Hover a node · click to focus · ←/→ steps
Promote only after eval. Canary a traffic slice on K8s; auto-rollback on SLO breach; GitOps is the audit trail.
When you need this
- Prefer Flux over Argo in an existing GitOps estate
- Want image update automation when registry promotes
- Need multi-tenant clusters with Flux Kustomizations
- Progressive delivery with Flagger + Flux
Prerequisites
- Flux installed (or plan to bootstrap)
- Git repo for cluster desired state
- Container registry webhooks / image policy
Tools
Assess as Argo CD peer; pick one GitOps engine and standardize.
Start with registry + experiment tracking before full deployment automation.
Steps
- 1
Bootstrap Flux for the AI namespace
GitRepository + Kustomization for staging/prod overlays. Keep secrets out of Git (SOPS / External Secrets).
- 2
Image automation from registry
ImageRepository + ImagePolicy watch the model service repo. On MLflow-linked build, Flux opens/updates the image tag in Git.
- 3
Progressive delivery
Optional Flagger canaries on the Flux-managed Deployment. Promote on metrics; rollback on SLO breach.
- 4
Operate and drill
Suspend/resume for incidents. Quarterly rollback by Git revert. Document who can reconcile prod.
Adoption pitfalls
- Image automation writing to prod without PR in locked-down setups
- Drift from kubectl edit
- No link between Flux commit and MLflow version
Adoption checklist
- Desired state only in Git
- Image tags updated by automation, not humans in cluster
- Prod change requires PR
- Rollback drill completed
SEER REAL assessment / sprint
Assessment maps GitOps tool choice and gaps. Sprint bootstraps Flux for one model service with image automation and a rollback drill.