Merge pull request 'Reconcile zone apex NS off the pod IP placeholder' (#22) from benvin/zone-nameservers into main

Reviewed-on: #22

api/v1alpha1/bindzone_types.go

@@ -64,6 +64,17 @@ type BindZoneSpec struct {
 	// +optional
 	DefaultTTL int32 `json:"defaultTTL,omitempty"`
 
+	// Nameservers are the names published in the zone's apex NS RRset, kept in
+	// sync on every reconcile. Each entry is a full domain name, never relative
+	// to the zone. Prefer out-of-zone names glued by the parent: an
+	// in-zone name needs an address record in the zone, and the seed can only
+	// supply the primary pod's (unstable) IP for it. When empty the operator
+	// leaves the apex NS alone and a newly seeded zone gets the primary's stable
+	// in-cluster DNS name; clearing the field later does not retract what it
+	// published.
+	// +optional
+	Nameservers []string `json:"nameservers,omitempty"`
+
 	// Records are static record sets seeded into a primary zone.
 	// +optional
 	Records []Record `json:"records,omitempty"`

api/v1alpha1/zz_generated.deepcopy.go

@@ -970,6 +970,11 @@ func (in *BindZoneList) DeepCopyObject() runtime.Object {
 // DeepCopyInto is an autogenerated deepcopy function, copying the receiver, writing into out. in must be non-nil.
 func (in *BindZoneSpec) DeepCopyInto(out *BindZoneSpec) {
 	*out = *in
+	if in.Nameservers != nil {
+		in, out := &in.Nameservers, &out.Nameservers
+		*out = make([]string, len(*in))
+		copy(*out, *in)
+	}
 	if in.Records != nil {
 		in, out := &in.Records, &out.Records
 		*out = make([]Record, len(*in))

config/crd/bases/bind.unkin.net_bindzones.yaml

@@ -94,6 +94,19 @@ spec:
                 items:
                   type: string
                 type: array
+              nameservers:
+                description: |-
+                  Nameservers are the names published in the zone's apex NS RRset, kept in
+                  sync on every reconcile. Each entry is a full domain name, never relative
+                  to the zone. Prefer out-of-zone names glued by the parent: an
+                  in-zone name needs an address record in the zone, and the seed can only
+                  supply the primary pod's (unstable) IP for it. When empty the operator
+                  leaves the apex NS alone and a newly seeded zone gets the primary's stable
+                  in-cluster DNS name; clearing the field later does not retract what it
+                  published.
+                items:
+                  type: string
+                type: array
               primaries:
                 description: Primaries lists source servers for a secondary/stub-type
                   zone.

config/crd/install.yaml

@@ -2760,6 +2760,19 @@ spec:
                 items:
                   type: string
                 type: array
+              nameservers:
+                description: |-
+                  Nameservers are the names published in the zone's apex NS RRset, kept in
+                  sync on every reconcile. Each entry is a full domain name, never relative
+                  to the zone. Prefer out-of-zone names glued by the parent: an
+                  in-zone name needs an address record in the zone, and the seed can only
+                  supply the primary pod's (unstable) IP for it. When empty the operator
+                  leaves the apex NS alone and a newly seeded zone gets the primary's stable
+                  in-cluster DNS name; clearing the field later does not retract what it
+                  published.
+                items:
+                  type: string
+                type: array
               primaries:
                 description: Primaries lists source servers for a secondary/stub-type
                   zone.

config/samples/01-authoritative.yaml

@@ -71,10 +71,11 @@ spec:
     - key transfer-key
   updateKeyRef: transfer-key
   dynamicUpdate: true
+  # Published apex NS, kept in sync on every reconcile. Full names only; an
+  # in-zone name (as here) needs its address record below.
+  nameservers:
+    - ns1.internal.example.com.
   records:
-    - name: "@"
-      type: NS
-      values: ["ns1.internal.example.com."]
     - name: ns1
       type: A
       values: ["10.0.0.53"]

internal/bind/consts.go

@@ -23,6 +23,7 @@ const (
 	NamedBin    = "/usr/sbin/named"
 	RndcBin     = "/usr/sbin/rndc"
 	NsupdateBin = "/usr/bin/nsupdate"
+	DigBin      = "/usr/bin/dig"
 )
 
 // Config file paths derived from ConfigDir.

internal/bind/nsupdate.go

@@ -19,35 +19,87 @@ type RecordUpdate struct {
 	Type   string   // RR type
 	TTL    int32    // record TTL
 	Values []string // RDATA entries
-	Delete bool     // when true, delete the RRset instead of replacing it
+	Delete bool     // when true, delete instead of add
+	// PerValue operates on individual records rather than the whole RRset: adds
+	// leave existing records in place, deletes remove only the listed Values.
+	// Required at a zone apex, where BIND silently ignores an RRset-wide delete
+	// of NS or SOA and would turn a replace into an append.
+	PerValue bool
 }
 
 // NSUpdate applies a set of record changes to zone by executing nsupdate on the
 // primary pod, targeting the local server and authenticating with creds. All
 // changes are sent in a single atomic transaction.
 func (e *Executor) NSUpdate(ctx context.Context, namespace, pod, zone string, creds TSIGCreds, updates []RecordUpdate) error {
-	var b strings.Builder
-	b.WriteString("server 127.0.0.1\n")
-	b.WriteString(fmt.Sprintf("zone %s\n", dot(zone)))
-	for _, u := range updates {
-		// Replace semantics: clear the RRset first, then add the desired values.
-		b.WriteString(fmt.Sprintf("update delete %s %s\n", dot(u.FQDN), u.Type))
-		if u.Delete {
-			continue
-		}
-		for _, v := range u.Values {
-			b.WriteString(fmt.Sprintf("update add %s %d %s %s\n", dot(u.FQDN), u.TTL, u.Type, v))
-		}
-	}
-	b.WriteString("send\n")
-
 	cmd := []string{NsupdateBin, "-y", fmt.Sprintf("%s:%s:%s", creds.Algorithm, creds.Name, creds.Secret)}
-	if out, err := e.Exec(ctx, namespace, pod, cmd, b.String()); err != nil {
+	if out, err := e.Exec(ctx, namespace, pod, cmd, nsupdateScript(zone, updates)); err != nil {
 		return fmt.Errorf("nsupdate zone %s: %w (out: %s)", zone, err, out)
 	}
 	return nil
 }
 
