Open this when you need the exact tool invocation and the number to report. Verified 2026-09 against https://angular.dev/tools/devtools, https://angular.dev/tools/cli/build and https://web.dev/articles/inp.
1. Rule zero: production build, stated conditions
ng build # production configuration by default
npx http-server dist/<app>/browser -p 4300 -c-1
Dev-mode runs change detection twice per cycle to surface NG0100, ships unminified code and skips optimizations — a dev measurement is not evidence about production. Record with the number: build mode, CPU throttle, network profile, browser version, and whether extensions were disabled.
2. Angular DevTools (browser extension)
- Profiler tab → record → perform exactly one interaction → stop. Read: how many change-detection cycles that one interaction caused, and the bar chart of per-component time. Two cycles for a click is normal; twenty means something re-enters.
- Components tab → select a node to inspect its inputs and signal values live; useful for confirming that an input reference changes when you did not expect it to.
- Signal graph → shows which signals a component reads and what depends on them. This is the fastest way to find a
computed()that depends on more than it needs and therefore recomputes on unrelated writes.
What to report: "one keystroke in the filter box triggers 14 CD cycles; OrderRow accounts for 62 ms of the 80 ms frame."
3. Bundle analysis
ng build --stats-json
npx esbuild-visualizer --metadata dist/<app>/stats.json --open
The application builder emits an esbuild metafile, so any esbuild metafile viewer works (esbuild-visualizer, esbuild's own analyzer, source-map-explorer on the emitted maps).
Read it top down and ask, per entry over ~30 kB: does the initial chunk need this at all? Common answers:
| Finding | Fix |
|---|---|
| A whole date library for one format call | Intl.DateTimeFormat, or date-fns with per-function imports |
| All of lodash | Import the single function, or write the ten lines |
| An icon font or full icon package | Inline the used SVGs |
| Chart/editor library in the initial chunk | @defer it, or lazy-load the route that uses it |
| Material imported in the root component | Import per component that actually uses it |
| Two copies of the same library | npm ls <pkg> — a transitive dependency pinned a different major |
Budgets in angular.json:
"budgets": [
{ "type": "initial", "maximumWarning": "500kB", "maximumError": "1MB" },
{ "type": "anyComponentStyle", "maximumWarning": "4kB", "maximumError": "8kB" }
]
Set maximumError to today's size plus a small margin rather than a round number: a budget that is 40 % above reality never fires.
4. Core Web Vitals, and INP in particular
Lab (Lighthouse) tells you whether a page can be fast; field data tells you whether it is. Collect both.
// main.ts (or a small reporting service) — npm i web-vitals
import { onCLS, onINP, onLCP, type Metric } from 'web-vitals';
const report = (metric: Metric) => navigator.sendBeacon('/rum', JSON.stringify(metric));
onLCP(report);
onINP(report);
onCLS(report);
onINP reports the interaction that produced the worst delay, with the target element in attribution when you use web-vitals/attribution. That attribution is what turns "INP is 480 ms" into "the filter input's keyup handler is the problem", which is the only form a fix can start from.
Thresholds to judge against: LCP ≤ 2.5 s, INP ≤ 200 ms, CLS ≤ 0.1 at the 75th percentile.
Lighthouse in CI — lhci autorun with assertions committed next to the app, failing the build on regression:
{ "ci": { "assert": { "assertions": {
"categories:performance": ["error", { "minScore": 0.85 }],
"largest-contentful-paint": ["error", { "maxNumericValue": 2500 }],
"total-blocking-time": ["error", { "maxNumericValue": 300 }]
} } } }
Run it against the built output on a fixed runner size; Lighthouse scores on shared CI hardware are noisy, so assert on a handful of stable metrics rather than the overall score alone.
5. Long tasks and the Performance panel
Record in the Chrome Performance panel with 4–6× CPU throttling. Anything over 50 ms on the main thread is a long task and is what INP measures. In an Angular trace, look for:
ApplicationRef.tick/refreshViewframes — change detection; go back to the Angular DevTools profiler for the component.- A single long
evaluateframe in app code — synchronous work that belongs in a worker or should be chunked. - Repeated forced reflow warnings — a DOM read after a write in the same frame; move the read into
afterNextRender({ read: … }). - Third-party frames — measure with the script blocked to get the honest cost before arguing about it.
6. Memory
- Open the app, navigate to the suspect route, navigate away, force GC (the bin icon), take snapshot A.
- Repeat the same navigation five times, force GC, take snapshot B.
- Compare B to A, filtered to "Objects allocated between A and B". Component class names with a growing count are retained instances.
- Expand a retained instance and read the retainer chain. It almost always ends in a
Subscription, anEventTargetlistener, or a root-provided service holding an array.
The two structural causes: a subscribe() without takeUntilDestroyed(), and state kept in a root-provided service that should have been provided on the route.
7. Reporting template
Baseline: INP p75 = 460 ms (field, 7 days, Chrome, all devices)
LCP p75 = 3.1 s
Bucket: template work during change detection
Evidence: Angular DevTools — one keystroke = 14 CD cycles; OrderRow 62 ms of an 80 ms frame.
web-vitals attribution names the filter input as the INP target on 78 % of samples.
Fix: computed() view model per row + debounceTime(300) on the filter signal.
Verify: one keystroke = 2 CD cycles; INP p75 = 140 ms after 7 days.