Performance Testing

<objective>

Measure, assert, and protect application performance with budgets enforced in CI, not subjective "feels fast enough" assessments. This skill covers two domains: load testing (can the backend handle traffic?) and web performance (is the frontend fast for users?). A report that says "LCP is 3.2s" is information; a CI gate that fails the build at 2.5s is accountability.

</objective>

Quick Route

Discovery Questions

Check .agents/qa-project-context.md first — if it exists, use it and skip questions already answered there.

What to Measure

• Web performance or load testing? Web performance measures user-perceived speed (Core Web Vitals, page load). Load testing measures server capacity (RPS, latency under load). Most products need both.
• Which user journeys are performance-critical? Not every endpoint needs load testing. Focus on high-traffic, revenue-critical, or latency-sensitive flows.
• Existing performance budgets? If yes, what are the targets? If no, this skill helps establish them — measure baseline first.

Current State

• What is the current baseline? You need numbers before you can set targets. Measure first, then define budgets.
• What broke due to performance before? Slow pages that lost users, endpoints that timed out under load, queries that locked up — these point at where to focus.
• Existing monitoring? APM (Datadog, New Relic), RUM, or synthetic probes give you field data to model realistic load.

Infrastructure

• Where does load testing run? Target staging or a dedicated load-test environment, never production without explicit operations coordination.
• Expected traffic pattern? Steady, daily peaks, seasonal spikes (Black Friday), or event-driven bursts — this picks the load profile.
• Rate limits, WAF rules, or auto-scaling in the path? These distort results and must be accounted for in test design.

Core Principles

1. Measure before optimizing

Performance intuition is unreliable; developers routinely optimize the wrong thing. Profile first, identify the actual bottleneck, then optimize. A profiled 50ms win in the right place beats an assumed 500ms win in the wrong one.

2. Budgets are only real if CI enforces them

A budget documented in a wiki but not checked in CI is violated within weeks. Wire budgets into the pipeline as k6 thresholds and Lighthouse assertions so regressions fail the build, not a quarterly review.

3. Realistic load beats maximum stress

A stress test to 10x traffic tells you the breaking point. A load test at 1.5x expected traffic tells you whether tomorrow's real users have a good experience. Both have value, but realistic load runs more often and catches regressions earlier.

4. Core Web Vitals are what users actually feel

Server-side metrics (latency, throughput) matter, but users experience performance through the browser. LCP, INP, and CLS measure perceived speed. A fast API that renders slowly is still slow to users.

5. Performance is a feature

It does not happen by accident. It needs dedicated test infrastructure, budgets, and monitoring, and the same continuous attention as functional correctness. Treat a performance regression with the same urgency as a functional bug.

k6 Load Testing

k6 is an open-source load testing tool that uses JavaScript/TypeScript for scripts, runs from the CLI, and integrates with CI. Current stable: k6 v2.0.0 (final shipped 2026-05). v2 has breaking changes from v1 — see the migration callout below.

A load test is built from three pieces: a load profile (the stages/scenarios shape), checks (per-request assertions), and thresholds (pass/fail budgets that set the exit code). Always drive the base URL from __ENV so the same script runs against local, staging, and CI.

See references/recipes.md for the full basic load test, custom metrics, and scenarios. A minimal threshold block:

export const options = {  stages: [    { duration: '1m', target: 20 },    { duration: '3m', target: 20 },    { duration: '1m', target: 0 },  ],  thresholds: {    http_req_duration: ['p(95)<500', 'p(99)<1000'],    http_req_failed: ['rate<0.01'],  },};

Load Profiles

A spike test is not done until "does it recover?" is an assertion, not a comment. Tag the post-spike window (e.g. phase:recovery) and scope a threshold to it so the run fails if p95 stays elevated. See the recovery-detection recipe in references/recipes.md.

Custom Metrics and Scenarios

k6 has four metric types: Counter (cumulative count), Rate (proportion of non-zero/true values, 0..1), Trend (statistical distribution — p50/p95/p99), Gauge (latest value). Tag requests with { tags: { name: 'endpoint' } } to filter metrics per endpoint, scenario, or flow.