+// ApexNS returns the zone's currently published apex NS names, so the operator
+// can converge the RRset rather than append to it. The query is TSIG-signed with
+// the same creds as an update: a zone behind a view whose match-clients is a key
+// is unreachable to an unsigned query, which named answers REFUSED (with an empty
+// body and a zero exit status), and the caller must not read that as "no NS".
+func (e *Executor) ApexNS(ctx context.Context, namespace, pod, zone string, creds TSIGCreds) ([]string, error) {
+	cmd := []string{
+		DigBin, "-y", fmt.Sprintf("%s:%s:%s", creds.Algorithm, creds.Name, creds.Secret),
+		"+short", "+time=5", "+tries=1", "@127.0.0.1", dot(zone), "NS",
+	}
+	out, err := e.Exec(ctx, namespace, pod, cmd, "")
+	if err != nil {
+		return nil, fmt.Errorf("query apex NS of %s: %w (out: %s)", zone, err, out)
+	}
+	return parseDigNames(out), nil
+}
+
+// parseDigNames picks the answers out of `dig +short` output: one fully-qualified
+// name per line. Anything without a trailing dot is not a name, and dig prefixes
+// its diagnostics (a missing or mismatched TSIG key among them) with ';'.
+func parseDigNames(out string) []string {
+	var names []string
+	for _, line := range strings.Split(out, "\n") {
+		line = strings.TrimSpace(line)
+		if strings.HasPrefix(line, ";") || !strings.HasSuffix(line, ".") {
+			continue
+		}
+		names = append(names, line)
+	}
+	return names
+}
+
+// nsupdateScript renders the nsupdate input for a set of changes.
+func nsupdateScript(zone string, updates []RecordUpdate) string {
+	var b strings.Builder
+	b.WriteString("server 127.0.0.1\n")
+	fmt.Fprintf(&b, "zone %s\n", dot(zone))
+	for _, u := range updates {
+		switch {
+		case u.PerValue && u.Delete:
+			for _, v := range u.Values {
+				fmt.Fprintf(&b, "update delete %s %s %s\n", dot(u.FQDN), u.Type, v)
+			}
+		case u.PerValue:
+			for _, v := range u.Values {
+				fmt.Fprintf(&b, "update add %s %d %s %s\n", dot(u.FQDN), u.TTL, u.Type, v)
+			}
+		default:
+			// Replace semantics: clear the RRset first, then add the values.
+			fmt.Fprintf(&b, "update delete %s %s\n", dot(u.FQDN), u.Type)
+			if u.Delete {
+				continue
+			}
+			for _, v := range u.Values {
+				fmt.Fprintf(&b, "update add %s %d %s %s\n", dot(u.FQDN), u.TTL, u.Type, v)
+			}
+		}
+	}
+	b.WriteString("send\n")
+	return b.String()
+}
+
 // dot ensures a name is fully qualified with a trailing dot.
 func dot(name string) string {
 	if name == "" || name == "@" {

internal/bind/nsupdate_test.go

@@ -0,0 +1,64 @@
+package bind
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestNSUpdateScriptReplaceSemantics(t *testing.T) {
+	got := nsupdateScript("acme.unkin.net", []RecordUpdate{
+		{FQDN: "www", Type: "A", TTL: 60, Values: []string{"10.0.0.1", "10.0.0.2"}},
+		{FQDN: "old.acme.unkin.net.", Type: "TXT", Delete: true},
+	})
+	want := `server 127.0.0.1
+zone acme.unkin.net.
+update delete www. A
+update add www. 60 A 10.0.0.1
+update add www. 60 A 10.0.0.2
+update delete old.acme.unkin.net. TXT
+send
+`
+	if got != want {
+		t.Errorf("got:\n%s\nwant:\n%s", got, want)
+	}
+}
+
+// At the apex BIND ignores an RRset-wide delete of NS, so the apex sync must add
+// the new names and delete the old ones record by record, adds first: named
+// refuses to leave an apex with no NS record.
+func TestNSUpdateScriptPerValueApexNS(t *testing.T) {
+	got := nsupdateScript("acme.unkin.net", []RecordUpdate{
+		{FQDN: "acme.unkin.net.", Type: "NS", TTL: 60, Values: []string{"acme-ns1.unkin.net."}, PerValue: true},
+		{FQDN: "acme.unkin.net.", Type: "NS", Values: []string{"ns1.acme.unkin.net."}, PerValue: true, Delete: true},
+		{FQDN: "ns1.acme.unkin.net.", Type: "A", Delete: true},
+	})
+	want := `server 127.0.0.1
+zone acme.unkin.net.
+update add acme.unkin.net. 60 NS acme-ns1.unkin.net.
+update delete acme.unkin.net. NS ns1.acme.unkin.net.
+update delete ns1.acme.unkin.net. A
+send
+`
+	if got != want {
+		t.Errorf("got:\n%s\nwant:\n%s", got, want)
+	}
+}
+
+func TestParseDigNames(t *testing.T) {
+	cases := []struct {
+		name, out, want string
+	}{
+		{"answers", "a.ns.unkin.net.\nb.ns.unkin.net.\n", "a.ns.unkin.net.,b.ns.unkin.net."},
+		{"REFUSED, SERVFAIL and NXDOMAIN all answer empty", "", ""},
+		// A zone behind a key-matched view answers an unsigned query REFUSED, and
+		// dig reports the key problem on a ';' line that happens to end in a dot.
+		{"dig diagnostics are not answers", ";; WARNING -- TSIG key was not used.\n", ""},
+		{"relative or partial lines are not names", "10.0.0.1\nns1\n", ""},
+		{"whitespace is trimmed", "  ns1.unkin.net.  \n\n", "ns1.unkin.net."},
+	}
+	for _, c := range cases {
+		if got := strings.Join(parseDigNames(c.out), ","); got != c.want {
+			t.Errorf("%s: parseDigNames(%q) = %q; want %q", c.name, c.out, got, c.want)
+		}
+	}
+}

internal/bind/seed.go

@@ -26,36 +26,65 @@ func (e *Executor) ZoneExists(ctx context.Context, namespace, pod, zone, view st
 	return err == nil
 }
 
