Manual steps, scripts, tribal knowledge. Cluster setup is slow and error-prone.
Teams wait for ops to provision infrastructure. No API, no automation, no guardrails.
Virtual machines and Kubernetes clusters managed separately with different tools.
"Is the platform actually working?" — no deterministic, testable proof of readiness.
A cluster isn't "provisioned" until its contract test passes. make new CLUSTER=ok-ai TYPE=talos
What every cluster must provide
Stable guarantees, fixed in an ADR
The reference implementation
Per-cluster configuration
Readiness gate — pass/fail
Immutable, API-driven
Installed automatically, dedicated LB IP
Metrics, logs, dashboards, alerts
Crossplane-ready (ADR-013)
No management-cluster SPOF
Cluster, PVs, secrets — idempotent
ServiceMonitor ingestion in Prometheus
Grafana pre-provisioned
Searchable in OpenSearch
Delivered to a receiver
Air-gap, no SPOF
ok CLI