Benchmark — Performance Baseline & Regression Detection#

Measure performance impact of code changes and set baselines for your project.

Quick Start#

Required: Bring your own benchmark tool (e.g. k6 for load testing, autocannon for API benchmarks, hyperfine for build speed).

1. Run baseline: benchmark baseline — captures current metrics
2. Make changes to your code
3. Run comparison: benchmark compare — measures impact against baseline
4. Review report: tracks deltas for Web Vitals, bundle size, build time, API latency

⚠️ Safety Warning: Load tests should not be run against production systems without explicit authorization from your team. Test against staging or local environments only.

Key Concepts#

• Page Performance: Core Web Vitals (LCP, CLS, INP, TTFB) and resource sizes
• API Performance: p50/p95/p99 latency under load
• Build Performance: cold build, hot reload (HMR), test suite, type-check, lint times
• Before/After: Compare metrics against stored baseline to detect regressions

Common Usage#

Page Performance Metrics#

text

Measures on real browser:
- LCP (Largest Contentful Paint) — target < 2.5s
- CLS (Cumulative Layout Shift) — target < 0.1
- INP (Interaction to Next Paint) — target < 200ms
- TTFB (Time to First Byte) — target < 800ms
- Total page weight (target < 1MB)
- JS bundle (target < 200KB gzipped)
- Network requests count

API Performance#

text

Hit each endpoint 100 times:
- p50, p95, p99 latency
- Response size, status codes
- Load test: 10 concurrent requests
- Compare against SLA targets

Build Performance#

text

Measures development feedback loop:
- Cold build time
- Hot reload time (HMR)
- Test suite duration
- TypeScript check time
- Lint time
- Docker build time

Output Format#

Reports saved to .benchmarks/ (git-tracked):

MetricBeforeAfterDeltaVerdictLCP1.2s1.4s+200ms⚠ WARNBundle180KB175KB-5KB✓ BETTER

References#

• references/integration.md — CI/CD setup, pairing with canary-watch and browser-qa
• references/modes.md — detailed explanation of each measurement mode

---

Adapted from everything-claude-code by @affaan-m (MIT)