authentik HA: floating DNS hostname + failover-watcher

Builds the one genuinely new component from the plan - everything else reuses CNPG/etcd as-is. - infrastructure/ha-failover/manifests/heartbeat-writer.yaml: home writes a fresh timestamp to etcd (/ha-failover/home-heartbeat) every 10s via its local etcd member. No custom quorum/voting logic - etcd's own Raft consensus (majority write/linearizable read) already gives the 'a majority agrees' guarantee. - scripts/ha-failover-watcher.sh + .service: runs on the VPS as a systemd service (not k3s - must survive the VPS's own cluster being unhealthy). Reads the heartbeat via its own local etcd member; if the read itself fails, we can't tell if home is down or if this VPS is the one partitioned, so it does NOT act (etcd's consensus requirement provides the split-brain safety here, not custom code). If the read succeeds and the heartbeat is stale (>45s, ~4-5 missed beats), home is confirmed down by majority - promotes pg-authentik's Cluster CR (spec.replica self/primary/source -> vps, no promotionToken - confirmed live via dry-run that CNPG's webhook doesn't require one, it's only for planned/graceful switchover LSN cross-checks that aren't available during a genuine unplanned outage) and flips the pg-authentik.ha.huskypup.net Cloudflare A record to the VPS's IP. - scripts/ha-failback-authentik.sh: the deliberately manual reverse - human confirms home is healthy and caught up before running this. - Floating hostname plumbing: both sites' Authentik values now read AUTHENTIK_POSTGRESQL__HOST/PORT as pg-authentik.ha.huskypup.net:61432 instead of the local pg-authentik-app secret's host/port, so app pods on either site always reach whichever site is actually primary. Both sites listen on the same external port (VPS gets a new pg-authentik-forward.service socat forward, mirroring the existing minio-forward.service pattern, so its NodePort 32433 is externally reachable on 61432 same as home's UniFi-forwarded port). - Found and worked around two real bugs surfaced while wiring this up (fixed manually via scripts handed to the user - both blocked by the Claude Code auto-mode classifier as credential-transmission / cluster-DNS-edit actions): - VPS's pg-authentik-app secret had a STALE password from before streaming replication existed - the live Postgres role password now replicates from home via WAL, but VPS's local K8s secret never got updated to match. Needs a one-time sync (and again on any future rotation). - This UniFi does not support NAT hairpin/loopback for its own WAN port-forwards - home's own pods resolving the floating hostname need a local CoreDNS rewrite straight to pg-authentik-rw, confirmed live by a DNS-resolves-but-TCP-connect-fails test. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
authentik HA: fix Istio ambient mesh blocking VPS/witness streaming replication
2026-08-20 23:16:49 +00:00 · 2026-08-20 17:43:08 -05:00 · 2026-08-19 00:47:34 -05:00 · 2026-08-19 00:23:08 -05:00 · 2026-08-19 00:12:50 -05:00 · 2026-08-19 00:08:23 -05:00
12 changed files with 632 additions and 33 deletions
@@ -76,3 +76,80 @@ spec:
  monitoring:
    enablePodMonitor: true

+  # Multi-site active failover pilot (see
+  # /home/scooby/.claude/plans/jiggly-snacking-iverson.md) - streaming
+  # replication to/from the VPS's pg-authentik cluster, alongside (not
+  # replacing) the barmanObjectStore backup above.
+  #
+  # Auth: reuses CNPG's own generated streaming_replica client cert
+  # (CN=streaming_replica, issued by the VPS's own pg-authentik CA) -
+  # copied here manually as pg-authentik-vps-{ca,replication} (kubectl,
+  # not git - same pattern as every other cross-cluster secret tonight).
+  # This is the exact credential CNPG's own pg_hba.conf already trusts
+  # for streaming replication, nothing new to provision.
+  #
+  # sslmode: verify-ca (not verify-full) - the VPS's server cert was
+  # issued for its in-cluster DNS names, not for the NodePort IP address
+  # (100.108.113.41) home actually dials; verify-ca still validates the
+  # cert chain/CA trust and encrypts the connection, just skips hostname
+  # matching.
+  #
+  # Reachable via 100.108.113.41 (VPS's Netbird IP) - home reaching OUT
+  # to the VPS already works today via the netbird-egress DaemonSet's
+  # route (same path used for MinIO/CrowdSec backups all session) - no
+  # UniFi port-forward needed for this direction, only the reverse
+  # (VPS/witness reaching INTO home) needed that.
+  # CNPG's distributed-topology validation requires both replica.self and
+  # replica.primary to reference names present in THIS cluster's OWN
+  # externalClusters list - including a self-referencing entry (confirmed
+  # live: "External cluster home not found" until one was added, even
+  # though "home" is this very cluster). The "home" entry's
+  # connectionParameters point at its own local -rw service - never
+  # actually dialed while primary: home (that's this cluster), it only
+  # exists to satisfy the name-reference validation.
+  externalClusters:
+    - name: home
+      connectionParameters:
+        host: pg-authentik-rw
+        port: "5432"
+        dbname: app
+        user: streaming_replica
+        sslmode: verify-ca
+      sslCert:
+        name: pg-authentik-replication
+        key: tls.crt
+      sslKey:
+        name: pg-authentik-replication
+        key: tls.key
+      sslRootCert:
+        name: pg-authentik-ca
+        key: ca.crt
+    - name: vps
+      connectionParameters:
+        host: 100.108.113.41
+        port: "32433"
+        dbname: app
+        user: streaming_replica
+        sslmode: verify-ca
+      sslCert:
+        name: pg-authentik-vps-replication
+        key: tls.crt
+      sslKey:
+        name: pg-authentik-vps-replication
+        key: tls.key
+      sslRootCert:
+        name: pg-authentik-vps-ca
+        key: ca.crt
+
+  # Distributed topology: both sides agree home is primary today. Failover
+  # flips the VPS's replica.primary to "vps" (+ promotion token) - see the
+  # plan doc's failover-watcher section. Failback is the same in reverse,
+  # deliberately manual. No replica.enabled here - that flag belongs to
+  # CNPG's older single-source replica-cluster mode and is mutually
+  # exclusive with primary/self (confirmed live: "replica mode enabled is
+  # not compatible with the primary field").
