This document describes how to configure pod anti-affinity for Longhorn’s CSI components. This feature enhances the resilience of the storage system, particularly in smaller clusters, by preventing multiple replicas of a CSI component from running on the same node and thereby avoiding a single point of failure.

For more information on pod anti-affinity, refer to the official Kubernetes documentation on Inter-pod affinity and anti-affinity.

Configuring Pod Anti-Affinity for Longhorn CSI Components

You can configure the pod anti-affinity for the following Longhorn CSI components:

• csi-attacher
• csi-provisioner
• csi-resizer
• csi-snapshotter

The podAntiAffinityPreset setting has two available values:

• soft (default): This is a “best-effort” rule using preferredDuringSchedulingIgnoredDuringExecution. The scheduler will try to prevent multiple CSI component replicas from running on the same node, but it is not guaranteed.
• hard: This is a “strict” rule using requiredDuringSchedulingIgnoredDuringExecution. The scheduler will not schedule a pod on a node if it violates the anti-affinity rule. This may result in pods remaining in a Pending state if no suitable nodes are available.

During Longhorn Installation

You can set the pod anti-affinity for the CSI components during the initial installation of Longhorn using one of the following methods.

Using Rancher

When installing Longhorn through Rancher, add the following parameters to the YAML on the Rancher UI (click Edit as YAML during the installation).

csi:
    podAntiAffinityPreset: "hard"

Using Helm

If installing with Helm, set the csi.podAntiAffinityPreset value in your values.yaml file, and then install the chart as usual.

Using Kubectl

If installing Longhorn with kubectl and the deployment YAML, you must manually edit the longhorn-driver-deployer deployment. Add the following environment variable to the container specification:

- name: CSI_POD_ANTIAFFINITY_PRESET
  value: hard

After Longhorn has been installed

Warning:

• longhorn-driver-deployer and the CSI pods will be redeployed.

Manually edit the longhorn-driver-deployer deployment using kubectl. Add the following environment variable to the container specification:

- name: CSI_POD_ANTIAFFINITY_PRESET
  value: hard

History

• Original feature request