Rustyload.

After building FlashKV, I had an obvious question with no easy answer: how fast is it, actually? redis-benchmark worked, but I wanted to understand what a load generator does under the hood — how you saturate a server without lying to yourself about the numbers. So I built one.

RustyLoad is a concurrent load-testing CLI in Rust with a dual-protocol engine: it stress-tests HTTP/HTTPS services (any method, custom headers, request bodies) and speaks raw TCP to FlashKV-style key-value stores — PING, GET/SET, mixed workloads, and randomized key distribution across a configurable key space. It’s published on crates.io with prebuilt binaries for Linux, macOS, and Windows, built and released automatically through GitHub Actions.

The statistics engine is the heart of the tool. Averages flatter a service; percentiles tell the truth. If the mean latency is 100ms but p99 is 2000ms, one in a hundred requests is twenty times slower than the number on the dashboard — and that’s the request users remember. RustyLoad reports min / max / average alongside p50, p95, and p99, computed with linear interpolation over the collected samples.

Concurrency is semaphore-limited: one async Tokio task per request, with a permit gate controlling how many run at once. That keeps pressure constant and measured — a load tester that overwhelms its own client machine produces numbers about the client, not the server.

| p50 (median):        210 ms |
| p95:                 445 ms |
| p99:                 823 ms |
| Requests/sec:        8.23   |

Load testers are usually flag soup. RustyLoad has two modes: a quick mode for power users (rustyload -u https://api.example.com/health -n 500 -c 50) and an interactive mode that walks through protocol, target, workload, concurrency, and timeout with validated prompts — run it with no arguments and the wizard just starts. A live progress bar (indicatif) and a boxed configuration summary make long runs legible.

The stack is deliberately boring and solid: clap for argument parsing, reqwest for HTTP, dialoguer for prompts, anyhow for error handling — with unit tests covering the percentile math, statistics aggregation, and configuration builder. It was inspired by tools like wrk, hey, and bombardier; building my own taught me what those tools have to get right.

Source on GitHub · crates.io