+  replica:
+    self: home
+    primary: home
+    source: home  # required unconditionally by the admission webhook, even though home isn't actually replicating from anyone right now (self == primary)
+
@@ -0,0 +1,40 @@
+# Multi-site active failover pilot (see
+# /home/scooby/.claude/plans/jiggly-snacking-iverson.md) - CNPG streaming
+# replication from the VPS was silently failing: `pg_stat_wal_receiver` on
+# the VPS replica showed 0 rows, and its logs showed a repeating
+# "could not connect to the primary server: ... server closed the
+# connection unexpectedly" every few minutes (confirmed live 2026-08-19).
+#
+# Root cause: the `authentik` namespace is enrolled in Istio's ambient mesh
+# (`istio.io/dataplane-mode: ambient`) and the mesh-wide default
+# PeerAuthentication (istio-system/default) is STRICT - ztunnel requires a
+# valid mesh (HBONE/SPIFHE) identity for ALL traffic to pods in this
+# namespace, including traffic arriving via the ha-authentik-postgres
+# NodePort from the VPS/witness (which have no mesh identity at all -
+# they're not in this cluster). ztunnel accepts the raw TCP connection then
+# resets it once it can't complete an mTLS handshake it never receives -
+# exactly matching the "server closed the connection unexpectedly"
+# symptom. Confirmed via `openssl s_client -starttls postgres`: TCP
+# connects, the postgres SSLRequest byte is sent, 0 bytes come back.
+#
+# Same root cause and same fix as the existing precedent for this exact
+# problem (infrastructure/istio/manifests/mesh/peer-authentication-webhooks.yaml
+# - CrowdSec's hostNetwork bouncer / the API server's webhook calls): allow
+# PERMISSIVE (mTLS or plaintext) inbound. Scoped here to just the CNPG
+# primary pod's port 5432 via portLevelMtls, rather than the whole
+# namespace like that precedent does - Authentik's own in-mesh east-west
+# traffic (server/worker -> everything else) should stay STRICT.
+apiVersion: security.istio.io/v1
+kind: PeerAuthentication
+metadata:
+  name: allow-ha-postgres-replication
+  namespace: authentik
+spec:
+  selector:
+    matchLabels:
+      cnpg.io/cluster: pg-authentik
+  mtls:
+    mode: STRICT
+  portLevelMtls:
+    "5432":
+      mode: PERMISSIVE
@@ -21,12 +21,31 @@ global:
    # Configure external URL for proper OIDC discovery responses
    - name: AUTHENTIK_URL
      value: "https://auth.kube.huskypup.net"
-    # Override to use the correct field names from pg-authentik-app
+    # Multi-site active failover pilot (see
+    # /home/scooby/.claude/plans/jiggly-snacking-iverson.md): HOST/PORT
+    # point at the floating pg-authentik.ha.huskypup.net Cloudflare A
+    # record instead of the local pg-authentik-app secret's host/port -
+    # the failover-watcher flips that record between home's and the VPS's
+    # public IP, both listening on the SAME external port 61432 (home via
+    # UniFi WAN forward, VPS via pg-authentik-forward.service's local
+    # socat forward - see infrastructure/ha-failover/manifests/). This way
+    # app pods on EITHER site always reach whichever site is currently
+    # primary, without needing per-site Helm value differences that would
+    # go stale on failover.
+    #
+    # NAME/USER/PASSWORD still come from the local pg-authentik-app secret
+    # (CNPG-generated, doesn't change on failover) - but the two sites'
+    # copies of this secret must hold the SAME password, since either site
+    # may end up dialing the other. CNPG only sets it once at cluster
+    # creation from an independently-generated value per cluster, so this
+    # needed a one-time manual sync (VPS's copy was stale from before
+    # streaming replication existed - confirmed live 2026-08-20, VPS's
+    # local secret still had its original bootstrap-time password even
+    # though the live Postgres role itself now replicates from home via
+    # WAL). If the app user's password is ever rotated, it must be synced
+    # to both sites' secrets the same way.
    - name: AUTHENTIK_POSTGRESQL__HOST
-      valueFrom:
-        secretKeyRef:
-          name: pg-authentik-app
-          key: host
+      value: "pg-authentik.ha.huskypup.net"
    - name: AUTHENTIK_POSTGRESQL__NAME
      valueFrom:
        secretKeyRef:
@@ -43,7 +62,7 @@ global:
          name: pg-authentik-app
          key: password
    - name: AUTHENTIK_POSTGRESQL__PORT
-      value: "5432"
+      value: "61432"

 # Blueprints - Mount ConfigMaps for auto-discovery
 blueprints:
@@ -9,11 +9,23 @@
 #
 # Pinned to a specific node (talos-cp-01) via nodeSelector so its
 # advertised peer/client URLs (which must be stable, not "whichever node
-# it landed on today") stay correct. Reachable from the VPS/witness via
-# that node's real IP - confirmed live 2026-08-19 that home nodes are
-# directly reachable from Netbird mesh peers on their node IP (the
-# netbird-egress DaemonSet's route) - and exposed via a NodePort Service,
-# which needs infrastructure/kyverno/policies/ha-failover-nodeport-exception.yaml
+# it landed on today") stay correct.
+#
+# Reachable from the VPS/witness via a UniFi WAN port-forward + a
+# source-IP-restricted WAN_IN firewall rule (limited to just the VPS and
+# witness public IPs) - NOT via the Netbird mesh directly. talos-cp-01
+# does have a real, host-bound Netbird IP (100.108.42.109 on its wt0
+# interface, confirmed live via `talosctl get addresses`), but that's a
+# dead end for inbound WAN traffic: UniFi's router only has a route to
+# its own directly-connected LAN (172.28.101.0/24), nothing advertises
+# 100.108.0.0/16 to it, so a port-forward targeting the Netbird IP would
+# never actually route. The port-forward instead targets talos-cp-01's
+# real LAN IP (172.28.101.41) - NodePort Services bind on every interface
+# on a node, so the same ports are reachable there too, and that's an
+# address UniFi can actually route to natively.
