Internet weather

Every figure on this page is a percentile of response times our own probes measured from 4 named fleet regions (US-East, US-West, Europe, Asia-Pacific) against the public, unauthenticated endpoints of the 192 services in our public outage catalog. It is a picture of how the internet looked from each vantage point, not a reading of any one service.

Not customer traffic. Customer monitoring data never enters this aggregate: the job reads one table, the catalog's own raw probe results, and nothing else. There is nothing to anonymize because nothing private goes in.

Only successful readings (HTTP 2xx with a recorded latency) count; failed and filtered readings carry no response time worth aggregating. Hourly and daily buckets under the sample floor are left out rather than drawn. A region reads slower than usual only after 3 consecutive hours with a median above 1.5x its 7-day baseline, and clears on the first hour at or under 1.25x, so a region hovering around the line does not flap. The 7-day baseline is the median of the region's hourly medians over the 7 days before the hour being judged.