d1431bed09
Closes #921 — bp-cluster-autoscaler-hcloud chart shipped without HCLOUD_CLUSTER_CONFIG / HCLOUD_CLOUD_INIT, so cluster-autoscaler 1.32.x FATALs at startup with "HCLOUD_CLUSTER_CONFIG or HCLOUD_CLOUD_INIT is not specified" on every Sovereign (otech112 evidence). HelmRelease reports Ready=True (Helm install succeeded) but the Pod CrashLoopBackOffs invisibly behind the False-positive condition. Closes #916 — wizard let operators dispatch unbuildable topologies (otech109: cpx32 worker in `ash`) because PROVIDER_NODE_SIZES did not encode regional orderability. Hetzner rejected the worker creation 41s into `tofu apply` after Phase-0 had already created the CP + network + LB + firewall. Chart fix (issue #921): - Add `clusterAutoscalerHcloud.{clusterConfig,cloudInit}` values to the umbrella chart (base64-encoded per upstream contract). - Render `hetzner-node-config` Secret unconditionally with both keys so the upstream Deployment's secretKeyRef references resolve cleanly during `helm template` AND in the live cluster regardless of overlay state. - Wire HCLOUD_CLUSTER_CONFIG + HCLOUD_CLOUD_INIT extraEnvSecrets onto the upstream chart's deployment. - Tofu Phase 0 base64-encodes the Phase-0 worker cloud-init and stamps it under `flux-system/cloud-credentials.hcloud-cloud-init`; the bootstrap-kit overlay lifts that key via Flux `valuesFrom` into `clusterAutoscalerHcloud.cloudInit`. Autoscaler-spawned workers thus receive the IDENTICAL bootstrap as the Phase-0 worker fleet. - Bump bp-cluster-autoscaler-hcloud chart 1.0.0 → 1.1.0. - Chart-test smoke gate (chart/tests/hetzner-node-config.sh) verifies Secret + env var wiring + no-regression of HCLOUD_TOKEN — runs in CI's blueprint-release "Run chart integration tests" step. Wizard fix (issue #916): - Add `availableRegions?: string[]` to NodeSize interface; encode cpx32 = ['fsn1','nbg1','hel1'], cpx21/cpx31 = [] (orderable nowhere new) per Hetzner /v1/server_types vs POST /v1/servers gap. - Add `isSkuAvailableInRegion()` + `suggestAlternativeSkus()` helpers. - StepProvider filters SKU dropdowns by selected region; auto-swaps current SKU to recommended default when region change drops it out of orderability. - Mirror the matrix Go-side in sku_availability.go; gate `provisioner.Request.Validate()` with same predicate so a stale wizard build OR direct API caller bypassing the UI cannot dispatch otech109's failure mode. - Two-sided enforcement covers both r.Regions[] (multi-region) and the legacy singular path. Tests: 13 vitest cases on the wizard side + 38 Go subtests on the API side. Chart smoke renders + helm template gates the env wiring at publish time. Co-authored-by: hatiyildiz <hati.yildiz@openova.io> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1463 lines
76 KiB
Plaintext
1463 lines
76 KiB
Plaintext
#cloud-config
|
||
# Catalyst Sovereign control-plane bootstrap.
|
||
# Sovereign: ${sovereign_fqdn}
|
||
# Provisioned by: catalyst-provisioner (https://console.openova.io/sovereign)
|
||
#
|
||
# This script:
|
||
# 1. Installs OS hardening (SSH password-auth off, fail2ban, unattended-upgrades).
|
||
# 2. Installs k3s with --flannel-backend=none (Cilium replaces it).
|
||
# 3. Installs Flux + bootstraps the GitRepository pointing at the shared
|
||
# clusters/_template/ tree in the public OpenOva monorepo. The
|
||
# Sovereign's FQDN is interpolated into the template manifests via
|
||
# Flux postBuild.substitute ($${SOVEREIGN_FQDN}) at apply time, so
|
||
# no per-Sovereign directory needs to be committed before
|
||
# provisioning. From this point Flux is the GitOps reconciler and
|
||
# installs the 11-component bootstrap kit (Cilium → cert-manager →
|
||
# Crossplane → ... → bp-catalyst-platform) in dependency order via
|
||
# Kustomizations the _template directory ships.
|
||
# 4. Touches /var/lib/catalyst/cloud-init-complete so the catalyst-api
|
||
# provisioner can detect cloud-init has finished.
|
||
|
||
package_update: true
|
||
package_upgrade: false
|
||
packages:
|
||
- curl
|
||
- iptables
|
||
- jq
|
||
- ca-certificates
|
||
- git
|
||
%{ if enable_fail2ban ~}
|
||
- fail2ban
|
||
%{ endif ~}
|
||
%{ if enable_unattended_upgrades ~}
|
||
- unattended-upgrades
|
||
- apt-listchanges
|
||
%{ endif ~}
|
||
|
||
write_files:
|
||
- path: /var/lib/catalyst/sovereign.json
|
||
permissions: '0644'
|
||
content: |
|
||
{
|
||
"sovereignFQDN": "${sovereign_fqdn}",
|
||
"sovereignSubdomain": "${sovereign_subdomain}",
|
||
"orgName": ${jsonencode(org_name)},
|
||
"orgEmail": ${jsonencode(org_email)},
|
||
"region": "${region}",
|
||
"haEnabled": ${ha_enabled},
|
||
"workerCount": ${worker_count},
|
||
"k3sVersion": "${k3s_version}",
|
||
"gitopsRepoUrl": "${gitops_repo_url}",
|
||
"gitopsBranch": "${gitops_branch}"
|
||
}
|
||
|
||
# ── Kernel inotify limits — k3s + Flux + CNPG + bao + Helm exhaust Ubuntu defaults ──
|
||
# Default Hetzner Ubuntu 24.04 ships fs.inotify.max_user_instances=128
|
||
# and fs.inotify.max_user_watches=524288 — but every Helm controller,
|
||
# CNPG operator, k3s kubelet, file-watching admin tool grabs an
|
||
# instance slot. On a 35-component bootstrap-kit the slots run out
|
||
# mid-install and the next process to ask gets:
|
||
# failed to create fsnotify watcher: too many open files
|
||
# Diagnosed live during otech35 — bp-openbao's `bao operator init`
|
||
# crash-looped 4× with that exact error, which Flux escalated to
|
||
# InstallFailed/RetriesExceeded — masking the real OS-level root cause.
|
||
#
|
||
# Bump well above k8s/k3s production guidance so future blueprint
|
||
# additions don't tickle the same wall.
|
||
- path: /etc/sysctl.d/99-catalyst-inotify.conf
|
||
permissions: '0644'
|
||
content: |
|
||
fs.inotify.max_user_instances = 8192
|
||
fs.inotify.max_user_watches = 1048576
|
||
fs.inotify.max_queued_events = 16384
|
||
|
||
# ── OS hardening: SSH daemon ──────────────────────────────────────────
|
||
# Drop-in overrides /etc/ssh/sshd_config defaults. Per Catalyst's threat
|
||
# model the operator's only valid path in is the Hetzner-project SSH key
|
||
# injected via cloud-init authorized_keys. Password auth, KbdInteractive,
|
||
# and root password login are all off.
|
||
- path: /etc/ssh/sshd_config.d/99-catalyst-hardening.conf
|
||
permissions: '0644'
|
||
content: |
|
||
# Managed by Catalyst Sovereign cloud-init — do not edit by hand.
|
||
PasswordAuthentication no
|
||
KbdInteractiveAuthentication no
|
||
ChallengeResponseAuthentication no
|
||
PermitRootLogin prohibit-password
|
||
PermitEmptyPasswords no
|
||
UsePAM yes
|
||
X11Forwarding no
|
||
AllowAgentForwarding no
|
||
AllowTcpForwarding no
|
||
ClientAliveInterval 300
|
||
ClientAliveCountMax 2
|
||
MaxAuthTries 3
|
||
LoginGraceTime 30
|
||
|
||
%{ if enable_unattended_upgrades ~}
|
||
# ── Unattended security upgrades ──────────────────────────────────────
|
||
# Ubuntu's stock unattended-upgrades, restricted to the security pocket.
|
||
# Runs daily, reboots automatically at 02:30 if a kernel upgrade requires
|
||
# it (k3s tolerates single-node restarts on a solo Sovereign within the
|
||
# ~60s window the Hetzner LB health-check covers).
|
||
- path: /etc/apt/apt.conf.d/20auto-upgrades
|
||
permissions: '0644'
|
||
content: |
|
||
APT::Periodic::Update-Package-Lists "1";
|
||
APT::Periodic::Unattended-Upgrade "1";
|
||
APT::Periodic::AutocleanInterval "7";
|
||
- path: /etc/apt/apt.conf.d/52unattended-upgrades-catalyst
|
||
permissions: '0644'
|
||
content: |
|
||
Unattended-Upgrade::Allowed-Origins {
|
||
"$${distro_id}:$${distro_codename}-security";
|
||
"$${distro_id}ESMApps:$${distro_codename}-apps-security";
|
||
"$${distro_id}ESM:$${distro_codename}-infra-security";
|
||
};
|
||
Unattended-Upgrade::Automatic-Reboot "true";
|
||
Unattended-Upgrade::Automatic-Reboot-Time "02:30";
|
||
Unattended-Upgrade::Remove-Unused-Kernel-Packages "true";
|
||
Unattended-Upgrade::Remove-Unused-Dependencies "true";
|
||
%{ endif ~}
|
||
|
||
%{ if enable_fail2ban ~}
|
||
# ── fail2ban: sshd jail ───────────────────────────────────────────────
|
||
# Even though SSH is firewalled to ssh_allowed_cidrs (or fully closed at
|
||
# the firewall), fail2ban remains a defence-in-depth layer for the case
|
||
# where the firewall rule is widened by an operator post-bootstrap.
|
||
- path: /etc/fail2ban/jail.d/catalyst-sshd.local
|
||
permissions: '0644'
|
||
content: |
|
||
[sshd]