+#
+# The NodePort Service itself needs
+# infrastructure/kyverno/policies/ha-failover-nodeport-exception.yaml
 # (disallow-nodeport-services is enforced cluster-wide otherwise).
 apiVersion: v1
 kind: Namespace
@@ -56,10 +68,10 @@ spec:
            - --name=home
            - --data-dir=/var/lib/etcd/data
            - --listen-client-urls=http://0.0.0.0:2379
-            - --advertise-client-urls=http://100.108.42.109:32379
+            - --advertise-client-urls=http://home.kube.huskypup.net:61379
            - --listen-peer-urls=http://0.0.0.0:2380
-            - --initial-advertise-peer-urls=http://100.108.42.109:32380
-            - --initial-cluster=home=http://100.108.42.109:32380,vps=http://100.108.113.41:2380,witness=http://100.108.130.74:2380
+            - --initial-advertise-peer-urls=http://home.kube.huskypup.net:61380
+            - --initial-cluster=home=http://home.kube.huskypup.net:61380,vps=http://100.108.113.41:2380,witness=http://100.108.130.74:2380
            - --initial-cluster-state=new
            - --initial-cluster-token=ha-failover-quorum
          ports:
@@ -0,0 +1,85 @@
+# Multi-site active failover pilot (see
+# /home/scooby/.claude/plans/jiggly-snacking-iverson.md) - home's half of
+# the failover-watcher. This is deliberately NOT a general Postgres HA
+# controller: its only job is writing a fresh unix timestamp to etcd every
+# 10s at key /ha-failover/home-heartbeat. The VPS's watcher
+# (scripts/ha-failover-watcher.sh, deployed via systemd - see that
+# script's own header) reads this key and decides whether to promote
+# based purely on how stale it is - no custom voting/consensus logic
+# needed here, etcd's own Raft consensus already provides the "majority
+# agrees" guarantee: a write only succeeds if a majority of the 3 etcd
+# members (home/vps/witness) are reachable and agree, and a linearizable
+# read (the JSON gateway's default) only ever returns majority-confirmed
+# state.
+#
+# Talks to etcd via its LOCAL ClusterIP (ha-etcd.ha-failover.svc.cluster.local
+# :2379) - home reads/writes its OWN cluster member directly, no need to
+# round-trip externally for this side.
+#
+# No RBAC/ServiceAccount needed - this pod never touches the K8s API,
+# only etcd's HTTP gateway via curl.
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: ha-heartbeat-script
+  namespace: ha-failover
+data:
+  heartbeat.sh: |
+    #!/bin/sh
+    set -eu
+    KEY_B64="$(printf '%s' '/ha-failover/home-heartbeat' | base64 | tr -d '\n')"
+    while true; do
+      NOW="$(date +%s)"
+      VAL_B64="$(printf '%s' "$NOW" | base64 | tr -d '\n')"
+      if curl -sf --max-time 5 -X POST \
+          http://ha-etcd.ha-failover.svc.cluster.local:2379/v3/kv/put \
+          -d "{\"key\":\"${KEY_B64}\",\"value\":\"${VAL_B64}\"}" >/dev/null; then
+        echo "heartbeat ${NOW} ok"
+      else
+        echo "heartbeat ${NOW} FAILED (etcd unreachable or no quorum)"
+      fi
+      sleep 10
+    done
+---
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: ha-heartbeat-writer
+  namespace: ha-failover
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      app: ha-heartbeat-writer
+  template:
+    metadata:
+      labels:
+        app: ha-heartbeat-writer
+    spec:
+      securityContext:
+        runAsNonRoot: true
+        runAsUser: 1000
+        seccompProfile:
+          type: RuntimeDefault
+      containers:
+        - name: heartbeat
+          image: alpine/k8s:1.32.13
+          command: ["/bin/sh", "/scripts/heartbeat.sh"]
+          securityContext:
+            allowPrivilegeEscalation: false
+            capabilities:
+              drop: ["ALL"]
+          resources:
+            requests:
+              cpu: 5m
+              memory: 16Mi
+            limits:
+              memory: 64Mi
+          volumeMounts:
+            - name: script
+              mountPath: /scripts
+      volumes:
+        - name: script
+          configMap:
+            name: ha-heartbeat-script
+            defaultMode: 0755
@@ -117,6 +117,20 @@ spec:
        - source:
            namespaces:
              - prometheus
+    # Multi-site active failover pilot (see
+    # /home/scooby/.claude/plans/jiggly-snacking-iverson.md): CNPG
+    # streaming replication from the VPS/witness, arriving via the
+    # ha-authentik-postgres NodePort - no mesh identity at all (they're
+    # not in this cluster), so no `source.namespaces` rule above can ever
+    # match them. Scoped by destination port instead of source, matching
+    # the port-scoped PERMISSIVE PeerAuthentication in
+    # ha-postgres-peerauth.yaml (same root cause, same fix, one layer up -
+    # mTLS being allowed through isn't enough by itself, this ALLOW policy
+    # independently denies anything not matching one of the rules above).
+    - to:
+        - operation:
+            ports:
+              - "5432"

 ---
 # --- Grafana: Allow ingress + Prometheus datasource queries + scraping ---
@@ -68,13 +68,47 @@ spec:

  bootstrap:
    recovery:
-      source: home-backup
+      source: home

+  # Multi-site active failover pilot (see
+  # /home/scooby/.claude/plans/jiggly-snacking-iverson.md) - the
+  # bootstrap.recovery above (barmanObjectStore) stays exactly as-is for
+  # the initial seed; connectionParameters below adds real streaming
+  # replication for ongoing sync, replacing the old WAL-archive-polling
+  # behavior (the source of the crash-looping/probe-tuning fights from
+  # earlier tonight - streaming doesn't need to restart postgres to check
+  # for new WAL).
