openova/platform/flux
e3mrah 238c6d2010
fix(bp-flux): mitigate helm-controller leader-election loss + stuck-HR recovery (#925) (#960)
* fix(bp-flux): mitigate helm-controller leader-election loss + recovery CronJob (#925)

On otech113.omani.works the bp-vpa HelmRelease became stuck Ready=Unknown
forever after a transient kube-apiserver blip caused helm-controller to
lose its leader-election lease mid-install. The Helm release secret was
already committed (Status=deployed) by the previous leader, but its last
write to the HR's Ready condition was Unknown and the new leader's
"release in storage?" short-circuit never re-evaluates that. The HR
blocked bootstrap-kit → sovereign-tls → cilium-gateway, breaking every
HTTPRoute on the Sovereign.

Fix is two-pronged:

1) PRIMARY (prevent the trigger). Stretch leader-election lease durations
   on the three Catalyst-critical controllers (helm/kustomize/source) from
   the upstream defaults of lease=35s renew=30s retry=5s to lease=60s
   renew=40s retry=5s, and bump memory limits from 256Mi to 512Mi (helm)
   / 384Mi (kustomize, source) so OOMKills during 35-HR fan-out installs
   don't themselves trigger leadership handoffs. Costs ~50s extra failover
   time on a real controller crash; that's acceptable since CP HA is a
   Phase 2 concern and we'd much rather avoid spurious flips during
   transient API pressure.

2) RECOVERY (handle the residual case). New CronJob bp-flux-stuck-hr-recovery
   runs every 2 minutes, scans every HelmRelease cluster-wide, and for each
   HR stuck in Ready=Unknown for >5 minutes whose underlying Helm release
   secret already has status=deployed, force-toggles spec.suspend (the only
   known workaround per #925). Guardrail: refuses to act if more than 10
   HRs would be touched in a single run (signals a cluster-wide outage).
   Operator-disablable via .Values.catalyst.stuckHelmReleaseRecovery.enabled=false.

Lock-in tests: tests/leader-election-and-recovery.sh covers all three
flag/memory bumps, CronJob render, RBAC presence, disable-toggle, and
threshold operator override. version-pin-replay + observability-toggle
still green.

Chart bumped 1.1.4 → 1.2.0.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(bp-flux): bump blueprint.yaml spec.version to 1.2.0 to match Chart.yaml (#925)

The bootstrap-kit static validation gate (Chart.yaml version ==
blueprint.yaml spec.version) caught the missed bump on PR #960.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: hatiyildiz <hatiyildiz@users.noreply.github.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-05 16:05:38 +04:00
..
chart fix(bp-flux): mitigate helm-controller leader-election loss + stuck-HR recovery (#925) (#960) 2026-05-05 16:05:38 +04:00
blueprint.yaml fix(bp-flux): mitigate helm-controller leader-election loss + stuck-HR recovery (#925) (#960) 2026-05-05 16:05:38 +04:00
README.md refactor(platform): remove k8gb — replaced by PowerDNS lua-records (#171) 2026-04-29 08:51:09 +02:00

Flux

GitOps delivery engine. Per-host-cluster infrastructure (see docs/PLATFORM-TECH-STACK.md §3.2). Catalyst runs one Flux instance per vcluster (lightweight: source + kustomize + helm controllers) plus a host-level Flux on each host cluster for the Catalyst control plane itself. Each per-vcluster Flux pulls from the single per-Sovereign Gitea instance (see docs/ARCHITECTURE.md §4).

Status: Accepted | Updated: 2026-04-27


Overview

Flux provides GitOps-based continuous delivery with Gitea as the internal Git provider and External Secrets Operator for secrets management.


Architecture

flowchart TB
    subgraph Gitea["Gitea (Internal Git)"]
        Components[Component Repos]
        Organization[Organization Repos]
    end

    subgraph K8s["Kubernetes Cluster"]
        subgraph Flux["Flux Controllers"]
            Source[source-controller]
            Kustomize[kustomize-controller]
            Helm[helm-controller]
            Notify[notification-controller]
        end

        subgraph ESO["Secrets"]
            ESOp[External Secrets Operator]
            PS[PushSecrets]
        end

        Resources[K8s Resources]
    end

    subgraph External["External"]
        OpenBao[OpenBao]
    end

    Components --> Source
    Organization --> Source
    Source --> Kustomize
    Source --> Helm
    Kustomize --> Resources
    Helm --> Resources
    Notify --> Gitea
    ESOp --> OpenBao
    ESOp --> Resources

Why Flux

Factor Flux ArgoCD
Memory overhead ~200MB ~500-800MB
Architecture Kubernetes-native CRDs Separate UI/API
Secrets Via ESO (PushSecrets) Via ESO (PushSecrets)
CLI workflow Excellent UI-focused

Key Decision Factors:

  • Lower resource overhead
  • CLI-focused fits single-developer workflow
  • Kubernetes-native CRDs
  • Works well with External Secrets Operator
  • Integrates seamlessly with Gitea

Components

Controller Memory Purpose
source-controller 64MB Git/Helm repo sync
kustomize-controller 64MB Kustomization apply
helm-controller 64MB HelmRelease management
notification-controller 32MB Alerts

