v0.2.5 (PR #14) added an options-scope allow-notify enumerating the primary
pod IP on secondaries. Options-scope config feeds the config-hash annotation
that rolls the StatefulSet, so any config change rolled the pods, the primary
came back on a new pod IP, the operator re-rendered with the new IP, the hash
changed, the pods rolled again — an infinite roll loop across every
BindCluster. The prod deployment was reverted to v0.2.4.
Replace the pod-IP allow-notify with TSIG-authenticated NOTIFY:
- Secondaries render `allow-notify { key "<name>"; };` — a static key element
with NO IPs. It depends only on the key name, so pod-IP churn can never
change the render, the config-hash, or trigger a restart.
- The primary signs its outgoing NOTIFYs: the zone-scope also-notify entries
(already enumerating replica pod IPs, applied via rndc addzone/modzone with
NO restart) now carry `key "<name>"`.
- Key choice: reuse the cluster's catalog transfer TSIG key (TransferKeyRef).
Secondaries already present it for AXFR and it is in keys.conf on every pod,
so no new key plumbing is needed.
Add a permanent regression guard for the loop class:
- controller: reconcile the ConfigMap with the primary pod on two different
IPs and assert the config-hash is byte-identical.
- render: render restart-scoped input and assert no pod IP appears in
allow-notify; RenderInput no longer has any pod-IP field.
Zone-scope also-notify (rndc, no restart) legitimately still lists pod IPs;
only restart-scoped config must be pod-IP-independent.
Secondaries transfer catalog member and plain secondary zones from the
primary Service ClusterIP (stable across primary pod restarts), and BIND
derives a zone's implicit allow-notify from its primaries list. But the
primary pod's NOTIFYs egress with its *pod* IP as source — k8s Services
NAT only the inbound direction — so BIND refuses them as "refused notify
from non-primary" and replication falls back to the SOA refresh timer, a
1-hour propagation delay on every dynamic zone (external-dns RFC2136 and
dns-updater nsupdates alike).
Render an options-scope allow-notify on secondaries covering the primary
pod IP (and the transfer address, since an explicit allow-notify replaces
the primaries-derived default). The cluster controller resolves the
primary pod IP the same way it already does for seeding/also-notify, and
the existing Pod watch re-renders the ConfigMap when the pod IP changes.
Dynamically-updated primary zones were only reaching the secondary pods on
the hardcoded 1h SOA refresh: the operator emitted no NOTIFY, and a zone's
only apex NS is the primary itself, so default 'notify yes' reached no one.
Queries load-balanced across the serve VIP hit stale secondaries and
returned NXDOMAIN (negatively cached downstream for the 300s SOA minimum),
so records flapped for up to an hour after every update.
Add 'notify explicit' + 'also-notify' with the secondary pod IPs to primary
zone stanzas so an update NOTIFYs the secondaries for an immediate IXFR.
Applied via modzone, so existing zones pick it up on the next reconcile.
Also shorten the seed SOA refresh/retry/minimum as a fallback for missed
NOTIFYs and to shrink stale-NXDOMAIN negative caching.
The operator-generated TSIG Secret previously carried only the managed-by
label, so it could not be mirrored to another namespace by emberstack
reflector (which requires reflection-allowed annotations on the source).
Add spec.secretTemplate.{annotations,labels}, applied both when the Secret
is first generated and reconciled onto the existing Secret when the CR
changes (imported secrets are left untouched so we don't fight their
external manager). This lets the external-dns TSIG key be managed in
bind-internal and reflected into the externaldns namespace.
Pods copy config from the projected volume into an emptyDir once at
startup, so a ConfigMap or keys.conf change never reaches a running pod:
rndc reconfig re-reads the stale startup copy, and a manual
`kubectl rollout restart` is reverted because the operator overwrites the
pod template every reconcile. The only thing that applies new config is a
restart, and nothing triggered one.
Stamp a hash of the projected config (rendered ConfigMap + keys.conf
Secret) onto the pod template as bind.unkin.net/config-hash. When config
changes the hash flips, the template changes, and the StatefulSet does a
normal rolling restart so every pod re-copies fresh config. The operator
owns the template, so the restart is operator-driven and not reverted; a
stable hash means no spurious restarts.
Covers named.conf changes (ACLs, views, forwarders, validate-except,
primary address) and TSIG key rotation.
Secondaries never replicated any member zone: the master's catalog zone
requires key-authenticated AXFR (allow-transfer { key "transfer-key"; }),
but the rendered secondary config transferred without presenting the key,
so every catalog transfer was REFUSED and no member zones provisioned.
Two further gaps compounded it: member zones had no allow-transfer at all,
and secondaries pointed at the primary's pod IP, which dies on restart.
- Render the catalog transfer key into the secondary catalog-zones
default-primaries and the secondary catalog zone primaries, so
key-authenticated AXFR from the primary is accepted.
- Add allow-transfer { key "<transfer-key>"; } to catalog member primary
zones (when the zone does not set an explicit allow-transfer), so
secondaries can pull them; applied to existing zones via modzone.
- Point secondaries at the stable primary Service ClusterIP instead of the
primary pod IP, so replication survives primary pod restarts (falls back
to the pod IP when no primary Service exists).
The vault-plugin-secrets-bind-tsig plugin needs an HTTP endpoint that
creates, reads, rotates and deletes TSIG keys on its behalf, decoupling
Vault from direct Kubernetes API access. This adds that companion API and
lets the operator deploy it declaratively.
