
- A DNS name, and port 443 reachable by your people and systems.
- Outbound HTTPS only to destinations you configure. Onshore never calls home; a linked install needs outbound HTTPS to its primary.
- Swap disabled or encrypted.
- Three to five key holders.
- A backup target for the database, in the same country.
Measured performance
On a 2019 laptop with everything in Docker, through TLS (re-measure on real servers before relying on these):
| Measure | p95 | Throughput |
|---|---|---|
| Tokenize, batch of 10, one client | 18–20 ms | about 70 batches/s |
| Detokenize, batch of 10, one client | 7–8 ms | about 160 batches/s |
| Tokenize, batch of 10, 10 clients | 37–43 ms | 300–335 batches/s |
| Detokenize, batch of 10, 10 clients | 22–33 ms | 525–620 batches/s |
| Log processor, through the edge | — | about 29,000 lines/s |
| Detection, 1 KB on one core | 0.61 ms | — |
Last updated on