Rework jellyfin-ha into a true-HA StatefulSet deployment

Turn the single-replica jellyfin-ha app into a proper high-availability
deployment so the fork's Redis-coordinated distributed transcoding and
PostgreSQL main database can actually be exercised.

- Replace the Deployment with a 2-replica StatefulSet for stable pod
  identity; set JELLYFIN_INSTANCE_ID from metadata.name (the fork's Redis
  transcode-lease owner id), add soft podAntiAffinity and a PDB
  minAvailable 1.
- Move the main Jellyfin DB to PostgreSQL via a CloudNativePG trio
  (3-instance Cluster, PgBouncer Pooler, Ceph RGW barman backups) mirroring
  the litellm pattern; an init container writes database.xml selecting the
  fork's Jellyfin-PostgreSQL provider and the DSN is composed from the
  CNPG-generated app secret pointed at the pooler.
- Share /config on an RWX cephfs PVC across replicas; keep /cache per-pod
  via a volumeClaimTemplate.
- Fix the transcode mount to the fork's real path /config/transcodes on the
  RWX PVC (raid5) so a surviving pod can resume the segments of the pod it
  takes over.
- Add Intel iGPU hardware transcoding via the gpu.intel.com/i915 device
  plugin resource plus render/video supplemental groups.
- Switch the Service to sessionAffinity ClientIP to reduce transcode churn.
- Disable UDP auto-discovery. Library scans still run on every replica; a
  single-scanner leader election is a planned follow-up.

apps/base/artifactapi/api-deployment.yaml

@@ -36,7 +36,7 @@ spec:
               mountPath: /combined-certs
       containers:
         - name: api
-          image: git.unkin.net/unkin/artifactapi:v3.7.7
+          image: git.unkin.net/unkin/artifactapi:v3.8.0
           imagePullPolicy: IfNotPresent
           ports:
             - containerPort: 8000

apps/base/artifactapi/ui-deployment.yaml

@@ -22,7 +22,7 @@ spec:
       automountServiceAccountToken: true
       containers:
         - name: ui
-          image: git.unkin.net/unkin/artifactapi-ui:v3.7.7
+          image: git.unkin.net/unkin/artifactapi-ui:v3.8.0
           imagePullPolicy: IfNotPresent
           ports:
             - containerPort: 80

apps/base/jellyfin/cnpg_backup.yaml

@@ -0,0 +1,45 @@
+---
+# Ceph RGW (S3) backup target for the jellyfin CNPG cluster, provisioned by the
+# in-estate cephrgw-operator: one dedicated bucket + owner user. CNPG reads the
+# S3 credential Secret from its own namespace.
+apiVersion: ceph.unkin.net/v1alpha1
+kind: ObjectStoreUser
+metadata:
+  name: cnpg-jellyfin-backup
+  namespace: jellyfin
+spec:
+  displayName: "CNPG backup owner (jellyfin)"
+  uid: cnpg-jellyfin-backup
+  maxBuckets: 5
+  secretName: cnpg-jellyfin-backup-s3
+  retainOnDelete: true
+---
+apiVersion: ceph.unkin.net/v1alpha1
+kind: Bucket
+metadata:
+  name: cnpg-jellyfin
+  namespace: jellyfin
+spec:
+  placementTarget: ec
+  bucketName: cnpg-jellyfin
+  ownerRef: cnpg-jellyfin-backup
+  versioning: false
+  tags:
+    app: jellyfin
+    purpose: cnpg-backup
+  retainOnDelete: true
+---
+# Nightly base backup on top of always-on WAL archiving. Scheduled off-peak and
+# staggered from the other CNPG clusters (6-field cron, seconds first).
+apiVersion: postgresql.cnpg.io/v1
+kind: ScheduledBackup
+metadata:
+  name: cnpg-jellyfin-nightly
+  namespace: jellyfin
+spec:
+  schedule: "0 35 3 * * *"
+  immediate: false
+  backupOwnerReference: self
+  method: barmanObjectStore
+  cluster:
+    name: jellyfin-postgres

