Alerts
On a cluster running Alertmanager, Alerts is what is wrong right now — grouped by state, with the silences over them, and a sentence at the top that says how current the list actually is.
What the page shows#
Alerts are grouped by the rule that raised them, and the rules by state:
- Firing — with a severity, how long it has been firing, and one line per instance with its labels.
- Pending — the condition is true but has not been true for long enough yet. Prometheus has these; Alertmanager has not been told about them.
- Silenced — still firing, but muted: by whom, until when, and the comment they left. Expire ends the silence early.
The badge on Alerts in the sidebar counts what is firing and not silenced — a silence is a decision that something is known, so counting it would make the badge mean "alerts" rather than "things to look at".
Keeping the list current#
Every other cluster page follows a watch stream and changes as the cluster does. Alertmanager has no such stream, so this one page has to ask again on a timer, and how often is a choice between old-but-cheap and current-but-chatty that is yours rather than ours.
Settings › General › Alerts is where it is made:
The picker offers Off, Every 15 seconds, Every 30 seconds, Every minute and Every 5 minutes. The default is every 30 seconds. Off is the first entry in that same picker rather than a separate switch beside it — a toggle and an interval can disagree, and there is no useful answer to "off, at 30 seconds".
The poll lives exactly as long as the Alerts page is on screen. Kontena never polls a cluster nobody is watching, and switching to another page stops it.
This is one setting for the whole installation, not one per cluster: the same interval applies wherever you open the page. Turn it off and the list is read when you open the page and when you press refresh — and it still tells you when that was.
When a refresh fails#
A failed read does not clear the page. The rows that are on screen stay — they are the only picture of the cluster anyone has — and the notice turns amber to say they are older than the clock suggests:
The refresh at 14:21:07 failed: <what came back> These alerts were read at 14:18:37 and may be out of date.
If the very first read never landed there is nothing to keep, and the notice says that instead — because an empty list would otherwise read as an all-clear that was never given.