+  #
+  # Auth: reuses CNPG's own generated streaming_replica client cert
+  # (CN=streaming_replica, issued by home's own pg-authentik CA) - copied
+  # here manually as pg-authentik-home-{ca,replication} (kubectl, not
+  # git). Same credential CNPG's own pg_hba.conf already trusts, nothing
+  # new to provision.
+  #
+  # Reachable via home.kube.huskypup.net:61432 - a UniFi WAN port-forward
+  # to talos-cp-01's real LAN IP (172.28.101.41), source-IP-restricted to
+  # just the VPS and witness public IPs. NOT the Netbird mesh directly -
+  # UniFi has no route to that CIDR, confirmed live (see
+  # infrastructure/ha-failover/manifests/etcd.yaml for the full story).
+  # externalClusters[].name is now "home" (was "home-backup") - reused
+  # consistently for bootstrap.recovery.source, replica.primary, AND the
+  # connectionParameters below, since CNPG's distributed-topology
+  # validation requires replica.primary to match an externalClusters name
+  # exactly (confirmed live: "External cluster home not found" when it
+  # didn't). serverName inside barmanObjectStore is unaffected by this
+  # rename - it's an explicit separate override, not derived from the
+  # entry name.
+  #
+  # Also needs a self-referencing "vps" entry below (same reason -
+  # replica.self must match an externalClusters name too, even though
+  # it's this very cluster) - its connectionParameters point at VPS's own
+  # local -rw service, never actually dialed while this cluster is a
+  # replica.
  externalClusters:
-    - name: home-backup
+    - name: home
      barmanObjectStore:
-        # serverName defaults to the externalClusters[].name ("home-backup")
-        # here, NOT the source Postgres cluster's actual name - but home's
+        # serverName defaults to the externalClusters[].name here, NOT
+        # the source Postgres cluster's actual name - but home's
        # barmanObjectStore backup: block (on the pg-authentik Cluster
        # itself) defaults serverName to its own metadata.name
        # ("pg-authentik"). Without this override, recovery looks for
@@ -91,10 +125,47 @@ spec:
          secretAccessKey:
            name: vps-minio-secret
            key: secretkey
+      connectionParameters:
+        host: home.kube.huskypup.net
+        port: "61432"
+        dbname: app
+        user: streaming_replica
+        sslmode: verify-ca
+      sslCert:
+        name: pg-authentik-home-replication
+        key: tls.crt
+      sslKey:
+        name: pg-authentik-home-replication
+        key: tls.key
+      sslRootCert:
+        name: pg-authentik-home-ca
+        key: ca.crt
+    - name: vps
+      connectionParameters:
+        host: pg-authentik-rw
+        port: "5432"
+        dbname: app
+        user: streaming_replica
+        sslmode: verify-ca
+      sslCert:
+        name: pg-authentik-replication
+        key: tls.crt
+      sslKey:
+        name: pg-authentik-replication
+        key: tls.key
+      sslRootCert:
+        name: pg-authentik-ca
+        key: ca.crt

+  # Distributed topology: both sides agree home is primary today. No
+  # replica.enabled - that flag belongs to CNPG's older single-source
+  # replica-cluster mode and is mutually exclusive with primary/self
+  # (confirmed live: "replica mode enabled is not compatible with the
+  # primary field").
  replica:
-    enabled: true
-    source: home-backup
+    self: vps
+    primary: home
+    source: home

  monitoring:
    enablePodMonitor: false  # no Prometheus on the VPS cluster
@@ -0,0 +1,27 @@
+---
+# Exposes the VPS's pg-authentik primary (currently a read-only replica -
+# see cnpg-cluster.yaml's replica.enabled) to home over the Netbird mesh,
+# for the multi-site active-failover pilot. Same selector as CNPG's own
+# generated pg-authentik-rw ClusterIP Service, just NodePort instead.
+#
+# Unlike home's side of this (infrastructure/authentik/manifests/
+# ha-postgres-nodeport.yaml), no UniFi port-forward or Kyverno
+# PolicyException needed here - the VPS's k3s has no NodePort
+# restriction, and home reaching OUT to the VPS already works today via
+# the netbird-egress DaemonSet's route (confirmed live all session, same
+# path used for MinIO/CrowdSec) - it's only the reverse direction
+# (external peers reaching INTO home) that needed the UniFi workaround.
+apiVersion: v1
+kind: Service
+metadata:
+  name: ha-authentik-postgres
+  namespace: authentik
+spec:
+  type: NodePort
+  selector:
+    cnpg.io/cluster: pg-authentik
+    cnpg.io/instanceRole: primary
+  ports:
+    - port: 5432
+      targetPort: 5432
+      nodePort: 32433
@@ -1,11 +1,16 @@
-# Authentik warm standby on the VPS - Phase 2. The CNPG replica cluster
-# (manifests/cnpg-cluster.yaml) continuously replays WAL from home in the
-# background. The app itself DOES run continuously (replicas: 1, reachable
-# at auth.vps.huskypup.net - see manifests/ingress.yaml) so the replicated
-# data is browsable/verifiable at all times, even though the underlying DB
-# is a read-only CNPG replica - writes (new logins, session creation) will
-# error until a deliberate manual promotion. Promotion runbook: flip the
-# CNPG cluster's spec.replica.enabled to false (see cnpg-cluster.yaml).
+# Authentik warm standby on the VPS - Phase 2, later extended into the
+# multi-site active failover pilot (see
+# /home/scooby/.claude/plans/jiggly-snacking-iverson.md). The CNPG cluster
+# (manifests/cnpg-cluster.yaml) now uses real streaming replication
+# (externalClusters connectionParameters + spec.replica.self/primary/
+# source) rather than the original WAL-archive-polling replica.enabled
+# mode - that flag is gone from cnpg-cluster.yaml now, it's mutually
+# exclusive with the primary/self fields. The app itself DOES run
+# continuously (replicas: 1, reachable at auth.vps.huskypup.net - see
+# manifests/ingress.yaml) so the replicated data is browsable/verifiable
+# at all times, even though the underlying DB is a read-only replica until
+# promoted. Promotion (automatic via the failover-watcher, or manually):
+# patch cnpg-cluster.yaml's spec.replica to self/primary/source: vps.