apps/base/jellyfin/cnpg_cluster.yaml

@@ -0,0 +1,119 @@
+---
+# Main Jellyfin database. The jellyfin-ha fork's experimental EF Core provider
+# moves the entire Jellyfin DB (incl. library items) off SQLite into PostgreSQL,
+# which is what makes a shared-nothing multi-replica deployment possible. No
+# bootstrap secret is given, so CNPG generates the jellyfin-postgres-app secret
+# (username/password/dbname) that the StatefulSet composes its DSN from.
+apiVersion: postgresql.cnpg.io/v1
+kind: Cluster
+metadata:
+  name: jellyfin-postgres
+  namespace: jellyfin
+spec:
+  affinity:
+    podAntiAffinityType: preferred
+  backup:
+    retentionPolicy: 30d
+    barmanObjectStore:
+      # Dedicated per-cluster Ceph RGW bucket (cephrgw-operator provisions it).
+      destinationPath: s3://cnpg-jellyfin
+      endpointURL: https://s3.ceph.unkin.net
+      endpointCA:
+        name: vault-ca-cert
+        key: ca.crt
+      s3Credentials:
+        accessKeyId:
+          name: cnpg-jellyfin-backup-s3
+          key: AWS_ACCESS_KEY_ID
+        secretAccessKey:
+          name: cnpg-jellyfin-backup-s3
+          key: AWS_SECRET_ACCESS_KEY
+      serverName: jellyfin
+      data:
+        compression: bzip2
+        jobs: 2
+      wal:
+        compression: zstd
+        maxParallel: 2
+  bootstrap:
+    initdb:
+      database: jellyfin
+      encoding: UTF8
+      localeCType: C
+      localeCollate: C
+      owner: jellyfin
+  enablePDB: true
+  enableSuperuserAccess: false
+  failoverDelay: 0
+  # PG 17 — accepted by the fork's Npgsql/EF Core provider (needs PG14+); the
+  # provider generates its own migrations on first start.
+  imageName: ghcr.io/cloudnative-pg/postgresql:17-system-trixie
+  instances: 3
+  logLevel: info
+  maxSyncReplicas: 0
+  minSyncReplicas: 0
+  monitoring:
+    customQueriesConfigMap:
+      - key: queries
+        name: cnpg-default-monitoring
+    disableDefaultQueries: false
+    enablePodMonitor: false
+  postgresql:
+    parameters:
+      archive_mode: "on"
+      archive_timeout: 5min
+      dynamic_shared_memory_type: posix
+      effective_cache_size: 256MB
+      full_page_writes: "on"
+      log_destination: csvlog
+      log_directory: /controller/log
+      log_filename: postgres
+      log_rotation_age: "0"
+      log_rotation_size: "0"
+      log_truncate_on_rotation: "false"
+      logging_collector: "on"
+      max_connections: "200"
+      max_parallel_workers: "16"
+      max_replication_slots: "16"
+      max_worker_processes: "16"
+      shared_buffers: 128MB
+      shared_memory_type: mmap
+      ssl_max_protocol_version: TLSv1.3
+      ssl_min_protocol_version: TLSv1.3
+      wal_keep_size: 256MB
+      wal_level: logical
+      wal_log_hints: "on"
+      wal_receiver_timeout: 5s
+      wal_sender_timeout: 5s
+    syncReplicaElectionConstraint:
+      enabled: false
+  primaryUpdateMethod: restart
+  primaryUpdateStrategy: unsupervised
+  probes:
+    liveness:
+      isolationCheck:
+        connectionTimeout: 1000
+        enabled: true
+        requestTimeout: 1000
+  replicationSlots:
+    highAvailability:
+      enabled: true
+      slotPrefix: _cnpg_
+    synchronizeReplicas:
+      enabled: true
+    updateInterval: 30
+  resources:
+    limits:
+      cpu: "1"
+      memory: 1Gi
+    requests:
+      cpu: 50m
+      memory: 512Mi
+  smartShutdownTimeout: 180
+  startDelay: 3600
+  stopDelay: 1800
+  storage:
+    resizeInUseVolumes: true
+    size: 10Gi
+    storageClass: cephrbd-fast-delete
+  switchoverDelay: 3600