|
||
enabled = true
|
||
port = ssh
|
||
filter = sshd
|
||
maxretry = 5
|
||
findtime = 10m
|
||
bantime = 1h
|
||
backend = systemd
|
||
%{ endif ~}
|
||
|
||
# ── flux-system/ghcr-pull Secret ─────────────────────────────────────
|
||
#
|
||
# Every HelmRepository CR in clusters/_template/bootstrap-kit/
|
||
# references `secretRef: name: ghcr-pull` because the bp-* OCI artifacts
|
||
# at `ghcr.io/openova-io/` are PRIVATE. Without this Secret, the
|
||
# source-controller logs:
|
||
#
|
||
# failed to get authentication secret 'flux-system/ghcr-pull':
|
||
# secrets "ghcr-pull" not found
|
||
#
|
||
# …and Phase 1 stalls at bp-cilium. The operator workaround (kubectl
|
||
# apply the Secret by hand after Flux installs) is not durable across
|
||
# re-provisioning of the same Sovereign — every fresh control-plane
|
||
# boots without the Secret.
|
||
#
|
||
# We write the Secret into flux-system at cloud-init time, BEFORE
|
||
# /var/lib/catalyst/flux-bootstrap.yaml is applied, so the GitRepository +
|
||
# Kustomization land into a cluster that already has working GHCR creds.
|
||
# The apply step is in runcmd: below; the manifest itself lives here.
|
||
#
|
||
# Token rotation policy: yearly, stored in 1Password under
|
||
# "Catalyst — GHCR pull token (catalyst-ghcr-pull-token)". See
|
||
# docs/SECRET-ROTATION.md. The token NEVER lives in git.
|
||
- path: /var/lib/catalyst/ghcr-pull-secret.yaml
|
||
permissions: '0600'
|
||
content: |
|
||
apiVersion: v1
|
||
kind: Secret
|
||
metadata:
|
||
name: ghcr-pull
|
||
namespace: flux-system
|
||
annotations:
|
||
# bp-reflector (slot 05a) mirrors this secret to every namespace
|
||
# so all workloads can pull from ghcr.io/openova-io without
|
||
# per-namespace manual creation. reflection-auto-enabled means
|
||
# Reflector creates the copy in new namespaces as they appear.
|
||
#
|
||
# ALLOWED + AUTO namespaces — explicitly enumerated.
|
||
# Issue #879 Bonus Bug 6: the previous values left both fields
|
||
# as empty strings, which Reflector interprets ambiguously
|
||
# depending on version. On otech103 (2026-05-05) catalyst-api
|
||
# POD failed to pull the SHA-pinned image until an operator
|
||
# manually created the Secret in the `catalyst-system`
|
||
# namespace. The fix here lists every namespace catalyst-api
|
||
# and SME services land in: sme, catalyst, catalyst-system,
|
||
# gitea, harbor — paired with `auto-namespaces` so a
|
||
# later-created namespace (the bp-* HelmReleases land in their
|
||
# own namespaces over time) still gets the mirror automatically
|
||
# the moment it appears. The list is the SUPERSET of what
|
||
# otech103 verified live. Future namespaces added to the
|
||
# bootstrap-kit (a new bp-* slot) only need an addition here
|
||
# plus a Pod restart to pick up the new mirror.
|
||
#
|
||
# Issue #952 (2026-05-05): added `newapi` — the bp-newapi
|
||
# bootstrap-kit slot 80 lands its Deployment in this namespace
|
||
# and pulls PRIVATE `newapi-mirror` + `services-metering-sidecar`
|
||
# images. Without `newapi` on this list the Pod stalls in
|
||
# ImagePullBackOff with 403 Forbidden, blocking alice signup
|
||
# gate 5 (LLM). Caught live on otech113.
|
||
reflector.v1.k8s.emberstack.com/reflection-allowed: "true"
|
||
reflector.v1.k8s.emberstack.com/reflection-allowed-namespaces: "sme,catalyst,catalyst-system,gitea,harbor,newapi"
|
||
reflector.v1.k8s.emberstack.com/reflection-auto-enabled: "true"
|
||
reflector.v1.k8s.emberstack.com/reflection-auto-namespaces: "sme,catalyst,catalyst-system,gitea,harbor,newapi"
|
||
type: kubernetes.io/dockerconfigjson
|
||
data:
|
||
.dockerconfigjson: ${base64encode(jsonencode({
|
||
auths = {
|
||
"ghcr.io" = {
|
||
username = ghcr_pull_username
|
||
password = ghcr_pull_token
|
||
auth = ghcr_pull_auth_b64
|
||
}
|
||
}
|
||
}))}
|
||
|
||
# ── flux-system/harbor-robot-token Secret (issue #557 follow-up) ─────
|
||
#
|
||
# The catalyst-api Pod template (products/catalyst/chart/templates/
|
||
# api-deployment.yaml) references a Secret named `harbor-robot-token`
|
||
# via a REQUIRED (non-optional) secretKeyRef on every Sovereign. The
|
||
# token authenticates pulls from the central harbor.openova.io
|
||
# proxy-cache (proxy-dockerhub, proxy-gcr, proxy-quay, proxy-k8s,
|
||
# proxy-ghcr) — the same value already interpolated into
|
||
# /etc/rancher/k3s/registries.yaml below.
|
||
#
|
||
# Without this Secret the catalyst-api Pod stays in
|
||
# CreateContainerConfigError indefinitely. Caught live on otech43,
|
||
# otech45, otech46 — the operator workaround was hand-creating a
|
||
# placeholder Secret on each iteration, which is a workaround, not
|
||
# a fix.
|
||
#
|
||
# Why this Secret lives in flux-system + uses Reflector auto-mirror:
|
||
# the same canonical pattern as `ghcr-pull` above — bp-reflector
|
||
# (slot 05a) propagates the Secret to every namespace via
|
||
# `reflector.v1.k8s.emberstack.com/reflection-auto-enabled: "true"`,
|
||
# so catalyst-system (and any other namespace that needs the token)
|
||
# picks it up event-driven on first reconcile.
|
||
#
|
||
# The Secret carries one key (`token`) which catalyst-api reads as
|
||
# CATALYST_HARBOR_ROBOT_TOKEN and re-stamps onto every grandchild
|
||
# Sovereign provision request — the Sovereign's own Sovereigns
|
||
# (post-handover) inherit the same central proxy-cache auth.
|
||
#
|
||
# Token rotation: yearly, see docs/SECRET-ROTATION.md. Rotation flows
|
||
# through `var.harbor_robot_token` → re-render cloud-init → re-apply
|
||
# this Secret. The plaintext NEVER lives in git.
|
||
- path: /var/lib/catalyst/harbor-robot-token-secret.yaml
|
||
permissions: '0600'
|
||
content: |
|
||
apiVersion: v1
|
||
kind: Secret
|
||
metadata:
|
||
name: harbor-robot-token
|
||
namespace: flux-system
|
||
annotations:
|
||
# bp-reflector (slot 05a) mirrors this Secret to every
|
||
# namespace so catalyst-api in catalyst-system (and any
|
||
# workload added later that needs the central Harbor robot
|
||
# auth) picks it up without per-namespace manual creation.
|
||
reflector.v1.k8s.emberstack.com/reflection-allowed: "true"
|
||
reflector.v1.k8s.emberstack.com/reflection-allowed-namespaces: ""
|
||
reflector.v1.k8s.emberstack.com/reflection-auto-enabled: "true"
|
||
reflector.v1.k8s.emberstack.com/reflection-auto-namespaces: ""
|
||
type: Opaque
|
||
data:
|
||
token: ${base64encode(harbor_robot_token)}
|
||
|
||
# ── cert-manager/powerdns-api-credentials Secret (PR #681 followup) ──────
|
||
#
|
||
# The bp-cert-manager-powerdns-webhook Pod reads X-API-Key from this
|
||
# Secret to authenticate against contabo's authoritative PowerDNS for
|
||
# the omani.works zone. Without this, DNS-01 challenges hang and the
|
||
# wildcard cert never issues — caught live on otech47.
|
||
#
|
||
# PR #681 dropped bp-cert-manager-dynadot-webhook in favour of the
|
||
# contabo-targeted webhook but left the cloud-init Secret block on the
|
||
# old shape. This block replaces it with the powerdns-api-credentials
|
||
# Secret pointed at by ClusterIssuer letsencrypt-dns01-prod-powerdns
|
||
# (apiKeySecretRef.name: powerdns-api-credentials, key: api-key).
|
||
#
|
||
# Namespace: cert-manager — same as the HelmRelease's targetNamespace.
|
||
# Source on contabo: openova-system/powerdns-api-credentials Secret;
|
||
# mirrored into the catalyst namespace via Reflector annotations
|
||
# (platform/powerdns/chart/templates/api-secret.yaml) so catalyst-api
|
||
# can mount it as env CATALYST_POWERDNS_API_KEY.
|
||
- path: /var/lib/catalyst/powerdns-api-credentials.yaml
|
||
permissions: '0600'
|
||
content: |
|
||
apiVersion: v1
|
||
kind: Secret
|
||
metadata:
|
||
name: powerdns-api-credentials
|
||
namespace: cert-manager
|
||
type: Opaque
|
||
data:
|
||
api-key: ${base64encode(powerdns_api_key)}
|
||
|
||
# ── flux-system/pdm-basicauth Secret (issue #879 Bug 2) ──────────────
|
||
#
|
||
# The Sovereign-side catalyst-api Pod (api-deployment.yaml) reads
|
||
# CATALYST_PDM_BASIC_AUTH_USER + CATALYST_PDM_BASIC_AUTH_PASS via
|
||
# secretKeyRef into `pdm-basicauth` (in the same namespace
|
||
# catalyst-api lives — catalyst-system). Reflector mirrors this
|
||
# Secret out of flux-system to sme,catalyst,catalyst-system,gitea,
|
||
# harbor (same canonical pattern flux-system/ghcr-pull and
|
||
# flux-system/harbor-robot-token already use).
|
||
#
|
||
# The Pod adds `Authorization: Basic …` to every PDM call so the
|
||
# Traefik basicAuth Middleware in front of pool.openova.io accepts
|
||
# the request — pdmFlipNS in parent_domains.go is the call site.
|
||
# Without this Secret + Reflector mirror, every Day-2 add-parent-
|
||
# domain POST returns 401 from PDM (caught live on otech103,
|
||
# 2026-05-05 — issue #879).
|
||
#
|
||
# optional=true on the secretKeyRef in api-deployment.yaml so:
|
||
# - Catalyst-Zero pods (contabo's catalyst-api) start cleanly
|
||
# when the Secret is absent. Contabo uses the in-cluster
|
||
# Service path which bypasses the ingress entirely.
|
||
# - CI / older Sovereigns that pre-date this provisioning seam
|
||
# start cleanly. POSTs without auth get 401 from PDM with a
|
||
# clear log line, instead of crashlooping on Pod start.
|
||
#
|
||
# Per Inviolable Principle #10: the credentials never enter a
|
||
# logged struct, a deployment record, or any committed git file.
|
||
# Plaintext only ever lives in the per-deployment OpenTofu workdir
|
||
# (mode 0600, wiped on tofu destroy) and inside the Sovereign's
|
||
# encrypted etcd.
|
||
- path: /var/lib/catalyst/pdm-basicauth-secret.yaml
|
||
permissions: '0600'
|
||
content: |
|
||
apiVersion: v1
|
||
kind: Secret
|
||
metadata:
|
||
name: pdm-basicauth
|
||
namespace: flux-system
|
||
annotations:
|
||
# bp-reflector (slot 05a) mirrors this Secret to every
|
||
# namespace listed below so catalyst-api in catalyst-system
|
||
# picks it up event-driven. List explicitly enumerates the
|
||
# known namespaces (issue #879 Bonus Bug 6 — empty-string
|
||
# ambiguity caused otech103 to require a manual mirror
|
||
# creation in catalyst-system).
|
||
reflector.v1.k8s.emberstack.com/reflection-allowed: "true"
|
||
reflector.v1.k8s.emberstack.com/reflection-allowed-namespaces: "sme,catalyst,catalyst-system,gitea,harbor"
|
||
reflector.v1.k8s.emberstack.com/reflection-auto-enabled: "true"
|
||
reflector.v1.k8s.emberstack.com/reflection-auto-namespaces: "sme,catalyst,catalyst-system,gitea,harbor"
|
||
type: Opaque
|
||
data:
