Sprint 3 unified-master-audit closure — two Helm-chart correctness
defects with overlapping CI-guard surface.
DEPL-003 — CERTCTL_MIGRATIONS_VIA_HOOK never rendered:
Pre-fix the env var was documented in values.yaml and the
migration-job.yaml comment but never made it into the server
Deployment env block. With migrations.viaHook=true the operator's
intent is 'the pre-install/pre-upgrade Helm Job owns migrations,'
but the server pods, missing the env, ran their own
cmd/server/migrations.go::runBootMigrations alongside the hook
Job, racing on the schema lock.
Fix: render '- name: CERTCTL_MIGRATIONS_VIA_HOOK / value: true'
in server-deployment.yaml under '{{- if .Values.migrations.viaHook }}'.
DEPL-006 — HA example missing rate-limit backend + sessionAffinity:
values-prod-ha.yaml sets replicas:3 but inherited the chart-wide
default rateLimiting.backend=memory (which gives each pod its
own bucket map, effectively tripling the cap on a 3-replica fleet)
AND the chart had no render path for server.service.sessionAffinity
even though docs/operator/runbooks/ha.md instructed operators to
set it for ClientIP-routed sticky sessions.
Fix:
- server-service.yaml gains a conditional sessionAffinity +
sessionAffinityConfig.clientIP.timeoutSeconds render.
- values.yaml grows the matching schema entries (default empty
so single-replica deploys are unaffected).
- values-prod-ha.yaml flips rateLimiting.backend=postgres and
service.sessionAffinity=ClientIP.
- NOTES.txt emits a loud warning when replicas>1 + either toggle
is still in the default state, so the misconfig surfaces at
helm install time instead of in a confused login-flow bug
report a week later.
CI:
scripts/ci-guards/B3-helm-chart-coherence.sh gains 'Check 7'
(DEPL-003 viaHook env render — both positive and negative —
the inverse case catches future drift that drops the {{- if }}
guard) and 'Check 8' (DEPL-006 sessionAffinity render). Both
helm-template through to assert the rendered YAML carries the
expected text.
Closes DEPL-003, DEPL-006.
certctl Helm Chart
Production-ready Helm chart for deploying certctl on Kubernetes. Wires up the certctl server (Deployment), PostgreSQL (StatefulSet with PVC), and the agent (DaemonSet — one per node) on a private cluster, with health probes, security contexts, and optional Ingress.
Quick install
helm install certctl deploy/helm/certctl/ \
--create-namespace --namespace certctl \
--set server.auth.apiKey="$(openssl rand -base64 32)" \
--set postgresql.auth.password="$(openssl rand -base64 24)"
This brings up:
<release>-serverDeployment (HTTPS-only on port 8443; TLS 1.3)<release>-postgresStatefulSet (PostgreSQL 16-alpine, 1 replica, 10Gi PVC by default)<release>-agentDaemonSet (polls server, generates ECDSA P-256 keys locally)- Service objects, optional Ingress, and ServiceAccount with RBAC
See values.yaml for the full configuration surface — issuer settings, target connectors, scheduler intervals, notifier credentials, and resource requests/limits all live there.
Operational notes
Postgres password rotation — read this before changing postgresql.auth.password
The trap. postgresql.auth.password is bound to pg_authid exactly once — when the StatefulSet's PVC is provisioned and initdb runs. The official postgres:16-alpine image only runs initdb when /var/lib/postgresql/data is empty, so on every subsequent rollout the POSTGRES_PASSWORD env var is read into the container but ignored by postgres itself. The certctl-server container also picks up the new value (via the database URL helper template), so the two halves diverge: server presents the new password, postgres still expects the old one.
Symptom. The certctl-server pod's startup log shows:
failed to ping database: postgres rejected the configured credentials
(SQLSTATE 28P01 — invalid_password). If you recently rotated POSTGRES_PASSWORD ...
That diagnostic is emitted by internal/repository/postgres/db.go::wrapPingError — it points operators at the two remediation paths below.
Remediation, non-destructive (preferred for any environment with real data):
# 1. Rotate the password in postgres directly
kubectl -n certctl exec -it <release>-postgres-0 -- \
psql -U certctl -c "ALTER ROLE certctl PASSWORD '<new-password>';"
# 2. Update the secret / Helm values to the same value
helm upgrade <release> deploy/helm/certctl/ \
--reuse-values \
--set postgresql.auth.password='<new-password>'
# 3. Bounce the certctl-server pod so it re-reads the secret
kubectl -n certctl rollout restart deployment/<release>-server
Remediation, destructive (DESTROYS ALL CERTCTL DATA — only acceptable on dev/demo clusters):
helm uninstall <release> -n certctl
kubectl -n certctl delete pvc -l \
app.kubernetes.io/name=certctl,app.kubernetes.io/component=postgres
helm install <release> deploy/helm/certctl/ \
--namespace certctl \
--set postgresql.auth.password='<new-password>'
The PVC re-creates empty, initdb runs on first boot of the new postgres pod, and pg_authid is seeded with the new password.
