Get the newsletterSubscribe

The $12 Server Challenge

Leaderboard

One small feed API on SQLite, any language, one 1 vCPU / 2 GB box. The community's fastest servers, benchmarked by me on the same machine.

Oct 25, 2026Synteq VPS: 1 vCPU (Xeon Platinum 8260), 2 GB RAM, Ubuntu 24.0428 implementations99 waiting to be benchmarked
Submit your implementation ↗GitHub repo ↗
RankImplementationUsers heldreq/sp95 msp99 msErrorsCPURAM
1
rust-epoll-scooter1337
Rust (io_uring) · @Scooter1337
264,000
25,1472594990%84%630 MB
2
asm-nasm-melroy89
x86-64 asm (io_uring) · @melroy89
171,000
16,777871420%78%358 MB
3
go-epoll-MegaPanchamZ
Go (raw epoll) · @MegaPanchamZ
165,000
16,1651942410%89%187 MB
4
asm-epoll-umitaltintas
x86-64 asm (epoll) · @umitaltintas
155,000
15,1822082940%89%211 MB
5
c-iouring-ssmurfgg04-gif
C (io_uring) · @ssmurfgg04-gif
150,000
14,6113855240%78%413 MB
6
go-epoll-champ3oy
Go (io_uring) · @champ3oy
128,000
12,6043230%77%385 MB
7
kotlin-nio-3x3xX3N0N
Kotlin (NIO, 2 JVMs) · @3x3xX3N0N
120,000
11,7592465240%100%1,380 MB
8
c-epoll-TheFirstIstari
C (epoll) · @TheFirstIstari
107,000
10,4932124090.94%78%169 MB
9
java-nio-stohirov
Java (FFM epoll) · @stohirov
102,000
10,023861280.85%66%555 MB
10
v-epoll-i098
V · @i098
100,000
9,8481.52.90.51%76%384 MB
Free with the newsletter

See all 28 implementations

Subscribe to The Dev Download and every leaderboard on this site opens up. Already subscribed? Enter the same email.

  • 18 more implementations on this leaderboard, every metric
  • Each implementation's write-up
  • The raw results as JSON
Unsubscribe anytime.

How it was run

The machine. A Synteq VPS with 1 vCPU (Xeon Platinum 8260) and 2 GB RAM, no swap, running Ubuntu 24.04 with no kernel tuning. The load generators are separate 16-core machines in the same datacenter.

How each entry is scored. Every entry is built from source on the box with its own install.sh and build.sh, started on a fresh copy of the seed database, and has to pass the 42-check test suite before any load. Virtual users follow the challenge's user model (feed → think → post → think → 15% like, 2% post). Two-minute runs double the users until one fails, then narrow down to within 5%. A 5-minute hold confirms the number, stepping down 5% until one passes. The score is the most users held for 5 minutes with p95 under 500 ms, p99 under 1 s and under 1% errors. Every run restarts the app on a fresh database.