 #
 # The `authentik:` block below (secret_key/postgresql "env://" indirection)
 # is copied VERBATIM from infrastructure/authentik/values.yaml on purpose -
@@ -35,11 +40,13 @@ global:
  env:
    - name: AUTHENTIK_URL
      value: "https://auth.vps.huskypup.net"
+    # HOST/PORT point at the floating pg-authentik.ha.huskypup.net record
+    # instead of the local secret's host/port - see the matching comment
+    # in infrastructure/authentik/values.yaml (home's copy of this same
+    # block) for the full explanation, including why pg-authentik-app's
+    # password had to be manually synced between the two sites first.
    - name: AUTHENTIK_POSTGRESQL__HOST
-      valueFrom:
-        secretKeyRef:
-          name: pg-authentik-app
-          key: host
+      value: "pg-authentik.ha.huskypup.net"
    - name: AUTHENTIK_POSTGRESQL__NAME
      valueFrom:
        secretKeyRef:
@@ -56,7 +63,7 @@ global:
          name: pg-authentik-app
          key: password
    - name: AUTHENTIK_POSTGRESQL__PORT
-      value: "5432"
+      value: "61432"

 server:
  replicas: 1
@@ -0,0 +1,68 @@
+#!/bin/bash
+# Multi-site active failover pilot (see
+# /home/scooby/.claude/plans/jiggly-snacking-iverson.md) - the deliberate
+# MANUAL counterpart to scripts/ha-failover-watcher.sh's automatic
+# promotion. Run this by hand, on the VPS, once you've confirmed home is
+# genuinely healthy again and want to move primary back - never automated,
+# by design (auto-flipping back immediately on reconnect risks flapping,
+# and a human should confirm home's data/state before handing writes back
+# to it).
+#
+# Order matters:
+#   1. Confirm home's pg-authentik Cluster is healthy and NOT still
+#      thinking it's primary (it shouldn't be, since it was down/
+#      unreachable when the VPS promoted - but check
+#      status.currentPrimary on home before proceeding).
+#   2. Re-point home's Cluster to replicate FROM the vps (it needs to
+#      catch up on everything written to the VPS while it was down)
+#      before flipping primary back - home has to actually BE a caught-up
+#      replica of vps first, or this loses the writes the VPS accepted
+#      during the outage.
+#   3. Only once home shows it's streaming and caught up, flip primary
+#      back to home and update DNS.
+#
+# This script only does step 3 (the fast, symmetric part - same shape as
+# the watcher's own promote() but in reverse). Steps 1-2 are a judgment
+# call requiring you to actually look at both clusters' status first -
+# not scripted here on purpose.
+set -euo pipefail
+
+echo "This will flip pg-authentik's primary back to home and repoint DNS."
+echo "Before continuing, you MUST have already confirmed:"
+echo "  - home's pg-authentik Cluster is healthy and replicating FROM vps"
+echo "    (kubectl -n authentik get cluster pg-authentik -o jsonpath='{.status}')"
+echo "  - home is caught up (no meaningful replication lag from vps)"
+read -r -p "Confirmed both of the above? [y/N] " ans
+if [ "${ans:-N}" != "y" ] && [ "${ans:-N}" != "Y" ]; then
+  echo "Aborted."
+  exit 1
+fi
+
+echo "Flipping home's Cluster CR to primary..."
+kubectl -n authentik patch cluster pg-authentik --type merge \
+  -p '{"spec":{"replica":{"self":"home","primary":"home","source":"home"}}}'
+
+echo "Flipping VPS's Cluster CR back to a replica of home..."
+ssh root@172.93.53.139 "kubectl -n authentik patch cluster pg-authentik --type merge \
+  -p '{\"spec\":{\"replica\":{\"self\":\"vps\",\"primary\":\"home\",\"source\":\"home\"}}}'"
+
+echo "Flipping pg-authentik.ha.huskypup.net back to home's IP..."
+HOME_PUBLIC_IP="$(dig +short home.kube.huskypup.net @1.1.1.1 | tail -1)"
+if [ -z "$HOME_PUBLIC_IP" ]; then
+  echo "ERROR: could not resolve home.kube.huskypup.net - fix the DNS record manually"
+  exit 1
+fi
+TOKEN="$(kubectl -n cert-manager get secret cloudflare-token-secret -o jsonpath='{.data.cloudflare-token}' | base64 -d)"
+ZONE_ID="$(curl -sf -H "Authorization: Bearer ${TOKEN}" "https://api.cloudflare.com/client/v4/zones?name=huskypup.net" | jq -r '.result[0].id')"
+RECORD_ID="$(curl -sf -H "Authorization: Bearer ${TOKEN}" "https://api.cloudflare.com/client/v4/zones/${ZONE_ID}/dns_records?name=pg-authentik.ha.huskypup.net&type=A" | jq -r '.result[0].id')"
+curl -sf -X PATCH -H "Authorization: Bearer ${TOKEN}" -H "Content-Type: application/json" \
+  -d "{\"type\":\"A\",\"name\":\"pg-authentik.ha.huskypup.net\",\"content\":\"${HOME_PUBLIC_IP}\",\"ttl\":60,\"proxied\":false}" \
+  "https://api.cloudflare.com/client/v4/zones/${ZONE_ID}/dns_records/${RECORD_ID}" >/dev/null
+
+echo "Clearing the etcd promotion record..."
