Failure modes

Each result names one failure mode. A predicted failure has not happened yet and carries etaSeconds. A detected failure is happening now.

A predicted saturation, release_regression or replica_divergence also carries etaEarlySeconds and etaLateSeconds: the earliest and latest failure, from the first and third quartiles of the growth rates the readings show. GET /api/v2/failures/series?workload=<namespace/kind/name> returns the memory readings behind the forecast, which the web console draws with the trend and the range.

Failure mode	Kind	Raised when
saturation	predicted	Memory of a container is on course to reach its limit within the horizon. Needs a memory limit and at least 10 minutes of samples.
release_regression	predicted	A new version of a Deployment grows toward its memory limit at least 3 times faster than the version it replaced, for 5 minutes past a 6 minute warm-up. Reported up to 2 hours ahead, with both rates and a roll back as action.
replica_divergence	predicted	One of at least 3 replicas of a version grows at least 3 times faster than the others, for 5 minutes. Reported up to 2 hours ahead and names the pod. Silent when all replicas grow together.
recurrent_failure	predicted	A workload failed at least 6 times at regular intervals; the next failure is expected one interval after the last.
oom_killed	detected	The last exit of a container was OOMKilled, in the last 30 minutes.
crash_loop	detected	Kubernetes reports a back-off restarting a failed container, or at least 3 restarts in the last 30 minutes. A loop of OOM kills with a regular rhythm is reported as recurrent_failure, predicted, with the next kill.
startup_stuck	detected	Kubernetes reports a pod that cannot start: image pull failure or FailedMount.
replicas_unavailable	detected	Deployment, StatefulSet or DaemonSet missing replicas for over 5 minutes.
rollout_stuck	detected	A Deployment passed its progress deadline.
cpu_throttling	detected	A container throttled at least 25% of the time for 10 minutes. Clears after 5 minutes below.

What a failure can take down

The engine reads which workloads call which from what the cluster declares: a container environment value that names a Service, and Routes or Ingresses that send traffic to one. GET /api/v2/callgraph returns the graph. Each failure lists in affects the workloads that call it, directly or not, adds an evidence line, and sends an impact annotation to Alertmanager. These are declared calls; calls seen in traces or network flows come later.

State

State	Meaning
critical	The failure is happening or close
degraded	The failure is expected, with time to act
calibrating	Not enough history yet to judge

What Ruptura cannot see

A workload without a memory limit cannot be forecast for saturation. GET /api/v2/coverage lists what each instance can and cannot measure on your cluster.