apps/base/jellyfin/cnpg_pooler.yaml

@@ -0,0 +1,36 @@
+---
+# PgBouncer pooler in front of the jellyfin-postgres cluster. Jellyfin connects
+# here (jellyfin-postgres-pooler:5432) rather than the -rw service so EF Core's
+# connection churn is absorbed by the pool.
+apiVersion: postgresql.cnpg.io/v1
+kind: Pooler
+metadata:
+  name: jellyfin-postgres-pooler
+  namespace: jellyfin
+spec:
+  cluster:
+    name: jellyfin-postgres
+  instances: 2
+  pgbouncer:
+    parameters:
+      default_pool_size: "50"
+      max_client_conn: "200"
+    paused: false
+    poolMode: session
+  template:
+    metadata:
+      labels:
+        app: jellyfin-pooler
+    spec:
+      affinity:
+        podAntiAffinity:
+          requiredDuringSchedulingIgnoredDuringExecution:
+            - labelSelector:
+                matchExpressions:
+                  - key: app
+                    operator: In
+                    values:
+                      - jellyfin-pooler
+              topologyKey: kubernetes.io/hostname
+      containers: []
+  type: rw

apps/base/jellyfin/deployment.yaml

@@ -1,97 +0,0 @@
----
-apiVersion: apps/v1
-kind: Deployment
-metadata:
-  name: jellyfin
-  namespace: jellyfin
-spec:
-  # Single-replica for now. The jellyfin-ha fork adds the Redis transcode store
-  # and RWX transcode scratch that make scaling to true HA a follow-up.
-  replicas: 1
-  strategy:
-    # Config PVC is RWO; Recreate avoids two pods contending for it.
-    type: Recreate
-  selector:
-    matchLabels:
-      app: jellyfin
-  template:
-    metadata:
-      labels:
-        app: jellyfin
-    spec:
-      securityContext:
-        fsGroup: 1000
-      nodeSelector:
-        feature.node.kubernetes.io/pci-0300_8086.present: "true"
-      containers:
-        - name: jellyfin
-          image: git.unkin.net/unkin/jellyfin-ha:v0.1.0
-          imagePullPolicy: IfNotPresent
-          ports:
-            - name: http
-              containerPort: 8096
-              protocol: TCP
-          env:
-            - name: TZ
-              value: Australia/Sydney
-            - name: PUID
-              value: "1000"
-            - name: PGID
-              value: "1000"
-            - name: JELLYFIN_PublishedServerUrl
-              value: https://jellyfin.k8s.syd1.au.unkin.net
-            # Distributed transcode session store (jellyfin-ha additions).
-            - name: Jellyfin__TranscodeStore__RedisConnectionString
-              value: "jellyfin-redis:6379,abortConnect=false"
-            - name: Jellyfin__TranscodeStore__LeaseDurationSeconds
-              value: "30"
-          resources:
-            requests:
-              cpu: 100m
-              memory: 1Gi
-            limits:
-              cpu: "4"
-              memory: 8Gi
-              gpu.intel.com/i915: "1"
-          livenessProbe:
-            httpGet:
-              path: /health
-              port: http
-            initialDelaySeconds: 30
-            periodSeconds: 30
-            timeoutSeconds: 5
-            failureThreshold: 3
-          readinessProbe:
-            httpGet:
-              path: /health
-              port: http
-            initialDelaySeconds: 10
-            periodSeconds: 10
-            timeoutSeconds: 5
-            failureThreshold: 3
-          volumeMounts:
-            - name: config
-              mountPath: /config
-            - name: cache
-              mountPath: /cache
-            - name: transcode
-              mountPath: /transcode
-            - name: media-library
-              mountPath: /mnt/movies
-              subPath: movies
-            - name: media-library
-              mountPath: /mnt/tvseries
-              subPath: tvseries
-      volumes:
-        - name: config
-          persistentVolumeClaim:
-            claimName: jellyfin-config
-        - name: cache
-          persistentVolumeClaim:
-            claimName: jellyfin-cache
-        - name: transcode
-          persistentVolumeClaim:
-            claimName: jellyfin-transcode
-        - name: media-library
-          persistentVolumeClaim:
-            claimName: media-library