|
||
username: ${base64encode(pdm_basic_auth_user)}
|
||
password: ${base64encode(pdm_basic_auth_pass)}
|
||
|
||
# ── flux-system/object-storage Secret (issue #371, vendor-agnostic since #425) ─
|
||
#
|
||
# The Sovereign's per-cluster S3 credentials, materialised as a stock
|
||
# Kubernetes Secret in the `flux-system` namespace. Harbor (#383) and
|
||
# Velero (#384) consume this Secret via the canonical `secretRef` field
|
||
# in their respective HelmRelease values blocks, e.g.
|
||
#
|
||
# harbor:
|
||
# persistence:
|
||
# imageChartStorage:
|
||
# type: s3
|
||
# s3:
|
||
# existingSecret: object-storage
|
||
#
|
||
# Per #425 the Secret name is vendor-AGNOSTIC (`object-storage`, no
|
||
# `hetzner-` prefix). A future AWS / Azure / GCP / OCI Sovereign
|
||
# provisions the same Secret name with the same key set via its own
|
||
# `infra/<provider>/` Tofu module — every existing chart Just Works
|
||
# without renaming.
|
||
#
|
||
# The Secret is namespace-bound to flux-system so the helm-controller can
|
||
# rewrite it into the workload namespaces at chart install time — that's
|
||
# the same boundary `ghcr-pull` already uses, so the apply ordering in
|
||
# runcmd: below stays a single sequenced step.
|
||
#
|
||
# Why pre-populated by cloud-init rather than a SealedSecret committed to
|
||
# git: ADR-0001 §9.2 forbids bespoke cloud-API calls and Hetzner exposes
|
||
# NO Cloud API for S3 credential issuance — they're operator-issued in
|
||
# the Hetzner Console exactly once. Therefore catalyst-api receives the
|
||
# plaintext from the wizard, validates it, and forwards it to the new
|
||
# Sovereign via the same encrypted-PVC + cloud-init channel as the GHCR
|
||
# pull token. The credentials never land in git; the only durable copies
|
||
# are the per-deployment OpenTofu workdir (mode 0600, wiped on tofu
|
||
# destroy) and inside the new Sovereign's etcd (encrypted at rest by
|
||
# k3s default).
|
||
#
|
||
# Token rotation policy: per Hetzner's docs, the secret half is shown
|
||
# exactly once at issue time. To rotate, the operator issues a fresh
|
||
# credential pair in the Hetzner Console, updates the wizard payload
|
||
# for the next provisioning, OR for an existing Sovereign uses a
|
||
# day-2 Crossplane XRC write (the Provider+ProviderConfig planted
|
||
# below makes this possible without a Tofu re-run; out of scope for
|
||
# the initial bootstrap).
|
||
- path: /var/lib/catalyst/object-storage-secret.yaml
|
||
permissions: '0600'
|
||
content: |
|
||
apiVersion: v1
|
||
kind: Secret
|
||
metadata:
|
||
name: object-storage
|
||
namespace: flux-system
|
||
type: Opaque
|
||
stringData:
|
||
# S3 endpoint URL — composed from object_storage_region in main.tf.
|
||
# Format: https://<region>.your-objectstorage.com per Hetzner docs.
|
||
s3-endpoint: ${object_storage_endpoint}
|
||
# S3 region — fsn1 / nbg1 / hel1 (Object Storage availability is
|
||
# European-only as of 2026-04). NOT the same as compute region.
|
||
s3-region: ${object_storage_region}
|
||
# Bucket name — deterministic per-Sovereign identifier composed by
|
||
# the catalyst-api from the Sovereign's FQDN slug. Created (or
|
||
# adopted if already present) by the minio_s3_bucket resource in
|
||
# main.tf earlier in this same `tofu apply`.
|
||
s3-bucket: ${object_storage_bucket_name}
|
||
# Operator-issued S3 access key + secret key. Hetzner's docs note
|
||
# the secret half is shown exactly once at credential creation
|
||
# time and is irretrievable thereafter — losing it means rotating.
|
||
s3-access-key: ${object_storage_access_key}
|
||
s3-secret-key: ${object_storage_secret_key}
|
||
|
||
# ── flux-system/cloud-credentials Secret (issue #425, OpenTofu→Crossplane) ─
|
||
#
|
||
# Bootstrap of the Crossplane Hetzner Cloud provider (planted further
|
||
# below in this cloud-init). Carries the operator's hcloud API token —
|
||
# the same token Tofu used to provision Phase 0 — under a single key
|
||
# `hcloud-token`. Per ADR-0001 §11.3 + INVIOLABLE-PRINCIPLES #3,
|
||
# Day-2 cloud-resource changes (additional Floating IPs, additional
|
||
# buckets, scaling LoadBalancers, firewall rule edits, ...) flow
|
||
# through Crossplane XRC writes against this Provider — NEVER through
|
||
# bespoke Go cloud-API calls in catalyst-api, NEVER through manual
|
||
# Tofu re-runs.
|
||
#
|
||
# The Secret name is vendor-agnostic (`cloud-credentials`); the
|
||
# `hcloud-token` key name encodes the cloud-specific shape of the
|
||
# credential. A future AWS Sovereign would write
|
||
# `aws-access-key-id`/`aws-secret-access-key` keys into the same
|
||
# Secret name; the matching Crossplane Provider/ProviderConfig
|
||
# (added in the same Tofu module's cloud-init) reads them.
|
||
- path: /var/lib/catalyst/cloud-credentials-secret.yaml
|
||
permissions: '0600'
|
||
content: |
|
||
apiVersion: v1
|
||
kind: Secret
|
||
metadata:
|
||
name: cloud-credentials
|
||
namespace: flux-system
|
||
type: Opaque
|
||
stringData:
|
||
hcloud-token: ${hcloud_token}
|
||
# Issue #921 — base64-encoded cloud-init payload for cluster-
|
||
# autoscaler's HCLOUD_CLOUD_INIT env var. The bp-cluster-autoscaler-
|
||
# hcloud HelmRelease's `valuesFrom` lifts this key into
|
||
# `clusterAutoscalerHcloud.cloudInit`; the chart's
|
||
# extraEnvSecrets.HCLOUD_CLOUD_INIT then maps it onto the autoscaler
|
||
# Pod's env. Without this, cluster-autoscaler 1.32.x's Hetzner
|
||
# provider exits at startup with FATAL "HCLOUD_CLUSTER_CONFIG or
|
||
# HCLOUD_CLOUD_INIT is not specified". Same content as the Phase-0
|
||
# worker servers' user_data so autoscaler-spawned workers join the
|
||
# cluster identically.
|
||
hcloud-cloud-init: ${worker_cloud_init_b64}
|
||
|
||
# ── Crossplane provider-hcloud + ProviderConfig (issue #425) ────────
|
||
#
|
||
# Phase 0→Day-2 handover. After Flux installs Crossplane core (via
|
||
# bp-crossplane in the bootstrap-kit), this Provider package + its
|
||
# ProviderConfig come up in the cluster and become the seam for ALL
|
||
# subsequent Hetzner Cloud mutations.
|
||
#
|
||
# Per ADR-0001 §11.3 + INVIOLABLE-PRINCIPLES #3:
|
||
# - OpenTofu provisions Phase 0 EXACTLY ONCE per Sovereign.
|
||
# - Crossplane is the only Day-2 cloud-API mutation seam.
|
||
# - Flux is the only GitOps reconciler.
|
||
# - Blueprints (`bp-<name>:<semver>` OCI) are the only install unit.
|
||
# - NEVER bespoke Go cloud-API calls. NEVER `exec.Command("helm",
|
||
# ...)`. NEVER direct `kubectl apply` of production manifests.
|
||
#
|
||
# Once `provider-hcloud` reaches `Healthy=True` (event the Catalyst
|
||
# control plane observes via the Crossplane status conditions), the
|
||
# catalyst-api's bespoke Hetzner-API calls for any RUNTIME-scaling
|
||
# concern (additional Floating IPs, additional buckets, scaling
|
||
# LoadBalancers, ...) MUST be retired in favour of XRC writes against
|
||
# this Provider. Provisioning Phase 0 (the very first server, network,
|
||
# firewall, LB, bucket) stays in this Tofu module by design — that's
|
||
# the bootstrap exception that lets the Provider exist in the first
|
||
# place.
|
||
#
|
||
# Package version pin: v0.4.0 of `crossplane-contrib/provider-hcloud`
|
||
# is the latest stable as of 2026-05. Per INVIOLABLE-PRINCIPLES #4
|
||
# (never hardcode), the version is operator-bumpable via PR; future
|
||
# rotations land here in the same commit that bumps the
|
||
# `bp-crossplane-claims` Composition referencing the new Provider
|
||
# types.
|
||
- path: /var/lib/catalyst/crossplane-provider-hcloud.yaml
|
||
permissions: '0644'
|
||
content: |
|
||
---
|
||
apiVersion: pkg.crossplane.io/v1
|
||
kind: Provider
|
||
metadata:
|
||
name: provider-hcloud
|
||
labels:
|
||
catalyst.openova.io/sovereign: ${sovereign_fqdn}
|
||
spec:
|
||
package: xpkg.upbound.io/crossplane-contrib/provider-hcloud:v0.4.0
|
||
packagePullPolicy: IfNotPresent
|
||
---
|
||
apiVersion: hcloud.crossplane.io/v1beta1
|
||
kind: ProviderConfig
|
||
metadata:
|
||
name: default
|
||
spec:
|
||
credentials:
|
||
source: Secret
|
||
secretRef:
|
||
namespace: flux-system
|
||
name: cloud-credentials
|
||
key: hcloud-token
|
||
|
||
# ── Handover JWT public key (issue #605, Phase-8b) ───────────────────
|
||
#
|
||
# RFC 7517 JWK JSON of the Catalyst-Zero RS256 public key. Written to
|
||
# /var/lib/catalyst/handover-jwt-public.jwk (mode 0600) on the new
|
||
# Sovereign control-plane. Agent-C (auth/handover) reads this file to
|
||
# verify the one-time handover JWT without a cross-cluster RPC.
|
||
#
|
||
# Per INVIOLABLE-PRINCIPLES.md #10 (credential hygiene): mode 0600 so
|
||
# the file is readable only by root. Even though RSA public keys are
|
||
# technically non-secret, tightening the permission costs nothing and
|
||
# avoids any future confusion about sensitivity.
|
||
- path: /var/lib/catalyst/handover-jwt-public.jwk
|
||
permissions: '0600'
