How the pressure is computed

For every signal and scope we keep a trailing baseline of 7 days, restricted to the same hour of day ±1 h once at least 3 days of history exist (normal daily patterns are not pressure). The most recent 600 s are excluded so an incident cannot baseline itself. Anomaly strength is the robust z-score

z = (x − median) / MAD, clipped to [-3, 8]

A (probe, target) pair is abnormal above z = 3. Signals with fewer than 12 samples are “weak coverage” and their weight is damped: no strong conclusions from thin data. Each component sums its weighted signal stresses and saturates into a 0–100 score with score = 100 × (1 − e^(−1.2 × stress)), then the index is the weighted sum of component scores. The engine runs every 10 s over a 120 s window (60 s for BGP rates).

Each target and ASN carries a Network Importance Score (1–5) reflecting centrality, prefixes, downstream dependency and known services; aggregates weight importance 5 about five times more than importance 1. Event pressure is anomaly strength × affected scope × network importance × confidence. Confidence combines probe count, geographic diversity, signal agreement, BGP corroboration, magnitude, duration and external corroboration. Alongside the index we publish velocity (points per hour), acceleration (per hour²) and volatility so that “high but recovering” is distinguishable from “moderate but worsening”.

A Pressure Front is declared when at least 3 source-region → destination-region pairs rise together above z = 2.5 with intensity ≥ 35; it is drawn on the map as a great-circle arc with a direction. Incidents open when a component or regional score exceeds 45 for 2 consecutive cycles (detected → developing), become active after 6, recover below 30 and resolve after 600 s continuously below. Correlation is rule-based first — BGP spike + latency + path change + failures raise confidence together — and hypotheses are always phrased as “possible”, “probable” or “high-confidence”.

If fewer than 2 probes are fresh (a probe is fresh if it reported within 180 s), if the BGP feed is older than 120 s, or if our stores are unhealthy, the engine freezes the index at its last value and flags internal_status = degraded. The interface then shows a visible “Instrument degraded” state with the freeze time instead of interpreting a frozen number. A probe whose targets fail at ≥ 0.8 at once is excluded — its own uplink is down, not the Internet.

Nothing on this site is animated without a measurement behind it: if one update arrives every 10 s, the numbers move every 10 s. We never randomise values or synthesise events, and pages for countries, ASNs, services and incidents exist only where we hold real data.