Cost optimization

Olympus's value prop is "no per-MAU pricing." But your own infrastructure costs scale with usage. Here's where the money goes and how to trim.

Baseline costs (single-host, ~50k MAU)

At 50k MAU, that's < $0.002 per user/month. Compare to Auth0 at $0.05+/MAU.

Where costs grow

MAU → DB size

Each identity is ~5KB in DB (encrypted traits, credentials, sessions).

• 100k identities = ~500 MB.
• 1M identities = ~5 GB.

Plus indexes (~30% overhead). Plus audit logs.

At 1M users with 1y audit retention, expect ~50 GB DB. Still fits a $50/mo VPS.

MAU → Email volume

Average verification + recovery + notifications: ~2 emails per active user per month.

• 50k MAU = 100k emails/mo (~$10-15).
• 1M MAU = 2M emails/mo (~$200-400).

Email becomes the biggest line item at scale.

MAU → Bandwidth

Each login session: ~50KB total (HTML, JS, CSS, API calls). Cached after first visit.

• 50k MAU at 4 logins/mo = 200k page loads × 50KB = 10 GB.
• VPS bandwidth allowance usually covers this.

Auth fraud

Bots probing your registration endpoint at 100 req/s 24/7. Costs:

• DB rows (if you don't filter): millions of failed registrations.
• Bandwidth.
• Compute (Argon2id on each).

Mitigation: aggressive rate limiting + CAPTCHA + bot detection.

Tactical savings

Right-size your VPS

Most Olympus deployments run at 5-15% CPU. Don't over-provision. Start small, monitor, scale up if needed.

Hetzner CAX21 (8 GB / 4 vCPU ARM): $9/mo. Plenty for < 50k MAU.

ARM instances

ARM is ~30% cheaper for the same performance on Hetzner Cloud. Olympus Go services (Kratos, Hydra) run native ARM. Node/Bun (Hera, Athena) also work on ARM.

# docker-compose.yml, image tags
image: oryd/kratos:v1.2.0  # multi-arch

Backup tier

S3 Standard: ~$0.023/GB/mo. S3 Glacier: ~$0.004/GB/mo.

For backups > 30 days, move to Glacier. Retrieval takes hours but you don't need fast access, these are emergency backups.

aws s3 cp olympus-backup.dump.gz s3://your-bucket/ \
  --storage-class GLACIER_IR

Email provider

Postmark: $1.25 per 1000 emails (consistent, reliable). Brevo (free tier): 300/day = ~9k/mo free. SendGrid (free tier): 100/day = ~3k/mo free.

For < 10k emails/mo, use free tiers. At scale: Postmark for reliability.

DNS

Cloudflare free tier: unlimited DNS, basic DDoS, proxying. Route 53: $0.50 per zone/mo + $0.40 per million queries.

For Olympus, Cloudflare free is plenty.

Monitoring

• Grafana Cloud free tier: 10k metrics + 50 GB logs/mo. Plenty.
• Prometheus self-host: $0 but operator overhead.
• Datadog: $$$. Don't unless you're already on it.

Storage

If DB > 100 GB and growing:

• Migrate to managed (Neon, RDS) for auto-scaling.
• OR archive old audit logs to S3 (cold storage).

Strategic savings

Lazy-delete inactive identities

If a user hasn't logged in 5 years, you probably don't need their data. GDPR requires data minimization anyway.

-- Identify inactive
SELECT id FROM identities
WHERE last_login < NOW() - INTERVAL '5 years'
  AND state = 'active';

-- Delete (with proper notification)

Cuts DB size, audit log volume, email volume.

Audit log retention

90 days for routine, 2 years for security events, then drop.

# Daily cron
psql -c "DELETE FROM security_audit WHERE event_type IN ('login', 'whoami') AND created_at < NOW() - INTERVAL '90 days'"

A million-MAU deployment without retention will have 100M+ rows in audit log within a year.

Session pruning

Same, old expired sessions accumulate:

podman exec ciam-kratos kratos janitor --keep-last 168h

CDN for assets

If you're paying for bandwidth, CDN the static assets (JS/CSS):

• Bunny.net: ~$0.005/GB.
• Cloudflare R2 / S3: ~$0.01/GB.

Frees VPS bandwidth for API calls.

Postgres tuning

Default Postgres config is conservative. Tune for your workload:

shared_buffers = 4GB           # 25% of RAM
effective_cache_size = 12GB    # 75% of RAM
work_mem = 16MB
maintenance_work_mem = 1GB

Reduces query times, lets you serve more on smaller instance.

See Postgres tuning.

When to scale up

Signals:

• CPU > 70% sustained → bigger instance.
• DB > 80% disk → bigger disk OR managed Postgres.
• Latency p99 > 500ms → bigger instance OR optimize queries.

Don't pre-scale. You'll know when it's needed.

When to scale out

Single-host is fine up to ~500k MAU. Beyond that:

• Add read replica.
• Add second app host with shared DB.
• Eventually: managed Postgres, multi-region.

See Multi-region deployment.

Compare to alternatives

Auth0: $0.05 per MAU (B2C) or $1.50 per MAU (B2B). At 100k MAU = $5k/mo or $150k/mo. Okta CIC: similar. AWS Cognito: ~$0.0055/MAU. At 100k MAU = $550/mo.

Olympus at 100k MAU: ~$80/mo infrastructure. Operator time: a few hours per quarter.

The savings are real and significant, but they're not free. You're trading dollars for operator time. Worth it if you have the time. Not worth it if you don't.

Related

• Operate, Postgres tuning
• Operate, Backup and restore
• Cookbook, Managed Postgres (Neon)
• Operate, Logs and observability