Skip to main content
In short
  • Headline = the worst metric on the page, made worse (never better) by an open incident or running maintenance.
  • A metric is healthy, degraded or unhealthy against your thresholds. Comparisons are strict: exactly on the line does not match.
  • Changes wait for a confirmation time, so one bad reading never flips the page.
  • History bars judge each hour by how long it was bad. A short blip stays green.
  • Uptime % = the share of measured time that was not an outage.

The headline

  1. Roll up the metrics. The page takes its worst metric: unhealthy, then degraded, then healthy, then no data. Averaging would hide problems.
  2. Fold in incidents and maintenance. An open incident or running maintenance can make the headline worse, never better. A minor incident reads as degraded; a major or critical one as an outage.
The same verdict appears everywhere: the page, the incident history, the embed, the badge, the JSON API and the feeds. A metric whose agent has gone quiet reads Delayed. Running maintenance reads Under maintenance.

Per-metric status

Each value is compared against your thresholds:
  • Matches the healthy rule: healthy.
  • Otherwise, matches the unhealthy rule: unhealthy.
  • Otherwise: degraded.
“Over 99” means greater than 99. A value of exactly 99 does not match. For agent metrics the comparison runs on the agent, inside your network. See Threshold operators.

Confirmation time

A new status must hold for a confirmation time before the page changes. Getting worse and getting better have separate windows, so you can flag an outage fast and declare recovery carefully. Until then the page keeps the old status. See Thresholds and confirmation.

No data

When a check can’t produce a value (the target is unreachable, the query returned nothing), the sample is no data, never a guess. Each metric has a no data treatment: keep it as no data, or count it as healthy, degraded or unhealthy. For example, an error-rate query that returns nothing usually means no errors, so count it as healthy. The treatment applies to the current status and to history, but never colors days before the metric existed or hours that haven’t happened.

Delayed

If an agent stops reporting, the page doesn’t keep showing its last verdict. Once the newest sample is older than the staleness cutoff, the metric reads Delayed:
Delayed metrics leave the live rollup. The headline keeps the verdict of the metrics still reporting and adds a note such as “2 checks delayed”. You get an agent-offline alert rather than a metric alert, because the pipeline broke, not the service.

History bars

Each metric shows one bar per day in the page’s time zone. A day takes its worst hour. How an hour gets its color. Each stored sample holds its status until the next sample (up to the staleness cutoff). Then: The threshold is two minutes for each minute of the check’s interval, at least 2 minutes and at most 30, and never more than the time the hour was measured. Examples with the default 1-minute check:
The threshold applies to the colors of the bars, the hourly breakdown and the uptime calendar only. The uptime percentage counts every minute of measured time, and the current status, incidents, alerts, SLOs and SLAs count every sample.
Group summary mode shows one bar per day for a group, taking the worst metric of each day.

Uptime percentage

Uptime = one minus the unhealthy time divided by the time that had a status, over the visible window. Each status holds from its sample until the next one, as for the bars, but the hour threshold does not apply: three minutes of outage take three minutes off. Degraded time counts as up. Time with no data counts neither way. For a service or the whole page, each hour counts the check that was unhealthy longest in that hour, so two checks down at the same time count once. Planned maintenance is not excluded: a check that was down during maintenance counts as down. The figure is truncated, never rounded up. Status.json and the badge use the same figure over 30 days.

SLOs on the page

An SLO is a promise: this metric is up (healthy or degraded) a set share of the time over a rolling window. The error budget is what the target leaves: 99.9% over 30 days allows about 43 minutes unhealthy. A new SLO shows “collecting” until it has a day of data. See SLOs and error budgets.

Incidents and maintenance

An open incident affects the headline until it ends: when it is resolved, or its actual end time is recorded. Incidents annotate the history bars on the day they were posted and stay on the history page. Maintenance shows as maintenance, not as an outage, while its window runs.

Why the headline and the bars can disagree

“All systems operational” above an amber bar is intentional. The headline answers “can I use it now?”; the bars answer “how has it behaved?”. A service that failed yesterday and recovered is operational today.