Allow min_node_count to be zero

[gtirloni]

The current behavior sets min_node_count to 1 if the user specifies it should be zero. Zero is supported by both Magnum and the Cluster Autoscaler.

This sets it to the default value 1 (same as Magnum when min_node_count is not specified) only when it was not specified.

Related:

• Magnum: https://review.opendev.org/c/openstack/magnum/+/737580
• Autoscaler: Enable magnum provider scale to zero kubernetes/autoscaler#3995

[gtirloni]

This unblocks the scaling down but scaling up later doesn't work because the autoscaler can't find a nodegroup to scale. That seems to happen because it iterates over all the existing nodes to find one that has a nodegroup it can scale up... since the nodegroup node_count is currently zero, there are not nodes (and it correctly ignores the controlplane node):

I0715 18:03:59.737912       1 filter_out_schedulable.go:63] Filtering out schedulables
I0715 18:03:59.738021       1 klogx.go:87] failed to find place for kube-system/csi-cinder-controllerplugin-cd7ffbdf9-ml59v: cannot put pod csi-cinder-controllerplugin-cd7ffbdf9-ml59v on any node
I0715 18:03:59.738092       1 klogx.go:87] failed to find place for kube-system/openstack-manila-csi-controllerplugin-0: cannot put pod openstack-manila-csi-controllerplugin-0 on any node
I0715 18:03:59.738149       1 klogx.go:87] failed to find place for test/force-autoscaler: cannot put pod force-autoscaler on any node
I0715 18:03:59.738168       1 filter_out_schedulable.go:120] 0 pods marked as unschedulable can be scheduled.
I0715 18:03:59.738191       1 filter_out_schedulable.go:83] No schedulable pods
I0715 18:03:59.738204       1 filter_out_daemon_sets.go:40] Filtering out daemon set pods
I0715 18:03:59.738218       1 filter_out_daemon_sets.go:49] Filtered out 0 daemon set pods, 3 unschedulable pods left

And later (correctly ignoring controplane nodes):

I0715 18:03:59.739693       1 pre_filtering_processor.go:57] Node kube-e9yct-t45jx-7g8rp should not be processed by cluster autoscaler (no node group config)
I0715 18:03:59.739796       1 static_autoscaler.go:623] Scale down status: lastScaleUpTime=2024-07-15 16:21:17.926417071 +0000 UTC m=-3596.573195991 lastScaleDownDeleteTime=2024-07-15 16:21:17.926417071 +0000 UTC m=-3596.573195991 lastScaleDownFailTime=2024-07-15 16:21:17.926417071 +0000 UTC m=-3596.573195991 scaleDownForbidden=false scaleDownInCooldown=false

Related:

• cluster autoscaler not scaling up the autoscaling group when already downscaled to 0 kubernetes/autoscaler#4893

[mnaser]

I don't think this is technically a valid fix, since scale out from zero is not possible, so that kind puts the user in a bad position, right?

[gtirloni]

I don't think this is technically a valid fix, since scale out from zero is not possible, so that kind puts the user in a bad position, right?

Itss supposed to work but there's an upstream bug about it. The cluster-autoscaler magnum driver was updated to allow for this situation.

In any case, Magnum accepts min_node_count=0 and MCAPI is overriding that.