node001: send the visitor's address to the home cluster

The stream server on 443 now starts each connection with a PROXY header.
nginx sends it to every upstream or none, so Stalwart (8443) and the
local nginx for Gatus and ntfy (8444) get theirs through a hop on
8445/8446 that takes it off again; nothing changes for them.

Deploy after the Gateway accepts the header, or the tunneled hosts stop
answering until it does.

gateway/proxy-protocol.yaml

@@ -0,0 +1,23 @@
+# node001 passes TLS for the tunneled hosts through without opening it, so
+# without this every request arrives from 46.224.26.65. Wordfence locks that
+# address out after 20 failed logins (2026-10-26), and with it everyone who
+# logs in to scoutingmbg.nl. node001 puts a PROXY header with the visitor's
+# address in front of each connection; Envoy takes the address from there
+# and adds it to X-Forwarded-For.
+#
+# Optional, because s2 and ntfy2 point straight at home and their clients
+# send no header, and neither does node001's proxy on port 80. The price:
+# whoever connects to the home address directly can send a header of their
+# own and claim any address.
+apiVersion: gateway.envoyproxy.io/v1alpha1
+kind: ClientTrafficPolicy
+metadata:
+  name: proxy-protocol
+  namespace: envoy-gateway-system
+spec:
+  targetRefs:
+  - group: gateway.networking.k8s.io
+    kind: Gateway
+    name: homelab
+  proxyProtocol:
+    optional: true

immich/helm-immich.yaml

@@ -127,6 +127,9 @@ spec:
                       TRANSFORMERS_CACHE: /cache
                       HF_XET_CACHE: /cache/huggingface-xet
                       MPLCONFIGDIR: /cache/matplotlib-config
+                      # The optimized models of v3.3.0: faster, less memory,
+                      # and the same results, so no jobs to run again.
+                      MACHINE_LEARNING_MODEL_REVISION: v2
                       # Trust store without the compromised eMudhra
                       # intermediates; requests would otherwise use certifi.
                       SSL_CERT_FILE: /etc/ssl/certs/ca-certificates.crt
@@ -139,11 +142,19 @@ spec:
                 globalMounts:
                   - path: /etc/ssl/certs
                     readOnly: true
+              # The models and their one-off preparation, kept across restarts.
+              # Downloaded again if lost, so no backups: the no-backup group
+              # keeps it out of Longhorn's default one (backup-daily), and
+              # Velero skips that label too (velero/skip-emptydir-policy.yaml).
               cache:
                 enabled: true
-                # Optional: to avoid downloading the ML models every start, use
-                # a persistentVolumeClaim with size, accessMode and storageClass
-                type: emptyDir
+                type: persistentVolumeClaim
+                storageClass: longhorn
+                accessMode: ReadWriteOnce
+                size: 10Gi
+                labels:
+                  recurring-job.longhorn.io/source: enabled
+                  recurring-job-group.longhorn.io/no-backup: enabled
 
   destination:
     server: https://kubernetes.default.svc

nix-infra-machine/nodes/node001.nix

@@ -173,12 +173,18 @@ in
     #   - mail.rubenhensen.nl (and anything else) → local stalwart on 8443
     #   - tunneled hosts → home cluster ingress on 443
     #   - s1 / ntfy1.rubenhensen.nl → local nginx on 8444 (Gatus, ntfy)
+    #
+    # Each connection starts with a PROXY header that carries the visitor's
+    # address, so the home cluster sees who is asking instead of this box
+    # (gateway/proxy-protocol.yaml in k8scd). nginx sends the header to
+    # every upstream or none, so the local ones go through a hop on
+    # 8445/8446 that takes it off again.
     streamConfig = ''
       map $ssl_preread_server_name $tunnel_upstream {
 ${sniMapEntries}
-        s1.rubenhensen.nl 127.0.0.1:8444;
-        ntfy1.rubenhensen.nl 127.0.0.1:8444;
-        default 127.0.0.1:8443;
+        s1.rubenhensen.nl 127.0.0.1:8446;
+        ntfy1.rubenhensen.nl 127.0.0.1:8446;
+        default 127.0.0.1:8445;
       }
 
       server {
@@ -186,6 +192,17 @@ ${sniMapEntries}
         listen [::]:443;
         proxy_pass $tunnel_upstream;
         ssl_preread on;
+        proxy_protocol on;
+      }
+
+      server {
+        listen 127.0.0.1:8445 proxy_protocol;
+        proxy_pass 127.0.0.1:8443;
+      }
+
+      server {
+        listen 127.0.0.1:8446 proxy_protocol;
+        proxy_pass 127.0.0.1:8444;
       }
     '';
 

velero/skip-emptydir-policy.yaml

@@ -1,6 +1,8 @@
 # Volume policy for the backup schedules. emptyDir volumes hold nothing
 # worth keeping, and backing them up fails when the pod is gone before its
 # volume is read, as happens with Velero's own repository maintenance jobs.
+# PVCs in Longhorn's no-backup group (caches, such as immich's ML models)
+# are skipped too.
 apiVersion: v1
 kind: ConfigMap
 metadata:
@@ -14,3 +16,8 @@ data:
         - emptyDir
       action:
         type: skip
+    - conditions:
+        pvcLabels:
+          recurring-job-group.longhorn.io/no-backup: enabled
+      action:
+        type: skip