-// renderSeedZone renders a minimal loadable zone (SOA + apex NS + glue). The
-// apex NS is the in-zone name ns1, and a glue A record pointing at primaryIP is
-// included so BIND's check-integrity accepts the zone (an in-zone NS without an
-// address record is a load error).
-func renderSeedZone(zone, primaryIP string, serial int64) string {
+// renderSeedZone renders a minimal loadable zone (SOA + apex NS). nameservers
+// are the names published in the apex NS RRset; when empty the in-zone name ns1
+// is used. A glue A pointing at primaryIP is emitted only for a nameserver that
+// falls inside the zone, because BIND refuses to load a zone whose in-zone NS
+// has no address record. Out-of-zone nameservers therefore keep pod IPs out of
+// the zone file entirely.
+func renderSeedZone(zone, primaryIP string, nameservers []string, serial int64) string {
 	origin := dot(zone)
-	ns := "ns1." + origin
+	ns := make([]string, 0, len(nameservers))
+	for _, n := range nameservers {
+		ns = append(ns, dot(n))
+	}
+	if len(ns) == 0 {
+		ns = []string{"ns1." + origin}
+	}
 	// Short refresh/retry so a secondary that misses a NOTIFY (e.g. its pod IP
 	// changed and the primary's also-notify was briefly stale) still converges
 	// in minutes, not the hour a 3600s refresh would impose. minimum is the
 	// negative-cache TTL: keep it low so a stale-secondary NXDOMAIN does not
 	// stick in downstream resolvers for long. NOTIFY (also-notify on the
 	// primary) remains the fast path; these are the fallback.
-	return fmt.Sprintf(`$TTL 3600
+	var b strings.Builder
+	fmt.Fprintf(&b, `$TTL 3600
 @ IN SOA %s hostmaster.%s (
     %d ; serial
     300        ; refresh
     60         ; retry
     1209600    ; expire
     60 )       ; minimum
-@ IN NS %s
-ns1 IN A %s
-`, ns, origin, serial, ns, primaryIP)
+`, ns[0], origin, serial)
+	for _, n := range ns {
+		fmt.Fprintf(&b, "@ IN NS %s\n", n)
+	}
+	for _, n := range ns {
+		if owner, ok := InZoneOwner(n, zone); ok {
+			fmt.Fprintf(&b, "%s IN A %s\n", owner, primaryIP)
+		}
+	}
+	return b.String()
+}
+
+// InZoneOwner reports whether name sits inside zone, and if so returns its owner
+// name relative to the apex ("@" for the apex itself).
+func InZoneOwner(name, zone string) (string, bool) {
+	name, origin := dot(name), dot(zone)
+	switch {
+	case name == origin:
+		return "@", true
+	case strings.HasSuffix(name, "."+origin):
+		return strings.TrimSuffix(name, "."+origin), true
+	}
+	return "", false
 }
 
 // EnsureSeedZone makes path loadable without discarding live data: it probes
 // the zone file and journal, moves aside whatever cannot load, and writes a
 // skeleton only when there is nothing to preserve. Every caller that needs a
 // zone database file on disk goes through here.
-func (e *Executor) EnsureSeedZone(ctx context.Context, namespace, pod, zone, path, primaryIP string) error {
+func (e *Executor) EnsureSeedZone(ctx context.Context, namespace, pod, zone, path, primaryIP string, nameservers []string) error {
 	state, err := e.ZoneDiskState(ctx, namespace, pod, path)
 	if err != nil {
 		return err
@@ -67,7 +96,7 @@ func (e *Executor) EnsureSeedZone(ctx context.Context, namespace, pod, zone, pat
 	if !plan.WriteSeed {
 		return e.Quarantine(ctx, namespace, pod, path, plan)
 	}
-	content := renderSeedZone(zone, primaryIP, plan.Serial)
+	content := renderSeedZone(zone, primaryIP, nameservers, plan.Serial)
 	cmd := []string{"sh", "-c", seedScript(path, plan, len(content))}
 	if out, err := e.Exec(ctx, namespace, pod, cmd, content); err != nil {
 		return fmt.Errorf("seed zone %s: %w (out: %s)", zone, err, out)

internal/bind/seed_nameservers_test.go

@@ -0,0 +1,75 @@
+package bind
+
+import (
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// checkZone runs named-checkzone with full integrity checking, the same check
+// named applies when loading a primary zone. Skips where the tool is absent.
+func checkZone(t *testing.T, zone, content string) {
+	t.Helper()
+	bin, err := exec.LookPath("named-checkzone")
+	if err != nil {
+		t.Skip("named-checkzone not installed")
+	}
+	path := filepath.Join(t.TempDir(), "db")
+	if err := os.WriteFile(path, []byte(content), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if out, err := exec.Command(bin, "-i", "full", zone, path).CombinedOutput(); err != nil {
+		t.Fatalf("zone not loadable: %v\n%s\n%s", err, out, content)
+	}
+}
+
+func TestRenderSeedZoneDeclaredNameservers(t *testing.T) {
+	got := renderSeedZone("acme.unkin.net", "10.42.6.38", []string{"ns1.unkin.net", "ns2.unkin.net."}, 7)
+	if strings.Contains(got, "10.42.6.38") {
+		t.Errorf("declared nameservers must not pull a pod IP into the zone:\n%s", got)
+	}
+	for _, want := range []string{"@ IN NS ns1.unkin.net.\n", "@ IN NS ns2.unkin.net.\n", "SOA ns1.unkin.net. hostmaster.acme.unkin.net."} {
+		if !strings.Contains(got, want) {
+			t.Errorf("missing %q in:\n%s", want, got)
+		}
+	}
+	checkZone(t, "acme.unkin.net", got)
+}
+
+// An in-zone nameserver has no address anywhere else, so the seed must glue it
+// or named refuses to load the zone.
+func TestRenderSeedZoneInZoneNameserverGetsGlue(t *testing.T) {
+	got := renderSeedZone("example.com", "10.0.0.1", []string{"ns.example.com"}, 1)
+	if !strings.Contains(got, "ns IN A 10.0.0.1\n") {
+		t.Errorf("in-zone nameserver needs glue:\n%s", got)
+	}
+	checkZone(t, "example.com", got)
+}
+
+func TestRenderSeedZoneFallbackLoads(t *testing.T) {
+	got := renderSeedZone("example.com", "10.0.0.1", nil, 1)
+	if !strings.Contains(got, "@ IN NS ns1.example.com.\n") || !strings.Contains(got, "ns1 IN A 10.0.0.1\n") {
+		t.Errorf("fallback seed changed shape:\n%s", got)
+	}
+	checkZone(t, "example.com", got)
+}
+
+func TestInZoneOwner(t *testing.T) {
+	cases := []struct {
+		name, zone, owner string
+		in                bool
+	}{
+		{"ns1.example.com.", "example.com", "ns1", true},
+		{"example.com", "example.com.", "@", true},
+		{"ns1.unkin.net", "acme.unkin.net", "", false},
+		{"notexample.com", "example.com", "", false},
+	}
+	for _, c := range cases {
+		owner, in := InZoneOwner(c.name, c.zone)
+		if in != c.in || owner != c.owner {
+			t.Errorf("InZoneOwner(%q, %q) = %q, %v; want %q, %v", c.name, c.zone, owner, in, c.owner, c.in)
+		}
+	}
+}

