AuthenticationSoftware Engineering

From a Rust Benchmark to an SSO Platform for 500,000 MAU

How AuthOS grew around a small Rust core, and what its measured budget-VM throughput means under an explicit 500,000-MAU traffic model.

Diagram connecting a Rust SSO core to web, mobile, and cloud clients
Lead image Diagram connecting a Rust SSO core to web, mobile, and cloud clients
On this page

Audio edition

0:00 / 6:30

Ready

The benchmark left me with a number I could use: on a one-vCPU, 1-GiB VM, AuthOS sustained 1,968 operations per second for five minutes with no failures, a 154 ms p99, and 68% average CPU.

From there, I had two jobs: give applications a clean boundary around the Rust API, and translate operations per second into a monthly-active-user scenario.

From Auth Service to Platform

The load-tested service began with a Rust API, SQLite storage, and OAuth device authorization for clients that cannot complete a browser redirect. I kept that core as one deployable process and built the rest of AuthOS around it.

AuthOS now hosts the login journey itself. It resolves the organization and service, presents the available identity providers, and handles password login, magic links, passkeys, MFA, recovery, and trusted devices. The account-security portal gives users one place to manage passkeys, MFA, backup codes, and active sessions.

The tenant layer includes organizations, teams, roles, permissions, services, plans, subscriptions, custom domains, branding, platform-owner administration, and server-side session revocation. Enterprise integrations add social and upstream OAuth/OIDC providers, SAML IdP SSO, SCIM provisioning, service credentials, OAuth device authorization, managed token exchange, webhooks, audit export, and platform analytics. Sensitive provider, TOTP, webhook, SAML, and connected-account material is stored in record- and field-bound AES-256-GCM envelopes with an explicit key-rotation path.

The SDK Became the Integration Boundary

The repository now exposes the platform through four main source areas:

AuthOS/
├── api/               # Rust API and database backends
├── lite-web-client/   # Hosted setup, login, and account-security UI
├── sso-sdk/           # Framework-agnostic TypeScript SDK
└── packages/          # React, Vue, Node.js, and CLI packages

The core SDK handles authentication flows, session persistence, token refresh, and the multi-tenant API. React and Vue adapters bring the same boundary into browser applications; the Node.js package and CLI cover server-side and operational work. Applications redirect users to the hosted AuthOS journey instead of rebuilding provider selection, recovery, passkeys, or MFA themselves.

This arrangement keeps authentication policy in the Rust service while giving each application a typed client.

From Operations per Second to 500,000 MAU

MAU is a count of people seen during a month, while the benchmark measures operations completed each second. Connecting the two requires assumptions about how often those people are active and how bursty their traffic becomes.

I used this scenario:

  • Each MAU is active for 10 minutes per day.
  • While active, the client causes one AuthOS operation every 10 seconds.
  • A month has 30 days.
  • Peak traffic reaches four times the monthly average.

That produces 60 operations per user per day, or 1,800 per month. At 500,000 MAU:

500,000 users × 1,800 operations = 900,000,000 operations/month
900,000,000 ÷ 2,592,000 seconds = 347 average operations/second
347 × 4 peak factor = 1,389 peak operations/second

At 1,389 operations per second, the 4× peak uses 70.6% of the VM’s sustained reference and leaves about 29% throughput headroom. Under those assumptions, 500,000 MAU produces a plausible traffic load for the budget-shaped VM. An 8× peak would require 2,778 operations per second—141% of the reference and inside the measured saturation range. A product with more frequent authentication, sharper bursts, or more write-heavy flows needs a lower estimate; 250,000 MAU at 8× produces the same 1,389 operations per second as 500,000 MAU at 4×.

The calculation models traffic from a five-minute reference run; it does not measure month-long reliability.

The earlier article also suggested a range reaching two million MAU. This budget-VM model does not support it: two million users at the same activity level and a 4× peak would require about 5,556 operations per second, nearly three times the sustained reference rate.

What the $10 Budget Buys

The guest used one vCPU, 1 GiB of RAM, and a 25 GiB virtual disk—the advertised dimensions of DigitalOcean’s $6 Basic Droplet on 17 July 2026. The benchmark still ran on my i7-1265U, local NVMe, and direct host networking, so the price supplies context rather than a DigitalOcean performance claim.

AuthOS brings more than the binary to that VM: the repository includes a systemd installer, Docker deployments, health and readiness endpoints, protected Prometheus metrics, backup and restore procedures, upgrade and rollback guidance, and application-key rotation tooling. The benchmark also caught one deployment setting worth carrying into production: high concurrency needs an explicit file-descriptor limit because the default 1,024-descriptor shell limit failed before CPU or memory became scarce.

Why the Small Architecture Matters

AuthOS ships the Rust API, hosted web client, SDK packages, standalone Linux bundles, and Docker images from one open-source repository . The API can be built for SQLite, PostgreSQL, or MySQL. The budget benchmark used the SQLite release binary: the same process that runs authentication, tenant administration, operational jobs, audit delivery, and the enterprise integration routes.

The workload measured subscription reads, user reads, and device authorization rather than every AuthOS journey. Even so, the full AuthOS process fit comfortably inside one small VM at the 2,000-VU operating point, giving me no performance reason to split the platform into more services as its feature set grew.

Under the stated activity assumptions, 500,000 MAU fits below the measured operating point at a 4× peak but exceeds it at an 8× peak.

Continue reading

Complete index →