Scenarios run distinct user flows concurrently, each with its own executor and per-scenario thresholds ('http_req_duration{scenario:checkout}': ['p(95)<500']). Use them to model a real mix — browsers + checkout + API-heavy load at once. Full custom-metric and scenario examples are in references/recipes.md.

k6 CI Integration

Install k6 with the official grafana/setup-k6-action@v1 — not a hand-rolled apt/gpg keyserver block (brittle, rots, no version pin). k6 exits non-zero when any threshold is breached, so a breached budget fails the job with no extra wiring. The full GitHub Actions workflow (checkout → setup-k6 → run → upload artifact) is in references/recipes.md.

k6 v1 → v2 migration (v2.0.0 final, 2026-05):

• k6/experimental/websockets → k6/websockets (drop the experimental/ prefix; stable now)
• k6/experimental/redis → k6/x/redis — NOT removed. The import auto-resolves the xk6-redis extension (auto-extension-resolution is on by default; JS usage unchanged). Do not hand-roll a Redis client.
• externally-controlled executor removed
• options.ext.loadimpact removed → use options.cloud (Grafana Cloud k6, formerly k6 Cloud / Load Impact)
• CLI: --no-summary → --summary-mode=disabled; --upload-only → k6 cloud upload script.js; k6 login/pause/resume/scale/status removed (use k6 cloud login, etc.); positional k6 cloud script.js removed
• Exit code 97 is new: a non-threshold cloud-side abort. Wire it into CI handling. Reference: https://grafana.com/docs/k6/latest/get-started/migrating-to-v2/

Web Performance

Lighthouse CI

Lighthouse CI (@lhci/cli) automates Google Lighthouse audits and enforces budgets in the pipeline via lighthouserc.js assertions. Current: @lhci/cli 0.15.x on the Lighthouse 12.6 engine. The project is in maintenance mode (last release ~a year ago) and does not yet support Lighthouse 13 (needs Node 22.19+); it remains the standard CI surface for Lighthouse, but watch upstream before adopting in greenfield projects.

The full lighthouserc.js (LCP/CLS/TBT/perf-score assertions) and the lhci autorun CI step are in references/recipes.md.

INP is field-only — gate TBT in the lab

This is the single most-misunderstood point in web perf, so be precise:

• INP (Interaction to Next Paint) is a field-only metric. A Lighthouse lab audit loads a page with no user interaction, so it cannot measure or score INP. Asserting interaction-to-next-paint in a standard lhci autorun run gates something Lighthouse never produces.
• In the lab, gate Total Blocking Time (TBT) as the proxy: assert 'total-blocking-time': ['error', { maxNumericValue: 200 }]. TBT correlates with INP but is not identical (a page can hit 0ms TBT and still fail field INP).
• Track real INP from the field — Chrome UX Report (CrUX) or your RUM tool. That is the number users actually experience and the one Google ranks on.
• Only if you must measure scripted-interaction latency, use Lighthouse user-flow / timespan mode with scripted clicks, or k6/browser with a PerformanceObserver on event entries. Treat that as a scripted lab proxy, still not field INP.

FID is gone: FID was deprecated and web-vitals v5+ removed it; INP became a Core Web Vital in March 2024. Do not assert on FID in any new code.

Measuring CWV in Playwright

For a per-page lab check inside your existing Playwright suite, use page.evaluate with a PerformanceObserver to capture LCP and CLS (both observe cleanly on page load), then assert the thresholds. Full test in references/recipes.md. For under-load CWV capture, use the k6/browser recipe in the same file.

Core Web Vitals

The three metrics Google uses for user-perceived performance.

Common fixes:

• LCP — slow server (cache/CDN, SSR/SSG the LCP content), render-blocking CSS/JS (defer, inline critical CSS), slow image load (WebP/AVIF, preload the LCP image).
• INP — long JS tasks (scheduler.yield(), requestIdleCallback), heavy handlers (debounce/throttle, Web Workers), layout thrashing (batch DOM reads/writes via requestAnimationFrame).
• CLS — set width/height on images, reserve space for injected content (aspect-ratio/min-height), font-display: swap with size-adjust, fixed-size containers for ads/embeds.

Field vs. Lab Data

Use lab data for CI gates and debugging; use field data to understand the real user experience. A page that scores 100 in Lighthouse but has poor CrUX data has a real problem — and INP only ever shows up in the field column.

