Symptom

A BudgetSourceUnavailable Warning event, the stageset_metric_source_errors_total counter rising, and — when a gate’s onSourceError is Hold — READY=False, REASON=BudgetSourceUnavailable. With onSourceError: Allow the rollout proceeds and only the event and metric fire, so the condition may not show the reason.

Cause

A metric source (the error-budget freeze or a promotion analysis check) was unreachable or returned no usable scalar: the Prometheus endpoint was down or returned a non-2xx, the query evaluated to an error, an empty or multi-sample vector, or NaN (the shape a wrong query takes). This is treated as transient — the controller keeps retrying.

What happens to the rollout depends on the gate’s onSourceError:

• The error-budget freeze defaults to Allow (fail-open): blocking it would stop every deploy, including the hotfix you need during the very outage that took the source down.
• A promotion analysis defaults to Hold (fail-closed): holding only parks the rollout at the current healthy stage, so it is safe to refuse to advance while behavior can’t be verified.

Diagnosis

kubectl --namespace <namespace> describe stageset <name>   # the event names the source error
kubectl --namespace <namespace> get stageset <name> --output jsonpath='{.status.stages[*].promotionState.lastAnalysis}'

Check the address and query in spec.errorBudget.source / spec.stages[].promotion.analysis.checks[].source. From a pod in the namespace, confirm the Prometheus endpoint is reachable and the query returns a single scalar:

curl -s '<address>/api/v1/query?query=<query>'

A NaN/empty result means the query’s labels match nothing — fix the query, not the gate. A connection failure points at NetworkPolicy or a wrong address. If a bearer token is configured, confirm the referenced Secret has a non-empty token key.

Remediation

Fix the source so the query returns a single numeric scalar. While debugging a freeze, spec.errorBudget.dryRun: true records what would happen without holding anything. If a service is genuinely safety-critical and you want it to stop deploying when its SLO source is down, set spec.errorBudget.onSourceError: Hold (accepting that a source outage then blocks its deploys).