Files
argocd-apps/apps/base/jellyfin/statefulset.yaml
T
unkin-agent 94da740a9d Add cheeztv kids Jellyfin instance + fafflix kids mounts (#413)
## Why
Kids need their own safe Jellyfin front (cheeztv.unkin.net) with a library that only ever exposes kids titles, while those same titles stay resumable in the existing adult instance (fafflix). The fafflix.unkin.net domain switch is deliberately out of scope.

## How
- **New `cheeztv` instance** (`apps/base/cheeztv`): mirrors the fafflix HA Jellyfin stack 1:1 - same `jellyfin-ha:v0.1.3` fork image, 2-replica StatefulSet, RWX transcode PVC + HA takeover, Intel iGPU transcode, CNPG Postgres + PgBouncer, Valkey transcode store, k8up config backup, VMPodScrape.
- **Kids-only media**: its own static PVs on the *same* shared movies/tv CephFS subvolumes, mounted with `subPath: kids` so it sees only `movies/kids` and `tvshows/kids`.
- **Fully separate state**: own namespace, config PVC, `cheeztv-postgres` cluster, `cheeztv-valkey`, and `cnpg-cheeztv` / `cheeztv-config-backup` RGW buckets - nothing shared with fafflix.
- **DNS/cert**: `cheeztv.unkin.net` via internal Traefik + external-dns (198.18.200.4) + cert-manager `cheeztv-tls`, matching logviewer.unkin.net; plus the cluster hostname `cheeztv.k8s.syd1.au.unkin.net` mirroring fafflix.
- **fafflix change**: ADDS `movies/kids` + `tvshows/kids` subPath mounts alongside its current media mounts so kids libraries are resumable there. Existing mounts, hostname and ingress untouched.
- Registers cheeztv in the media ApplicationSet generator and AppProject destinations.

## Validation
kubeconform clean: cheeztv 28/28, jellyfin 25/25 valid. No arr/watchstate/domain changes.

Note: the k8up restic repo password must be seeded at `kv/kubernetes/namespace/cheeztv/default/k8up-restic` (same manual step as fafflix); templated Vault policy already grants read, no terraform-vault change needed.
Reviewed-on: #413
Co-authored-by: unkin-agent <unkin-agent@unkin.net>
Co-committed-by: unkin-agent <unkin-agent@unkin.net>
2026-08-25 19:59:08 +10:00

258 lines
10 KiB
YAML

---
apiVersion: apps/v1
kind: StatefulSet
metadata:
name: jellyfin
namespace: jellyfin
spec:
# HA: two replicas coordinate transcode session ownership through Valkey 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:
# Seed the fork's PostgreSQL provider (database.xml) and Intel iGPU
# hardware transcode settings (encoding.xml) before Jellyfin starts.
# Runs as root to chown into the shared config volume; mirrors the fork
# Helm chart's inject-db-config. Each file is written only when absent so
# admin changes persisted to the shared RWX /config survive pod restarts.
- name: inject-config
image: busybox:1.37.0
command:
- sh
- -c
- |
mkdir -p /config/config
chown 1000:1000 /config/config
chmod 775 /config/config
if [ ! -e /config/config/database.xml ]; then
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
fi
# VAAPI on the Intel render node the device plugin injects
# (/dev/dri/renderD128 — ffmpeg's default DRM node, reachable via
# the render/video supplementalGroups). Without this the attached
# iGPU is idle and every transcode runs in software. Omitted
# elements fall back to the fork's EncodingOptions defaults.
if [ ! -e /config/config/encoding.xml ]; then
cat > /config/config/encoding.xml << 'ENCEOF'
<?xml version="1.0" encoding="utf-8"?>
<EncodingOptions xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:xsd="http://www.w3.org/2001/XMLSchema">
<EncodingThreadCount>-1</EncodingThreadCount>
<HardwareAccelerationType>vaapi</HardwareAccelerationType>
<VaapiDevice>/dev/dri/renderD128</VaapiDevice>
<EnableHardwareEncoding>true</EnableHardwareEncoding>
<AllowHevcEncoding>true</AllowHevcEncoding>
<AllowAv1Encoding>false</AllowAv1Encoding>
<EnableIntelLowPowerH264HwEncoder>false</EnableIntelLowPowerH264HwEncoder>
<EnableIntelLowPowerHevcHwEncoder>false</EnableIntelLowPowerHevcHwEncoder>
<EnableTonemapping>false</EnableTonemapping>
<EnableVppTonemapping>false</EnableVppTonemapping>
<HardwareDecodingCodecs>
<string>h264</string>
<string>hevc</string>
<string>vc1</string>
<string>vp9</string>
</HardwareDecodingCodecs>
</EncodingOptions>
ENCEOF
chown 1000:1000 /config/config/encoding.xml
chmod 664 /config/config/encoding.xml
fi
resources:
requests:
cpu: 10m
memory: 32Mi
limits:
cpu: 100m
memory: 64Mi
volumeMounts:
- name: config
mountPath: /config
containers:
- name: jellyfin
image: artifactapi.k8s.syd1.au.unkin.net/docker-internal/jellyfin-ha:v0.1.3
imagePullPolicy: IfNotPresent
ports:
- name: http
containerPort: 8096
protocol: TCP
env:
# Pod identity for the Valkey 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: "valkey-jellyfin-valkey: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. VA-API is
# pre-enabled via the seeded encoding.xml (see inject-config), so
# transcodes use the iGPU on first boot with no manual UI step.
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-tv
mountPath: /media/tv
readOnly: true
- name: media-movies
mountPath: /media/movies
readOnly: true
# Kids subtrees surfaced as their own paths (same media PVCs, subPath
# kids) so a "Kids TV"/"Kids Movies" library can be added here and its
# titles are browsable and resume in this instance's own DB. The full
# /media/{tv,movies} mounts above are unchanged.
- name: media-tv
mountPath: /media/tv-kids
subPath: kids
readOnly: true
- name: media-movies
mountPath: /media/movies-kids
subPath: kids
readOnly: true
volumes:
- name: config
persistentVolumeClaim:
claimName: jellyfin-config
- name: transcode
persistentVolumeClaim:
claimName: jellyfin-transcode
- name: media-tv
persistentVolumeClaim:
claimName: jellyfin-media-tv
- name: media-movies
persistentVolumeClaim:
claimName: jellyfin-media-movies
volumeClaimTemplates:
# Per-pod scratch cache — RWO, disposable, one PVC per replica.
- metadata:
name: cache
annotations:
# Exclude the per-pod cache PVCs from the jellyfin-config k8up Schedule
# (skipWithoutAnnotation is false cluster-wide). Cache is disposable and
# RWO — it would also fail to mount into the backup pod while in use.
k8up.io/backup: "false"
spec:
accessModes:
- ReadWriteOnce
resources:
requests:
storage: 30Gi
storageClassName: cephrbd-fast-delete
volumeMode: Filesystem