Suppose 99 requests finish in 12 ms and every hundredth request stalls behind a lock or cold connection for 2.4 s. The average latency across those 100 calls comes out to 35.8 ms, which looks fine on a status board even though users loading a page that makes five API calls keep hitting multi-second stalls.
I wrote Wiremark in Go because I wanted a small command-line benchmark tool that puts tail percentiles and before/after comparisons front and center.
wiremark compare \
--baseline https://staging-v1.example.com/search \
--candidate https://staging-v2.example.com/search \
--requests 2000 --concurrency 40 --format markdown
Details that keep benchmark numbers honest
Spawning one goroutine per request ends up benchmarking your laptop’s ephemeral port table rather than the server, so Wiremark runs a fixed worker pool at the exact --concurrency level you ask for.
Every run prints P50, P90, P95, and P99 next to min, mean, and max. Watching P99 drop after adding a database index (something I saw directly while testing SQLite Clone) tells you far more than watching the mean move by two milliseconds.
Failed requests also stay out of clean latency buckets. If a server starts returning instant 502 Bad Gateway errors under load, mixing those responses into overall timing makes the endpoint look faster right as it falls over. Wiremark splits timeouts, connection errors, and HTTP 5xx responses into their own counts.
Finally, because benchmark commands regularly pass Authorization or X-API-Key flags, Wiremark strips sensitive headers before writing Markdown, HTML, or JSON reports (the same default redaction I use in Wardline), and pairs well with front-end payload checks like my page weight study.