apps/base/jellyfin/httproute.yaml

@@ -26,7 +26,7 @@ spec:
 apiVersion: gateway.networking.k8s.io/v1
 kind: HTTPRoute
 metadata:
-  name: jellyfin
+  name: jellyfin-route
   namespace: jellyfin
 spec:
   hostnames:

apps/base/jellyfin/kustomization.yaml

@@ -4,15 +4,17 @@ kind: Kustomization
 
 resources:
   - namespace.yaml
-  - pv_media-library.yaml
-  - pvc_media-library.yaml
-  - pvc_config.yaml
-  - pvc_cache.yaml
-  - pvc_transcode.yaml
-  - deployment.yaml
+  - cnpg_cluster.yaml
+  - cnpg_pooler.yaml
+  - cnpg_backup.yaml
+  - pvc-config.yaml
+  - pvc-transcode.yaml
+  - pvc-media.yaml
+  - statefulset.yaml
+  - pdb.yaml
   - service.yaml
   - redis-deployment.yaml
-  - redis-service.yaml
   - redis-pvc.yaml
+  - redis-service.yaml
   - gateway.yaml
   - httproute.yaml

apps/base/jellyfin/pdb.yaml

@@ -0,0 +1,13 @@
+---
+# Keep at least one Jellyfin replica serving through voluntary disruptions
+# (node drains, rollouts) so active streams can fail over rather than drop.
+apiVersion: policy/v1
+kind: PodDisruptionBudget
+metadata:
+  name: jellyfin
+  namespace: jellyfin
+spec:
+  minAvailable: 1
+  selector:
+    matchLabels:
+      app: jellyfin

apps/base/jellyfin/pv_media-library.yaml

@@ -1,41 +0,0 @@
----
-# Static CephFS PersistentVolume bound to the pre-existing, ACTIVELY-USED
-# puppet media library (ceph filesystem "mediafs", mounted by the VM/incus
-# instances at /shared/media). ceph-csi only mounts this volume; staticVolume
-# tells it the storage pre-exists and it must never provision or delete it.
-#
-# reclaimPolicy MUST stay Retain: deleting this PV or its PVC must NEVER be able
-# to reclaim or destroy the underlying CephFS data that the VM instances use.
-apiVersion: v1
-kind: PersistentVolume
-metadata:
-  name: jellyfin-media-library
-spec:
-  accessModes:
-    - ReadWriteMany
-  capacity:
-    storage: 10Ti
-  # Load-bearing safety control. Do not change to Delete.
-  persistentVolumeReclaimPolicy: Retain
-  storageClassName: ""
-  volumeMode: Filesystem
-  # Pre-bind to the media-library claim so nothing else can grab this PV.
-  claimRef:
-    apiVersion: v1
-    kind: PersistentVolumeClaim
-    name: media-library
-    namespace: jellyfin
-  csi:
-    driver: cephfs.csi.ceph.com
-    volumeHandle: jellyfin-media-library-static
-    nodeStageSecretRef:
-      name: csi-cephfs-secret
-      namespace: csi-cephfs
-    volumeAttributes:
-      # clusterID maps (in the csi-cephfs ceph-csi-config) to the mon set that
-      # also serves mediafs; for a static volume only the mon lookup is used.
-      clusterID: cephfs_csi_ssd_ec_4_1
-      fsName: mediafs
-      staticVolume: "true"
-      # Filesystem-internal root of the library (mediafs root == /shared/media).
-      rootPath: /

apps/base/jellyfin/pvc-config.yaml

@@ -0,0 +1,17 @@
+---
+# Jellyfin config: metadata images, plugins, subtitles and config XML. Shared
+# ReadWriteMany across replicas (all pods read/write the same library metadata);
+# the main library DB now lives in PostgreSQL, not here. Retain — this is state.
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: jellyfin-config
+  namespace: jellyfin
+spec:
+  accessModes:
+    - ReadWriteMany
+  resources:
+    requests:
+      storage: 20Gi
+  storageClassName: cephfs-raid5-retain
+  volumeMode: Filesystem

