A $12 server, one API, every language, picked by load test

Most teams rent a cheap box as a staging site. They never find out how much it will take before it falls over. The twelve dollar server challenge is a public contest that answers that question with real numbers. Developers build the same small API, back it with SQLite, and run it on one DigitalOcean droplet. The droplet costs $12 a month. It has one vCPU and 2 GB of RAM. The repository has 535 stars and 266 forks after four weeks, with 62 pull requests open.

The rules are tight. The API has five endpoints. SQLite is the only database. There is no Redis, no outside cache, no second machine. The operating system is Ubuntu 24.04 on x86_64. An Nginx reverse proxy is optional, and so is exposing port 80 directly. The author, who posts as arjaythedev, runs each submission on the same box with k6 as the load tester.

How the pass bar is set

A run starts at 2,500 concurrent users and doubles until the server breaks. A submission passes when three numbers hold for five minutes. P95 latency stays below 500 ms. P99 latency stays below 1 second. The error rate stays under one percent. Those three cutoffs are a reasonable set of numbers for an API that fronts a web app. A pass at a given user count is also a budget line: the host of a $12 box can give a product team a real number. The number is how many simultaneous users the back end will take.

The leaderboard at benchmarks.arjaythedev.com/12-dollar-server tracks results by language and framework. Current pull requests include entries in Go, Rust, Python, Node.js, C, Zig, and others. Every submission runs on the same hardware against the same test. The methodology is in the repository as shell scripts. A reader who wants to repeat the test on their own hardware has the harness to do it.

Where the real finding is

The useful number for most readers is not which language won. It is the shape of the drop-off: how far past 5,000 users a given stack keeps p95 below 500 ms, and where each one falls over. A team planning capacity on a shared vCPU has had to guess at this. Vendor benchmarks come from much bigger boxes and never tell a buyer where the knee is on their own tier. This challenge publishes the knee.

The limits of the result are also worth being honest about. The challenge is a single API shape against one hardware configuration. A production back end has sessions, authentication, long queries, cold caches, and backups that compete for the same vCPU. A submission that passes here is not promised to pass under a real workload. What the leaderboard does give is a comparison, under the same floor, between many languages written by many developers. That is a better starting point than a vendor's own number.

The entries themselves are usually under 400 lines of code. A team picking a stack for a small internal service can read the top three submissions in a morning. Each one is a working reference back end with its own load test.

Source

GitHub: arjaythedev/twelve-dollar-server-challenge. Figures come from the project's own README and the public leaderboard.