kubernetes/infrastructure/netbird-ingress/daemonset.yaml

@@ -38,6 +38,16 @@ spec:
     rollingUpdate:
       # roll one node's peer at a time; the others keep serving the route.
       maxUnavailable: 1
+  # Pace the roll to NetBird's route re-convergence. A rolled peer is replaced
+  # by a freshly-enrolled peer (ephemeral state below) that must join the
+  # k8s-ingress-router route group and propagate to mesh clients before it
+  # actually forwards. Without this, the roll steps to the next node as soon as
+  # `netbird status --check ready` passes (seconds) — faster than the mesh
+  # re-converges — so the HA route runs degraded for the whole roll and public
+  # ingress (home.renner.uno) drops for ~1 min (observed 2026-07-11). Hold each
+  # new peer Ready this long before rolling the next, so >=2 peers are always
+  # actually serving the route.
+  minReadySeconds: 45
   template:
     metadata:
       labels:
@@ -86,6 +96,17 @@ spec:
               memory: 64Mi
             limits:
               memory: 256Mi
+          # Graceful drain: withdraw this peer from the mesh BEFORE the process
+          # is killed. Without a preStop the pod is only SIGTERM'd, so the mesh
+          # keeps the (now-dead) peer in the HA route group until a keepalive
+          # timeout (~25-30s) — traffic riding through it black-holes until
+          # then. `netbird down` disconnects the peer from the management server
+          # so it fails the route over to the other per-node peers immediately.
+          # Runs within terminationGracePeriodSeconds (default 30s).
+          lifecycle:
+            preStop:
+              exec:
+                command: ["netbird", "down"]
           livenessProbe:
             exec:
               command: ["netbird", "status", "--check", "live"]