|
||
content: |
|
||
${handover_jwt_public_key}
|
||
|
||
# ── catalyst-system/catalyst-handover-jwt-public Secret (issue #606 followup) ─
|
||
#
|
||
# On Phase-8b live (otech48, 2026-05-03) the Sovereign-side catalyst-api
|
||
# responded to GET /auth/handover with
|
||
# {"error":"server misconfiguration: public key unavailable"}.
|
||
# Root cause: the JWK above was written ONLY to host disk
|
||
# (/var/lib/catalyst/handover-jwt-public.jwk). The catalyst-api Pod's
|
||
# volume mount references K8s Secret `catalyst-handover-jwt-public`
|
||
# (key: public.jwk). That Secret was never created on the Sovereign,
|
||
# so the volume mount fell through (the Secret is `optional: true`)
|
||
# and the file was missing inside the container at the env-pinned
|
||
# path `/etc/catalyst/handover-jwt-public/public.jwk`.
|
||
#
|
||
# Fix: mirror the canonical pattern that flux-system/ghcr-pull (PR #543)
|
||
# and flux-system/harbor-robot-token (PR #680) already follow. Cloud-init
|
||
# writes the Secret manifest into catalyst-system NS and runcmd: applies
|
||
# it BEFORE flux-bootstrap, so the Secret exists by the time the
|
||
# bp-catalyst-platform HelmRelease lands and the catalyst-api Pod starts.
|
||
#
|
||
# Why catalyst-system rather than flux-system + Reflector: the Secret is
|
||
# mounted ONLY by the catalyst-api Pod (single workload, single namespace).
|
||
# Reflector auto-mirror would just create unused copies in every other
|
||
# namespace — extra blast radius for no benefit. The harbor-robot-token
|
||
# case is different: it has to be visible to flux-system (Flux pulls OCI
|
||
# charts) AND catalyst-system (catalyst-api re-stamps it onto child
|
||
# provisions), so Reflector is justified there.
|
||
#
|
||
# The catalyst-system namespace itself is created by the bp-catalyst-platform
|
||
# HelmRelease wrapper in clusters/_template/bootstrap-kit/13-bp-catalyst-platform.yaml,
|
||
# but Flux applies that namespace alongside the HelmRelease — racing with
|
||
# this Secret apply. Pre-creating the namespace here (idempotent dry-run/apply
|
||
# pattern, identical to the cert-manager pre-create at line ~1099) eliminates
|
||
# the race.
|
||
- path: /var/lib/catalyst/handover-jwt-public-secret.yaml
|
||
permissions: '0600'
|
||
content: |
|
||
apiVersion: v1
|
||
kind: Secret
|
||
metadata:
|
||
name: catalyst-handover-jwt-public
|
||
namespace: catalyst-system
|
||
type: Opaque
|
||
data:
|
||
public.jwk: ${base64encode(handover_jwt_public_key)}
|
||
|
||
# ── Cilium bootstrap values (issue #491) ─────────────────────────────
|
||
#
|
||
# The bootstrap helm install below MUST land the same effective values
|
||
# as the Flux bp-cilium HelmRelease (clusters/_template/bootstrap-kit/
|
||
# 01-cilium.yaml). Anything that differs becomes drift, and drift in
|
||
# this particular release is fatal because:
|
||
#
|
||
# 1. Flux applies bp-cilium with `helm upgrade --install`, which is
|
||
# a no-op when the in-cluster release already has the right values
|
||
# and a UPGRADE when it does not.
|
||
# 2. The bootstrap-kit Kustomization is `wait: true` (issue #492).
|
||
# Until cilium-agent is Ready, NO other HelmRelease in
|
||
# bootstrap-kit reconciles — including the bp-cilium upgrade
|
||
# itself, because Flux's source-controller will not pull a fresh
|
||
# GitRepository revision while the existing one is unhealthy.
|
||
# 3. cilium-agent waits for the operator to register
|
||
# `ciliumenvoyconfigs` + `ciliumclusterwideenvoyconfigs` CRDs.
|
||
# The upstream chart only registers them when
|
||
# `envoyConfig.enabled=true`. If the bootstrap install omits
|
||
# that flag, the CRDs are never registered, the agent never
|
||
# reaches Ready, the upgrade never fires, and Phase 1 deadlocks.
|
||
#
|
||
# Phase-8a bug #16 (otech8 2026-05-01): the prior bootstrap helm
|
||
# install used six --set flags (`kubeProxyReplacement`, `k8sService*`,
|
||
# `tunnelProtocol`, `bpf.masquerade`) and produced a release missing
|
||
# `envoyConfig.enabled`, `gatewayAPI.enabled`, `envoy.enabled`,
|
||
# `l7Proxy`, `encryption.*`, `hubble.*`, etc. Every fresh provision
|
||
# crash-looped cilium-agent.
|
||
#
|
||
# Canonical seam: this file IS the values overlay for the bootstrap
|
||
# install, and `clusters/_template/bootstrap-kit/01-cilium.yaml`'s
|
||
# `spec.values.cilium:` block IS the values overlay for the Flux HR.
|
||
# The umbrella chart wraps under `cilium:` (subchart key), the
|
||
# bootstrap install targets the upstream `cilium/cilium` chart
|
||
# directly so values land at top level. The merged effective set
|
||
# below mirrors `platform/cilium/chart/values.yaml`'s `cilium:`
|
||
# block PLUS the overlay in 01-cilium.yaml. A divergence test in
|
||
# `products/catalyst/bootstrap/api/internal/provisioner/
|
||
# cilium_values_parity_test.go` (issue #491) locks the two files
|
||
# together so a future operator cannot change one without the other.
|
||
#
|
||
# Per INVIOLABLE-PRINCIPLES.md #4 (never hardcode): the chart
|
||
# version is parameterised below via the helm install --version flag,
|
||
# and the values in this file are operator-overridable post-bootstrap
|
||
# via the Flux HR's `spec.values` block (which always wins on
|
||
# subsequent `helm upgrade`).
|
||
- path: /var/lib/catalyst/cilium-values.yaml
|
||
permissions: '0644'
|
||
content: |
|
||
# Catalyst bootstrap cilium values — MUST stay in lock-step with
|
||
# platform/cilium/chart/values.yaml `cilium:` block + the overlay
|
||
# in clusters/_template/bootstrap-kit/01-cilium.yaml. See the
|
||
# comment block immediately above this write_files entry, and
|
||
# cilium_values_parity_test.go for the regression guard.
|
||
kubeProxyReplacement: true
|
||
# k8sServiceHost: the CP's stable private IP on the 10.0.1.0/24
|
||
# subnet (cp1=10.0.1.2 per main.tf network block). Was previously
|
||
# 127.0.0.1 which works on the CP (k3s server listens on
|
||
# localhost:6443) but FAILS on workers (k3s agent does NOT expose
|
||
# apiserver on localhost — only the supervisor port on :6444).
|
||
# When worker_count>0 (issue #733 multi-node default), the Cilium
|
||
# DaemonSet on workers used to crashloop with "127.0.0.1:6443:
|
||
# connect: connection refused" forever. Routing every Cilium agent
|
||
# to the CP private IP fixes the worker path and is a no-op on
|
||
# the CP itself (10.0.1.2 IS its own private IP).
|
||
k8sServiceHost: 10.0.1.2
|
||
k8sServicePort: 6443
|
||
tunnelProtocol: vxlan
|
||
bpf:
|
||
masquerade: true
|
||
ipam:
|
||
mode: kubernetes
|
||
encryption:
|
||
enabled: true
|
||
type: wireguard
|
||
gatewayAPI:
|
||
enabled: true
|
||
gatewayClass:
|
||
# Force GatewayClass creation regardless of CRD presence at Helm
|
||
# render time. Default is "auto" which skips creation when the
|
||
# gateway.networking.k8s.io CRDs are not yet present — exactly
|
||
# what happens during bootstrap: the upstream Helm chart's
|
||
# Capabilities check fires BEFORE bp-gateway-api has run, so
|
||
# GatewayClass/cilium is never rendered into the release.
|
||
# Forcing "true" ensures the GatewayClass is always created.
|
||
# Fix for cilium-gateway-race (issue #503).
|
||
create: "true"
|
||
envoy:
|
||
enabled: true
|
||
# envoyConfig.enabled is the load-bearing flag from issue #491.
|
||
# Without it the upstream chart skips the CiliumEnvoyConfig and
|
||
# CiliumClusterwideEnvoyConfig CRD registrations, cilium-agent
|
||
# waits forever for them, and the node taint
|
||
# `node.cilium.io/agent-not-ready` never lifts.
|
||
envoyConfig:
|
||
enabled: true
|
||
l7Proxy: true
|
||
hubble:
|
||
enabled: true
|
||
relay:
|
||
enabled: false
|
||
ui:
|
||
enabled: false
|
||
metrics:
|
||
# `null` (not [] and not a populated list) is the exact value
|
||
# that makes the upstream chart skip the metrics ServiceMonitor
|
||
# template branch. See platform/cilium/chart/values.yaml.
|
||
enabled: null
|
||
serviceMonitor:
|
||
enabled: false
|
||
l2announcements:
|
||
enabled: false
|
||
operator:
|
||
replicas: 1
|
||
resources:
|
||
requests:
|
||
cpu: 100m
|
||
memory: 256Mi
|
||
limits:
|
||
memory: 1Gi
|
||
prometheus:
|
||
enabled: false
|
||
serviceMonitor:
|
||
enabled: false
|
||
|
||
# ── containerd pull-through mirror: harbor.openova.io (issue #557, Option A) ──
|
||
#
|
||
# k3s uses containerd. Containerd's mirror table is configured via
|
||
# /etc/rancher/k3s/registries.yaml BEFORE k3s starts — the file is
|
||
# read once at startup and cannot be hot-reloaded without a k3s restart.
|
||
#
|
||
# Each `mirrors:` key is the upstream registry hostname containerd
|
||
# intercepts. When a pod pulls `nats:2.10` (implicitly `docker.io/library/
|
||
# nats:2.10`), containerd tries `harbor.openova.io/proxy-dockerhub/library/
|
||
# nats:2.10` first; on a cache miss, harbor.openova.io fetches from
|
||
# DockerHub and caches the blob for subsequent pulls by other pods or
|
||
# Sovereign nodes.
|
||
#
|
||
# This eliminates DockerHub's anonymous rate limit (100 pulls/6h per IP)
|
||
# on fresh Sovereign IPs where the bootstrap-kit pulls 50+ images in the
|
||
# first 30 minutes.
|
||
#
|
||
# The `configs:` block supplies harbor.openova.io credentials so containerd
|
||
# can authenticate against the proxy project (Harbor proxy-cache projects
|
||
# are set Public but a robot account is provided for future private-project
|
||
# pulls and consistent audit logging).
|
||
#
|
||
# harbor_robot_token is interpolated from var.harbor_robot_token (added to
|
||
# infra/hetzner/variables.tf); the catalyst-api provisioner reads it from
|
||
# the `harbor-robot-token` K8s Secret in the openova-harbor namespace on
|
||
# contabo and passes it to each new Sovereign's cloud-init render at
|
||
# provisioning time. This keeps the token out of git.
|
||
#
|
||
# CRITICAL ORDERING: this file MUST be written to disk BEFORE k3s installs
|
||
# (the k3s install runcmd below). k3s reads registries.yaml at startup and
|
||
# configures containerd's mirror table; a missing file at startup means
|
||
# direct pulls from DockerHub for the entire lifetime of that node.
|
||
- path: /etc/rancher/k3s/registries.yaml
|
||
permissions: '0600'