apps/base/jellyfin/pvc-media.yaml

@@ -0,0 +1,17 @@
+---
+# Media library, shared read-many across replicas. Retain — this holds the
+# actual media and must survive PVC deletion. Empty on first deploy; populating
+# it is out of scope for this app.
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: jellyfin-media
+  namespace: jellyfin
+spec:
+  accessModes:
+    - ReadWriteMany
+  resources:
+    requests:
+      storage: 1Ti
+  storageClassName: cephfs-raid6-retain
+  volumeMode: Filesystem

apps/base/jellyfin/pvc-transcode.yaml

@@ -1,7 +1,8 @@
 ---
 # Shared transcode scratch. ReadWriteMany is the hard requirement for the HA
 # fork: a taking-over pod must read the in-flight HLS segments written by the
-# pod it replaces. Scratch data, so delete reclaim policy.
+# pod it replaces. Scratch data (delete reclaim); raid5 avoids the raid6
+# double-parity write penalty on the many small HLS segment writes.
 apiVersion: v1
 kind: PersistentVolumeClaim
 metadata:

apps/base/jellyfin/pvc_cache.yaml

@@ -1,15 +0,0 @@
----
-# Local transcode/image cache. Scratch, delete reclaim policy.
-apiVersion: v1
-kind: PersistentVolumeClaim
-metadata:
-  name: jellyfin-cache
-  namespace: jellyfin
-spec:
-  accessModes:
-    - ReadWriteOnce
-  resources:
-    requests:
-      storage: 200Gi
-  storageClassName: cephrbd-fast-delete
-  volumeMode: Filesystem

apps/base/jellyfin/pvc_config.yaml

@@ -1,15 +0,0 @@
----
-# Jellyfin config + SQLite library database. Single-writer, block storage.
-apiVersion: v1
-kind: PersistentVolumeClaim
-metadata:
-  name: jellyfin-config
-  namespace: jellyfin
-spec:
-  accessModes:
-    - ReadWriteOnce
-  resources:
-    requests:
-      storage: 10Gi
-  storageClassName: cephrbd-fast-retain
-  volumeMode: Filesystem

apps/base/jellyfin/pvc_media-library.yaml

@@ -1,19 +0,0 @@
----
-# Claim bound to the static mediafs PV. RWX so every app in the stack shares the
-# one library. storageClassName "" + volumeName pin it to the static PV (no
-# dynamic provisioning). Deleting this claim cannot reclaim the data (PV is
-# Retain).
-apiVersion: v1
-kind: PersistentVolumeClaim
-metadata:
-  name: media-library
-  namespace: jellyfin
-spec:
-  accessModes:
-    - ReadWriteMany
-  storageClassName: ""
-  volumeName: jellyfin-media-library
-  resources:
-    requests:
-      storage: 10Ti
-  volumeMode: Filesystem

apps/base/jellyfin/redis-deployment.yaml

@@ -2,21 +2,20 @@
 apiVersion: apps/v1
 kind: Deployment
 metadata:
-  name: jellyfin-redis
+  name: redis
   namespace: jellyfin
 spec:
   replicas: 1
-  strategy:
-    type: Recreate
   selector:
     matchLabels:
-      app: jellyfin-redis
+      app: redis
+  strategy:
+    type: Recreate
   template:
     metadata:
       labels:
-        app: jellyfin-redis
+        app: redis
     spec:
-      restartPolicy: Always
       containers:
         - name: redis
           image: redis:7-alpine
@@ -27,37 +26,40 @@ spec:
             - "20"
             - "1"
           ports:
-            - name: redis
-              containerPort: 6379
+            - containerPort: 6379
+              name: redis
               protocol: TCP
           livenessProbe:
             exec:
               command:
                 - redis-cli
                 - ping
+            failureThreshold: 3
             initialDelaySeconds: 30
             periodSeconds: 30
+            successThreshold: 1
             timeoutSeconds: 5
-            failureThreshold: 3
           readinessProbe:
             exec:
               command:
                 - redis-cli
                 - ping
+            failureThreshold: 3
             initialDelaySeconds: 5
             periodSeconds: 10
+            successThreshold: 1
             timeoutSeconds: 5
-            failureThreshold: 3
           resources:
-            requests:
-              cpu: 50m
-              memory: 128Mi
             limits:
               cpu: 500m
               memory: 512Mi
+            requests:
+              cpu: 50m
+              memory: 128Mi
           volumeMounts:
-            - name: data
-              mountPath: /data
+            - mountPath: /data
+              name: data
+      restartPolicy: Always
       volumes:
         - name: data
           persistentVolumeClaim:

