Bug 1868645

I've seen the same problem for 2 fluentd pods in the openshift-logging namespace and deleted them manually. After that they came up fine. Could this be some sort of a race condition between worker nodes coming up and pods already being created?

The must-gather is available at http://file.str.redhat.com/mkarg/bz1868645/must-gather.tgz

can you get us the yaml for the pod? I don't see the affinity spec here

or ideally dump it all with:
`oc get namespace nodevertical-rerun-10 -o yaml`
`oc get all -n nodevertical-rerun-10 -o yaml`

thanks

I've attached the outputs as namespace.yml and all.yml to this bz. 
Let me know if you need anything else.

Created attachment 1711685 [details]
the namespace definition

Created attachment 1711686 [details]
the entire contents of the namespace

Looks like those are bare pods without a controller to recreate them.

Unfortunately the dump isn't from the same state  - all the pods there are either Running or Pending.

The pods don't have explicit affinity but have a nodeSelector:
    nodeSelector:
      node-role.kubernetes.io/worker: ""
      nodevertical: "true"

and the default project node selector set on the namespace is
  openshift.io/node-selector: node-role.kubernetes.io/worker=

kube-scheduler thinks the node is fit to run the pod

  ./namespaces/openshift-kube-scheduler/pods/openshift-kube-scheduler-master-2/kube-scheduler/kube-scheduler/logs/current.log:1392:2020-08-13T09:02:51.308294392Z I0813 09:02:51.308256       1 scheduler.go:731] pod nodevertical-rerun-10/nodevert-pod-999 is bound successfully on node "worker044", 123 nodes evaluated, 100 nodes were found feasible.

and the nodeName is set
  Warning  NodeAffinity    35m   kubelet, worker044  Predicate NodeAffinity failed

snapshotted labels on worker044 are:
  labels:
    beta.kubernetes.io/arch: amd64
    beta.kubernetes.io/os: linux
    kubernetes.io/arch: amd64
    kubernetes.io/hostname: worker044
    kubernetes.io/os: linux
    node-role.kubernetes.io/worker: ""
    node.openshift.io/os_id: rhcos
    nodevertical: "true"
    placement: logtest

which seem to match as well. kube-scheduler evaluated the predicates as matching an assigned the node worker044 to the pod but kubelet on that node thinks the labels don't match and fails the pod.

It would be good to get kubelet logs from the worker044 to help debug the issue. Also the YAML dump for a failed pod would be useful.

As the above points to kubelet, I am sending it to the Node team to have a look.

The behaviour seem similar to https://github.com/kubernetes/kubernetes/issues/93338

this code looks suspicious btw. https://github.com/kubernetes/kubernetes/issues/92067#issuecomment-643711902

// getNodeAnyWay() must return a *v1.Node which is required by RunGeneralPredicates().
// The *v1.Node is obtained as follows:
// Return kubelet's nodeInfo for this node, except on error or if in standalone mode,
// in which case return a manufactured nodeInfo representing a node with no pods,
// zero capacity, and the default labels.
func (kl *Kubelet) getNodeAnyWay() (*v1.Node, error) {
	if kl.kubeClient != nil {
		if n, err := kl.nodeLister.Get(string(kl.nodeName)); err == nil {
			return n, nil
		}
	}
	return kl.initialNode(context.TODO())
}

Created attachment 1711707 [details]
worker044 logs (compressed)

created with 
oc adm node-logs worker044 --since=-6d > /tmp/worker044.logs

Forget about those log files please, the cluster was re-deployed in between so they are useless. Sorry about that.

Derek and I discussed the situation and, for this particular situation of bare pods running on hard downed nodes in a DR situation, there shouldn't be an expectation that these pods restart on the same nodes.

However, this can be an issue in less contrived situations, like the ones mentioned in upstream issues, and Derek is coding up a PR for upstream.

upstream PR https://github.com/kubernetes/kubernetes/pull/94087

The expectation was not that pods would get restarted on the same nodes as before the outage, but that they would all come up running at some point. 
Sorry if that wasn't clear.

Pods, once scheduled, can not move nodes.  So a bare pod, with no higher level controller to recreate a new pod from a pod template (i.e. ReplicaSet, DaemonSet, etc), will not restart.

(In reply to Seth Jennings from comment #13)
> Pods, once scheduled, can not move nodes.  So a bare pod, with no higher
> level controller to recreate a new pod from a pod template (i.e. ReplicaSet,
> DaemonSet, etc), will not restart.

So if all nodes come back up, all pods should have as well (no hard down nodes)?

I'm under the impression that there still is some confusion about what the expectation after a disaster outage is. For me it's like this:

--
Power outage for all nodes

Masters get powered on and form a quorate cluster

Masters start to power up worker nodes and schedule pods (it does not matter for the pods where they get started)

Nodes start the assigned pods

Ultimately all the pods that were running before the outage are running again, even on different hosts. 
--

If the pods are not running again, I would consider that a bug, and I'm sure our customers would see that the same way. If I understand the upstream PR correctly, pods should only get scheduled to a node when that node is synced and thus can actually run pods, is that correct?

Thanks!

Deferring to 4.7 as the test, expected behavior, and potential fix are all in dispute and there is no regression here.

TestBlocker removed

Upstream PR merged.

Upstream backports:
https://github.com/kubernetes/kubernetes/pull/97995
https://github.com/kubernetes/kubernetes/pull/97996
https://github.com/kubernetes/kubernetes/pull/97997

OpenShift PRs to follow.

Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.

For information on the advisory (Moderate: OpenShift Container Platform 4.7.0 security, bug fix, and enhancement update), and where to find the updated
files, follow the link below.

If the solution does not work for you, open a new bug report.

https://access.redhat.com/errata/RHSA-2020:5633

The needinfo request[s] on this closed bug have been removed as they have been unresolved for 500 days