kubernetes/infrastructure/coredns-config/configmap.yaml

@@ -15,9 +15,23 @@ data:
     # (The old Proxmox DNS VIP 10.2.40.53 is DECOMMISSIONED.) 10.96.0.11 is a
     # pinned ClusterIP of the k8s-gateway Service — avoids DNS-name
     # resolution during CoreDNS startup.
+    # Cache tuning mirrors k8s-gateway's (issue #393): the UDM's dnsmasq
+    # refuses connections in bursts, and because the public renner.uno zone
+    # has a *.renner.uno wildcard CNAME that matches at every depth, a local
+    # failure never returns NXDOMAIN — it resolves to the WAN proxy instead.
+    # Correct answers carry TTL 5 (dnsmasq lease records) while the wrong
+    # wildcard answer carries TTL 300, so one brief blip pins a wrong answer
+    # for minutes. `success ... 60` raises the 5s TTL to a 60s floor and
+    # `serve_stale 1h immediate` keeps serving through an upstream outage.
+    # NOTE: only cache/serve_stale — the `template` plugin is deliberately NOT
+    # reintroduced here (it broke cluster DNS in 7a09bad5).
     renner.uno:53 {
         errors
-        cache 30
+        cache 300 {
+            success 9984 300 60
+            denial 9984 300 30
+            serve_stale 1h immediate
+        }
         forward . 10.96.0.11
         loop
         reload

kubernetes/infrastructure/k8s-gateway/helmrelease.yaml

@@ -64,8 +64,33 @@ spec:
         parameters: ":8080"
       - name: ready
         parameters: ":8181"
+      # Survive the UDM's intermittent DNS outages (issue #393).
+      #
+      # The UDM's dnsmasq refuses connections in bursts — 330 "connection
+      # refused" errors in the last 3000 log lines here, on BOTH 10.2.40.1 and
+      # fd00::1. That alone would only be a blip, but it combines with a
+      # wildcard trap: the public renner.uno zone has a *.renner.uno CNAME to
+      # proxy.renner.uno that matches at EVERY depth, so renner.uno never
+      # returns NXDOMAIN. Any name that fails locally resolves publicly to the
+      # WAN proxy instead of failing honestly.
+      #
+      # The TTLs make it stick: the correct answer is a dnsmasq lease record
+      # with TTL 5, the wrong wildcard answer has TTL 300 — 60x more durable.
+      # One 5-second UDM blip therefore pins a wrong answer for 5 minutes.
+      #
+      # `success ... 60` clamps that 5s TTL up to a 60s floor, and
+      # `serve_stale 1h immediate` keeps answering from cache while the
+      # upstream is refusing connections, so a blip stops becoming a lookup
+      # failure that the wildcard can capture.
+      #
+      # Written on ONE line deliberately: the chart interpolates `parameters`
+      # into its ConfigMap template, and a multi-line value breaks the YAML
+      # render ("could not find expected ':'"). Verified with `helm template`
+      # plus a real `k8s_gateway:1.7.0` container (CoreDNS 1.14.2) before
+      # merging. For the same reason the forward block is left untouched —
+      # its options need newlines and cannot be expressed here.
       - name: cache
-        parameters: "30"
+        parameters: "300 { success 9984 300 60 denial 9984 300 30 serve_stale 1h immediate }"
       - name: loop
       - name: reload
       - name: forward