apps/base/jellyfin/redis-service.yaml

@@ -2,16 +2,16 @@
 apiVersion: v1
 kind: Service
 metadata:
-  name: jellyfin-redis
+  name: redis
   namespace: jellyfin
 spec:
-  type: ClusterIP
   internalTrafficPolicy: Cluster
-  sessionAffinity: None
-  selector:
-    app: jellyfin-redis
   ports:
     - name: redis
       port: 6379
-      targetPort: redis
       protocol: TCP
+      targetPort: redis
+  selector:
+    app: redis
+  sessionAffinity: None
+  type: ClusterIP

apps/base/jellyfin/service.yaml

@@ -5,13 +5,14 @@ metadata:
   name: jellyfin
   namespace: jellyfin
 spec:
-  type: ClusterIP
   internalTrafficPolicy: Cluster
-  sessionAffinity: None
-  selector:
-    app: jellyfin
   ports:
     - name: http
       port: 8096
-      targetPort: http
       protocol: TCP
+      targetPort: http
+  selector:
+    app: jellyfin
+  # Pin each client to one replica to reduce transcode-session churn/takeover.
+  sessionAffinity: ClientIP
+  type: ClusterIP

apps/base/jellyfin/statefulset.yaml

@@ -0,0 +1,199 @@
+---
+apiVersion: apps/v1
+kind: StatefulSet
+metadata:
+  name: jellyfin
+  namespace: jellyfin
+spec:
+  # HA: two replicas coordinate transcode session ownership through Redis and
+  # resume each other's HLS segments off the shared RWX transcode PVC. Stable
+  # pod names (jellyfin-0/1) are the lease owner identity, hence StatefulSet.
+  replicas: 2
+  serviceName: jellyfin
+  podManagementPolicy: Parallel
+  updateStrategy:
+    type: RollingUpdate
+  selector:
+    matchLabels:
+      app: jellyfin
+  template:
+    metadata:
+      labels:
+        app: jellyfin
+    spec:
+      securityContext:
+        # Group-write the shared RWX volumes and grant the render/video groups so
+        # the runAsUser 1000 process can open the Intel DRI render node injected
+        # by the device plugin.
+        fsGroup: 1000
+        supplementalGroups:
+          - 44
+          - 105
+          - 109
+        seccompProfile:
+          type: RuntimeDefault
+      affinity:
+        # Spread the two replicas across nodes for node-level HA. Soft so a
+        # single-GPU-node cluster still schedules both (i915 has 4 shared slots).
+        podAntiAffinity:
+          preferredDuringSchedulingIgnoredDuringExecution:
+            - weight: 100
+              podAffinityTerm:
+                labelSelector:
+                  matchLabels:
+                    app: jellyfin
+                topologyKey: kubernetes.io/hostname
+      initContainers:
+        # Select the fork's experimental PostgreSQL provider by writing
+        # database.xml before Jellyfin starts. Runs as root to chown into the
+        # shared config volume; mirrors the fork Helm chart's inject-db-config.
+        - name: inject-db-config
+          image: busybox:1.37.0
+          command:
+            - sh
+            - -c
+            - |
+              mkdir -p /config/config
+              chown 1000:1000 /config/config
+              chmod 775 /config/config
+              cat > /config/config/database.xml << 'DBEOF'
+              <?xml version="1.0" encoding="utf-8"?>
+              <DatabaseConfigurationOptions>
+                <DatabaseType>Jellyfin-PostgreSQL</DatabaseType>
+                <LockingBehavior>NoLock</LockingBehavior>
+              </DatabaseConfigurationOptions>
+              DBEOF
+              chown 1000:1000 /config/config/database.xml
+              chmod 664 /config/config/database.xml
+          resources:
+            requests:
+              cpu: 10m
+              memory: 32Mi
+            limits:
+              cpu: 100m
+              memory: 64Mi
+          volumeMounts:
+            - name: config
+              mountPath: /config
+      containers:
+        - name: jellyfin
+          image: git.unkin.net/unkin/jellyfin-ha:v0.1.0
+          imagePullPolicy: IfNotPresent
+          ports:
+            - name: http
+              containerPort: 8096
+              protocol: TCP
+          env:
+            # Pod identity for the Redis transcode lease owner. The fork reads
+            # JELLYFIN_INSTANCE_ID (falling back to MachineName); the stable
+            # StatefulSet pod name gives each replica a unique lease identity so
+            # takeover can target a dead replica. JELLYFIN_HA_POD_NAME is set for
