From 9450a706524fcb4c99feb1da6c289a7946649f7a Mon Sep 17 00:00:00 2001 From: unkin-agent Date: Sun, 16 Aug 2026 07:34:15 +1000 Subject: [PATCH] jellyfin: make Valkey HA (#375) Jellyfin runs 2 replicas that coordinate distributed transcoding through the operator-managed `jellyfin-valkey` ValkeyCluster, but it was single-instance (shards:1, replicas:0) with no failover, so a node/pod loss would drop the shared transcode-lease state. This promotes that store to an HA topology. Changes: - Set replicas:2 (shards:1 unchanged) so the shard group is one primary + two replicas; losing the primary triggers an automatic in-cluster failover to a replica. - Add scheduling.node.spread.shard.mode:Required so the three ValkeyNodes land on distinct hosts (kubernetes.io/hostname anti-affinity) and one host loss removes at most one node. - Set podDisruptionBudget.mode:Cluster so the operator manages a quorum-aware PDB. - Keep persistence omitted (emptyDir): lease state is ephemeral (30s TTLs), replication+failover already provide redundancy and a replaced node re-syncs from the primary, and an operator-managed PVC cannot carry the k8up.io/backup:"false" annotation the namespace k8up Schedule needs to skip in-use RWO volumes. - Bump per-node memory limit to 512Mi (unchanged) / requests unchanged; resources are per ValkeyNode. Connection string is unchanged: still cluster-mode via the single headless service `valkey-jellyfin-valkey:6379`; StackExchange.Redis seeds off it and auto-discovers topology and failovers. kustomize build --enable-helm apps/overlays/au-syd1/jellyfin renders cleanly; pre-commit passes. Reviewed-on: https://git.unkin.net/unkin/argocd-apps/pulls/375 Co-authored-by: unkin-agent Co-committed-by: unkin-agent --- apps/base/jellyfin/valkey.yaml | 36 +++++++++++++++++++++++----------- 1 file changed, 25 insertions(+), 11 deletions(-) diff --git a/apps/base/jellyfin/valkey.yaml b/apps/base/jellyfin/valkey.yaml index bd018a3..147895b 100644 --- a/apps/base/jellyfin/valkey.yaml +++ b/apps/base/jellyfin/valkey.yaml @@ -1,14 +1,21 @@ --- -# Single-instance Valkey managed by valkey-operator, replacing the self-managed -# redis Deployment as Jellyfin's distributed transcode-session store. shards:1 + -# replicas:0 is one primary with no replicas. This operator always runs Valkey -# cluster-mode-enabled with protected-mode off and leaves the built-in `default` -# user passwordless, so clients connect with no auth/TLS (parity with the old -# open redis); StackExchange.Redis auto-discovers the single node. HA (more -# shards/replicas) is a follow-up. Persistence is intentionally omitted so /data -# is an emptyDir: the lease state is ephemeral, and the operator-managed PVC -# cannot carry the k8up.io/backup:"false" annotation the namespace's cluster-wide -# k8up Schedule needs to skip an in-use RWO volume. +# HA Valkey managed by valkey-operator as Jellyfin's distributed transcode-session +# store. shards:1 + replicas:2 is one primary with two replicas in a single shard +# group (three ValkeyNodes total): losing the primary triggers an automatic +# failover to a replica, so a node/pod loss no longer drops the shared transcode +# lease state the two Jellyfin replicas coordinate through. The operator always +# runs Valkey cluster-mode-enabled with protected-mode off and leaves the built-in +# `default` user passwordless, so clients connect with no auth/TLS (parity with the +# old open redis); StackExchange.Redis seeds off the single headless service and +# auto-discovers the topology plus failovers, so the connection string is +# unchanged. scheduling.node.spread.shard:Required renders pod anti-affinity that +# keeps the three nodes on distinct hosts (kubernetes.io/hostname), so one host +# loss removes at most one node. podDisruptionBudget.mode:Cluster lets the operator +# manage a quorum-aware PDB. Persistence stays omitted so /data is an emptyDir: +# lease state is ephemeral (30s TTLs), replication+failover already provide +# redundancy and a replaced node re-syncs from the primary, and an operator-managed +# PVC cannot carry the k8up.io/backup:"false" annotation the namespace's +# cluster-wide k8up Schedule needs to skip in-use RWO volumes. apiVersion: valkey.io/v1alpha1 kind: ValkeyCluster metadata: @@ -16,10 +23,17 @@ metadata: namespace: jellyfin spec: shards: 1 - replicas: 0 + replicas: 2 image: artifactapi.k8s.syd1.au.unkin.net/dockerhub/valkey/valkey:9.0.0 exporter: enabled: false + scheduling: + node: + spread: + shard: + mode: Required + podDisruptionBudget: + mode: Cluster resources: requests: cpu: 50m