Connection events and uptime

Two RIPE Atlas built-in measurements underpin most of the health and outage analysis in Atlas. Both are ingested continuously and kept forever.

Connection events — measurement 7000

Every time a probe connects to or disconnects from RIPE's infrastructure it emits an event. Atlas pulls these every 15 minutes into a permanent archive, one row per event.

Stitched together, connect/disconnect pairs are the ground truth for a probe's uptime — and, aggregated by AS, country or time bucket, for the health of a whole network. When a country "goes dark" or an AS drops off, this is the series that says so.

Late arrivals

An event carries both when it happened and when RIPE accepted it, and the second can lag the first by seconds to minutes — occasionally hours during an incident. Rather than trying to track a moving watermark, ingest re-scans a generous overlap window on every pass and relies on each event having a stable identity, so re-importing the same event changes nothing.

The practical consequence: the most recent few minutes of any event-derived view can still fill in. Don't read a sharp drop at the right-hand edge of a chart as an outage until it has had time to settle.

Longer holes — a multi-hour storage delay upstream, or extended downtime on this side — are recovered by an explicit backfill rather than automatically.

Uptime — measurement 7001

Every probe reports its uptime roughly every three minutes. Storing that raw would be millions of near-identical samples a day, so Atlas stores change-points instead: a row is written only when something meaningful moves.

Four things count as meaningful:

• Firmware version changed
• Source IP changed — the probe was renumbered or re-homed
• A reboot — the uptime counter went backwards
• A clock resync — the probe's local time jumped far enough to mean a real re-sync rather than the sub-second drift of a healthy probe

A quiet fleet writes almost nothing. This is why probe history in Atlas reads as a list of events rather than a continuous trace: the flat stretches genuinely contained no news.

Coverage is tracked separately

Because a day with no change-points is indistinguishable from a day that was never processed, Atlas records which time windows were scanned independently of what was found. A quiet day reads as fully processed rather than as a gap — which is what the coverage calendars on the archive views are showing you.

Reading uptime from events

Uptime for a period is reconstructed by pairing connects with disconnects. Two cases need care:

• Open intervals. A probe still connected has no closing event; a probe that disconnected and never came back has no reopening one. Both are normal.
• Missed disconnects. Sometimes a probe goes away without a disconnect event being recorded — it lost power, or the event was lost upstream. Naive uptime then counts the probe as connected for the whole gap, inflating the number. There is a staff view (/statistics/missed-disconnects) dedicated to finding these; see Statistics.

Where it shows up

• Statistics → Connection-event stats — aggregate rates
• Trends → Connections and Uptimes — long-run series
• Probes — a single probe's connection history
• Deployments — countries going dark