internal/bind/seed_test.go

@@ -29,7 +29,7 @@ func inFlightZone(t *testing.T) (dir, path string) {
 	t.Helper()
 	dir = t.TempDir()
 	path = filepath.Join(dir, "db.example.com")
-	if err := os.WriteFile(path, []byte(renderSeedZone("example.com", "10.0.0.1", 1)), 0o600); err != nil {
+	if err := os.WriteFile(path, []byte(renderSeedZone("example.com", "10.0.0.1", nil, 1)), 0o600); err != nil {
 		t.Fatal(err)
 	}
 	if err := os.WriteFile(JournalPath(path), journalHeader(";BIND LOG V9.2\n", 10, 16), 0o600); err != nil {
@@ -134,7 +134,7 @@ func TestSeedScriptInterruptedWriteLeavesDiskUntouched(t *testing.T) {
 	if !plan.WriteSeed || !plan.QuarantineZoneFile || !plan.QuarantineJournal {
 		t.Fatalf("expected a reseed over both files, got %+v", plan)
 	}
-	content := renderSeedZone("example.com", "10.0.0.1", plan.Serial)
+	content := renderSeedZone("example.com", "10.0.0.1", nil, plan.Serial)
 
 	if err := runSeedScript(t, sh, path, plan, content, content[:len(content)/2]); err == nil {
 		t.Fatal("a truncated seed write must fail rather than install a torn zone file")
@@ -165,7 +165,7 @@ func TestSeedScriptInstallsOverQuarantinedFiles(t *testing.T) {
 	dir, path := inFlightZone(t)
 
 	plan := PlanSeed(probeState(t, sh, path))
-	content := renderSeedZone("example.com", "10.0.0.1", plan.Serial)
+	content := renderSeedZone("example.com", "10.0.0.1", nil, plan.Serial)
 	if err := runSeedScript(t, sh, path, plan, content, content); err != nil {
 		t.Fatalf("seed script: %v", err)
 	}
@@ -198,7 +198,7 @@ func TestSeedScriptFreshInstall(t *testing.T) {
 	path := filepath.Join(t.TempDir(), "zones", "db.example.com")
 
 	plan := PlanSeed(ZoneDiskState{})
-	content := renderSeedZone("example.com", "10.0.0.1", plan.Serial)
+	content := renderSeedZone("example.com", "10.0.0.1", nil, plan.Serial)
 	if err := runSeedScript(t, sh, path, plan, content, content); err != nil {
 		t.Fatalf("seed script: %v", err)
 	}
@@ -225,7 +225,7 @@ func TestSeedScriptFailedQuarantineAbortsInstall(t *testing.T) {
 exec `+realTool(t, "mv")+` "$@"`)
 
 	plan := PlanSeed(probeState(t, sh, path))
-	content := renderSeedZone("example.com", "10.0.0.1", plan.Serial)
+	content := renderSeedZone("example.com", "10.0.0.1", nil, plan.Serial)
 	if err := runSeedScriptWithPath(t, sh, path, plan, content, content, pathEnv); err == nil {
 		t.Fatal("a failed quarantine must fail the seed")
 	}
@@ -240,7 +240,7 @@ func TestSeedScriptUnmeasurableStagingAbortsInstall(t *testing.T) {
 	pathEnv := shimPath(t, "wc", "exit 127")
 
 	plan := PlanSeed(probeState(t, sh, path))
-	content := renderSeedZone("example.com", "10.0.0.1", plan.Serial)
+	content := renderSeedZone("example.com", "10.0.0.1", nil, plan.Serial)
 	if err := runSeedScriptWithPath(t, sh, path, plan, content, content, pathEnv); err == nil {
 		t.Fatal("an unmeasurable staging file must fail the seed")
 	}
@@ -256,7 +256,7 @@ func TestSeedScriptFailedMkdirAbortsInstall(t *testing.T) {
 	pathEnv := shimPath(t, "mkdir", "exit 1")
 
 	plan := PlanSeed(probeState(t, sh, path))
-	content := renderSeedZone("example.com", "10.0.0.1", plan.Serial)
+	content := renderSeedZone("example.com", "10.0.0.1", nil, plan.Serial)
 	if err := runSeedScriptWithPath(t, sh, path, plan, content, content, pathEnv); err == nil {
 		t.Fatal("a failed mkdir must fail the seed")
 	}

internal/bind/zonestate_test.go

@@ -48,7 +48,7 @@ k8s.syd1.au.unkin.net	IN SOA	ns1.k8s.syd1.au.unkin.net. hostmaster.k8s.syd1.au.u
 		ok      bool
 	}{
 		{"bind dump", bindDump, 16, true},
-		{"seed", renderSeedZone("example.com", "10.0.0.1", 42), 42, true},
+		{"seed", renderSeedZone("example.com", "10.0.0.1", nil, 42), 42, true},
 		{"single line", "@ IN SOA ns1.example.com. hostmaster.example.com. 7 300 60 1209600 60\n", 7, true},
 		{"glued paren", "@ IN SOA ns. host. (9 300 60 1209600 60)\n", 9, true},
 		{"no soa", "$TTL 3600\nwww IN A 192.0.2.1\n", 0, false},
@@ -308,7 +308,7 @@ func TestPlanSeedBlocksOnUnreadableOrphanJournal(t *testing.T) {
 }
 
