Load Testing and Baselines

Updated

September 8, 2026

Load Testing and Baselines

Overview

Microbenchmarks don’t replace load tests. Establish a baseline: throughput, p50/p99 latency, error rate, resource use—then change one variable.

Tools

Tool Role
hey / vegeta / k6 HTTP load
go test -bench Function-level
pprof under load See real hotspots 
hey -z 30s -c 50 http://127.0.0.1:8080/api/books

Baseline template

Scenario: GET /api/books  (cached / uncached)
Concurrency: 50
Duration: 60s
Result: QPS=… p50=… p99=… errors=…
CPU=… RSS=…  (from metrics or `ps`)
Commit: abc123

SLO-oriented load

Drive load until error budget or latency SLO breaks—know the capacity ceiling.

Avoiding test lies

  • Warm up before measuring
  • Separate client machine when possible
  • Watch coordinated omission (load tools that delay sending)
  • Fix payloads and auth

Rules of thumb

Do Don’t
Record env + commit with results Compare laptop vs prod blindly
Increase load gradually Jump to 10k connections first
Pair with profiles Optimize without a number

Try next

  1. Baseline one endpoint; save numbers in a markdown table.
  2. Add artificial time.Sleep and remeasure.
  3. Capture CPU pprof during hey.