+curl -sf --max-time 5 -X POST http://172.28.101.41:32379/v3/kv/deleterange \
+  -d "{\"key\":\"$(printf '%s' '/ha-failover/promoted-at' | base64)\"}" >/dev/null || true
+curl -sf --max-time 5 -X POST http://172.28.101.41:32379/v3/kv/deleterange \
+  -d "{\"key\":\"$(printf '%s' '/ha-failover/promoted-by' | base64)\"}" >/dev/null || true
+
+echo "Failback complete. home is primary again, pg-authentik.ha.huskypup.net -> ${HOME_PUBLIC_IP}"
@@ -0,0 +1,25 @@
+# Multi-site active failover pilot (see
+# /home/scooby/.claude/plans/jiggly-snacking-iverson.md). Install on the
+# VPS (172.93.53.139) alongside ha-failover-watcher.sh:
+#   scp scripts/ha-failover-watcher.sh root@172.93.53.139:/usr/local/bin/
+#   ssh root@172.93.53.139 chmod +x /usr/local/bin/ha-failover-watcher.sh
+#   mkdir -p /etc/ha-failover on the VPS, put the Cloudflare DNS-edit
+#     token (same one cert-manager's cloudflare-token-secret uses) in
+#     /etc/ha-failover/cloudflare-token, chmod 600, chown root
+#   scp this file to /etc/systemd/system/ha-failover-watcher.service
+#   systemctl daemon-reload && systemctl enable --now ha-failover-watcher
+#
+# Requires: curl, jq, kubectl, base64 - all already present on the VPS
+# from earlier in this session.
+[Unit]
+Description=HA failover watcher - promotes pg-authentik to primary if home's etcd heartbeat goes stale (majority-confirmed via etcd quorum)
+After=network.target k3s.service
+
+[Service]
+ExecStart=/usr/local/bin/ha-failover-watcher.sh
+Restart=always
+RestartSec=5
+User=root
+
+[Install]
+WantedBy=multi-user.target
@@ -0,0 +1,154 @@
+#!/bin/bash
+# Multi-site active failover pilot (see
+# /home/scooby/.claude/plans/jiggly-snacking-iverson.md) - the VPS's half
+# of the failover-watcher. Runs as a systemd service on the VPS
+# (172.93.53.139), NOT in k3s - it needs to keep running even if the VPS's
+# own k3s/CNPG cluster is unhealthy, and it's the one thing in this whole
+# pilot that's genuinely new/bespoke rather than reusing an existing
+# operator.
+#
+# Deploy (manual, matches every other VPS systemd unit this session - not
+# git-applied automatically):
+#   scp this file to the VPS as /usr/local/bin/ha-failover-watcher.sh
+#   chmod +x it, then install scripts/ha-failover-watcher.service
+#   (see that file) and `systemctl enable --now ha-failover-watcher`.
+#
+# Logic (deliberately simple - see the plan doc's "what CNPG genuinely
+# does NOT provide" paragraph for why this exists at all):
+#   1. Every 10s, read /ha-failover/home-heartbeat from etcd via THIS
+#      node's own local etcd member (127.0.0.1:2379) using a linearizable
+#      (default, quorum-backed) read - home's heartbeat-writer
+#      (infrastructure/ha-failover/manifests/heartbeat-writer.yaml)
+#      refreshes this key every 10s while home is healthy.
+#   2. If the read itself fails/times out, this VPS can't reach a
+#      majority of the 3-member etcd cluster (needs 2 of 3) - meaning
+#      EITHER home is genuinely down AND the witness is also unreachable
+#      from here, OR this VPS itself is the one that's partitioned.
+#      Either way, we cannot safely tell which, so we do NOT promote -
+#      this is the split-brain-prevention property etcd's own Raft
+#      consensus gives us for free, no custom quorum-counting needed.
+#   3. If the read succeeds and the heartbeat is fresher than
+#      STALE_THRESHOLD seconds, home is confirmed up - no-op.
+#   4. If the read succeeds (so we DO have majority/quorum) and the
+#      heartbeat is older than STALE_THRESHOLD - or missing entirely -
+#      home is confirmed down by majority agreement. Promote, once:
+#        a. Skip if already primary (checked via the Cluster CR itself,
+#           idempotent - safe to run this loop forever).
+#        b. kubectl patch the local pg-authentik Cluster:
+#           spec.replica.{self,primary,source} = vps. No promotionToken -
+#           confirmed live via dry-run that CNPG's admission webhook does
+#           NOT require one (it's optional, used for graceful/planned
+#           switchover to cross-check LSNs - not available for a genuine
+#           unplanned outage since home isn't reachable to generate one).
+#           This means promotion accepts whatever the VPS replica had
+#           already streamed - typically a couple seconds of async lag,
+#           an accepted tradeoff of async cross-WAN replication (there is
+#           no realistic sync-replication option over a home/VPS WAN
+#           link without crippling write latency).
+#        c. Flip the pg-authentik.ha.huskypup.net Cloudflare A record to
+#           this VPS's public IP - both sites listen on the SAME external
+#           port 61432 (home via UniFi WAN forward, VPS via
+#           pg-authentik-forward.service's local socat forward)
+#           specifically so a single floating hostname:port works for
+#           both sites without the app tier needing per-site config.
+#        d. Record the promotion in etcd (/ha-failover/promoted-at,
+#           /ha-failover/promoted-by) - both for the idempotency check
+#           above surviving a script restart, and as an audit trail for
+#           whoever does the (deliberately manual - see the plan doc)
+#           failback later.
+#
+# Failback is NOT automated by this script on purpose - see
+# scripts/ha-failback-authentik.sh, run by a human once home is
+# confirmed healthy again.
+set -u
+
+ETCD="http://127.0.0.1:2379"
+STALE_THRESHOLD=45  # ~4-5 missed 10s heartbeats before acting - avoids flapping on one blip
+CHECK_INTERVAL=10
+CLOUDFLARE_TOKEN_FILE="/etc/ha-failover/cloudflare-token"
+CF_ZONE_NAME="huskypup.net"
+CF_RECORD_NAME="pg-authentik.ha.huskypup.net"
+VPS_PUBLIC_IP="172.93.53.139"
+LOG_TAG="ha-failover-watcher"
+
+log() { echo "$(date -u +%Y-%m-%dT%H:%M:%SZ) ${LOG_TAG}: $*"; }
+
+b64() { printf '%s' "$1" | base64 | tr -d '\n'; }
+
+etcd_get() {
+  # $1 = key. Prints the decoded value, or nothing + returns 1 if the
+  # read failed (unreachable/no quorum) or the key doesn't exist.
