gke-alert-configuration

Define google_monitoring_alert_policy resources with condition_prometheus_query_language blocks containing PromQL queries. Declare variables for project_id, cluster_name, and namespace, and use grouping aggregations like by (cluster, namespace, service) so one policy dynamically covers all workloads.

It evaluates error budget consumption across two simultaneous lookback windows, such as 14.4x burn rate over 1 hour and 5 minutes for a 99% SLO. Both windows must exceed the threshold, eliminating false alarms from short bursts that simple error-rate ratios produce.

Yes. Managed Service for Prometheus bills per ingested sample, and unmodified kube-state-metrics scrapes state for every cluster object. Use a filtered PodMonitoring resource with metricRelabeling keep rules to ingest only the specific kube_* metrics your alerts require.

When traffic drops to zero, Prometheus stops emitting the time series, so rate(...) == 0 evaluates to an empty vector and never fires. Use sum(rate(...[5m])) default 0 == 0 or absent(...) == 1 to reliably detect total traffic loss.

No. It is strictly scoped to GKE workloads using Managed Service for Prometheus and PromQL. For standalone Compute Engine or Cloud Run, use standard google_monitoring_alert_policy with condition_threshold or MQL against metrics like compute.googleapis.com/instance/cpu/utilization.

Queries with aggregated lookback windows like increase(...[15m]) already smooth transient spikes, so adding duration = "300s" only inflates mean time to detect by up to 20 minutes. Set duration to "0s" or "60s" for lookback-based queries and reserve 300s for instantaneous gauges.