Performance — Faultsense

0ms

INP Impact

Across 12 stress configs (3 assertion counts × 2 churn rates × 2 CPU profiles)

1.7KB

Idle Heap

Observer, listeners, sweep timer

2.2ms

Worst P99

4% of the 50ms long-task threshold

Long Tasks

Never creates one, any scenario

MutationObserver Callback Scaling

The agent's hot path — stress benchmark measures callback duration directly. Most production deployments run 10–50 assertions per page; the tables below stress-test up to 1,000.

Unthrottled (Apple M4 Pro)

Assertions	Churn	P50	P95	P99
50	0/s	0ms	0.3ms	0.5ms
50	100/s	0ms	0.1ms	0.3ms
200	0/s	0ms	0.3ms	0.5ms
200	100/s	0ms	0.1ms	0.3ms
1000	0/s	0.1ms	0.6ms	0.8ms
1000	100/s	0.1ms	0.2ms	0.6ms

CPU 4x Throttle (simulating mid-tier mobile)

Assertions	Churn	P50	P95	P99
50	0/s	0ms	0.7ms	1.1ms
50	100/s	0ms	0.7ms	1.0ms
200	0/s	0ms	0.7ms	1.1ms
200	100/s	0ms	0.7ms	0.9ms
1000	0/s	0.3ms	1.1ms	2.2ms
1000	100/s	0.4ms	1.0ms	1.4ms

In most configurations, higher churn rates produce lower P95 values — the MutationObserver batches mutations, and more frequent small batches are cheaper to process than fewer large ones.

Heap Overhead

Sub-linear scaling. 20x more assertions costs 1.8x more memory.

Assertions	Churn	Heap (unthrottled)	Heap (CPU 4x)
Idle	—	+1.7KB	—
50	0/s	+76KB	+69KB
50	100/s	+92KB	+82KB
200	0/s	+115KB	+120KB
200	100/s	+150KB	+152KB
1000	0/s	+116KB	+117KB
1000	100/s	+140KB	+102KB

The agent's internal GC sweep (every 5 seconds) cleans up resolved assertions, keeping the working set bounded.

First Paint Impact

LCP impact is not statistically significant below 200 assertions. At 1,000 assertions — a deliberately extreme stress case — the initial attribute scan adds 6–12ms (one frame at 60fps).

For context: a typical instrumented page has 10–50 assertions. The 1,000-assertion test exercises the agent at 20–100x normal density to establish a scaling ceiling.

Caveats

What the benchmarks don't cover.

Chromium only. Firefox and Safari are not tested. The agent uses standard Web APIs (MutationObserver, querySelector, addEventListener) and is expected to work across modern browsers, but we have not validated performance empirically on other engines.
Lab INP, not field INP. INP is a lab estimate in headless Chromium with scripted interactions. Real-user INP requires actual interaction in the field.
Hardware-dependent. All benchmarks ran on Apple M4 Pro (14 cores, 48GB RAM), Chromium 147. Your numbers will differ. Run the tool yourself on comparable hardware.
Field validation in progress. We are validating real-user performance impact with design partners. Lab benchmarks are necessary but not a substitute for production telemetry.

Methodology

How we measure.

Each benchmark run launches a fresh Chromium instance per pair of measurements. Within each pair, condition A (page without Faultsense) and condition B (page with Faultsense) share the same browser process so V8 isolate noise cancels out in the differential.

Runs are strict A-B-A-B interleaved. The first pair is discarded as JIT warmup, leaving 29 usable pairs (demo) or 14 usable pairs per configuration (stress). Each measurement navigates to the target URL, waits for load, then soaks for 60 seconds.

Heap measurements use Chrome DevTools Protocol — not performance.memory which would pollute the target isolate. Two forced GCs run before each heap read. CPU profiler sampling interval is 100µs.

Core Web Vitals are captured via the web-vitals v4 library with explicit finalization via a synthetic visibilitychange event.

Statistical significance uses the Wilcoxon signed-rank test (two-tailed, non-parametric) on paired A-B differences. 95% confidence intervals use the Hodges-Lehmann estimator over Walsh averages.

An A-vs-A baseline (both conditions without the agent) validates measurement stability: 0ms LCP delta (p=0.351), +16 bytes heap (p=0.875). The apparatus does not create false positives.

Observer Integrity

The agent does not modify DOM nodes outside its own script element — it reads mutation records and attribute values, never writes to them. Events are dispatched via sendBeacon to reliably deliver without blocking navigation or user interactions.

The overhead numbers above confirm the observer does not perturb the signals it collects.

Run it yourself.

Clone the repo, run the benchmarks, get the same numbers.

# Demo benchmark (30 pairs, ~2 hours)
npm run benchmark:demo

# Stress benchmark (15 pairs, ~4 hours)
npm run benchmark:stress

# A-vs-A measurement validation
npm run benchmark:ava

Benchmark tool, stress harness, and raw data: github.com/Faultsense/faultsense-agent

Full statistical reports with p-values and confidence intervals: faultsense.org/docs