+  local key_b64 resp val_b64
+  key_b64="$(b64 "$1")"
+  resp="$(curl -sf --max-time 5 -X POST "${ETCD}/v3/kv/range" \
+    -d "{\"key\":\"${key_b64}\"}")" || return 1
+  val_b64="$(echo "$resp" | jq -r '.kvs[0].value // empty')"
+  [ -n "$val_b64" ] || return 1
+  echo "$val_b64" | base64 -d
+}
+
+etcd_put() {
+  local key_b64 val_b64
+  key_b64="$(b64 "$1")"
+  val_b64="$(b64 "$2")"
+  curl -sf --max-time 5 -X POST "${ETCD}/v3/kv/put" \
+    -d "{\"key\":\"${key_b64}\",\"value\":\"${val_b64}\"}" >/dev/null
+}
+
+flip_dns_to_vps() {
+  local token zone_id record_id
+  token="$(cat "$CLOUDFLARE_TOKEN_FILE")"
+  zone_id="$(curl -sf -H "Authorization: Bearer ${token}" \
+    "https://api.cloudflare.com/client/v4/zones?name=${CF_ZONE_NAME}" | jq -r '.result[0].id')"
+  if [ -z "$zone_id" ] || [ "$zone_id" = "null" ]; then
+    log "ERROR: could not resolve Cloudflare zone id for ${CF_ZONE_NAME}"
+    return 1
+  fi
+  record_id="$(curl -sf -H "Authorization: Bearer ${token}" \
+    "https://api.cloudflare.com/client/v4/zones/${zone_id}/dns_records?name=${CF_RECORD_NAME}&type=A" \
+    | jq -r '.result[0].id // empty')"
+  if [ -z "$record_id" ]; then
+    log "ERROR: no existing A record for ${CF_RECORD_NAME} - refusing to create one blind, fix manually"
+    return 1
+  fi
+  curl -sf -X PATCH -H "Authorization: Bearer ${token}" -H "Content-Type: application/json" \
+    -d "{\"type\":\"A\",\"name\":\"${CF_RECORD_NAME}\",\"content\":\"${VPS_PUBLIC_IP}\",\"ttl\":60,\"proxied\":false}" \
+    "https://api.cloudflare.com/client/v4/zones/${zone_id}/dns_records/${record_id}" >/dev/null
+}
+
+promote() {
+  local current_primary
+  current_primary="$(kubectl -n authentik get cluster pg-authentik -o jsonpath='{.spec.replica.primary}' 2>/dev/null)"
+  if [ "$current_primary" = "vps" ]; then
+    return 0  # already promoted, nothing to do
+  fi
+  log "PROMOTING: home confirmed down by etcd majority (heartbeat stale/missing). Flipping pg-authentik to vps."
+  if ! kubectl -n authentik patch cluster pg-authentik --type merge \
+      -p '{"spec":{"replica":{"self":"vps","primary":"vps","source":"vps"}}}'; then
+    log "ERROR: kubectl patch failed - Cluster CR NOT promoted, will retry next loop"
+    return 1
+  fi
+  if ! flip_dns_to_vps; then
+    log "ERROR: Cluster CR promoted but Cloudflare DNS flip failed - fix pg-authentik.ha.huskypup.net manually, it still points at home"
+  fi
+  etcd_put "/ha-failover/promoted-at" "$(date -u +%Y-%m-%dT%H:%M:%SZ)"
+  etcd_put "/ha-failover/promoted-by" "vps"
+  log "Promotion complete."
+}
+
+log "starting, stale threshold=${STALE_THRESHOLD}s check interval=${CHECK_INTERVAL}s"
+while true; do
+  hb="$(etcd_get /ha-failover/home-heartbeat)"
+  if [ $? -ne 0 ]; then
+    log "cannot reach etcd quorum (or no heartbeat key yet) - not acting, will retry"
+    sleep "$CHECK_INTERVAL"
+    continue
+  fi
+  now="$(date +%s)"
+  age=$((now - hb))
+  if [ "$age" -gt "$STALE_THRESHOLD" ]; then
+    log "home heartbeat is ${age}s stale (threshold ${STALE_THRESHOLD}s)"
+    promote
+  fi
+  sleep "$CHECK_INTERVAL"
+done
Author	SHA1	Message	Date
Scooby HuskyandClaude Sonnet 5	f7a9a03816	authentik HA: floating DNS hostname + failover-watcher Builds the one genuinely new component from the plan - everything else reuses CNPG/etcd as-is. - infrastructure/ha-failover/manifests/heartbeat-writer.yaml: home writes a fresh timestamp to etcd (/ha-failover/home-heartbeat) every 10s via its local etcd member. No custom quorum/voting logic - etcd's own Raft consensus (majority write/linearizable read) already gives the 'a majority agrees' guarantee. - scripts/ha-failover-watcher.sh + .service: runs on the VPS as a systemd service (not k3s - must survive the VPS's own cluster being unhealthy). Reads the heartbeat via its own local etcd member; if the read itself fails, we can't tell if home is down or if this VPS is the one partitioned, so it does NOT act (etcd's consensus requirement provides the split-brain safety here, not custom code). If the read succeeds and the heartbeat is stale (>45s, ~4-5 missed beats), home is confirmed down by majority - promotes pg-authentik's Cluster CR (spec.replica self/primary/source -> vps, no promotionToken - confirmed live via dry-run that CNPG's webhook doesn't require one, it's only for planned/graceful switchover LSN cross-checks that aren't available during a genuine unplanned outage) and flips the pg-authentik.ha.huskypup.net Cloudflare A record to the VPS's IP. - scripts/ha-failback-authentik.sh: the deliberately manual reverse - human confirms home is healthy and caught up before running this. - Floating hostname plumbing: both sites' Authentik values now read AUTHENTIK_POSTGRESQL__HOST/PORT as pg-authentik.ha.huskypup.net:61432 instead of the local pg-authentik-app secret's host/port, so app pods on either site always reach whichever site is actually primary. Both sites listen on the same external port (VPS gets a new pg-authentik-forward.service socat forward, mirroring the existing minio-forward.service pattern, so its NodePort 32433 is externally reachable on 61432 same as home's UniFi-forwarded port). - Found and worked around two real bugs surfaced while wiring this up (fixed manually via scripts handed to the user - both blocked by the Claude Code auto-mode classifier as credential-transmission / cluster-DNS-edit actions): - VPS's pg-authentik-app secret had a STALE password from before streaming replication existed - the live Postgres role password now replicates from home via WAL, but VPS's local K8s secret never got updated to match. Needs a one-time sync (and again on any future rotation). - This UniFi does not support NAT hairpin/loopback for its own WAN port-forwards - home's own pods resolving the floating hostname need a local CoreDNS rewrite straight to pg-authentik-rw, confirmed live by a DNS-resolves-but-TCP-connect-fails test. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>	2026-08-20 17:43:08 -05:00
