Web performance и Core Web Vitals: field distributions, не один score

Core Web Vitals измеряют разные части пользовательского опыта: LCP — загрузку основного контента, INP — responsiveness взаимодействий, CLS — неожиданную визуальную нестабильность. На 2026 год good thresholds: LCP не более 2.5 s, INP не более 200 ms и CLS не более 0.1.

Классификация использует 75-й percentile page views. Lab test полезен для диагностики, но не заменяет field population: устройство, сеть, cache state, route и реальные взаимодействия меняют распределение.

Проверяемые утверждения

• LCP, INP и CLS имеют разные units и причины; их нельзя усреднить в физически осмысленную одну метрику.
• LCP можно разложить на TTFB, resource load delay, resource load duration и element render delay для resource-backed candidate; оптимизировать надо фактический доминирующий компонент.
• INP отражает latency пользовательских взаимодействий и требует сокращать длинную работу main thread, а не только уменьшать initial bundle bytes.
• CLS — безразмерная score неожиданных shifts. Резервирование размеров media/ads и предсказуемое добавление контента предотвращают многие shifts.

Границы и компоненты

Компонент	Роль в модели
Real Browser Visit	document and assets; parse render interact; sampled metrics
Edge Delivery	document and assets; document miss; asset miss
Origin Renderer	document miss
Versioned Asset Origin	asset miss
Browser Main Thread	parse render interact
Privacy-scoped RUM Beacon	sampled metrics; aggregated measurements
Field Distribution Store	aggregated measurements

Сценарии

▶Walkthrough: decompose-lcp (not found)

Decompose the LCP path

A field trace separates document TTFB, discovery delay, hero transfer and element render delay. The team fixes the largest measured component instead of applying an unrelated bundle tweak.

Проверяемый исход: The LCP change has an explicit causal hypothesis and is compared at p75 by route and device class.

▶Walkthrough: protect-inp (not found)

Protect interaction responsiveness

A user interaction arrives while application work is queued. The app yields/splits non-critical work, updates visible feedback promptly and measures the full interaction-to-paint latency.

Проверяемый исход: Long tasks no longer monopolize the main thread, and p75 INP is evaluated from real interactions rather than inferred from page-load TTFB.

▶Walkthrough: prevent-cls (not found)

Prevent an unexpected layout shift

Critical media arrives after initial layout. Intrinsic dimensions/aspect ratio reserve its slot, while late banners use a preallocated region or user-initiated transition.

Проверяемый исход: The same content loads without moving existing controls unexpectedly, and field CLS stays attributable to concrete shift sources.

▶Walkthrough: evaluate-field-percentiles (not found)

Evaluate field percentiles correctly

Sampled RUM sends metric value, route template, device class and release context under a privacy policy. Aggregation computes p75 per metric instead of averaging percentiles or mixing units.

Проверяемый исход: A release decision reflects the target population and reports sample coverage, uncertainty and regressions separately.

Failure, concurrency и evolution checklist

• Segment regressions by route template, release, device class and connection quality; a global average can hide a broken critical cohort.
• Keep lab and field labels explicit. A synthetic run with no interaction cannot produce the same evidence as full field INP.
• Version metric attribution and sampling logic. A dashboard change can otherwise look like a product regression or improvement.
• Set performance budgets alongside correctness/accessibility gates; do not remove essential behavior merely to improve a score.

Security и privacy

• Minimize RUM payloads, strip URL query/fragment and user-entered content, apply retention/access control and respect the product consent/legal basis.
• Treat route/release metadata as potentially identifying when combined; sample and aggregate deliberately.
• Do not expose internal error bodies or credentials in PerformanceEntry names, custom marks or beacon payloads.

Метрики, формулы и допущения

• Good field thresholds: LCP_p75 ≤ 2.5 s, INP_p75 ≤ 200 ms, CLS_p75 ≤ 0.1; CLS is unitless and the time metrics use milliseconds/seconds.
• For a resource-backed LCP candidate: LCP = TTFB + resource_load_delay + resource_load_duration + element_render_delay by the diagnostic decomposition.
• p75 is the smallest observed value at or above 75% of the chosen population under the aggregation method; p75(A ∪ B) is not generally the average of subgroup p75 values.
• cache_hit_ratio = edge_hits / cacheable_requests; report bypasses and uncacheable requests separately rather than inflating the denominator ambiguously.

Числа и bounds выше действуют только при названных units, population и assumptions. Ни паттерн, ни browser API сами по себе не задают SLA, capacity или correctness.

Решения для production review

• Define target population, route criticality and release guardrails before selecting dashboards.
• Use field data to prioritize and lab traces/profilers to explain causes; require both for high-confidence fixes.
• Optimize the measured critical path while preserving content, security and accessibility semantics.

Первичные и официальные источники

• https://web.dev/articles/defining-core-web-vitals-thresholds
• https://web.dev/articles/optimize-lcp
• https://web.dev/articles/inp
• https://web.dev/articles/cls

Scope note

Диаграмма отражает текущие Core Web Vitals thresholds на 2026-08-25. Метрики и thresholds могут эволюционировать; они не являются SLA и не заменяют task-success research или accessibility testing.