fix(probes): skip probes for mirror and standby pods

internal/controller/databasecluster_probes.go

@@ -27,18 +27,27 @@ import (
 // addDatabaseProbes attaches startup, readiness, and liveness probes
 // to the database container.
 //
-// The probe is a single pg_isready check against the role's local port.
-// Port discovery:
+// Probe strategy by role:
 //
-//	coordinator, standby → fixed 5432
-//	segment              → 6000 + ordinal (derived from $HOSTNAME)
-//	mirror               → 7000 + ordinal (derived from $HOSTNAME)
+//	coordinator, segment → pg_isready against the local port (5432 / 6000).
+//	                       These roles accept normal SQL connections, so
+//	                       pg_isready returns "accepting connections" once
+//	                       postgres is up.
+//	mirror, standby      → NO probes. These roles run postgres in continuous
+//	                       recovery mode (replaying WAL from a primary) and
+//	                       deliberately do NOT accept normal connections.
+//	                       pg_isready returns non-zero against them, which
+//	                       triggers restart loops in Kubernetes. Mirror and
+//	                       standby health is detected by the operator via
+//	                       gp_segment_configuration (status='u', mode='s'),
+//	                       so K8s-level probes are both unnecessary and
+//	                       actively harmful here.
 //
-// 3-probe layout:
+// 3-probe layout for coordinator/segment:
 //
 //	StartupProbe   — generous boot window (~10 min) to absorb gpinitsystem
 //	                 on coord-0 and the wait-for-coordinator phase on
-//	                 segments/mirrors. Once it succeeds, readiness + liveness
+//	                 segments. Once it succeeds, readiness + liveness
 //	                 take over.
 //	ReadinessProbe — tight. Pod only appears in Service endpoints when
 //	                 postgres truly accepts connections.
@@ -48,14 +57,18 @@ import (
 // All probes pass cleanly through StatefulSet rollouts and operator restarts
 // because they read state from the pod itself, not from operator status.
 func addDatabaseProbes(c *corev1.Container, suffix string, image *keldoniov1alpha1.DatabaseImage) {
+	// Skip probes for replicas in recovery mode. See doc comment above.
+	if suffix == "mirror" || suffix == "mirrors" ||
+		suffix == "standby" {
+		return
+	}
+
 	var portExpr string
 	switch suffix {
-	case "coordinator", "standby":
+	case "coordinator":
 		portExpr = "5432"
 	case "segment", "segments":
 		portExpr = "6000"
-	case "mirror", "mirrors":
-		portExpr = "7000"
 	default:
 		portExpr = "5432"
 	}