Web vitals observatory (3D)

A Blender-built monitoring room: page loads fall into a 3D histogram with threshold zones and a p75 pole, and a 3D row of per-minute 404 bars trips a siren.

Switch to 2D

Presets

Playback

RUM sample (Core Web Vitals)

404 monitoring

Camera

Layers

Quality

Model check

The science

Core Web Vitals are 75th percentiles

Google assesses a page on three field metrics, each judged at the 75th percentile of real-user page loads: LCP (Largest Contentful Paint, good up to 2.5 s, poor above 4 s), INP (Interaction to Next Paint, good up to 200 ms, poor above 500 ms) and CLS (Cumulative Layout Shift, good up to 0.1, poor above 0.25). A page passes when all three p75 values are good. This page draws a synthetic Real User Monitoring (RUM) sample; each metric is lognormal, which is a common shape for timing data: ln X ~ Normal(mu, s2), median = exp(mu) and true p75 = exp(mu + 0.6745 s).

Sampling error of a percentile

A sample p75 from n loads is not the true p75. For a continuous distribution with density f the sample quantile is asymptotically normal with standard error SE = sqrt(p (1 - p) / n) / f(q), p = 0.75. The 95 % interval comes from order statistics, x(n p - 1.96 sqrt(n p (1 - p))) to x(n p + 1.96 sqrt(n p (1 - p))). Because the label (good, needs improvement, poor) depends on which side of a threshold the estimate falls, the probability of a wrong label is P(p-hat > t) when the truth is below t, computed from the same normal approximation and checked by simulation. The number of loads needed to clear a threshold with 95 % confidence is n = p (1 - p) z2 / (f(q) |t - q|)2: it grows without bound as the true p75 approaches the threshold.

404s as a Poisson process

If broken URLs are hit independently at a constant average rate lambda, the number in a window of W minutes is Poisson with mean lambda W: P(N = j) = e^-m m^j / j!. An alert fires when the count reaches k, so the false-alarm probability per calm window is the upper tail P(N >= k) and false alarms per day are (1440 / W) times that for non-overlapping windows. During an incident the rate rises to lambda + extra, which gives the detection probability per window and, for tumbling windows, a mean delay of W (1 / P_det - 1/2) minutes. Raising k lowers false alarms and lowers detection power; the ROC-style graph shows that trade for every k.

What is simplified

Real RUM data are not independent lognormal draws: they are mixtures of devices, networks and pages, with cache effects, and CLS has an excess of zeros. Real 404 traffic has bots, daily cycles and bursts, so it is overdispersed relative to Poisson and the independent-window false-alarm numbers are optimistic. The page uses the published metric thresholds but is an educational model, not a monitoring tool.

What you see in 3D

Each falling sphere is one page load; green, amber and red mark good, needs-improvement and poor loads. Bars are the histogram of the loads that have landed, the white pole is the sample p75 and the translucent block its 95 % interval. The back row of bars shows 404s per minute, the red glass plate is the alert level k / W, and the siren flashes when a window reaches k. The sixth check compares the index of dispersion (variance / mean) of the per-minute 404 counts shown with the Poisson value of 1.

Loading model...
Loading Blender model...
Running p75 vs sample size with the 95 % band
Chance of a wrong label vs sample size
404 hits per window: calm vs incident
Made with 🟠 Blender · Three.js