+            # parity with the fork Helm chart (nothing currently reads it).
+            - name: JELLYFIN_INSTANCE_ID
+              valueFrom:
+                fieldRef:
+                  fieldPath: metadata.name
+            - name: JELLYFIN_HA_POD_NAME
+              valueFrom:
+                fieldRef:
+                  fieldPath: metadata.name
+            # Multiple replicas must not each answer UDP auto-discovery.
+            - name: JELLYFIN_Network__AutoDiscovery
+              value: "false"
+            # Config dir must differ from the data root (Jellyfin sanity check).
+            - name: JELLYFIN_CONFIG_DIR
+              value: /config/config
+            # Distributed transcode session store (jellyfin-ha additions).
+            - name: Jellyfin__TranscodeStore__RedisConnectionString
+              value: "redis:6379,abortConnect=false"
+            - name: Jellyfin__TranscodeStore__LeaseDurationSeconds
+              value: "30"
+            # PostgreSQL main DB via the CNPG-generated app secret, routed through
+            # the PgBouncer pooler. Composed with $(VAR) expansion so the password
+            # is never rendered into the manifest; CNPG passwords are URL-safe.
+            - name: PGUSER
+              valueFrom:
+                secretKeyRef:
+                  name: jellyfin-postgres-app
+                  key: username
+            - name: PGPASSWORD
+              valueFrom:
+                secretKeyRef:
+                  name: jellyfin-postgres-app
+                  key: password
+            - name: PGDB
+              valueFrom:
+                secretKeyRef:
+                  name: jellyfin-postgres-app
+                  key: dbname
+            - name: POSTGRES_CONNECTION_STRING
+              value: "postgresql://$(PGUSER):$(PGPASSWORD)@jellyfin-postgres-pooler:5432/$(PGDB)"
+            - name: DATABASE_URL
+              value: "postgresql://$(PGUSER):$(PGPASSWORD)@jellyfin-postgres-pooler:5432/$(PGDB)"
+          livenessProbe:
+            httpGet:
+              path: /health
+              port: http
+            initialDelaySeconds: 30
+            periodSeconds: 30
+            timeoutSeconds: 5
+            failureThreshold: 3
+          readinessProbe:
+            httpGet:
+              path: /health
+              port: http
+            initialDelaySeconds: 10
+            periodSeconds: 10
+            timeoutSeconds: 5
+            failureThreshold: 3
+          resources:
+            requests:
+              cpu: "1"
+              memory: 1Gi
+              gpu.intel.com/i915: "1"
+            limits:
+              cpu: "4"
+              memory: 6Gi
+              # Intel iGPU (QSV/VA-API) slot. Requesting it pins the pod to a
+              # GPU-labelled node and injects /dev/dri/renderD* automatically, so
+              # no /dev/dri hostPath or privileged container is needed. Enable
+              # QSV/VA-API once in the Jellyfin admin UI; it persists to /config.
+              gpu.intel.com/i915: "1"
+          securityContext:
+            runAsUser: 1000
+            runAsGroup: 1000
+          volumeMounts:
+            - name: config
+              mountPath: /config
+            - name: transcode
+              # Fork's real transcode temp path. RWX so a surviving pod reads the
+              # in-flight .ts/.m3u8 segments of the pod it takes over. A per-pod
+              # volume here silently breaks HA takeover.
+              mountPath: /config/transcodes
+            - name: cache
+              mountPath: /cache
+            - name: media
+              mountPath: /media
+              readOnly: true
+      volumes:
+        - name: config
+          persistentVolumeClaim:
+            claimName: jellyfin-config
+        - name: transcode
+          persistentVolumeClaim:
+            claimName: jellyfin-transcode
+        - name: media
+          persistentVolumeClaim:
+            claimName: jellyfin-media
+  volumeClaimTemplates:
+    # Per-pod scratch cache — RWO, disposable, one PVC per replica.
+    - metadata:
+        name: cache
+      spec:
+        accessModes:
+          - ReadWriteOnce
+        resources:
+          requests:
+            storage: 30Gi
+        storageClassName: cephrbd-fast-delete
+        volumeMode: Filesystem

