Performance Budget

Performance Budget Skill

Produce a complete, actionable performance budget document for a web service or application. A performance budget is not a wishlist — it is a set of measurable, enforced constraints that define what "acceptable performance" means and who is responsible when those constraints are violated.

A good performance budget answers: what are the targets, how are they measured, what triggers an investigation, and what happens when a budget is breached.

Required Inputs

Ask for these if not already provided:

• Service name and type — web app, API service, mobile app, or combination
• Key user journeys — the 3–5 most important flows users take (e.g. "search → product page → checkout")
• Current baseline metrics — P50/P95/P99 latency, LCP, CLS, INP if available (state "no baseline" if not collected yet)
• Tech stack — frontend framework, backend language/framework, CDN, database
• Deployment environment — cloud provider, region(s), edge/CDN configuration
• Cost constraints — any budget or infrastructure limits that affect headroom

Output Format

Performance Budget: [Service Name]

Service: [Name] | Team: [Team name] Last updated: [Date] | Owner: [Name / role] Environment: [Production / Staging baseline] | Review cadence: [Quarterly / per-sprint]

Overview

[2–3 sentences describing the service, its user-facing performance requirements, and why performance is a priority. Reference the business impact of latency — e.g. conversion rate, user retention, SLA obligations.]

Performance philosophy: [e.g. "Performance is a feature. Every engineer is responsible for keeping the service within budget. Regressions must be caught in CI before they reach production."]

Key User Journeys

Define the critical paths that the performance budget is designed to protect.

Frontend Performance Budget

Complete this section for web and mobile applications. Skip for API-only services.

Core Web Vitals Targets

Targets apply to the 75th percentile of real user sessions (field data), measured on a mid-range Android device on a 4G connection unless otherwise stated.

Page Weight Budget

Per-Journey Frontend Targets

Backend Performance Budget

API Latency SLOs

Targets measured at the service boundary (not including client-side network latency).

Overall service SLOs:

Database Query Budget

Measurement Methodology

Real User Monitoring (RUM)

Tool: [e.g. Google CrUX, SpeedCurve, Datadog RUM, Sentry Performance, custom] Data source: [Field data from real users / Lab data from synthetic tests / Both] Sample rate: [X% of sessions] How to access: [Dashboard URL or tool access instructions]

What is measured:

• Core Web Vitals (LCP, INP, CLS) per page and journey
• Custom performance marks for business-critical interactions
• Resource timing for key assets
• Long tasks (>50ms on main thread)

Synthetic Monitoring

Tool: [e.g. Lighthouse CI, WebPageTest, k6, Artillery, Playwright with performance assertions] Frequency: [Every X minutes / on every deploy / nightly] Test location(s): [e.g. eu-west-1, us-east-1] Device profile: [Desktop 10Mbps / Mobile 4G Moto G4 / both]

Synthetic test suite location: [Link to test files]

Backend Observability

APM tool: [e.g. Datadog, Grafana + Prometheus, New Relic, AWS X-Ray] Metrics collected:

• Request rate, error rate, duration (RED metrics) per endpoint
• Database query duration and connection pool utilisation
• Cache hit/miss rates
• Background job queue depth and processing latency

Dashboard: [Link to primary performance dashboard]

CI/CD Performance Enforcement

Performance budgets are enforced at two gates:

Gate 1 — Build-time Bundle Analysis

Tool: [e.g. bundlesize, size-limit, webpack-bundle-analyzer with CI assertion] Config file: [[.bundlesizerc / .size-limit.js / etc.]] Trigger: Every PR targeting main Blocking: Yes — PR cannot merge if bundle size budget is exceeded

// Example .size-limit.js
[
  {
    "path": "dist/js/*.js",
    "limit": "200 KB"
  },
  {
    "path": "dist/css/*.css",
    "limit": "50 KB"
  }
]

Gate 2 — Synthetic Performance Tests in CI

Tool: [e.g. Lighthouse CI, k6, Artillery] Trigger: On deploy to staging Blocking: Yes — production deploy is blocked if thresholds fail Thresholds checked:

• LCP ≤ [Xs]
• CLS ≤ [0.X]
• P95 API latency ≤ [Xms]
• Error rate < [X%]

CI config location: [[.github/workflows/perf.yml / ci/performance.yaml]]

How to run locally:

# Run Lighthouse CI against local build
[command — e.g. lhci autorun --config=lighthouserc.js]

# Run load test locally
[command — e.g. k6 run load-tests/api-smoke.js]

Budget Breach Response Process

A budget breach is when a measured metric exceeds its target for [X consecutive measurements / X minutes sustained / a single deploy].

Breach Severity Levels

Breach Investigation Checklist

When a breach is detected, work through this checklist in order:

1. Identify the regression commit

# Compare performance across recent deploys
[command — e.g. datadog metrics query, lighthouse-ci compare, git bisect]

2. Classify the breach

• Is this a code change? (new feature, refactor, dependency bump)
• Is this an infrastructure change? (new instance type, config change)
• Is this an external factor? (CDN issue, DNS, upstream dependency)
• Is this a measurement anomaly? (test environment issue, sample size)

3. Immediate actions

• If P1/P2 in production and a code cause is confirmed: roll back or disable the feature flag
• If cause is unknown: do not roll back immediately — gather more data first
• Notify [#performance / #incidents Slack channel] with: metric name, current value, budget target, suspected cause

4. Resolution

• Fix the root cause — do not just adjust the budget threshold
• Budget thresholds should only change after a team discussion and explicit approval from [tech lead / EM]
• Document the breach in the [performance log / incident record]

Budget change policy: Budget thresholds may only be relaxed if: (a) the feature delivering the regression has measurable business value that outweighs the performance cost, and (b) the change is reviewed and approved by [tech lead].

Performance Review Cadence

Next scheduled review: [Date] Review owner: [Name / role]

Quality Checks

• Every budget threshold is a specific number — not a range or "TBD"
• Both frontend (if applicable) and backend targets are defined — not just one or the other
• Measurement tooling is named with a link to the dashboard or config file
• CI enforcement is configured for at least one gate (build-time or deploy-time)
• Budget breach response process names specific Slack channels and owners
• Budget thresholds are anchored to baseline measurements or a justified target — not pulled from thin air
• Per-journey targets are defined for critical user journeys, not just global averages