 func TestSeedZoneRoundTripsThroughParser(t *testing.T) {
-	content := renderSeedZone("200.18.198.in-addr.arpa", "198.18.200.8", 17)
+	content := renderSeedZone("200.18.198.in-addr.arpa", "198.18.200.8", nil, 17)
 	got, ok := ParseZoneSerial(content)
 	if !ok || got != 17 {
 		t.Fatalf("seed zone serial = (%d,%v) want (17,true)", got, ok)
@@ -418,12 +418,12 @@ func TestZoneStateProbeRoundTrip(t *testing.T) {
 		{name: "fresh install"},
 		{
 			name: "zone file only",
-			zone: renderSeedZone("example.com", "10.0.0.1", 42),
+			zone: renderSeedZone("example.com", "10.0.0.1", nil, 42),
 			want: ZoneDiskState{ZoneFile: true, ZoneSerial: 42, ZoneSerialOK: true},
 		},
 		{
 			name: "zone file and journal",
-			zone: renderSeedZone("example.com", "10.0.0.1", 12),
+			zone: renderSeedZone("example.com", "10.0.0.1", nil, 12),
 			jnl:  journalHeader(";BIND LOG V9.2\n", 10, 16),
 			want: ZoneDiskState{
 				ZoneFile: true, ZoneSerial: 12, ZoneSerialOK: true,
@@ -452,7 +452,7 @@ func TestZoneStateProbeRoundTrip(t *testing.T) {
 		},
 		{
 			name:    "live zone beside old quarantine evidence",
-			zone:    renderSeedZone("example.com", "10.0.0.1", 42),
+			zone:    renderSeedZone("example.com", "10.0.0.1", nil, 42),
 			orphans: []string{".orphaned-16"},
 			want: ZoneDiskState{
 				ZoneFile: true, ZoneSerial: 42, ZoneSerialOK: true,