|
||
content: |
|
||
# Harbor proxy-cache projects use the URL form
|
||
# https://harbor.openova.io/v2/<project>/<image>/manifests/<tag>
|
||
# NOT
|
||
# https://harbor.openova.io/<project>/v2/<image>/manifests/<tag>
|
||
# which is what containerd would naively build from
|
||
# `endpoint: ["https://harbor.openova.io/<project>"]`.
|
||
# Harbor returns its UI HTML (status 200, content-type text/html)
|
||
# for the wrong-shape URL — containerd then surfaces:
|
||
# "unexpected media type text/html for sha256:..."
|
||
# and cilium / coredns / pause-image pulls all fail forever.
|
||
#
|
||
# k3s registries.yaml supports a per-mirror `rewrite` map:
|
||
# containerd builds `<endpoint>/v2/<repo>/...` (host-only endpoint),
|
||
# then rewrite() transforms the repo path before the request goes out.
|
||
# Mapping `(.*)` → `proxy-<flavor>/$1` produces the correct
|
||
# Harbor-project-prefixed path. Diagnosed live during otech25.
|
||
mirrors:
|
||
"docker.io":
|
||
endpoint:
|
||
- "https://harbor.openova.io"
|
||
rewrite:
|
||
"(.*)": "proxy-dockerhub/$1"
|
||
"quay.io":
|
||
endpoint:
|
||
- "https://harbor.openova.io"
|
||
rewrite:
|
||
"(.*)": "proxy-quay/$1"
|
||
"gcr.io":
|
||
endpoint:
|
||
- "https://harbor.openova.io"
|
||
rewrite:
|
||
"(.*)": "proxy-gcr/$1"
|
||
"registry.k8s.io":
|
||
endpoint:
|
||
- "https://harbor.openova.io"
|
||
rewrite:
|
||
"(.*)": "proxy-k8s/$1"
|
||
"ghcr.io":
|
||
endpoint:
|
||
- "https://harbor.openova.io"
|
||
rewrite:
|
||
"(.*)": "proxy-ghcr/$1"
|
||
configs:
|
||
"harbor.openova.io":
|
||
auth:
|
||
username: "robot$openova-bot"
|
||
password: "${harbor_robot_token}"
|
||
|
||
# Flux GitRepository + Kustomizations that take over after k3s is up.
|
||
#
|
||
# ── Per-Sovereign tree vs. shared _template (issue #218) ─────────────
|
||
#
|
||
# Earlier revisions of this template selected a per-FQDN cluster tree
|
||
# (`!/clusters/${sovereign_fqdn}`) and pointed the Kustomization
|
||
# `spec.path` at `./clusters/${sovereign_fqdn}/bootstrap-kit`. That
|
||
# required a per-Sovereign directory to be committed to the public
|
||
# openova repo BEFORE provisioning, which the wizard does NOT do —
|
||
# only `clusters/_template/` is canonical. Result on every fresh
|
||
# Sovereign was Phase-1 stall:
|
||
# kustomization path not found:
|
||
# stat /tmp/kustomization-…/clusters/<fqdn>/bootstrap-kit:
|
||
# no such file or directory
|
||
# (live evidence: otech.omani.works deployment ce476aaf80731a46.)
|
||
#
|
||
# Canonical fix: GitRepository selects the shared `_template/` tree,
|
||
# Kustomization paths point at `clusters/_template/{bootstrap-kit,
|
||
# infrastructure}`, and Flux's `postBuild.substitute` interpolates
|
||
# `$${SOVEREIGN_FQDN}` into the template manifests at apply time. The
|
||
# per-FQDN copy that prior provisioning depended on becomes a no-op:
|
||
# one shared tree serves every Sovereign, with the Sovereign's FQDN
|
||
# injected by Flux on the cluster instead of by sed in the repo.
|
||
- path: /var/lib/catalyst/flux-bootstrap.yaml
|
||
permissions: '0644'
|
||
content: |
|
||
apiVersion: source.toolkit.fluxcd.io/v1
|
||
kind: GitRepository
|
||
metadata:
|
||
name: openova
|
||
namespace: flux-system
|
||
spec:
|
||
interval: 1m
|
||
url: ${gitops_repo_url}
|
||
ref:
|
||
branch: ${gitops_branch}
|
||
ignore: |
|
||
/*
|
||
!/clusters/_template
|
||
!/platform
|
||
!/products
|
||
---
|
||
# Two Flux Kustomizations with dependsOn so Crossplane CRDs land
|
||
# before any resource that uses them is dry-run-applied.
|
||
#
|
||
# bootstrap-kit installs the 11 HelmReleases (Cilium, cert-manager,
|
||
# Flux, Crossplane core, sealed-secrets, SPIRE, NATS-JetStream,
|
||
# OpenBao, Keycloak, Gitea, bp-catalyst-platform). bp-crossplane
|
||
# registers the Crossplane core CRDs (Provider, ProviderConfig…)
|
||
# AND the bp-catalyst-platform umbrella reconciles the rest.
|
||
#
|
||
# infrastructure-config applies the cluster's Provider package +
|
||
# ProviderConfig + Compositions. Because it dependsOn bootstrap-kit
|
||
# AND uses wait: true, Flux waits until bootstrap-kit's HelmReleases
|
||
# are Ready (Crossplane core + provider-hcloud installed,
|
||
# hcloud.crossplane.io/v1beta1 CRDs registered) before dry-running
|
||
# ProviderConfig — which is the exact ordering the prior single-
|
||
# Kustomization model tripped over with:
|
||
# no matches for kind "ProviderConfig" in version
|
||
# "hcloud.crossplane.io/v1beta1"
|
||
#
|
||
# postBuild.substitute (issue #218): Flux's envsubst runs over the
|
||
# rendered manifests after kustomize build, replacing $${SOVEREIGN_FQDN}
|
||
# with the Sovereign's FQDN that this cloud-init was rendered for.
|
||
# The template manifests in clusters/_template/bootstrap-kit/*.yaml
|
||
# use $${SOVEREIGN_FQDN} as the substitution token.
|
||
apiVersion: kustomize.toolkit.fluxcd.io/v1
|
||
kind: Kustomization
|
||
metadata:
|
||
name: bootstrap-kit
|
||
namespace: flux-system
|
||
spec:
|
||
interval: 5m
|
||
path: ./clusters/_template/bootstrap-kit
|
||
prune: true
|
||
sourceRef:
|
||
kind: GitRepository
|
||
name: openova
|
||
wait: true
|
||
# timeout: 5m (issue #492). Phase-8a iteration discipline: when
|
||
# the FIRST apply of bootstrap-kit is unhealthy (e.g. cilium
|
||
# crash-loop from issue #491), kustomize-controller holds the
|
||
# revision lock for the FULL timeout window and refuses to pick
|
||
# up new GitRepository revisions, even fixes that have already
|
||
# landed on main. The 30m default deadlocked otech8 deployment
|
||
# 1bfc46347564467b: fix `66ea39f0` was on main 1m after the bad
|
||
# SHA, but bootstrap-kit's `lastAttemptedRevision` stayed pinned
|
||
# to the old SHA waiting for HRs to become Ready (which they
|
||
# never would, because of #491). Operator wiped + reprovisioned.
|
||
# 5m matches the GitRepository poll interval — failed reconciles
|
||
# release the revision lock fast (~6m worst case) so a fresh fix
|
||
# gets applied on the next poll. We KEEP `wait: true` to preserve
|
||
# the consolidated "Kustomization Ready=True ⇒ every HR Ready"
|
||
# contract that downstream `dependsOn: bootstrap-kit` relies on.
|
||
timeout: 5m
|
||
postBuild:
|
||
substitute:
|
||
SOVEREIGN_FQDN: ${sovereign_fqdn}
|
||
# SOVEREIGN_LB_IP — Hetzner load balancer's public IPv4 (issue
|
||
# #900). Threaded into bp-catalyst-platform's
|
||
# `global.sovereignLBIP` so catalyst-api can pre-register glue
|
||
# records at the customer's Dynadot account before flipping NS
|
||
# for Day-2 multi-domain adds. Sourced from
|
||
# hcloud_load_balancer.main.ipv4 in main.tf.
|
||
SOVEREIGN_LB_IP: ${load_balancer_ipv4}
|
||
# Marketplace mode toggle (issue #710). Default "false" so a
|
||
# zero-touch provision lands a non-marketplace Sovereign. The
|
||
# operator (or the wizard's "Enable Marketplace" component
|
||
# checkbox) flips this to "true" via per-Sovereign overlay
|
||
# patch — bp-catalyst-platform 1.3.0 then renders the
|
||
# marketplace + tenant-wildcard HTTPRoutes.
|
||
MARKETPLACE_ENABLED: "${marketplace_enabled}"
|
||
# Multi-domain Sovereign (issue #827, parent epic #825).
|
||
# Parent-zone list as a YAML inline-array literal. Threads
|
||
# into bootstrap-kit slot 11 (bp-powerdns) for zone creation
|
||
# AND slot 13 (bp-catalyst-platform) for per-zone wildcard
|
||
# Certificate rendering — both consumers see the same string
|
||
# so they stay in lockstep on what the Sovereign considers a
|
||
# parent zone. Default when the wizard doesn't supply an
|
||
# explicit list: single-zone array derived from
|
||
# sovereign_fqdn (rendered in main.tf via coalesce()).
|
||
PARENT_DOMAINS_YAML: '${parent_domains_yaml}'
|
||
---
|
||
apiVersion: kustomize.toolkit.fluxcd.io/v1
|
||
kind: Kustomization
|
||
metadata:
|
||
name: sovereign-tls
|
||
namespace: flux-system
|
||
spec:
|
||
# Carries the cert-manager Certificate that backs Cilium Gateway's
|
||
# wildcard-TLS listener. Split out of bootstrap-kit so its
|
||
# `cert-manager.io/v1` CRD-dependent dry-run only runs AFTER
|
||
# bp-cert-manager is Ready (Phase-8a bug #13). dependsOn
|
||
# bootstrap-kit's `wait: true` semantics: bootstrap-kit reaches
|
||
# Ready iff every HelmRelease + Kustomization in it Ready=True.
|
||
interval: 5m
|
||
path: ./clusters/_template/sovereign-tls
|
||
prune: true
|
||
sourceRef:
|
||
kind: GitRepository
|
||
name: openova
|
||
dependsOn:
|
||
- name: bootstrap-kit
|
||
wait: true
|
||
# timeout: 5m (issue #492) — same rationale as bootstrap-kit.
|
||
# Releases the revision lock quickly so iterative fixes during
|
||
# Phase-8a get applied on the next GitRepository poll instead
|
||
# of stalling 30 minutes per failed apply.
|
||
timeout: 5m
|
||
postBuild:
|
||
substitute:
|
||
SOVEREIGN_FQDN: ${sovereign_fqdn}
|
||
# SOVEREIGN_LB_IP — Hetzner load balancer's public IPv4 (issue
|
||
# #900). Threaded into bp-catalyst-platform's
|
||
# `global.sovereignLBIP` so catalyst-api can pre-register glue
|
||
# records at the customer's Dynadot account before flipping NS
|
||
# for Day-2 multi-domain adds. Sourced from
|
||
# hcloud_load_balancer.main.ipv4 in main.tf.
|
||
SOVEREIGN_LB_IP: ${load_balancer_ipv4}
|
||
# Marketplace mode toggle (issue #710). Default "false" so a
|
||
# zero-touch provision lands a non-marketplace Sovereign. The
|
||
# operator (or the wizard's "Enable Marketplace" component
|
||
# checkbox) flips this to "true" via per-Sovereign overlay
|
||
# patch — bp-catalyst-platform 1.3.0 then renders the
|
||
# marketplace + tenant-wildcard HTTPRoutes.
|
||
MARKETPLACE_ENABLED: "${marketplace_enabled}"
|
||
# Multi-domain Sovereign (issue #827) — same value threaded
|
||
# in here so a future cert resource that gets moved out of
|
||
# bp-catalyst-platform into clusters/_template/sovereign-tls/
|
||
# has access to the parent-zone list without a config copy.
|
||
PARENT_DOMAINS_YAML: '${parent_domains_yaml}'
|
||
---
|
||
apiVersion: kustomize.toolkit.fluxcd.io/v1
|
||
kind: Kustomization
|
||
metadata:
|
||
name: infrastructure-config
|
||
namespace: flux-system
|
||
spec:
|
||
interval: 5m
|
||
path: ./clusters/_template/infrastructure
|
||
prune: true
|
||
sourceRef:
|
||
kind: GitRepository
|
||
name: openova
|
||
dependsOn:
