Delivery as reconciliation

GitOps is most useful when desired state, reconciliation order, and ownership boundaries are explicit enough to operate under failure.

  • FluxCD
  • Helm
  • Kustomize
  • GitOps
Site plan · K8S-031

A controller changes the delivery model

A pipeline usually performs an action and exits. A GitOps controller keeps comparing desired state with observed state. That shift from imperative delivery to reconciliation is the real architectural change: drift, retries, dependencies, and partial failure become continuous concerns rather than exceptional pipeline branches.

Git is useful here because it provides reviewable intent and history, not because every operational fact belongs in a repository. Secrets, generated state, and runtime observations still need purpose-built systems.

Repository boundaries are ownership boundaries

Bases, overlays, Helm releases, and Kustomizations are easy to compose technically. The harder question is who may change each layer and who responds when reconciliation fails. Platform-owned infrastructure, shared capabilities, and application releases usually benefit from separate scopes even if one controller observes them all.

Dependencies should also be visible. Namespaces, controllers, custom resource definitions, and workloads do not become ordered merely because their YAML files sort neatly. Explicit dependency edges and health checks make rollout behavior understandable.

Design for the failure loop

A good GitOps design answers what happens when source is unavailable, a release is unhealthy, a human makes an emergency change, or two controllers claim the same object. Reconciliation works best when ownership is singular, promotion is deliberate, and operators can suspend, inspect, and resume the loop without bypassing its source of truth.

Further down

Keep digging.

  1. EDGE-012

    Ingress is a routing problem

    Ingress failures become tractable when DNS, TLS, edge reachability, controller routing, service discovery, and workload health are tested as one path.

    • Gateway API
    • Istio
    • Traefik
    • DNS
    2 min read
  2. CI-024

    One pipeline primitive, many languages

    Reusable CI should standardize security and delivery behavior across languages while leaving language-specific build logic explicit.

    • CI/CD
    • Reusable workflows
    • Supply chain
    2 min read
  3. OPS-044

    Restore is not recovery until the cluster agrees

    Restoring etcd data is one step in Kubernetes recovery; control-plane identity, manifests, endpoints, certificates, and reconciliation still have to agree.

    • etcd
    • Kubernetes
    • Recovery
    2 min read