This message was deleted.

Oh it looks like it's because nodeName is used which bypasses the configured scheduler and therefore I guess bypasses pre-emption.

There are lots of legacy issues for deploying instance manager, including using

nodeName

for each instance manager pods on all nodes rather than simply deploying a daemonset. Notice that each worker node (that will use Longhorn volume) needs a instance manager pods. Can avoiding using

nodeName

help instance manager be launched on the node that has no enough CPU? If you don't want to change your existing workloads but still need Longhorn on this node, you can set

node.spec.instanceManagerCPURequest

(by executing

kubectl -n longhorn-system edit lhn <node name>

) for this individual node.

Thank you for the reply. I was wondering if nodeName had some advantages but it sounds like it's just tech debt! We got it scheduled by evicting other pods from the node so no issues anymore.