Metrics and SLOs

This sensor queries the numbers a running service reports about itself (error rates, latency, resource use) and compares each with a limit, a service level objective (SLO) or the previous version. Loops read it after a release, where a regression often shows first as a number that moves.

What it sees

Aggregates over a window of time: how often requests fail, how long they take, how much the service uses. Canary analysis compares the new version with the old one under the same traffic, so a difference belongs to the change and not to the time of day. A metric hides the requests behind it: it says that something got worse, not which request or which line of code made it so.

Reading the signal

• Ok: the metric stayed within its limit over the window read.
• Over: the metric passed its limit, or the new version did worse than the old one under the same traffic.
• Burn rate over its threshold: the SLO's error budget is being spent faster than it allows. A fast burn spends it in hours; a slow burn, left alone, spends it before the period ends.
• No data is not ok. A metric that reported nothing has observed nothing.

No data usually points at the run, not the work: the service did not start, its metrics were not collected, or a metric was renamed. Check that the series exists for the new version before trusting its value.