internal/controller/apex_ns_test.go

@@ -0,0 +1,165 @@
+package controller
+
+import (
+	"strings"
+	"testing"
+
+	bindv1alpha1 "git.unkin.net/unkin/bind-operator/api/v1alpha1"
+	"git.unkin.net/unkin/bind-operator/internal/bind"
+)
+
+func zoneWith(spec bindv1alpha1.BindZoneSpec) *bindv1alpha1.BindZone {
+	spec.ZoneName = "acme.unkin.net"
+	return &bindv1alpha1.BindZone{Spec: spec}
+}
+
+func testCluster() *bindv1alpha1.BindCluster {
+	c := &bindv1alpha1.BindCluster{}
+	c.Name, c.Namespace = "auth", "bind-internal"
+	return c
+}
+
+const stableNS = "auth-0.auth-headless.bind-internal.svc.cluster.local."
+
+func TestZoneNameservers(t *testing.T) {
+	ttl60 := int32(60)
+	cases := []struct {
+		name     string
+		spec     bindv1alpha1.BindZoneSpec
+		want     []string
+		ttl      int32
+		declared bool
+	}{
+		{"fallback is out-of-zone, so it needs no glue", bindv1alpha1.BindZoneSpec{}, []string{stableNS}, 3600, false},
+		{"declared wins", bindv1alpha1.BindZoneSpec{Nameservers: []string{"ns1.unkin.net"}}, []string{"ns1.unkin.net."}, 3600, true},
+		{"spec.defaultTTL applies", bindv1alpha1.BindZoneSpec{Nameservers: []string{"ns1.unkin.net."}, DefaultTTL: 60}, []string{"ns1.unkin.net."}, 60, true},
+		// An apex NS in spec.records cannot converge on its own: BIND ignores an
+		// RRset-wide delete at the apex, so it has to go through the apex path.
+		{"apex NS in records counts as declared", bindv1alpha1.BindZoneSpec{Records: []bindv1alpha1.Record{
+			{Name: "@", Type: "ns", Values: []string{"a.ns.unkin.net.", "b.ns.unkin.net."}},
+			{Name: "www", Type: "A", Values: []string{"10.0.0.1"}},
+		}}, []string{"a.ns.unkin.net.", "b.ns.unkin.net."}, 3600, true},
+		// A TTL on the apex NS record itself must survive the fold.
+		{"record TTL beats the zone default", bindv1alpha1.BindZoneSpec{DefaultTTL: 3600, Records: []bindv1alpha1.Record{
+			{Name: "@", Type: "NS", TTL: &ttl60, Values: []string{"a.ns.unkin.net."}},
+		}}, []string{"a.ns.unkin.net."}, 60, true},
+		{"spec.nameservers beats records", bindv1alpha1.BindZoneSpec{
+			Nameservers: []string{"ns1.unkin.net."},
+			Records:     []bindv1alpha1.Record{{Name: "@", Type: "NS", Values: []string{"other.unkin.net."}}},
+		}, []string{"ns1.unkin.net."}, 3600, true},
+	}
+	for _, c := range cases {
+		got, ttl, declared := zoneNameservers(zoneWith(c.spec), testCluster())
+		if declared != c.declared || ttl != c.ttl || strings.Join(got, ",") != strings.Join(c.want, ",") {
+			t.Errorf("%s: got %v/%d/%v; want %v/%d/%v", c.name, got, ttl, declared, c.want, c.ttl, c.declared)
+		}
+	}
+}
+
+// A query that cannot see the zone returns nothing, which must not be read as an
+// empty apex: retracting blind means deleting the last NS record, which named
+// rejects, leaving the zone stuck.
+func TestApexNSUpdatesUnreadableLiveSetIsAdditive(t *testing.T) {
+	zone := zoneWith(bindv1alpha1.BindZoneSpec{Nameservers: []string{"ns1.unkin.net."}})
+	got := apexNSUpdates(zone, []string{"ns1.unkin.net."}, nil, 3600)
+	assertUpdates(t, got, []bind.RecordUpdate{
+		{FQDN: "acme.unkin.net.", Type: "NS", TTL: 3600, Values: []string{"ns1.unkin.net."}, PerValue: true},
+	})
+}
+
+// The apex NS RRset must be converged per record: an RRset-wide delete at the
+// apex is ignored by BIND, which would leave the placeholder published alongside
+// the real nameservers. Retracting an in-zone name takes its glue with it.
+func TestApexNSUpdatesConverges(t *testing.T) {
+	zone := zoneWith(bindv1alpha1.BindZoneSpec{Nameservers: []string{"ns1.unkin.net."}, DefaultTTL: 60})
+	got := apexNSUpdates(zone, []string{"ns1.unkin.net."}, []string{"ns1.acme.unkin.net."}, 60)
+	assertUpdates(t, got, []bind.RecordUpdate{
+		{FQDN: "acme.unkin.net.", Type: "NS", TTL: 60, Values: []string{"ns1.unkin.net."}, PerValue: true},
+		{FQDN: "acme.unkin.net.", Type: "NS", Values: []string{"ns1.acme.unkin.net."}, PerValue: true, Delete: true},
+		{FQDN: "ns1.acme.unkin.net.", Type: "A", Delete: true},
+	})
+}
+
+// Glue retirement is not special-cased to the name ns1: the seed glues every
+// declared in-zone nameserver.
+func TestApexNSUpdatesRetiresAnyInZoneGlue(t *testing.T) {
+	zone := zoneWith(bindv1alpha1.BindZoneSpec{Nameservers: []string{"a.ns.unkin.net."}})
+	got := apexNSUpdates(zone, []string{"a.ns.unkin.net."}, []string{"dns.acme.unkin.net."}, 3600)
+	assertUpdates(t, got, []bind.RecordUpdate{
+		{FQDN: "acme.unkin.net.", Type: "NS", TTL: 3600, Values: []string{"a.ns.unkin.net."}, PerValue: true},
+		{FQDN: "acme.unkin.net.", Type: "NS", Values: []string{"dns.acme.unkin.net."}, PerValue: true, Delete: true},
+		{FQDN: "dns.acme.unkin.net.", Type: "A", Delete: true},
+	})
+}
+
+func TestApexNSUpdatesNoopWhenConverged(t *testing.T) {
+	zone := zoneWith(bindv1alpha1.BindZoneSpec{Nameservers: []string{"ns1.unkin.net"}})
+	// Case differs: DNS names compare case-insensitively, so this is converged.
+	if got := apexNSUpdates(zone, []string{"ns1.unkin.net."}, []string{"NS1.Unkin.Net."}, 3600); len(got) != 0 {
+		t.Fatalf("expected no updates, got %+v", got)
+	}
+}
+
+// Changing the declared set replaces only what changed, keeping the overlap.
+func TestApexNSUpdatesPartialChange(t *testing.T) {
+	zone := zoneWith(bindv1alpha1.BindZoneSpec{Nameservers: []string{"a.ns.unkin.net.", "c.ns.unkin.net."}})
+	got := apexNSUpdates(zone, []string{"a.ns.unkin.net.", "c.ns.unkin.net."}, []string{"a.ns.unkin.net.", "b.ns.unkin.net."}, 3600)
+	assertUpdates(t, got, []bind.RecordUpdate{
+		{FQDN: "acme.unkin.net.", Type: "NS", TTL: 3600, Values: []string{"c.ns.unkin.net."}, PerValue: true},
+		{FQDN: "acme.unkin.net.", Type: "NS", Values: []string{"b.ns.unkin.net."}, PerValue: true, Delete: true},
+	})
+}
+
+// A declared in-zone nameserver owns its glue; removing it would fail named's
+// post-update nameserver sanity check.
+func TestApexNSUpdatesKeepsNeededGlue(t *testing.T) {
+	zone := zoneWith(bindv1alpha1.BindZoneSpec{Nameservers: []string{"ns1.acme.unkin.net."}})
+	if got := apexNSUpdates(zone, []string{"ns1.acme.unkin.net."}, []string{"ns1.acme.unkin.net."}, 3600); len(got) != 0 {
+		t.Fatalf("expected no updates, got %+v", got)
+	}
+}
+
+// spec.records owning the ns1 address means the glue is real data, not the seed
+// placeholder.
+func TestApexNSUpdatesLeavesRecordOwnedGlue(t *testing.T) {
+	zone := zoneWith(bindv1alpha1.BindZoneSpec{
+		Nameservers: []string{"ns1.unkin.net."},
+		Records:     []bindv1alpha1.Record{{Name: "ns1", Type: "a", Values: []string{"10.0.0.53"}}},
+	})
+	got := apexNSUpdates(zone, []string{"ns1.unkin.net."}, []string{"ns1.acme.unkin.net."}, 3600)
+	assertUpdates(t, got, []bind.RecordUpdate{
+		{FQDN: "acme.unkin.net.", Type: "NS", TTL: 3600, Values: []string{"ns1.unkin.net."}, PerValue: true},
+		{FQDN: "acme.unkin.net.", Type: "NS", Values: []string{"ns1.acme.unkin.net."}, PerValue: true, Delete: true},
+	})
+}
+
+// The apex NS is converged by the apex path, so it must not also be emitted as a
+// record: an RRset-wide delete there is ignored and the add would append.
+func TestRecordsToUpdatesSkipsApexNS(t *testing.T) {
+	records := []bindv1alpha1.Record{
+		{Name: "@", Type: "NS", Values: []string{"a.ns.unkin.net."}},
+		{Name: "sub", Type: "NS", Values: []string{"d.ns.unkin.net."}},
+		{Name: "@", Type: "MX", Values: []string{"10 mail.unkin.net."}},
+	}
+	got := recordsToUpdates("acme.unkin.net", records, 3600)
+	if len(got) != 2 || got[0].FQDN != "sub.acme.unkin.net." || got[1].Type != "MX" {
+		t.Fatalf("got %+v", got)
+	}
+}
+
+func assertUpdates(t *testing.T, got, want []bind.RecordUpdate) {
+	t.Helper()
+	if len(got) != len(want) {
+		t.Fatalf("got %d updates %+v; want %d %+v", len(got), got, len(want), want)
+	}
+	for i := range want {
+		g, w := got[i], want[i]
+		if g.FQDN != w.FQDN || g.Type != w.Type || g.TTL != w.TTL || g.Delete != w.Delete || g.PerValue != w.PerValue {
+			t.Errorf("update %d = %+v; want %+v", i, g, w)
+			continue
+		}
+		if strings.Join(g.Values, ",") != strings.Join(w.Values, ",") {
+			t.Errorf("update %d values = %v; want %v", i, g.Values, w.Values)
+		}
+	}
+}