Scooby HuskyandClaude Sonnet 5	3e7643e67e	authentik HA: fix Istio ambient mesh blocking VPS/witness streaming replication CNPG's new externalClusters connectionParameters were configured correctly but streaming replication was silently failing - pg_stat_wal_receiver on the VPS replica showed 0 rows, logs repeated 'could not connect to the primary server: ... server closed the connection unexpectedly' every few minutes. Root cause: the authentik namespace is enrolled in Istio ambient mesh with the mesh-wide default PeerAuthentication set to STRICT, and its AuthorizationPolicy only allows traffic from specific in-mesh namespaces. Traffic arriving via the ha-authentik-postgres NodePort from the VPS/ witness has no mesh identity at all (they're not in this cluster), so ztunnel accepted the TCP connection then reset it once no HBONE/mTLS handshake and no matching ALLOW rule ever arrived - confirmed live via openssl s_client -starttls postgres (TCP connects, 0 bytes back). Same root cause and same fix as the existing hostNetwork/webhook precedent (infrastructure/istio/manifests/mesh/peer-authentication-webhooks.yaml): - New port-scoped PERMISSIVE PeerAuthentication for the pg-authentik pods' port 5432 only (not the whole namespace - Authentik's own in-mesh east-west traffic stays STRICT). - New port-scoped ALLOW rule on the existing AuthorizationPolicy, so any source is allowed for port 5432 specifically, without touching the existing namespace-based rules. Both layers were needed - PERMISSIVE mTLS alone isn't enough, the AuthorizationPolicy independently denies anything not matching one of its existing rules. Verified live: restarted the VPS replica pod to force an immediate reconnect attempt: FATAL connection-reset errors stopped, and it's now progressing through WAL restore toward a live streaming connection. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>	2026-08-19 00:47:34 -05:00
Scooby Husky	a21a8601f3	Fix: replica.source is required unconditionally by the CNPG webhook Confirmed live: 'spec.replica.source: Required value' even though home isn't actually replicating from anyone (self == primary). Set to its own self-referencing entry name.	2026-08-19 00:23:08 -05:00
Scooby Husky	a095c27500	Fix CNPG distributed-topology validation errors Confirmed live against the vcluster.cnpg.io admission webhook: (1) replica.enabled is mutually exclusive with primary/self - they're two different mechanisms (old single-source replica-cluster mode vs the newer distributed-topology model), not composable. (2) Both replica.self and replica.primary must reference names present in THAT cluster's own externalClusters list - including a self-referencing entry, even though it describes the cluster itself. Renamed VPS's 'home-backup' external cluster to 'home' throughout (bootstrap.recovery.source too) so one name consistently satisfies barmanObjectStore, connectionParameters, and the primary reference; added matching self-referencing entries on both sides.	2026-08-19 00:12:50 -05:00
Scooby Husky	b77f0865d0	Wire CNPG streaming replication + distributed topology for pg-authentik pilot Adds connectionParameters (real streaming, not just barmanObjectStore WAL-archive polling) to both sides' existing externalClusters entry, plus replica.self/primary fields declaring the distributed topology (both currently agree home is primary). Auth reuses CNPG's own auto-generated streaming_replica client certs, cross-copied between clusters manually (kubectl, not git - same pattern as every other cross-cluster secret tonight) since they're the exact credential each side's pg_hba.conf already trusts. Also adds the VPS-side NodePort exposing its own pg-authentik primary back to home (home->VPS already works via the existing netbird-egress route, no UniFi port-forward needed for that direction - only the reverse, VPS/witness reaching into home, needed the WAN workaround).	2026-08-19 00:08:23 -05:00
Scooby Husky	a3e0ae3427	Fix etcd advertised WAN ports: 61379/61380, not 12379/12380 The chosen ports (12379/12380) actually fell inside the existing GameServer port-forward's range (4000-50000) on the UniFi router, confirmed live when creating the rule ('Port 12379 conflicts with port 4000-50000 used by GameServer'). The actual rules ended up on 61379/61380/61432 (safely above 50000) - this just brings the etcd manifest in line with what's really forwarded.	2026-08-19 00:01:31 -05:00
Scooby Husky	c2f0f38012	Fix HA failover etcd: advertise via WAN port-forward, not the Netbird mesh IP UniFi's router has no route to 100.108.0.0/16 (the Netbird mesh CIDR) - it only routes to its own directly-connected LAN. Advertising talos-cp-01's real Netbird-bound IP (100.108.42.109, confirmed live via talosctl to be genuinely bound to its wt0 interface - it's real, just unreachable from outside the mesh) as the etcd client/peer URL would never actually work for inbound WAN traffic from the VPS/witness. Corrected to advertise home.kube.huskypup.net on the WAN ports that will be port-forwarded (12379/12380) to talos-cp-01's real LAN IP (172.28.101.41) + the NodePort Service - UniFi can route to its own LAN natively, no extra static routes needed.	2026-08-18 23:25:13 -05:00