|
||
- name: bootstrap-kit
|
||
- name: sovereign-tls
|
||
wait: true
|
||
# timeout: 5m (issue #492) — same rationale as bootstrap-kit.
|
||
# Releases the revision lock quickly so iterative fixes during
|
||
# Phase-8a get applied on the next GitRepository poll instead
|
||
# of stalling 30 minutes per failed apply.
|
||
timeout: 5m
|
||
postBuild:
|
||
substitute:
|
||
SOVEREIGN_FQDN: ${sovereign_fqdn}
|
||
# SOVEREIGN_LB_IP — Hetzner load balancer's public IPv4 (issue
|
||
# #900). Threaded into bp-catalyst-platform's
|
||
# `global.sovereignLBIP` so catalyst-api can pre-register glue
|
||
# records at the customer's Dynadot account before flipping NS
|
||
# for Day-2 multi-domain adds. Sourced from
|
||
# hcloud_load_balancer.main.ipv4 in main.tf.
|
||
SOVEREIGN_LB_IP: ${load_balancer_ipv4}
|
||
# Marketplace mode toggle (issue #710). Default "false" so a
|
||
# zero-touch provision lands a non-marketplace Sovereign. The
|
||
# operator (or the wizard's "Enable Marketplace" component
|
||
# checkbox) flips this to "true" via per-Sovereign overlay
|
||
# patch — bp-catalyst-platform 1.3.0 then renders the
|
||
# marketplace + tenant-wildcard HTTPRoutes.
|
||
MARKETPLACE_ENABLED: "${marketplace_enabled}"
|
||
|
||
runcmd:
|
||
- swapoff -a
|
||
- sed -i '/swap/d' /etc/fstab
|
||
- update-alternatives --set iptables /usr/sbin/iptables-legacy || true
|
||
- update-alternatives --set ip6tables /usr/sbin/ip6tables-legacy || true
|
||
|
||
# Apply inotify-limit bumps written by write_files. sysctl --system
|
||
# picks up /etc/sysctl.d/*.conf so future blueprints + bao init never
|
||
# hit "too many open files" again.
|
||
- sysctl --system
|
||
|
||
# Activate hardened sshd config (cloud-init may have written authorized_keys
|
||
# already from Hetzner ssh_keys[]; we never touch that file).
|
||
- systemctl reload ssh || systemctl reload sshd || true
|
||
%{ if enable_fail2ban ~}
|
||
- systemctl enable --now fail2ban
|
||
%{ endif ~}
|
||
%{ if enable_unattended_upgrades ~}
|
||
- systemctl enable --now unattended-upgrades
|
||
%{ endif ~}
|
||
|
||
# k3s control-plane. Flags per docs/SOVEREIGN-PROVISIONING.md §3 and
|
||
# docs/PLATFORM-TECH-STACK.md §8.1:
|
||
# --cluster-init Initialise embedded etcd (HA-ready).
|
||
# --flannel-backend=none Cilium replaces flannel.
|
||
# --disable=traefik Cilium Gateway replaces traefik.
|
||
# --disable=servicelb Hetzner LB handles ingress.
|
||
# --disable-network-policy Cilium handles NetworkPolicy.
|
||
# --tls-san=${sovereign_fqdn} API server cert valid for the sovereign FQDN.
|
||
#
|
||
# ── kube-apiserver OIDC flags (issue #326) ─────────────────────────────
|
||
# --kube-apiserver-arg=oidc-issuer-url=https://auth.<sovereign_fqdn>/realms/sovereign
|
||
# --kube-apiserver-arg=oidc-client-id=kubectl
|
||
# --kube-apiserver-arg=oidc-username-claim=preferred_username
|
||
# --kube-apiserver-arg=oidc-username-prefix=oidc:
|
||
# --kube-apiserver-arg=oidc-groups-claim=groups
|
||
# --kube-apiserver-arg=oidc-groups-prefix=oidc:
|
||
# Wire k3s api-server's OIDC validator to the per-Sovereign Keycloak
|
||
# realm (`sovereign`), shipped by bp-keycloak's keycloakConfigCli realm
|
||
# import (platform/keycloak/chart/values.yaml). After the Sovereign's
|
||
# bootstrap kit lands, customer admins authenticate kubectl against
|
||
# Keycloak (see docs/omantel-handover-wbs.md §11 "kubectl OIDC for
|
||
# customer admins"). The username/groups prefixes prefix every
|
||
# OIDC-issued subject with `oidc:` so RoleBindings reference them as
|
||
# e.g. `subjects[0].name=oidc:alice@org` — distinct from any local
|
||
# ServiceAccount or x509 subject. Per INVIOLABLE-PRINCIPLES #4 the
|
||
# issuer URL is composed from sovereign_fqdn — never hardcoded.
|
||
#
|
||
# Trust-chain note: the per-Sovereign Keycloak is exposed via the
|
||
# `cilium-gateway` Gateway (kube-system), whose serving Certificate is
|
||
# issued by Let's Encrypt via bp-cert-manager. k3s's kube-apiserver
|
||
# reaches the in-cluster Keycloak Service over plain HTTPS using the
|
||
# node's system trust store; LE roots are present by default on the
|
||
# Ubuntu 24.04 control-plane image, so no `--oidc-ca-file` is needed
|
||
# in this configuration. Air-gapped Sovereigns (deferred Phase 9+)
|
||
# add a CA-file flag here when their Keycloak fronts a private CA.
|
||
#
|
||
# NOTE: --disable=local-storage is intentionally NOT passed. k3s ships a
|
||
# built-in local-path-provisioner (Rancher) and registers a `local-path`
|
||
# StorageClass. That is the canonical solo-Sovereign StorageClass:
|
||
# PVCs (bp-spire data dir, bp-keycloak postgres, bp-openbao raft store,
|
||
# bp-nats-jetstream, bp-gitea, bp-catalyst-platform postgres) bind to
|
||
# node-local storage on the single CPX21/CPX31 control-plane node and
|
||
# come up immediately. Operators upgrading to multi-node migrate to
|
||
# hcloud-csi (Hetzner Cloud Volumes) as a separate, deliberate step —
|
||
# see docs/RUNBOOK-PROVISIONING.md §"StorageClass missing".
|
||
#
|
||
# Architectural background: the prior version of this template passed
|
||
# `--disable=local-storage` with the intent that Crossplane would
|
||
# install hcloud-csi day-2 and register the StorageClass after
|
||
# bp-crossplane reconciled. That created a circular dependency: the
|
||
# 11-component bootstrap kit (bp-spire / bp-keycloak / bp-openbao / …)
|
||
# all carry PVCs whose bind step blocks waiting for a StorageClass that
|
||
# would only exist AFTER bp-crossplane had finished installing AND
|
||
# provisioned hcloud-csi. Result on a fresh Sovereign: every PVC stuck
|
||
# Pending forever, bootstrap-kit deadlocked. Keeping local-path solves
|
||
# the circularity by giving the cluster a default StorageClass at boot.
|
||
# ── k3s server install ───────────────────────────────────────────────────
|
||
#
|
||
# --node-taint node-role.kubernetes.io/control-plane=true:NoSchedule
|
||
# ─────────────────────────────────────────────────────────────────────────
|
||
# CONDITIONAL: applied ONLY when worker_count > 0 (multi-node Sovereign).
|
||
#
|
||
# Why: by k3s default, the server (control-plane) node is fully
|
||
# schedulable — the kube-scheduler distributes pods by resource fit. On
|
||
# a 1-CP + N-worker Sovereign with the 37-HelmRelease bootstrap-kit +
|
||
# guest workloads (bp-keycloak, bp-cnpg, bp-harbor, bp-catalyst-platform,
|
||
# SME microservices), the scheduler happily lands guest workloads on the
|
||
# CP. They eat its memory, crowd kubelet/etcd/apiserver, and the whole
|
||
# cluster degrades: kubectl flakes, Helm post-install hooks time out,
|
||
# HelmReleases get stuck mid-reconcile. This is the root cause of the
|
||
# "apiserver flake / cpx22 too small / 8 stuck HRs" symptom chain
|
||
# (issue #751). The taint reserves the CP for system + bootstrap
|
||
# controllers (cilium agent + cilium-operator + flux + cert-manager +
|
||
# kube-system pods, all of which tolerate this taint by upstream chart
|
||
# convention), and pushes guest workloads to workers where they belong.
|
||
#
|
||
# Why CONDITIONAL on worker_count > 0: a Catalyst-Zero / solo Sovereign
|
||
# (worker_count=0) has only the CP — tainting NoSchedule there leaves
|
||
# NO node for any deployment to schedule onto, the cluster never
|
||
# becomes ready. Skip the taint when there are no workers; fall back
|
||
# to k3s default (CP fully schedulable) so the solo node carries
|
||
# everything.
|
||
- 'curl -sfL https://get.k3s.io | INSTALL_K3S_VERSION=${k3s_version} K3S_TOKEN=${k3s_token} INSTALL_K3S_EXEC="server --cluster-init --flannel-backend=none --disable-network-policy --disable=traefik --disable=servicelb --tls-san=${sovereign_fqdn} --tls-san=10.0.1.2 --kube-apiserver-arg=oidc-issuer-url=https://auth.${sovereign_fqdn}/realms/sovereign --kube-apiserver-arg=oidc-client-id=kubectl --kube-apiserver-arg=oidc-username-claim=preferred_username --kube-apiserver-arg=oidc-username-prefix=oidc: --kube-apiserver-arg=oidc-groups-claim=groups --kube-apiserver-arg=oidc-groups-prefix=oidc: --node-label catalyst.openova.io/role=control-plane ${worker_count > 0 ? "--node-taint node-role.kubernetes.io/control-plane=true:NoSchedule " : ""}--write-kubeconfig-mode=0644" sh -'
|
||
|
||
# Wait for the API server to be reachable. Cilium needs to come up before
|
||
# nodes Ready, so we wait specifically for the API endpoint.
|
||
- 'until kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml get --raw /healthz; do sleep 5; done'
