Monitoring metrics

Uptime, response time, incidents and recovery: what each figure counts.

The measured figures on a provider's listing and status page, and precisely what each one counts.

Uptime

The proportion of measurements over the window in which the provider's address space was reachable, according to the consensus rather than to any single worker.

It carries its window (30, 60 or 90 days) and, on the status page, what it was measured across: countries, networks, and independent workers. A figure is only published for a window the record actually reaches back over; where it does not, the page says how much of the window exists instead of averaging what it has under a label that claims more.

Average response

The median round-trip time across measured targets. Median rather than mean, because RTT distributions have a long tail and one badly routed path would drag a mean somewhere unrepresentative.

This is network distance from the workers, not machine speed. A provider whose workers are mostly nearby will show a lower figure than one measured from further away, which is why the number is most useful compared against the same provider over time, or against providers measured from a similar spread.

Incidents

Periods the consensus recorded as degraded or out. An incident is not a customer report and not a provider announcement: it is a stretch of time during which independent measurement said something was wrong.

Counted over the same window as the uptime figure beside it.

Average recovery

How long incidents took to resolve, across resolved incidents in the window. An ongoing incident is not counted, because its duration is not known yet, and including a partial duration would drag the average down and make an ongoing problem look like a fast recovery.

The availability timeline

One cell per day, coloured by that day's worst state. Worst rather than average, for the reason given in How consensus is calculated: a day with a two-hour outage is not an operational day.

Weekly and monthly views aggregate the same way. A day with no measurement is drawn as unmeasured, distinct from a day the network measured and could not reach a verdict on.

What none of these say

Nothing here measures a specific server, a specific service, or a specific customer's experience. It measures whether a provider's network answers the rest of the internet, which is a real and independently checkable thing, and a smaller thing than "is this provider good".