🐮 longhorn-scheduler

This is a fork of linstor-scheduler-extender modified to work with Longhorn volumes. It uses Stork to schedule pods alongside existing replicas for optimal performance.

This is in contrast to Longhorn's built-in data locality functionality, which only affects pods that have already been scheduled, and which re-allocates replicas to achieve locality.

Behavior

Scheduling is best-effort -- if no replica nodes have capacity the pod will be scheduled without locality.

In situations where a pod has multiple Longhorn volumes attached, preference is given to nodes with more replicas.

RWX volumes will attempt to schedule pods on the same node as the NFS share manager, but the share manager may not have data locality unless the admission controller is also deployed (see below).

Deployment

kubectl apply -f deploy/longhorn-scheduler.yaml

Usage

Use schedulerName: longhorn in your pod spec to take advantage of the scheduler.

Admission controller

You can also deploy the admission controller to automatically assign schedulerName: longhorn to all pods with Longhorn volumes:

kubectl apply -f deploy/longhorn-scheduler-admission.yaml