|
||
|
||
# ── Default StorageClass: local-path (k3s built-in) ─────────────────────
|
||
#
|
||
# k3s ships local-path-provisioner (deployment in kube-system,
|
||
# `app=local-path-provisioner`) and registers a `local-path`
|
||
# StorageClass on first boot. We need the StorageClass to exist AND
|
||
# be marked default BEFORE Flux applies the bootstrap-kit Kustomization
|
||
# below — otherwise the 11-component bootstrap kit (bp-spire,
|
||
# bp-keycloak postgres, bp-openbao, bp-nats-jetstream, bp-gitea,
|
||
# bp-catalyst-platform postgres) ships HelmReleases with PVCs that
|
||
# have no `storageClassName` set, expecting the cluster default to
|
||
# take over. Without a default, every one of those PVCs sits Pending
|
||
# waiting on a class that nobody nominates, and the bootstrap-kit
|
||
# Kustomization deadlocks.
|
||
#
|
||
# Sequence (#207 — fix the circular wait that blocked every fresh provision):
|
||
# 1. Poll until the `local-path` StorageClass object is registered by
|
||
# k3s. We CANNOT wait for the local-path-provisioner POD to be
|
||
# Ready here — k3s runs with --flannel-backend=none so the node
|
||
# stays Ready=False until Cilium installs (further down). Waiting
|
||
# on the Pod creates a circular deadlock and 60s timeout. The SC
|
||
# object itself is registered by k3s manifests independently of CNI
|
||
# (verified live: SC creationTimestamp 3s after k3s start).
|
||
# 2. Patch the `local-path` StorageClass with the
|
||
# `storageclass.kubernetes.io/is-default-class: "true"` annotation.
|
||
# 3. Verify (the poll already implies presence; the explicit grep stays
|
||
# as defensive belt-and-braces, identical exit semantics).
|
||
- 'until kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml get sc local-path >/dev/null 2>&1; do sleep 2; done'
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml patch storageclass local-path -p ''{"metadata":{"annotations":{"storageclass.kubernetes.io/is-default-class":"true"}}}'''
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml get sc -o name | grep -q "^storageclass.storage.k8s.io/local-path$" || { echo "FATAL: local-path StorageClass missing after k3s install — see docs/RUNBOOK-PROVISIONING.md StorageClass missing section" >&2; exit 1; }'
|
||
|
||
%{ if deployment_id != "" && kubeconfig_bearer_token != "" && catalyst_api_url != "" ~}
|
||
# ── Cloud-init kubeconfig postback (issue #183, Option D) ───────────────
|
||
#
|
||
# The k3s install above wrote /etc/rancher/k3s/k3s.yaml with the API
|
||
# server URL pinned to https://127.0.0.1:6443 — kubectl's default for a
|
||
# local single-node install. catalyst-api lives off-cluster (Catalyst-Zero
|
||
# franchise console on contabo-mkt) and cannot reach 127.0.0.1 on this
|
||
# node, so we MUST rewrite that field before sending the kubeconfig
|
||
# back.
|
||
#
|
||
# Issue #542: kubeconfig server: must be the control-plane's PUBLIC IPv4,
|
||
# NOT the load balancer's. The Hetzner LB only forwards 80/443 (Cilium
|
||
# Gateway ingress); 6443 is exposed directly on the CP via firewall rule
|
||
# main.tf:51-56 (0.0.0.0/0 → CP:6443). Earlier code rewrote to the LB IP
|
||
# which silently fails with "connect: connection refused" — wizard jobs
|
||
# page stuck PENDING for 50+ minutes after install completes.
|
||
#
|
||
# Plaintext: we read from /etc/rancher/k3s/k3s.yaml (mode 0644 written
|
||
# by k3s), apply the rewrite via sed, write the result to
|
||
# /etc/rancher/k3s/k3s.yaml.public (mode 0600 explicitly), then
|
||
# curl --data-binary the file content to catalyst-api with the bearer
|
||
# token. The .public file is removed at the end of the runcmd block
|
||
# so the bearer-protected kubeconfig only lives on this node for the
|
||
# few seconds it takes to PUT.
|
||
#
|
||
# --retry 60 --retry-delay 10 --retry-all-errors handles the case
|
||
# where catalyst-api is briefly unreachable (image roll, ingress
|
||
# reconciliation) — the cloud-init runcmd budget is bounded by the
|
||
# systemd cloud-final timeout (~30 minutes).
|
||
# control_plane_ipv4 resolved at runtime via Hetzner metadata service
|
||
# (rather than templated by Tofu — that would create a dependency cycle:
|
||
# cloud-init → control_plane.ipv4_address → control_plane.user_data → cloud-init).
|
||
# 169.254.169.254 is the standard cloud metadata endpoint; Hetzner exposes
|
||
# public-ipv4 at /hetzner/v1/metadata/public-ipv4 — single line, no auth.
|
||
- install -m 0600 /dev/null /etc/rancher/k3s/k3s.yaml.public
|
||
- 'CP_PUBLIC_IPV4=$(curl -fsSL --retry 30 --retry-delay 2 http://169.254.169.254/hetzner/v1/metadata/public-ipv4) && sed "s|https://127.0.0.1:6443|https://$${CP_PUBLIC_IPV4}:6443|g" /etc/rancher/k3s/k3s.yaml > /etc/rancher/k3s/k3s.yaml.public'
|
||
- chmod 0600 /etc/rancher/k3s/k3s.yaml.public
|
||
- |
|
||
curl -fsSL --retry 60 --retry-delay 10 --retry-all-errors \
|
||
-X PUT \
|
||
-H "Authorization: Bearer ${kubeconfig_bearer_token}" \
|
||
-H "Content-Type: application/x-yaml" \
|
||
--data-binary @/etc/rancher/k3s/k3s.yaml.public \
|
||
${catalyst_api_url}/api/v1/deployments/${deployment_id}/kubeconfig
|
||
- rm -f /etc/rancher/k3s/k3s.yaml.public
|
||
%{ endif ~}
|
||
|
||
# ── Cilium FIRST (before Flux) ───────────────────────────────────────────
|
||
#
|
||
# k3s started with --flannel-backend=none, so the cluster has NO CNI yet.
|
||
# If we apply Flux install.yaml at this point, the Flux controller pods
|
||
# stay Pending forever — kubelet rejects them with
|
||
# "container runtime network not ready: cni plugin not initialized"
|
||
# Flux is then unable to reconcile bp-cilium, so Cilium is never
|
||
# installed → bootstrap deadlock that we hit in production at
|
||
# omantel.omani.works deployment 5cd1bceaaacb71f6 (25 min stuck Pending).
|
||
#
|
||
# Bootstrap chicken-and-egg: Cilium IS the install unit (bp-cilium), but
|
||
# Flux needs a CNI to run, and Cilium IS the CNI. Resolution: install
|
||
# Cilium ONCE here via Helm with the same chart + values bp-cilium would
|
||
# apply later. When Flux reconciles bp-cilium, it adopts the existing
|
||
# release (Helm release-name match), so there is no churn.
|
||
#
|
||
# Per INVIOLABLE-PRINCIPLES.md #3 the GitOps engine is Flux — this Helm
|
||
# install is the one-shot bootstrap exception explicitly authorised by
|
||
# the same principle's "everything ELSE" qualifier. Both the chart
|
||
# version AND the values must match `platform/cilium/blueprint.yaml`
|
||
# + `clusters/_template/bootstrap-kit/01-cilium.yaml` so the bootstrap
|
||
# install and the reconciled HelmRelease are byte-identical — issue
|
||
# #491. The values come from /var/lib/catalyst/cilium-values.yaml
|
||
# written via cloud-init `write_files:` above; chart version stays
|
||
# inline as a --version flag because OpenTofu's `var.k3s_version`
|
||
# parameterisation wires through to it (per INVIOLABLE-PRINCIPLES
|
||
# #4 — never hardcode).
|
||
# ── Gateway API CRDs BEFORE Cilium ──────────────────────────────────────
|
||
#
|
||
# Cilium 1.16.x operator checks for gateway.networking.k8s.io CRDs at
|
||
# startup. If the CRDs are absent the operator disables its gateway
|
||
# controller entirely and never re-checks — a static decision made once
|
||
# at boot. This creates a race when Gateway API CRDs are installed AFTER
|
||
# k3s/Cilium, which is the normal Flux GitOps order (bp-gateway-api
|
||
# reconciles minutes after bp-cilium). Result: every fresh Sovereign has
|
||
# no GatewayClass/cilium, all HTTPRoutes are orphaned, no routing.
|
||
#
|
||
# Fix: pre-install the Gateway API experimental CRDs here, before the
|
||
# Cilium helm install below. The experimental channel is required because
|
||
# Cilium 1.16.x references tlsroutes.gateway.networking.k8s.io (v1alpha2)
|
||
# at startup; the standard channel does not ship TLSRoute.
|
||
#
|
||
# Version choice — v1.1.0 NOT v1.2.0:
|
||
# Gateway API v1.2.0 changed status.supportedFeatures from an array of
|
||
# strings to an array of objects ({name: string}). Cilium 1.16.5 still
|
||
# writes the old string format; the v1.2.0 CRD rejects its status patch
|
||
# with "must be of type object: string", leaving GatewayClass/cilium
|
||
# permanently in status=Unknown/Pending. v1.1.0 retains the string
|
||
# format and is fully compatible with Cilium 1.16.x.
|
||
#
|
||
# bp-gateway-api Flux blueprint becomes a no-op on first reconcile
|
||
# (CRDs already present, kubectl apply is idempotent); it is kept as the
|
||
# GitOps record and handles CRD upgrades when Cilium is bumped.
|
||
#
|
||
# Incident reference: otech22 2026-05-02 — all 8 HTTPRoutes orphaned,
|
||
# cilium-operator log: "Required GatewayAPI resources are not found …
|
||
# tlsroutes.gateway.networking.k8s.io not found". Fix: issue #503.
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f https://github.com/kubernetes-sigs/gateway-api/releases/download/v1.1.0/experimental-install.yaml'
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml wait --for=condition=Established crd/tlsroutes.gateway.networking.k8s.io --timeout=60s'
|
||
|
||
- 'curl -sSL https://raw.githubusercontent.com/helm/helm/main/scripts/get-helm-3 | bash'
|
||
- 'helm repo add cilium https://helm.cilium.io/'
|
||
- 'helm repo update'
|
||
- |
|
||
KUBECONFIG=/etc/rancher/k3s/k3s.yaml helm install cilium cilium/cilium \
|
||
--version 1.16.5 \
|
||
--namespace kube-system \
|
||
-f /var/lib/catalyst/cilium-values.yaml
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml -n kube-system rollout status ds/cilium --timeout=240s'
|
||
|
||
# Install Flux core. Cilium is now the cluster's CNI, so Flux pods will
|
||
# actually start. Flux then reconciles clusters/_template/ (with
|
||
# SOVEREIGN_FQDN substituted via postBuild — issue #218) which
|
||
# adopts the Helm release above as bp-cilium and continues with
|
||
# bp-cert-manager, bp-flux (which ADOPTS this Flux install rather than
|
||
# reinstalls — see version-pin invariant below), bp-crossplane, etc.
|
||
#
|
||
# CRITICAL VERSION-PIN INVARIANT — DO NOT CHANGE IN ISOLATION
|
||
# -----------------------------------------------------------
|
||
# The version pinned in the URL below MUST match the upstream Flux
|
||
# release that `platform/flux/chart/Chart.yaml`'s `flux2` subchart
|
||
# bundles, otherwise bp-flux's HelmRelease runs `helm install` on top
|
||
# of THIS Flux installation with a different upstream version, the
|
||
# CRD `status.storedVersions` mismatches, Helm install fails, rollback
|
||
# fires, and rollback DELETES the running Flux controllers — leaving
|
||
# the cluster with no GitOps engine, unrecoverable in-place.
|
||
#
|
||
# Live verified on omantel.omani.works on 2026-04-29 — every Sovereign
|
||
# provisioned without this pin in sync was destroyed minutes after
|
||
# bp-flux's first reconcile. See docs/RUNBOOK-PROVISIONING.md
|
||
# §"bp-flux double-install".
|
||
#
|
||
# Mapping (cloud-init install.yaml -> chart subchart -> appVersion):
|
||
# v2.4.0 -> flux2 2.14.1 -> appVersion 2.4.0 <- CURRENT
|
||
# v2.3.0 -> flux2 2.13.0 -> appVersion 2.3.0
|
||
#
|
||
# CI gate `platform/flux/chart/tests/version-pin-replay.sh` rejects
|
||
# divergence between this URL's version and the chart's subchart pin.
|
||
- 'curl -fsSL https://github.com/fluxcd/flux2/releases/download/v2.4.0/install.yaml | kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f -'
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml -n flux-system wait --for=condition=Available --timeout=300s deployment --all'
