pod init errors not reflected in VMI

[Nir Dothan]

Hello,

I've come across kubelet pod init issues that emit Warning events in the virt-luancher pod context, but nothing bubbles up to the VM/VMI and it simply remains in "Scheduling" state.

This has been witnessed with different types of CNI errors that fail the runtime in creation of the network sandbox.  

One example of provoking this, is if I rename the CNI type in the NAD to a name that does not exist: e.g. I added an xx- prefix to the passt cni type to xx-kubevirt-passt-binding.

In the pod context an event is emitted:

Warning  FailedCreatePodSandBox  5s (x5 over 54s)  kubelet  (combined from similar events): Failed to create pod sandbox: rpc error: code = Unknown desc = failed to create ... error adding container to network "primary-udn-kubevirt-binding": plugin type="xx-kubevirt-passt-binding" failed (add): failed to find plugin "xx-kubevirt-passt-binding" in path [/opt/cni/bin]

However the VMI knows nothing about this:

$ k get vmi
NAME   AGE   PHASE        IP    NODENAME   READY
vm-a   15m   Scheduling                    False

And its events:

 
Events:
  Type    Reason            Age   From                       Message
  ----    ------            ----  ----                       -------
  Normal  SuccessfulCreate  15m   virtualmachine-controller  Created virtual machine pod virt-launcher-vm-a-cj2bz

This seems like a bug that is affecting UX. 

Please advise if you think that we can fix this.

Regards

Nir Dothan

SIG Network

[Alex Kalenyuk]

If I'm not mistaken kubevirt propagates the conditions from the pod as is,
so unless this message is in e.g. the "k8sv1.PodScheduled" condition reason, it won't propagate.
I am not aware of anything in kubevirt propagating events from the pod->vmi but it might make sense for us to do this

--
You received this message because you are subscribed to the Google Groups "kubevirt-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kubevirt-dev...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/kubevirt-dev/CAFfhveVE4%3DxYCoKyaxgYQx45Q1SVG5SS73-s2-OCgzVmYw6Mmg%40mail.gmail.com.

[Nir Dothan]

Hello,

FYI I've run a similar test using a k8s Deployment which runs a regular pod rather than a VM/VMI, and fed it with a CNI that does not exist.

Results are the same: i.e. only pod level events are visible, rs and deployment seem to know nothing about it. 

Deployment is showing a condition:  message: ReplicaSet "netshoot-7b4956476d" is progressing.

At least our UX is not worse than vanilla k8s.