- Add BindTSIGAPI CRD: creating one makes the operator reconcile a
Deployment, Service, ConfigMap (env vars), token Secret and namespaced
RBAC for the companion API. Spec covers image, replicas, port,
targetNamespace, tokenSecretName, extra env, service exposure and
resources.
- Generate the master access token Secret only when absent, so a
VaultStaticSecret may pre-seed/overwrite it; the operator does not own it.
- Add the companion API server (internal/tsigapi): bearer-auth HTTP
contract POST /v1/keys, GET/DELETE /v1/keys/{name}, POST
/v1/keys/{name}/rotate, backed by BindTSIGKey custom resources the
operator reconciles into key material.
- Add cmd/tsigapi entrypoint and Dockerfile.tsigapi (distroless).
- Wire the reconciler into setup, regenerate CRDs/RBAC/deepcopy, and add
Woodpecker build (PR dry-run) and release (tag push) steps for the
bind-tsig-api image.
- Cover the API server with auth and key-lifecycle unit tests.
Secondaries reject RFC2136/nsupdate writes, but the read Service round-
robins across all pods. Add an optional per-cluster write endpoint that
targets only the primary pod (ordinal 0) via the StatefulSet pod-name
label. Reads keep using the all-pods Service.
- api: BindCluster.spec.primaryService (*ClusterServiceSpec) — ClusterIP
for in-cluster writers (external-dns) or LoadBalancer for external
- reconcilePrimaryService creates <cluster>-primary selecting pod-0 when
set, deletes it when unset
- regenerate CRDs + install.yaml
Adds BindCluster.spec.service.externalTrafficPolicy so DNS LoadBalancers
can preserve client source IPs (Local), which the source-IP ACLs on the
authoritative/resolver need to actually restrict external clients (Cluster
SNATs everything to node IPs).
- api: ClusterServiceSpec.externalTrafficPolicy (enum Cluster;Local)
- set it on the client Service for LoadBalancer/NodePort types
- regenerate CRDs + install.yaml
type:forward zones are pure configuration (conditional forwarding), not
replicated data, so a resolver needs them on all pods. They were being
rndc-addzone'd on the primary only, so queries hitting a secondary pod
missed the forwarding. Render them into named.conf instead.
- render forward zones inside their view (or top-level when no views)
- BindCluster lists type:forward zones and watches BindZone to re-render
- BindZone controller skips forward zones (config-managed, no addzone)
- unit test for forward-zone-in-view rendering
TSIG keys were included in every cluster's keys.conf namespace-wide. When
multiple clusters share a namespace, that leaks keys across clusters. Add
spec.clusterRef so a key can target a specific BindCluster; empty keeps
the shared (all-clusters-in-namespace) behaviour.
- api: BindTSIGKey.spec.clusterRef (optional)
- BindCluster keys.conf now includes only keys with matching or empty
clusterRef
- regenerate CRDs + install.yaml bundle
Two bugs made every provisioned zone fail to load:
1. The seed zone's apex NS (ns1.<zone>) is in-zone but had no address
record, so BIND check-integrity refused to load it and rndc addzone
reverted. Add a glue A record pointing at the primary pod IP.
2. Secondaries rendered primaries/default-primaries with the primary's
DNS name, but BIND only accepts IP addresses there (it read the name
as a remote-servers list and failed config load, crash-looping the
secondary). Render the primary pod IP instead, and watch Pods so the
config re-renders when that IP appears or changes.
- bind.WriteSeedZone writes 'ns1 IN A <primaryIP>' glue
- controllers resolve primaryPodIP and pass it to the seed (requeue if
the primary has no IP yet)
- BindCluster renders PrimaryAddress from pod-0's IP and watches Pods
- render omits catalog primaries when the IP is unknown (no empty list)
Uses internetsystemsconsortium/bind9 as the default base image instead of
a self-hosted one, verified against internetsystemsconsortium/bind9:9.20
(runs as root; named/rndc/nsupdate at /usr/sbin,/usr/sbin,/usr/bin).
- project operator config at /etc/bind-operator instead of overmounting
the image's /etc/bind (keeps bind.keys / base config intact)
- reference named/rndc/nsupdate by absolute path (exec PATH may exclude
/usr/sbin)
- centralise filesystem + binary paths in internal/bind/consts.go
- default spec.image to internetsystemsconsortium/bind9:9.20
Implements a Kubernetes operator that manages fleets of BIND9 servers
declaratively, using controller-runtime (matching forgebot conventions).
- add BindCluster reconciler: StatefulSet (pod-0 primary, secondaries),
headless + client Services, rendered named.conf ConfigMap, TSIG keys
Secret and rndc control Secret; watches dependent CRs to re-render
- add BindTSIGKey reconciler that generates key material into a Secret
- add BindZone/DNSRecord reconcilers using fully-dynamic delivery
(rndc addzone + TSIG nsupdate against the primary pod)
- add BindCatalogZone reconciler so secondaries auto-provision zones
- add BindPolicy (RPZ), BindDNSSECPolicy, BindView, BindACL reconcilers
- render primary/secondary named.conf variants selected by pod ordinal
- generate CRDs, deepcopy and RBAC; add samples mapping the three Puppet
roles (authoritative/resolver/external-dns) to three BindClusters
- add Makefile, Dockerfile.operator, Woodpecker CI and kind manifests
unkinben