Skip to content
Features
Anti-affinity

Anti-affinity

Configure it onDockerDocker SwarmKubernetesPodmanProxmox LXCSystemdNomadAWS ECS

Make an instance back off automatically while another group is in use, with the anti-affinity setting. It lists the groups the instance yields to.

Use it when heavy services compete for a resource that cannot be shared, most often GPU memory or RAM, and running them together causes an out-of-memory crash or a severe slowdown. For example, a Nextcloud instance with anti-affinity set to streaming frees the GPU whenever the streaming group, Plex, has an active session.

    flowchart LR
    stream["streaming session<br/>becomes active"]
    idle["nextcloud forced idle<br/>stopped or throttled"]
    restored["nextcloud restored<br/>started or active profile"]
    stream -->|anti-affinity: streaming| idle
    idle -->|streaming session expires| restored

How it works:

  • When a session for a listed group becomes active, Sablier forces the instance to its idle state:
    • a plain instance is stopped (paused with the Docker pause strategy);
    • an instance in scale mode gets its idle resource profile instead.
  • When that session expires and no other listed group is active, Sablier restores the instances it forced idle: it starts them again, or applies their active profile.
  • You can list several groups. The instance stays idle while any of them is active.

Behavior notes:

  • Only instances Sablier actually forced idle while they were running are restored. An instance that was already idle is left untouched.
  • The relationship is one-directional: the instance yields to the group, not the reverse.
  • A request for a held instance does not start it. While a listed group is active, the instance is reported as not-ready with the message “paused while group "streaming" is active (anti-affinity)”, shown on the waiting page (with show_details) and in the API response. A blocking request keeps waiting and times out if the other group outlasts its timeout. The dynamic waiting page keeps refreshing and starts the instance once the other group’s session expires.
  • The instance must be managed by Sablier (its enable setting is true), so Sablier can stop and start it.
  • Sablier reconciles on session start and expiry, with a periodic safety net.
GroupsReady after