← Status

kemek-monitoringmethod

How uptime is measured

The short version: maintenance we announced — and any spell when our own monitoring was blind — is left out of the headline number, and the number including everything is published too.

Questions

What do the ninety small bars mean?
One bar per day, oldest on the left, for the same ninety days the percentage beside them covers.
A bar is green when every check that day got a usable answer, amber when some did not, and red when none did.
A pale outlined bar is planned maintenance — we were watching, and the work was declared in advance, so it is neither an outage nor counted as uptime.
A faint bar means no data: either nothing was recorded that day, or the component did not exist yet. It is not a zero, and it is deliberately drawn differently from a bad day — "we were not watching" and "it was broken" are different things to tell you.
Hover a bar for the exact figure and the number of checks behind it.
What does the uptime percentage mean?
It counts checks, not minutes. Every check that ran in the last 90 days is one data point, and the percentage is the share of them that got a usable answer.
A slow-but-working response counts as available. A check that failed counts against it.
Does planned maintenance count against it?
No. When we declare a maintenance window in advance, the checks that run inside it are removed from both sides of the division — they are not counted as successes and not counted as failures.
90 days · one check every 5 minutes · 25,920 checks answered normally ......................... 25,538 failed ....................................     94 failed during declared maintenance ........    288
published 25,538 ÷ 25,632  =  99.63% (maintenance removed from both sides) including maintenance 25,538 ÷ 25,920  =  98.53% (published as uptime_90d_raw)
Isn't excluding your own downtime just flattering yourself?
It is a fair question, and it is a judgement rather than a fact. There are three things we could do with a maintenance window, and all three are defensible:
Count it as up. That claims we were serving when we were not, and it gets better the longer the maintenance runs. We think that one is dishonest.
Count it against us. That treats a deliberate, announced withdrawal as a failure.
Leave it out. That asserts nothing about the window either way, which is the only one of the three that does not say something untrue. It is what we do.
We would rather you did not have to take our word for it, so the other number is published too — see below.
Where is the number that includes maintenance?
On /status.json, beside the published one. Every component carries uptime_90d (what this page shows), uptime_90d_raw (everything included), maintenance_runs_90d and uplink_runs_90d (how many checks each exclusion removed).
Nothing is deleted to produce the headline figure. The raw record is what we keep; the published number is a view of it.
What if your monitoring is the thing that broke?
Then those checks come out too, and they are counted separately so you can see it happened.
When this service loses its own connection every component fails at once, and every one of those failures is ours rather than yours. We do not take the fleet going dark as proof on its own — it is indistinguishable from a shared host or an expired wildcard certificate genuinely taking everything down. Before concluding anything we try to reach two services we do not monitor, run by unrelated providers, and we only call it our outage if all of them are unreachable too.
Probes carry on running throughout, and the results are kept. What changes is that they stop counting as evidence about your service: they leave uptime_90d from both sides, exactly as maintenance does, and the count appears as uplink_runs_90d.
It is a narrower claim than it sounds. If we can reach the outside world and your component cannot be reached, that is recorded against the component and the alert fires — which is the direction it has to err in.
While it is happening, this page says so at the top. The status badges below that notice are still whatever the last check returned, and during one of our outages that check failed for our reasons — so a component may read "Outage" when it was serving the whole time. We leave the badges alone rather than blanking them, because blanking would also hide anything that is genuinely broken: we stop opening new incidents while we are blind, so an incident still listed during one started before we lost connectivity and is real.
What does a maintenance window actually stop?
It stops us raising an incident and paging ourselves. That is the whole of it.
It does not stop the checking. Probes keep running on their normal schedule for the entire window and every result is recorded, so the history has no hole in it and we can see the moment service comes back.
A recovery during a window still clears this page. And if something was already broken before the window opened, it stays on this page as an outage — planned work does not explain a problem that started before it.
If the service is still down when the window ends, the next failed check raises an incident normally.
Can a maintenance window be declared after the fact?
Yes, within 30 days, and it is labelled as such wherever it appears.
We allow it because real maintenance is often retroactive — a certificate renewed at 3am, a provider migration we learn about afterwards. A rule that only permitted windows declared in advance would mostly go unused, and unused means the alerts cry wolf during planned work.
What bounds it is evidence, not policy: we keep raw check results for 30 days, so past that there is nothing left to apply a window to. A window that old cannot be created, and an existing one cannot be deleted — removing it would change the uptime figure with no declaration left to explain why.
Why does a component say "no data"?
Because there were no checks to judge in the window — usually a component added recently.
It also covers a period spent entirely under maintenance, or one spent entirely inside an outage of our own. With every check excluded there is nothing left to divide, and we would rather show nothing than a 100% we did not measure.
How current is this page?
It is generated when you load it, from the most recent check results. The timestamp on the status page is the moment it was rendered.