internal/controller/bindcatalogzone_controller.go

@@ -54,7 +54,7 @@ func (r *BindCatalogZoneReconciler) Reconcile(ctx context.Context, req ctrl.Requ
 		if primaryIP == "" {
 			return r.fail(ctx, &catalog, "PrimaryNoIP", "waiting for primary pod IP")
 		}
-		if err := r.Exec.EnsureSeedZone(ctx, catalog.Namespace, primaryPod, catalog.Spec.ZoneName, bind.CatalogFilePath(catalog.Spec.ZoneName), primaryIP); err != nil {
+		if err := r.Exec.EnsureSeedZone(ctx, catalog.Namespace, primaryPod, catalog.Spec.ZoneName, bind.CatalogFilePath(catalog.Spec.ZoneName), primaryIP, clusterNameservers(cluster)); err != nil {
 			return r.fail(ctx, &catalog, "SeedFailed", err.Error())
 		}
 	}

internal/controller/bindpolicy_controller.go

@@ -67,7 +67,7 @@ func (r *BindPolicyReconciler) Reconcile(ctx context.Context, req ctrl.Request)
 		if primaryIP == "" {
 			return r.fail(ctx, &policy, "PrimaryNoIP", "waiting for primary pod IP")
 		}
-		if err := r.Exec.EnsureSeedZone(ctx, policy.Namespace, primaryPod, policy.Spec.ZoneName, bind.ZoneFilePath(policy.Spec.ZoneName), primaryIP); err != nil {
+		if err := r.Exec.EnsureSeedZone(ctx, policy.Namespace, primaryPod, policy.Spec.ZoneName, bind.ZoneFilePath(policy.Spec.ZoneName), primaryIP, clusterNameservers(cluster)); err != nil {
 			return r.fail(ctx, &policy, "SeedFailed", err.Error())
 		}
 	}

internal/controller/bindzone_controller.go

@@ -99,6 +99,8 @@ func (r *BindZoneReconciler) Reconcile(ctx context.Context, req ctrl.Request) (c
 		return r.setPhase(ctx, &zone, "Error", "ConfigError", err.Error())
 	}
 
+	nameservers, nsTTL, nsDeclared := zoneNameservers(&zone, cluster)
+
 	created := !r.Exec.ZoneExists(ctx, zone.Namespace, primaryPod, zone.Spec.ZoneName, zone.Spec.ViewRef)
 	if created && (zone.Spec.Type == bindv1alpha1.ZonePrimary || zone.Spec.Type == "") {
 		primaryIP := primaryPodIP(ctx, r.Client, cluster)
@@ -108,7 +110,7 @@ func (r *BindZoneReconciler) Reconcile(ctx context.Context, req ctrl.Request) (c
 		// The zone is absent from named's memory, but its database file and
 		// journal may still be on the PVC from a previous incarnation.
 		path := bind.ZoneFilePath(zone.Spec.ZoneName)
-		if err := r.Exec.EnsureSeedZone(ctx, zone.Namespace, primaryPod, zone.Spec.ZoneName, path, primaryIP); err != nil {
+		if err := r.Exec.EnsureSeedZone(ctx, zone.Namespace, primaryPod, zone.Spec.ZoneName, path, primaryIP, nameservers); err != nil {
 			return r.setPhase(ctx, &zone, "Error", "SeedFailed", err.Error())
 		}
 	}
@@ -116,18 +118,42 @@ func (r *BindZoneReconciler) Reconcile(ctx context.Context, req ctrl.Request) (c
 		return r.setPhase(ctx, &zone, "Error", "AddZoneFailed", err.Error())
 	}
 
-	// Seed static records (primary zones only).
+	// Records are applied before the apex NS: an in-zone nameserver's address
+	// record has to exist first, or named rejects the apex transaction with a
+	// post-update nameserver sanity check failure.
 	recordCount := 0
-	if isPrimaryType(zone.Spec.Type) && len(zone.Spec.Records) > 0 {
+	records := recordsToUpdates(zone.Spec.ZoneName, zone.Spec.Records, zone.Spec.DefaultTTL)
+	if isPrimaryType(zone.Spec.Type) && (len(records) > 0 || nsDeclared) {
 		creds, err := r.zoneUpdateCreds(ctx, &zone)
 		if err != nil {
 			return r.setPhase(ctx, &zone, "Error", "NoUpdateKey", err.Error())
 		}
-		updates := recordsToUpdates(zone.Spec.ZoneName, zone.Spec.Records, zone.Spec.DefaultTTL)
-		if err := r.Exec.NSUpdate(ctx, zone.Namespace, primaryPod, zone.Spec.ZoneName, creds, updates); err != nil {
-			return r.setPhase(ctx, &zone, "Error", "RecordUpdateFailed", err.Error())
+		if len(records) > 0 {
+			if err := r.Exec.NSUpdate(ctx, zone.Namespace, primaryPod, zone.Spec.ZoneName, creds, records); err != nil {
+				return r.setPhase(ctx, &zone, "Error", "RecordUpdateFailed", err.Error())
+			}
+			recordCount = len(records)
+		}
+		// Converge the apex NS on every pass, not only at seed time, so a zone
+		// seeded with the placeholder moves onto its real nameservers. Only for a
+		// zone that declared them: otherwise the operator would fight whoever else
+		// manages the RRset.
+		if nsDeclared {
+			live, err := r.Exec.ApexNS(ctx, zone.Namespace, primaryPod, zone.Spec.ZoneName, creds)
+			if err != nil {
+				return r.setPhase(ctx, &zone, "Error", "ApexNSQueryFailed", err.Error())
+			}
+			if len(live) == 0 {
+				logger.Info("apex NS not readable, publishing without retracting", "zone", zone.Spec.ZoneName)
+			}
+			if apex := apexNSUpdates(&zone, nameservers, live, nsTTL); len(apex) > 0 {
+				if err := r.Exec.NSUpdate(ctx, zone.Namespace, primaryPod, zone.Spec.ZoneName, creds, apex); err != nil {
+					return r.setPhase(ctx, &zone, "Error", "ApexNSSyncFailed",
+						fmt.Sprintf("%s (a nameserver inside the zone needs an address record here)", err))
+				}
+				logger.Info("apex NS converged", "zone", zone.Spec.ZoneName, "nameservers", nameservers)
+			}
 		}
-		recordCount = len(updates)
 	}
 