|
||
|
||
# ── flux-system/ghcr-pull Secret (applied BEFORE GitRepository) ──────
|
||
#
|
||
# Apply the docker-registry pull secret rendered above. This MUST land
|
||
# before the GitRepository + Kustomization in flux-bootstrap.yaml,
|
||
# because the bootstrap-kit Kustomization includes HelmRepository CRs
|
||
# that reference this Secret by name; the source-controller resolves
|
||
# them on its first reconciliation tick and a missing Secret propagates
|
||
# as a Ready=False/AuthError state that has been observed to persist
|
||
# for 5+ minutes even after the Secret is later applied.
|
||
#
|
||
# Idempotent: `kubectl apply` against an existing Secret is a no-op
|
||
# when the manifest's bytes match. A reprovision (same Sovereign FQDN)
|
||
# rewrites this with the same content; a token rotation propagates
|
||
# through here on the next cloud-init render.
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f /var/lib/catalyst/ghcr-pull-secret.yaml'
|
||
|
||
# ── flux-system/harbor-robot-token Secret (issue #557 follow-up) ─────
|
||
#
|
||
# Apply the central-Harbor robot-account Secret BEFORE Flux reconciles
|
||
# the bootstrap-kit. bp-catalyst-platform's catalyst-api Pod has a
|
||
# REQUIRED (non-optional) secretKeyRef to `harbor-robot-token` in its
|
||
# own namespace; bp-reflector (slot 05a) mirrors this Secret from
|
||
# flux-system into catalyst-system on first reconcile so the Pod can
|
||
# start cleanly. Same idempotency property as ghcr-pull above.
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f /var/lib/catalyst/harbor-robot-token-secret.yaml'
|
||
|
||
# ── flux-system/pdm-basicauth Secret (issue #879 Bug 2) ──────────────
|
||
#
|
||
# Apply the PDM basic-auth credentials BEFORE Flux reconciles the
|
||
# bootstrap-kit. bp-reflector (slot 05a) mirrors this Secret from
|
||
# flux-system into catalyst-system on first reconcile so the
|
||
# catalyst-api Pod can mount it via secretKeyRef (optional=true so
|
||
# Pod start is not blocked when this is absent). Same idempotency
|
||
# property as ghcr-pull above — re-running cloud-init rewrites the
|
||
# bytes; a token rotation propagates through here on the next render.
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f /var/lib/catalyst/pdm-basicauth-secret.yaml'
|
||
|
||
# ── cert-manager/powerdns-api-credentials Secret (PR #681 followup) ──
|
||
#
|
||
# Apply the contabo PowerDNS credentials BEFORE Flux reconciles
|
||
# bp-cert-manager-powerdns-webhook: the webhook reads
|
||
# apiKeySecretRef.name=powerdns-api-credentials at startup. A missing
|
||
# Secret causes the DNS-01 challenge to hang indefinitely with
|
||
# "secrets powerdns-api-credentials not found" and the wildcard cert
|
||
# never issues — the Sovereign Console TLS handshake then fails
|
||
# (caught live on otech47).
|
||
#
|
||
# cert-manager namespace is created by bp-cert-manager via Flux — we
|
||
# pre-create it idempotently here so the Secret apply does not fail.
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml create namespace cert-manager --dry-run=client -o yaml | kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f -'
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f /var/lib/catalyst/powerdns-api-credentials.yaml'
|
||
|
||
# ── catalyst-system/catalyst-handover-jwt-public Secret (issue #606 followup) ─
|
||
#
|
||
# Apply the Sovereign-side Catalyst handover-JWT public-key Secret BEFORE
|
||
# Flux reconciles bp-catalyst-platform. Without this Secret the catalyst-api
|
||
# Pod's optional Secret-volume mount falls through, the JWK file is absent
|
||
# at /etc/catalyst/handover-jwt-public/public.jwk, and GET /auth/handover
|
||
# returns "server misconfiguration: public key unavailable" (caught live on
|
||
# otech48, 2026-05-03).
|
||
#
|
||
# Pre-create the catalyst-system namespace (the bp-catalyst-platform
|
||
# HelmRelease wrapper also declares it, but Flux applies the namespace
|
||
# alongside the HelmRelease — racing this Secret apply). Same idempotency
|
||
# pattern as the cert-manager pre-create above.
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml create namespace catalyst-system --dry-run=client -o yaml | kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f -'
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f /var/lib/catalyst/handover-jwt-public-secret.yaml'
|
||
|
||
# ── OpenBao auto-unseal seed Secret (issue #316) ─────────────────────
|
||
#
|
||
# Generate a one-shot 32-byte recovery seed during cloud-init and
|
||
# write it to a K8s Secret `openbao-recovery-seed` in the `openbao`
|
||
# namespace. The bp-openbao chart (v1.2.0+) renders a post-install
|
||
# Job (templates/init-job.yaml, Helm hook weight 5) that:
|
||
# 1. Reads this seed Secret.
|
||
# 2. Calls `bao operator init -recovery-shares=1 -recovery-threshold=1`.
|
||
# 3. Persists the recovery key inside OpenBao's auto-unseal config
|
||
# (so subsequent pod restarts unseal automatically).
|
||
# 4. Deletes this seed Secret on success.
|
||
#
|
||
# The seed is single-use — once consumed by the init Job, it never
|
||
# exists again. The recovery key + root token live ONLY inside
|
||
# OpenBao's Raft state (acceptance criterion #6 of issue #316).
|
||
#
|
||
# Why a fresh /dev/urandom value (NOT a value baked into Terraform):
|
||
# the recovery seed must NEVER be readable from outside the
|
||
# control-plane node, NEVER appear in tfstate, NEVER appear in any
|
||
# cloud-init render audit log. Generating it here at provision time
|
||
# means the only window of plaintext exposure is the few seconds
|
||
# between this Secret apply and the Helm post-install Job consuming
|
||
# it — bounded by the bootstrap-kit reconcile cadence (1m max).
|
||
#
|
||
# Why we create the namespace here: the bp-openbao HelmRelease in
|
||
# clusters/_template/bootstrap-kit/08-openbao.yaml ships a Namespace
|
||
# manifest, but Flux applies that Namespace + the HelmRelease
|
||
# together. The Helm post-install hook would race the seed Secret
|
||
# apply if we waited for Flux to create the namespace. Pre-creating
|
||
# the namespace at cloud-init time eliminates the race.
|
||
#
|
||
# Idempotency: `kubectl apply` of the namespace and `kubectl create
|
||
# secret --dry-run=client -o yaml | kubectl apply -f -` of the
|
||
# Secret are both safe to re-run. A re-provision (same Sovereign
|
||
# FQDN) regenerates a fresh seed and re-applies — at which point the
|
||
# init Job has either already consumed the previous seed (so the new
|
||
# one becomes a no-op the next time the Helm hook runs) OR sees
|
||
# OpenBao already initialised and exits idempotently without
|
||
# touching the new seed.
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml create namespace openbao --dry-run=client -o yaml | kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f -'
|
||
- |
|
||
OPENBAO_SEED=$(head -c 32 /dev/urandom | base64 | tr -d '\n')
|
||
kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml -n openbao create secret generic openbao-recovery-seed \
|
||
--from-literal=recovery-seed="$OPENBAO_SEED" \
|
||
--dry-run=client -o yaml \
|
||
| kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml annotate --local -f - \
|
||
openbao.openova.io/single-use=true -o yaml \
|
||
| kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f -
|
||
unset OPENBAO_SEED
|
||
|
||
# ── flux-system/object-storage Secret (issue #371, vendor-agnostic since #425) ─
|
||
#
|
||
# Apply the operator-issued Object Storage credentials so they're in
|
||
# the cluster BEFORE Flux reconciles bp-harbor (#383) and bp-velero
|
||
# (#384). Both Blueprints reference `secretRef: name: object-storage`
|
||
# in their HelmRelease values; without this Secret the install reports
|
||
# NoSuchKey at chart-install probe time and Phase 1 stalls.
|
||
#
|
||
# Same idempotency property as ghcr-pull above — re-running cloud-init
|
||
# against an existing Sovereign overwrites the manifest with the same
|
||
# bytes (or rotated bytes when the operator has issued fresh keys); a
|
||
# missing-bucket scenario is impossible by construction because main.tf's
|
||
# minio_s3_bucket resource creates the bucket in the same `tofu apply`
|
||
# run that renders this user_data.
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f /var/lib/catalyst/object-storage-secret.yaml'
|
||
|
||
# ── flux-system/cloud-credentials Secret + Crossplane Provider (issue #425) ─
|
||
#
|
||
# Apply the Hetzner Cloud API token Secret + the Crossplane Provider
|
||
# package + ProviderConfig BEFORE Flux's bootstrap-kit lands
|
||
# bp-crossplane. The Provider package itself is installed by
|
||
# Crossplane core (which bp-crossplane brings up); applying the
|
||
# Provider CR here just registers the package install request — it
|
||
# transitions Healthy=True a few minutes later once the bootstrap-
|
||
# kit's Crossplane core controllers come online. The ProviderConfig
|
||
# sits in waiting state until the Provider's CRDs are registered, at
|
||
# which point it goes Ready=True and the Sovereign is ready to accept
|
||
# Day-2 XRC writes.
|
||
#
|
||
# Per ADR-0001 §11.3 + INVIOLABLE-PRINCIPLES #3 this is the OpenTofu
|
||
# → Crossplane handover seam. Tofu provisions Phase 0 exactly once;
|
||
# everything else flows through XRC writes against this Provider.
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f /var/lib/catalyst/cloud-credentials-secret.yaml'
|
||
# IMPORTANT: the Provider CRD (`pkg.crossplane.io/v1`) is registered by
|
||
# bp-crossplane via Flux AFTER flux-bootstrap.yaml is applied below.
|
||
# Apply with `|| true` so the failure here doesn't propagate into
|
||
# cloud-init's exit status. The Provider CR is re-applied below in a
|
||
# background retry once Crossplane core CRDs land.
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f /var/lib/catalyst/crossplane-provider-hcloud.yaml || true'
|
||
|
||
# Apply the Flux bootstrap GitRepository + Kustomization. From here, Flux
|
||
# owns the cluster: pulls clusters/_template/ (with $${SOVEREIGN_FQDN}
|
||
# substituted to ${sovereign_fqdn} via postBuild), installs Cilium
|
||
# via bp-cilium, cert-manager via bp-cert-manager, etc., then bp-catalyst-platform.
|
||
- 'kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f /var/lib/catalyst/flux-bootstrap.yaml'
|
||
|
||
# Background retry of the Crossplane Provider CR apply. Polls every
|
||
# 30s up to 30m for the Crossplane Provider CRD to be registered
|
||
# (i.e. bp-crossplane reconciled by Flux), then `kubectl apply`
|
||
# succeeds and the loop exits. Detached via `&` so cloud-init runcmd
|
||
# completes without waiting for Crossplane to be Ready.
|
||
- 'nohup bash -c "deadline=\$((SECONDS+1800)); while [ \$SECONDS -lt \$deadline ]; do kubectl --kubeconfig=/etc/rancher/k3s/k3s.yaml apply -f /var/lib/catalyst/crossplane-provider-hcloud.yaml >/dev/null 2>&1 && break; sleep 30; done" >/dev/null 2>&1 &'
|
||
|
||
# Marker for the catalyst-api provisioner to detect cloud-init is done.
|
||
- mkdir -p /var/lib/catalyst
|
||
- touch /var/lib/catalyst/cloud-init-complete
|
||
|
||
final_message: "Catalyst control-plane bootstrap complete after $UPTIME seconds — Flux is now reconciling clusters/_template/ for ${sovereign_fqdn}"
|