Unikernels at Scale: 1 Million Databases, 1 Server
summary
Claim: one million Postgres databases served from a single $750/month server by packing 1GB microVMs into 256GB RAM with snapshot/restore and NVMe reads ~15GB/s to hit sub-10ms resume latency. Implementation uses Firecracker microVMs, Unicraft unikernels to halve image size, memory snapshotting to pause/resume ~250 concurrent DBs (~180k DB-hours per hour), and a stateful Go proxy that routes and hands off live connections. Result: production Prisma Postgres runs at that density and enables a free tier.
AI-written summary · about
chapters (14)
- 0:00Introduction to Prisma
- 0:20The Prisma ecosystem
- 0:49Understanding user needs
- 1:20Defining the product goal
- 3:21Infrastructure challenges
- 3:33Moving beyond AWS
- 4:53High-density architecture
- 5:06Implementing Firecracker
- 11:59Unikernel optimization
- 13:40Seamless proxying
- 13:51Global routing and access
- 15:29Business and market impact
- 18:30Modern platform integrations
- 18:57Conclusion
related talks
54mPostgres: From Cloud to Hybrid and On Prem Again
28mFrom VMs to Ku...
15mLeveraging Unikernels and Kubernetes to (Transparently)...
15mStuffing 10K Pods on a Commodity Server - And Starting Them in Milliseconds via Kubernetes
16mPower Hour, Managing Databases at Scale – Bala Kuchibhotla
25mPostgres on Kubernetes for the Reluctant DBA
50mFrom VMs to Cloud-Native PostgreSQL in Kubernetes...
51mPostgres on OpenStack