backup: allow live Longhorn policies and enforce exclusions

internal/server/backup_handlers.go

@@ -206,10 +206,19 @@ func (s *Server) executeBackup(ctx context.Context, req api.BackupRequest, reque
 
 	switch s.cfg.BackupDriver {
 	case "longhorn":
-		volumeName, _, _, err := s.client.ResolvePVCVolume(ctx, req.Namespace, req.PVC)
+		volumeName, pvc, _, err := s.client.ResolvePVCVolume(ctx, req.Namespace, req.PVC)
 		if err != nil {
 			return api.BackupResponse{}, "validation_error", err
 		}
+		storageClass := ""
+		if pvc != nil && pvc.Spec.StorageClassName != nil {
+			storageClass = *pvc.Spec.StorageClassName
+		}
+		// Apply data-location exclusions to manual and namespace requests as
+		// well as scheduled backups, before invoking the Longhorn backend.
+		if excluded, reason := s.pvcExcluded(req.Namespace, req.PVC, storageClass); excluded {
+			return api.BackupResponse{}, "validation_error", errors.New(reason)
+		}
 
 		backupID := backupName("backup", req.Namespace+"-"+req.PVC)
 		response := api.BackupResponse{

internal/server/longhorn_live_policy_test.go

@@ -0,0 +1,54 @@
+package server
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"scm.bstein.dev/bstein/soteria/internal/api"
+	"scm.bstein.dev/bstein/soteria/internal/config"
+	"scm.bstein.dev/bstein/soteria/internal/k8s"
+)
+
+// TestLiveRWOLonghornPolicy protects live workloads without a second mount.
+func TestLiveRWOLonghornPolicy(t *testing.T) {
+	client := &policyCycleTestKubeClient{
+		inventoryTestKubeClient: &inventoryTestKubeClient{
+			fakeKubeClient: &fakeKubeClient{
+				pvcs:      []k8s.PVCSummary{{Namespace: "apps", Name: "data", VolumeName: "vol-data", Phase: "Bound", AccessModes: []string{"ReadWriteOnce"}}},
+				pvcMounts: map[string][]k8s.PVCMount{"apps/data": {{PodName: "database-0", NodeName: "worker", Phase: "Running"}}},
+			},
+		},
+	}
+	backend := &fakeLonghornClient{}
+	srv := &Server{
+		cfg:    &config.Config{BackupDriver: "longhorn", BackupMaxAge: 24 * time.Hour, PolicyBackupsPerCycle: 1},
+		client: client, longhorn: backend, metrics: newTelemetry(),
+		policies: map[string]api.BackupPolicy{"apps__all": {ID: "apps__all", Namespace: "apps", IntervalHours: 1, Enabled: true, Dedupe: true}},
+	}
+	srv.runPolicyCycle(context.Background())
+	if backend.createSnapshotName == "" {
+		t.Fatal("live RWO workload did not receive a Longhorn snapshot")
+	}
+	if len(client.backupRequests) != 0 {
+		t.Fatal("Longhorn policy attempted to launch a filesystem-mount job")
+	}
+	if metricCount(srv.metrics.policyBackups, map[string]string{"result": "success"}) != 1 {
+		t.Fatal("expected one successful policy backup")
+	}
+}
+
+// TestLonghornExclusionAppliesToManualRequests prevents bypassing local-only policy.
+func TestLonghornExclusionAppliesToManualRequests(t *testing.T) {
+	backend := &fakeLonghornClient{}
+	srv := &Server{
+		cfg:    &config.Config{BackupDriver: "longhorn", ExcludedPVCs: []string{"hermes/*"}},
+		client: &fakeKubeClient{}, longhorn: backend,
+	}
+	for _, dryRun := range []bool{false, true} {
+		_, code, err := srv.executeBackup(context.Background(), api.BackupRequest{Namespace: "hermes", PVC: "workspace", DryRun: dryRun}, "test")
+		if err == nil || code != "validation_error" || backend.createSnapshotName != "" {
+			t.Fatal("excluded Longhorn PVC reached the backup backend")
+		}
+	}
+}

internal/server/policy_runtime.go

@@ -103,15 +103,19 @@ func (s *Server) runPolicyCycle(ctx context.Context) {
 			s.metrics.RecordPolicyBackup("excluded")
 			continue
 		}
-		blocked, _, err := s.liveExclusivePVCMounted(runCtx, pvc.Namespace, pvc.PVC, pvc.AccessModes)
+		// Only restic needs a second filesystem mount. Longhorn snapshots use
-		if err != nil {
+		// the existing engine and must protect live RWO workloads too.
-			log.Printf("policy cycle live PVC mount lookup failed for %s/%s: %v", pvc.Namespace, pvc.PVC, err)
+		if s.cfg.BackupDriver == "restic" {
-			s.metrics.RecordPolicyBackup("active_lookup_error")
+			blocked, _, err := s.liveExclusivePVCMounted(runCtx, pvc.Namespace, pvc.PVC, pvc.AccessModes)
-			continue
+			if err != nil {
-		}
+				log.Printf("policy cycle live PVC mount lookup failed for %s/%s: %v", pvc.Namespace, pvc.PVC, err)
-		if blocked {
+				s.metrics.RecordPolicyBackup("active_lookup_error")
-			s.metrics.RecordPolicyBackup("live_rwo_mount")
+				continue
-			continue
+			}
+			if blocked {
+				s.metrics.RecordPolicyBackup("live_rwo_mount")
+				continue
+			}
 		}
 		// Never enqueue a new policy backup while one is already active for this PVC.
 		// This prevents runaway job storms when a backup is stuck Pending/Running.