Skip to content

Flux reconciliation deep dive

Source artifacts and revisions

Source-controller fetches external content and publishes an artifact plus revision. Downstream controllers consume that artifact. Separate intervals mean source fetch, manifest application, and Helm reconciliation are related but distinct loops.

Kustomization behavior

Important fields interviewers expect you to distinguish:

  • path: directory built from the source artifact;
  • prune: garbage-collect managed objects removed from desired state;
  • wait: health-assess all reconciled resources;
  • healthChecks: assess selected resources when wait is not used;
  • dependsOn: block until another Flux Kustomization is Ready;
  • serviceAccountName: reconcile with constrained Kubernetes RBAC;
  • decryption: decrypt SOPS-encrypted Secrets before apply;
  • timeout and retryInterval: bound health waits and failed retries.

Pruning is ownership-sensitive. Removing YAML from Git can delete the corresponding live object; suspension and deletion policies change lifecycle behavior.

HelmRelease behavior

Beyond chartRef and inline values, know:

  • valuesFrom merges ConfigMaps/Secrets in order, then inline values override them;
  • dependsOn sequences releases by readiness;
  • install and upgrade remediation can retry, uninstall, or roll back;
  • tests can participate in release success and remediation;
  • driftDetection.mode: enabled compares live state with Helm storage and corrects differences;
  • serviceAccountName limits what the release can create;
  • targetNamespace and storageNamespace solve different problems;
  • suspend pauses future reconciliations but does not cancel an action already running.

Event-driven versus interval-driven reconciliation

Flux reconciles periodically, when watched resources change, and when explicitly requested. Webhook Receivers can reduce Git polling latency, but intervals remain the eventual consistency safety net.

Multi-environment reasoning

A common pattern separates:

clusters/<cluster>/        bootstrap and cluster entrypoint
infrastructure/            controllers, CRDs, policies
apps/base/                 reusable app definitions
apps/staging|production/   environment overlays and pinned versions

Promote by pull request and version/digest change, not by editing the live cluster. Avoid branches as environments unless the team accepts harder cross-environment diffing and promotion.