How uptime is calculated

Uptime is measured in time, to the second: what counts as downtime, as available, as planned maintenance and as not monitored, and what the colours of the bars mean.

Heimdall measures uptime in time, not in checks. It keeps a record of the state each monitor was in and when, and every figure — in the app and on status pages — is read from that record.

The formula

uptime         = available time ÷ monitored time
monitored time = the period − planned maintenance − time not monitored

When nothing was monitored in a period there is no figure: you see a dash, never 0% or 100%. A figure has two decimals and is cut, not rounded: 99.9996% reads 99.99%, and only a period without a second of downtime reads 100.00%.

What counts as what

Counts as	The monitor was	In the figure
Available	`up`, or `degraded`: a region failing below the quorum, or slow answers	Uptime. Degraded time is also shown as its own duration and colour
Downtime	`down` by its quorum, or `flapping`	The only thing that lowers uptime
Planned maintenance	Inside a maintenance window	Left out, from the moment the window began to the moment it ended
Not monitored	Paused, new and not yet checked, `unknown` (too few regions could check it), not created yet — or Heimdall itself was not watching	Left out: neither uptime nor downtime

Downtime is what the quorum decided, not a failed check: one region failing while the others pass makes the monitor degraded, and a degraded monitor is available. See How quorum decides.

The availability card of a monitor over the last 90 days: the uptime figure; downtime, planned maintenance and not-monitored time side by side; one bar for each day; and the monitor’s one outage with its start and end to the second. — The availability card on a monitor’s page: the figure, the three kinds of time behind it, a bar for each day, and every outage.

Measured to the second

A monitor changes state when its quorum decides so, and the change is booked at the time the deciding check was measured — not when its result reached us. An outage is as long as it was measured to be, to the second; a result that arrives late neither shortens nor lengthens it.

Durations are written in their two largest units: 45 s, 4 min 30 s, 2 h 5 min, 3 d 4 h.

Planned maintenance

A maintenance window is taken out exactly, from the instant it began to the instant it ended — not the hour around it, and not the day.

A window from 14:10 to 14:20 takes ten minutes out of the monitored time
An outage at 14:40 in that same hour is downtime, all of it
An outage that began before a window and is still there after it is one outage; what of it lay inside the window does not count
A monitor that was down only inside a window had no outage

Covered monitors are still checked while a window runs, so you see the service go away and come back. See Planned maintenance.

Not monitored

Time for which there is no evidence either way is left out like maintenance, and shown as its own duration. A paused monitor collects neither uptime nor downtime, and a new one does not start with a day of 100%.

The bars

The monitor list shows the last 30 days, a monitor’s page 24 hours to a year under Availability, a status page 90 days. Each day — or hour, or month — has the colour of its level:

Colour	Level	When
Green	Operational	No downtime, and at most 0.1% of the monitored time degraded
Yellow	Disrupted	Any downtime at all, or more than 0.1% of the monitored time degraded
Red	Major outage	Uptime below 95%
Blue	Planned maintenance	Nothing but planned maintenance is left of the day
Grey	No data	Nothing was monitored

A day keeps the colour of what happened on it. Planned maintenance that ran on that day is a blue slice at the foot of its bar, as large as its share of the day and never too thin to see; only a day of which nothing but maintenance is left is blue altogether. Point at a day, or tap it, for its date, its level and its figures.

The components of a status page, each with 90 bars: most days green, a few yellow for a disrupted day, one red for a major outage, and a blue slice at the foot of the days on which planned maintenance ran. Under each row are the uptime, the downtime and the maintenance time. — The bars of a status page. The blue slices are planned maintenance: shown, and not counted as an outage.

Days and time zones

In the app a day is a calendar day of your time zone (Profile), today included; the last 24 hours come in hours and the year in calendar months. Status pages are public and use UTC days. How far back the app shows depends on the history of your plan: a longer period is cut to it, and the app says where it starts.

The charts on a monitor’s page are something else

The response-time chart and the strip under it show what each region measured: response times, and the share of checks that succeeded. That share is not uptime. A check that fails in one region is not downtime, and a monitor can be 100.00% available while a region fails some of its checks.