Visibility transitions per day
Spotlight day 2026-10-10 — per 5-minute slot
AS visibility curves (5-minute resolution, click chips to toggle)
Cliff episodes (visible-prefix count below 90% of the 60-min baseline, depth ≥ 20%, loss ≥ 50 prefixes)
| AS | name | window (UTC) | baseline | min | depth | minutes down | prefixes gone | of which multi-collector | recovery |
|---|
Method & limits
What is tracked. Per (collector, peer) session we keep only the set of announced
prefixes (membership), never paths or attributes. A prefix's visibility set is the set of
collectors where at least one session announces it; n_collectors is the size of
that set. A transition row is written only when that set changes in a slot.
What a cliff episode is here. Per origin AS, the visible-prefix count is rebuilt at 5-minute resolution; an episode is a run of slots where the count stays below 90% of its 60-minute baseline, at least 20% deep and at least 50 prefixes below it. "2→0 flips" counts prefixes that dropped out of both collectors inside the episode: the strongest evidence that a change is origin-side rather than a single collector's session reset.
Known limits. Three collectors, nine days, one fold pass (cold start on day 1). Session-level events (a peer resets and re-announces its table) show up as large reveal/loss bursts that are not outages: the 2026-10-10 09:10 slot is the calibration example, 988 ASNs hit, 97% of them in a single collector only. The origin ASN is kept once per prefix (the last announced origin), not per collector or per session, so when an origin attribute changes a still-visible prefix silently moves between ASN buckets; the prefixes gone column counts prefixes that really left every collector, and episodes with no such disappearances are filtered out as attribution moves. No peer-level attribution is stored.
Data. RIPE NCC RIS updates files (5-minute slots) for ;
fold output loaded into PostgreSQL (visibility.transitions, visibility.as_census)
and exported to the static JSON behind this page.