apps/base/puppet/deployment_puppetserver-master.yaml

@@ -4,6 +4,8 @@ metadata:
   annotations:
     configmap.reloader.stakater.com/auto: "true"
     secret.reloader.stakater.com/reload: "vault-ca-cert"
+    # Replace clears the stale, API-server-defaulted spec.strategy.rollingUpdate that SSA cannot drop, which otherwise makes Recreate invalid.
+    argocd.argoproj.io/sync-options: Replace=true
   labels:
     app.kubernetes.io/component: puppetserver
     app.kubernetes.io/instance: puppetserver

argocd/applicationsets/media.yaml

@@ -6,12 +6,11 @@ metadata:
   namespace: argocd
 spec:
   generators:
-    - git:
-        repoURL: https://git.unkin.net/unkin/argocd-apps
-        revision: HEAD
-        directories:
-          # jellyfin only for now; downloads/managers apps join in later PRs.
-          - path: apps/overlays/*/jellyfin
+  - git:
+      repoURL: https://git.unkin.net/unkin/argocd-apps
+      revision: HEAD
+      directories:
+      - path: apps/overlays/*/jellyfin
   template:
     metadata:
       name: 'media-{{path[3]}}'
@@ -23,10 +22,10 @@ spec:
         path: '{{path}}'
       destination:
         server: https://kubernetes.default.svc
-        namespace: jellyfin
+        namespace: '{{path[3]}}'
       syncPolicy:
         automated:
           prune: true
           selfHeal: true
         syncOptions:
-          - ServerSideApply=true
+        - ServerSideApply=true

argocd/projects/media.yaml

@@ -5,17 +5,15 @@ metadata:
   name: media
   namespace: argocd
 spec:
-  description: Media services (jellyfin; downloads/managers namespaces to follow)
+  description: Media services
   sourceRepos:
     - https://git.unkin.net/unkin/argocd-apps
   destinations:
-    - namespace: jellyfin
+    - namespace: 'jellyfin'
       server: https://kubernetes.default.svc
   clusterResourceWhitelist:
     - group: ''
       kind: Namespace
-    - group: ''
-      kind: PersistentVolume
   namespaceResourceWhitelist:
     - group: '*'
       kind: '*'

clusters/au-syd1/bootstrap/argocd-cm-patch.yaml

@@ -6,6 +6,16 @@ metadata:
   namespace: argocd
 data:
   kustomize.buildOptions: "--enable-helm"
+  # Kubernetes defaults apiVersion/kind onto every StatefulSet
+  # volumeClaimTemplates entry, but neither the raw manifests nor the helm
+  # charts emit them, so live StatefulSets carry TypeMeta that git lacks.
+  # volumeClaimTemplates are immutable on an existing StatefulSet, so ArgoCD
+  # can never reconcile the removal and the resource stays perpetually
+  # OutOfSync. Ignore the defaulted TypeMeta fleet-wide.
+  resource.customizations.ignoreDifferences.apps_StatefulSet: |
+    jqPathExpressions:
+      - '.spec.volumeClaimTemplates[]?.apiVersion'
+      - '.spec.volumeClaimTemplates[]?.kind'
   # External URL ArgoCD serves on (TLS terminated at the traefik-internal gateway).
   url: https://argocd.k8s.syd1.au.unkin.net
   # OIDC login via Authentik. The client secret is seeded in Vault out of band