Bottleneck Identification

When a budget is breached, investigate in this order — never guess, never optimize before profiling:

1. Identify the slow endpoint. Per-endpoint Trend metrics in k6 give p50/p95/p99 per API. The slowest is the first target.
2. Database. Check missing indexes (EXPLAIN), N+1 patterns (JOIN/batch), lock contention on write-heavy tables (optimize transactions), unbounded result sets (paginate).
3. CDN/caching. Verify cache-control on static assets (public, max-age=31536000, immutable) and the x-cache: HIT rate.
4. Third-party scripts. Run Lighthouse with blockedUrlPatterns for analytics/chat/tracking; compare scores with and without to quantify the cost.
5. Correlate with server metrics. Client-side latency alone hides the cause — check CPU, memory, disk I/O, connection-pool saturation, and query execution time during the run.

Anti-Patterns

1. Load testing production without coordination

Load tests can trigger auto-scaling (expensive), rate limiting (test fails), alerts (unnecessary pages), or outages. Always coordinate with operations and target staging or a dedicated load-test environment.

2. Unrealistic load scenarios

Testing 10,000 concurrent users when the product has 500 DAU, or uniform traffic when real traffic has peaks. Model load from analytics; absent that, start at 2x estimated peak and increase.

3. Ignoring server-side metrics

Looking only at k6's client-side response times. The server may be at 95% CPU, the connection pool exhausted, or memory leaking. Correlate load results with server metrics.

4. Performance testing only before releases

A quarterly pre-release load test lets dozens of regressions accumulate with untraceable root causes. Run in CI on every merge to main with budgets that catch regressions immediately.

5. No performance budgets

"LCP is 3.2s" is information; "LCP must be under 2.5s" is a gate. Define budgets, enforce them as k6 thresholds and Lighthouse assertions, and treat violations as bugs.

6. Asserting INP in standard Lighthouse CI

Standard lhci autorun lab mode cannot produce an INP audit — no user interaction. Gating interaction-to-next-paint there gates a number Lighthouse never measures. Gate total-blocking-time as the lab proxy and track real INP from CrUX/RUM.

7. Optimizing without profiling

Spending days on a function that is 2% of response time. Profile, find the actual bottleneck, then optimize. The 200ms query beats the 5ms JS function every time.

8. Testing with empty databases

Load testing against 100 seeded rows when production has 10 million. Query performance is radically different at scale. Seed the load-test environment with production-scale (anonymized) data first.

Verification

Prove the artifacts actually work before claiming done:

• k6 script: k6 run --summary-mode=disabled load-tests/api-load.js exits 0 and the end-of-test summary shows every thresholds line green (✓). A red threshold line means a breached budget and a non-zero exit.
• Spike recovery: run the spike script and confirm the {phase:recovery} threshold appears in the summary and passes — proving recovery is asserted, not just commented.
• Lighthouse CI: lhci autorun exits 0 with all ['error', ...] assertions passing; a breached LCP/CLS/TBT assertion exits non-zero.

Done When

• k6 scripts cover the target load profiles: baseline (constant), stress (ramp to breaking point), and soak (sustained), each with __ENV-driven base URL.
• Spike test asserts recovery via a threshold scoped to the post-spike window (e.g. http_req_duration{phase:recovery}), not just a // recovery comment.
• Performance budgets encoded as k6 thresholds (e.g. p(95)<500, http_req_failed rate<0.01) that fail the CI job when exceeded.
• Lighthouse CI lighthouserc.js gates merges on largest-contentful-paint, cumulative-layout-shift, and total-blocking-time (the lab proxy for INP) as ['error', ...] assertions; real INP tracked from CrUX/RUM, not asserted in lab.
• Core Web Vitals baselines documented for each key page (home, checkout, dashboard) with Good/Needs Improvement/Poor classification.
• Test results include p95 and p99 latency, error rate, and throughput per scenario, stored as CI artifacts.

Reference Files (in references/)

• recipes.md — runnable artifacts: basic k6 load test, spike-with-recovery detection, custom metrics, scenarios, k6 CI workflow (grafana/setup-k6-action), lighthouserc.js, the k6/browser CWV-under-load example, and the Playwright LCP/CLS test.

Related Skills

• ci-cd-integration — pipeline wiring for k6 and Lighthouse CI, scheduling nightly runs, gating deployments on budgets.
• qa-metrics — LCP/INP/CLS and p95 latency as part of the broader QA metrics dashboard.
• release-readiness — performance benchmarks in the go/no-go checklist.
• synthetic-monitoring — scheduled production CWV/uptime probes after release; this skill is pre-release lab gating.
• observability-driven-testing — when prod telemetry is the input to designing new perf tests.
• qa-project-context — captures performance budgets, traffic patterns, and critical flows to test.