Why we don't fix this in the chart. The env-vs-pg_authid divergence is intrinsic to how the upstream postgres image bootstraps — initdb is run-once-per-empty-data-dir, and there is no upstream-supported way to make subsequent boots re-seed pg_authid from POSTGRES_PASSWORD. The ergonomic answer is the runtime diagnostic plus this operational note.
Cross-references. Same root cause is documented for the docker-compose path in docs/quickstart.md (Warning callout after the cp .env.example .env block) and in deploy/ENVIRONMENTS.md (Stateful volume — first-boot password binding section). The runtime diagnostic itself lives in internal/repository/postgres/db.go::wrapPingError with regression coverage in internal/repository/postgres/db_test.go.
Server API key rotation
Unlike the postgres password, server.auth.apiKey accepts a comma-separated list, so zero-downtime rotation is straightforward:
# 1. Add the new key alongside the old
helm upgrade <release> deploy/helm/certctl/ \
--reuse-values \
--set server.auth.apiKey='new-key,old-key'
# 2. Roll your agents / clients over to the new key
# 3. Remove the old key
helm upgrade <release> deploy/helm/certctl/ \
--reuse-values \
--set server.auth.apiKey='new-key'
JWT / OIDC via authenticating gateway
certctl's in-process auth surface is intentionally narrow: server.auth.type=api-key for production deployments and server.auth.type=none for development. There is no in-process JWT, OIDC, mTLS, or SAML middleware. (server.auth.type=jwt was accepted pre-G-1 but silently routed every request through the api-key bearer middleware — silent auth downgrade. The chart now fails at helm install/helm upgrade template time via the certctl.validateAuthType helper if you set it. See ../../../docs/upgrade-to-v2-jwt-removal.md if you previously had this in your values.)
For deployments that need JWT/OIDC, the canonical Kubernetes-flavored shape is to put oauth2-proxy in front of the certctl Service, attach an authenticating Ingress middleware, and run certctl with server.auth.type=none:
# 1. Install oauth2-proxy (or any OIDC-terminating sidecar) in the same namespace
helm install oauth2-proxy oauth2-proxy/oauth2-proxy \
--namespace certctl \
--set config.clientID="$OIDC_CLIENT_ID" \
--set config.clientSecret="$OIDC_CLIENT_SECRET" \
--set config.cookieSecret="$(openssl rand -base64 32)" \
--set config.configFile='|
provider = "oidc"
oidc_issuer_url = "https://your-issuer/"
upstreams = ["http://<release>-server.certctl.svc.cluster.local:8443"]
pass_authorization_header = true
set_authorization_header = true
email_domains = ["*"]
'
# 2. Install certctl with type=none (gateway terminates auth)
helm install certctl deploy/helm/certctl/ \
--namespace certctl \
--set server.auth.type=none \
--set postgresql.auth.password="$(openssl rand -base64 24)"
# 3. Attach an Ingress that routes through oauth2-proxy
# (Traefik ForwardAuth, nginx auth_request, Envoy ext_authz, etc.)
Same root pattern works with Pomerium, Authelia, Caddy forward_auth, Apache mod_auth_openidc, or any service-mesh ext_authz. See ../../../docs/architecture.md "Authenticating-gateway pattern" for the full design rationale and ../../../docs/upgrade-to-v2-jwt-removal.md for the migration walkthrough.
TLS certificate sourcing
By default the chart provisions a self-signed cert via the same init-container pattern as the docker-compose deploy. For production, supply an operator-managed Secret (cert-manager, internal CA, etc.) — see docs/tls.md for the full provisioning matrix and docs/upgrade-to-tls.md for upgrade-from-HTTP procedures.
Disabling embedded postgres
If you have an existing PostgreSQL cluster, disable the embedded one and point at it directly:
helm install certctl deploy/helm/certctl/ \
--set postgresql.enabled=false \
--set server.databaseUrl='postgres://certctl:<pw>@my-pg-host:5432/certctl?sslmode=require'
The volume-trap section above does not apply to this configuration — your postgres operator (or cloud DB) handles password rotation, and you control pg_authid directly.
Uninstall
helm uninstall <release> -n certctl
# Optional — also delete the postgres PVC (DESTROYS DATA):
kubectl -n certctl delete pvc -l \
app.kubernetes.io/name=certctl,app.kubernetes.io/component=postgres
By default helm uninstall retains the StatefulSet's PVCs, so reinstalling with the same release name preserves the database. If you've changed postgresql.auth.password in your values between uninstall and reinstall, you'll hit the trap on the reinstall — apply the non-destructive remediation above, or also delete the PVC.