This page documents the list of known issues and possible work arounds/solutions.
The latest Ubuntu 22.04 AMI (released on 30th July) prevents creating new EBS volumes on some AWS clusters.
The latest Ubuntu 22.04 AMI have IMDSv2 (the AWS instance metadata API) enabled as a default, while the previous Ubuntu AMIs allowed both IMDS v2 and v1. By default, the limit of hops of PUT requests to metadata service is 2. However, the default settings provided by AWS are not compatible with the containerized environments (i.e. Kubernetes) since the instance metadata service is not reachable from the container network in case if further hops are required. At this time, it appears that only Cilium clusters are affected, and only if the CSI components are running on nodes that are using the latest Ubuntu AMI.
A new machine-controller (MC) version (v1.58.6) has been released that can be configured into KubermaticConfiguration:
apiVersion: kubermatic.k8c.io/v1
kind: KubermaticConfiguration
metadata:
name: kubermatic
namespace: kubermatic
spec:
userCluster:
machineController:
imageTag: v1.58.6
[...]
The updated MC version will be included in the next patch release (2.24.12).
When using a recent (beginning of May 2024) Ubuntu 22.04 image provided by Azure, user cluster nodes provisioned by machine-controller and operating-system-manager fail to bootstrap and never join the cluster. Instead, the bootstrap.service systemd unit is constantly looping.
A recent change to the Ubuntu 22.04 image has modified the configuration for cloud-init and how it accesses its datasource in Azure. cloud-init clean (which is used to prepare the machine for configuration as Kubernetes node) removes files crucial to this new way of communicating with the datasource and requires an additional step to recreate the missing configuration files.
A new operating-system-manager (OSM) version (v1.4.3) has been released that can be configured into KubermaticConfiguration:
apiVersion: kubermatic.k8c.io/v1
kind: KubermaticConfiguration
metadata:
name: kubermatic
namespace: kubermatic
spec:
userCluster:
operatingSystemManager:
imageTag: v1.4.3
[...]
The updated OSM version will be included in the next patch release (2.24.9).
For custom OSPs the following change is relevant:
diff --git a/deploy/osps/default/osp-ubuntu.yaml b/deploy/osps/default/osp-ubuntu.yaml
index 85e8942..353bded 100644
--- a/deploy/osps/default/osp-ubuntu.yaml
+++ b/deploy/osps/default/osp-ubuntu.yaml
@@ -102,6 +102,11 @@ spec:
curl -s -k -v --header 'Authorization: Bearer {{ .Token }}' {{ .ServerURL }}/api/v1/namespaces/cloud-init-settings/secrets/{{ .SecretName }} | jq '.data["cloud-config"]' -r| base64 -d > /etc/cloud/cloud.cfg.d/{{ .SecretName }}.cfg
cloud-init clean
+ {{- /* Azure's cloud-init provider integration has changed recently (end of April 2024) and now requires us to run this command below once to set some files up that seem required for another cloud-init run. */}}
+ {{- if (eq .CloudProviderName "azure") }}
+ cloud-init init --local
+ {{- end }}
+
{{- /* The default cloud-init configurations files have a bug on Digital Ocean that causes the machine to be in-accessible on the 2nd cloud-init and in case of Hetzner, ipv6 addresses are missing. Hence we disable network configuration. */}}
{{- if (or (eq .CloudProviderName "digitalocean") (eq .CloudProviderName "hetzner")) }}
rm /etc/netplan/50-cloud-init.yaml
MachineDeployments using the affected OSP need to be restarted after updating OSM or the custom OSP.
When upgrading to or installing KKP in version 2.24.0 or 2.24.1, kube-apiserver components fail to start with log output similar to the snippet below:
{"level":"error","time":"2023-11-30T10:43:46.518Z","caller":"etcd-launcher/main.go:116","msg":"Operation failed: failed to initialize etcd cluster configuration: failed to get API group resources: unable to retrieve the complete list of server APIs: kubermatic.k8c.io/v1: Get \"https://10.96.0.1:443/apis/kubermatic.k8c.io/v1\": dial tcp 10.96.0.1:443: i/o timeout."}
This problem arises when Cilium is used as CNI for the underlying seed cluster.
It was fixed in KKP 2.24.2, see relevant issue: kubermatic/kubermatic#12874.
A bug in Cilium tracked as cilium/cilium#12277 and cilium/cilium#20550.
Kubernetes NetworkPolicy documentation implies that allowing node access via CIDRs is supposed to work:
Node specific policies (you can use CIDR notation for these, but you cannot target nodes by their Kubernetes identities specifically).
But Cilium does not properly allow traffic even though NetworkPolicies with node CIDRs are in place.
For installations with KKP versions 2.24.0 or 2.24.1, there are two options available:
CiliumClusterwideNetworkPolicy object (this is a cluster-scoped resource, i.e. global) to allow the erroneously blocked traffic:apiVersion: cilium.io/v2
kind: CiliumClusterwideNetworkPolicy
metadata:
name: cilium-seed-apiserver-allow
spec:
egress:
- toEntities:
- kube-apiserver
endpointSelector:
matchLabels:
app: apiserver
For KKP 2.24.2 and higher, this policy is automatically created.
The issue arises in Ubuntu 22.04 cloud image OVAs starting from version 20230602 when they are run on VMware Cloud Director. This problem disrupts the provisioning of new Kubernetes nodes using machine-controller due to interruptions caused by reboots.
The root cause of this issue can be traced back to a change in the default settings of open-vm-tools. These changes, in turn, affect the behavior of cloud-init during startup, leading to the disruptive behavior observed when provisioning new Kubernetes nodes. Specifically, the open-vm-tools.service starts before cloud-init, and it runs with the default timeout (30 seconds).
One interim solution in this scenario is to create a custom Ubuntu 22.04 image with the following setting preconfigured in /etc/vmware-tools/tools.conf file.
[deployPkg]
wait-cloudinit-timeout=0
This adjustment will help ensure that the issue no longer disrupts the provisioning of new Kubernetes nodes on the affected Ubuntu 22.04 cloud images running on VMware Cloud Director provider.
For additional details and discussions related to this issue, you can refer to the following GitHub issues: