mirror of
https://github.com/cozystack/cozystack.git
synced 2026-11-09 14:02:56 +09:00
Neither the Keycloak StatefulSet nor the keycloak-kms-proxy Deployment exposed any placement, so the scheduler was free to put both replicas of each in one failure domain. For Keycloak that is worse than it looks: a StatefulSet does not replace a pod stuck on an unreachable node, so losing that domain takes every replica for as long as it stays lost. Expose affinity and topologySpreadConstraints for both, empty by default so nothing changes unless an installation opts in, and give the proxy a PodDisruptionBudget once it runs more than one replica. Assisted-By: LLM Signed-off-by: Kirill Ilin <stitch14@yandex.ru>