 	// Register in the catalog so secondaries auto-provision.

internal/controller/zone_helpers.go

@@ -42,6 +42,11 @@ func fqdn(name, zone string) string {
 func recordsToUpdates(zone string, records []bindv1alpha1.Record, defaultTTL int32) []bind.RecordUpdate {
 	updates := make([]bind.RecordUpdate, 0, len(records))
 	for _, rec := range records {
+		// The apex NS RRset is converged by apexNSUpdates: an RRset-wide delete
+		// here is ignored by BIND and would only append to the live set.
+		if strings.EqualFold(rec.Type, "NS") && fqdn(rec.Name, zone) == fqdn("@", zone) {
+			continue
+		}
 		ttl := defaultTTL
 		if rec.TTL != nil {
 			ttl = *rec.TTL
@@ -114,3 +119,119 @@ func alsoNotifyList(addrs []string, key string) string {
 	}
 	return strings.Join(parts, " ")
 }
+
+// absolute qualifies a nameserver name. Unlike record owner names, a
+// spec.nameservers entry is always a full domain name, never relative to the
+// zone: an in-zone nameserver is spelled out in full.
+func absolute(name string) string { return strings.TrimSuffix(name, ".") + "." }
+
+// zoneNameservers resolves the names to publish in a zone's apex NS RRset, the
+// TTL to publish them with, and whether the zone declared them. An apex NS in
+// spec.records counts as a declaration: BIND ignores an RRset-wide delete at the
+// apex, so records alone can only append to what the zone was seeded with, never
+// replace it. Undeclared zones fall back to the primary's stable in-cluster name,
+// which is deliberately out-of-zone so no pod IP is needed as glue.
+func zoneNameservers(zone *bindv1alpha1.BindZone, cluster *bindv1alpha1.BindCluster) (names []string, ttl int32, declared bool) {
+	ttl = zone.Spec.DefaultTTL
+	for _, ns := range zone.Spec.Nameservers {
+		names = append(names, absolute(ns))
+	}
+	for _, rec := range zone.Spec.Records {
+		if len(names) > 0 {
+			break
+		}
+		if !strings.EqualFold(rec.Type, "NS") || fqdn(rec.Name, zone.Spec.ZoneName) != fqdn("@", zone.Spec.ZoneName) {
+			continue
+		}
+		for _, v := range rec.Values {
+			names = append(names, absolute(v))
+		}
+		if rec.TTL != nil {
+			ttl = *rec.TTL
+		}
+	}
+	if ttl <= 0 {
+		ttl = 3600
+	}
+	if len(names) > 0 {
+		return names, ttl, true
+	}
+	return clusterNameservers(cluster), ttl, false
+}
+
+// apexNSUpdates moves a zone's apex NS RRset from live onto desired, and retires
+// the glue of any in-zone name it retracts. Adds come first: BIND refuses to
+// leave an apex with no NS record, so the replacement must exist before the old
+// name goes, and deleting glue still referenced by an in-zone NS fails named's
+// post-update nameserver sanity check.
+//
+// An empty live set means the query could not see the zone, not that the apex has
+// no NS records — a primary always has one. Nothing is retracted in that case:
+// retracting blind is what turns a delete into "delete the last NS", which named
+// rejects outright.
+// ponytail: names already published are diffed by name only, so an edit to just
+// the TTL never republishes them (dig +short cannot report a TTL). Re-add the
+// whole desired set each pass if TTL edits need to converge.
+func apexNSUpdates(zone *bindv1alpha1.BindZone, desired, live []string, ttl int32) []bind.RecordUpdate {
+	apex := fqdn("@", zone.Spec.ZoneName)
+	add := missing(desired, live)
+
+	var updates []bind.RecordUpdate
+	if len(add) > 0 {
+		updates = append(updates, bind.RecordUpdate{FQDN: apex, Type: "NS", TTL: ttl, Values: add, PerValue: true})
+	}
+	// missing() yields nothing against an empty live set, so an unreadable RRset
+	// retracts nothing on its own.
+	del := missing(live, desired)
+	if len(del) == 0 {
+		return updates
+	}
+	updates = append(updates, bind.RecordUpdate{FQDN: apex, Type: "NS", Values: del, PerValue: true, Delete: true})
+	// The seed glues an in-zone nameserver to the primary pod's IP, which goes
+	// stale on the first reschedule. Drop that address with the name, unless
+	// spec.records owns it (then it is real data, not the placeholder).
+	for _, ns := range del {
+		owner, in := bind.InZoneOwner(ns, zone.Spec.ZoneName)
+		if in && owner != "@" && !recordsOwn(zone, owner, "A") {
+			updates = append(updates, bind.RecordUpdate{FQDN: fqdn(owner, zone.Spec.ZoneName), Type: "A", Delete: true})
+		}
+	}
+	return updates
+}
+
+// clusterNameservers is the apex NS for the operator's own internal zones
+// (catalog, policy): the primary's stable in-cluster name, never a pod IP.
+func clusterNameservers(cluster *bindv1alpha1.BindCluster) []string {
+	return []string{primaryAddress(cluster.Name, cluster.Namespace) + "."}
+}
+
+// missing returns the names in want with no match in have. DNS names compare
+// case-insensitively.
+func missing(want, have []string) (out []string) {
+	for _, w := range want {
+		if !containsName(have, w) {
+			out = append(out, w)
+		}
+	}
+	return out
+}
+
+func containsName(names []string, name string) bool {
+	for _, n := range names {
+		if strings.EqualFold(absolute(n), absolute(name)) {
+			return true
+		}
+	}
+	return false
+}
+
+// recordsOwn reports whether spec.records already manages an owner/type pair, in
+// which case the apex sync must leave it alone.
+func recordsOwn(zone *bindv1alpha1.BindZone, name, typ string) bool {
+	for _, rec := range zone.Spec.Records {
+		if strings.EqualFold(rec.Type, typ) && strings.EqualFold(fqdn(rec.Name, zone.Spec.ZoneName), fqdn(name, zone.Spec.ZoneName)) {
+			return true
+		}
+	}
+	return false
+}