Repository Structure

flux/
├── clusters/
│   └── <region>/
│       ├── flux-system/       # Flux controllers
│       ├── network/           # cilium, stunner, powerdns
│       ├── security/          # kyverno, external-secrets, cert-manager
│       ├── database/          # cnpg, ferretdb, valkey
│       ├── middleware/        # strimzi
│       ├── storage/           # seaweedfs, velero
│       ├── observability/     # grafana (LGTM stack)
│       ├── autoscaling/       # keda
│       ├── workplace/         # stalwart
│       └── organizations/     # per-Organization Application deployments

Categories

Category Components Purpose
network cilium, stunner, powerdns CNI + Service Mesh, TURN, authoritative DNS + lua-records GSLB
security kyverno, external-secrets, cert-manager Policy, secrets, TLS
database cnpg, ferretdb, valkey Database operators
middleware strimzi Apache Kafka streaming
storage seaweedfs, velero Object storage, backup
observability grafana LGTM stack
autoscaling keda Event-driven scaling
workplace stalwart Email server
organizations <org> Per-Organization Application deployments

Configuration

GitRepository for Gitea

apiVersion: source.toolkit.fluxcd.io/v1
kind: GitRepository
metadata:
  name: <component>
  namespace: flux-system
spec:
  interval: 5m
  url: https://gitea.<location-code>.<sovereign-domain>/<org>/<component>.git
  ref:
    branch: main
  secretRef:
    name: gitea-token

Kustomization

apiVersion: kustomize.toolkit.fluxcd.io/v1
kind: Kustomization
metadata:
  name: <component>
  namespace: flux-system
spec:
  interval: 10m
  targetNamespace: <namespace>
  sourceRef:
    kind: GitRepository
    name: <component>
  path: ./deploy/prod
  prune: true
  wait: true
  healthChecks:
    - apiVersion: apps/v1
      kind: Deployment
      name: <component>
      namespace: <namespace>

Multi-Region GitOps

Catalyst runs one Gitea per Sovereign on the management cluster (see docs/PLATFORM-TECH-STACK.md §2.3). Each workload region's Flux pulls from that single Gitea over the cross-region network. Per-region HA comes from Gitea's intra-cluster replicas + CNPG-backed metadata, not cross-region bidirectional mirror.

flowchart TB
    subgraph Mgt["Management cluster (per Sovereign)"]
        Gitea[Gitea — single instance, K8s-replicated for HA]
    end

    subgraph Region1["Workload region 1 (rtz)"]
        Flux1[Per-vcluster Flux]
        K8s1[Org vcluster workloads]
    end

    subgraph Region2["Workload region 2 (rtz)"]
        Flux2[Per-vcluster Flux]
        K8s2[Org vcluster workloads]
    end

    Gitea --> Flux1
    Gitea --> Flux2
    Flux1 --> K8s1
    Flux2 --> K8s2
  • One Gitea per Sovereign (on the mgt cluster). HA via intra-cluster replicas, not cross-region mirror.
  • Each vcluster runs its own lightweight Flux that pulls from this single Gitea.
  • Per-region configuration is encoded in the Environment's manifests via Kustomize overlays / Placement metadata.

Secrets Management

Flux uses External Secrets Operator (ESO) with PushSecrets pattern:

  • No SOPS: SOPS has been eliminated from the architecture
  • PushSecrets: 100% PushSecrets pattern for multi-region
  • K8s Secrets as source of truth: Apps read from K8s Secrets only

Release Management

Release Flow

Feature Branch → PR → main → Flux Sync → Staging → Promote → Production
                                            ↓
                                      Canary Analysis (Flagger)
                                            ↓
                                      Auto-Rollback (on failure)

Environment Strategy

Environment Sync Interval
Staging 1m
Production 10m

Rollback Strategy

Trigger Action
Flagger metric failure Auto-rollback
Manual intervention Git revert + sync
Emergency kubectl rollout undo

Bootstrap

# Initial cluster bootstrap (Gitea)
flux bootstrap git \
  --url=https://gitea.<location-code>.<sovereign-domain>/openova/flux \
  --branch=main \
  --path=clusters/<region> \
  --token-auth

Key Commands

# Status overview
flux get all

# Force reconciliation
flux reconcile kustomization organizations

# View logs
flux logs --all-namespaces

# Suspend (manual gate)
flux suspend kustomization <org>-prod

Sync Intervals

Resource Interval
GitRepository 1 minute
Kustomization 10 minutes
HelmRelease 10 minutes

Gitea Actions Integration

Gitea Actions can trigger Flux reconciliation:

# .gitea/workflows/notify-flux.yaml
name: Notify Flux
on:
  push:
    branches: [main]

jobs:
  notify:
    runs-on: ubuntu-latest
    steps:
      - name: Trigger Flux reconciliation
        run: |
          curl -X POST \
            -H "Authorization: Bearer ${{ secrets.FLUX_WEBHOOK_TOKEN }}" \
            https://flux-webhook.<location-code>.<sovereign-domain>/hook/...          

Part of OpenOva