[gardener]
Most notable changes
-
[USER] The
type
field has been removed from theStorageClass
deployed in OpenStack Shoots. Moreover, theStorageClass
es deployed by Gardener are now reconciled and not only created once anymore. (#395, @afritzler) -
[USER] Initializers for cloud-specific
PersistentVolume
s are now added by Gardener instead of the Kubernetes initializer admission plugin (to temporarily mitigate a runtime panic that would otherwise happen in the Kubernetes cloud-controller-manager (see kubernetes/kubernetes#68996 for more details)). (#392, @rfranzke) -
[USER] When deleting a Shoot cluster Gardener does now forcefully delete
CustomResourceDefinition
s andAPIService
s if they have not been cleaned up regularly by their responsible controllers within 5 minutes. (#391, @Adracus) - [USER] Gardener does now delete pods stuck in terminating state longer than their deletion grace period allows. (#383, @rfranzke)
- [USER] Failed Shoot operations are now retried when a different Gardener version takes over without requiring user interaction. (#382, @zanetworker)
- [USER] All control plane components other than kube-controller-manager and cloud-controller-manager are now blocked via network policies from accessing the cloud’s metadata service. (#381, @zanetworker)
- [USER] System component pods as well as addons managed by Gardener are now scheduled with a high priority using the Kubernetes pod priority feature. They might preempt pods created by users causing their eviction/deletion. Users should consider scaling up their cluster (either vertically or horizontally) in such cases. (#379, @rfranzke)
-
[USER] Gardener does now replace
kube-dns
withCoreDNS
in all Shoot clusters.kube-dns
is becoming deprecated, andCoreDNS
is the new default DNS plugin for Kubernetes (as it is more pluggable and provides more features). (#364, @zanetworker) -
[OPERATOR] The horizontal pod autoscaler downscale delay for shooted Seed clusters has been increased to
24h
. (#397, @amshuman-kr) -
[OPERATOR] The automatic-updates daemon on CoreOS is now disabled permanently, ensuring that the OS image version configured in
Shoot
resources is fixed. (8950d68
)
Improvements
- [USER] The metrics-server deployment in Shoot cluster has been added to the deletion whitelist, i.e. Gardener won’t delete it when deleting a Shoot. This is due to the fact that the Kubernetes namespace controller requires to discover all APIs (including aggregated APIs) in order to recursively clean up all object in the namespace. (#390, @rfranzke)
-
[USER] The delays for the horizontal pod autoscaler have been modified in the following way to achieve a less oscillating behaviour: Downscale:
5m
->15m
and Upscale:3m
->1m
. (#389, @rfranzke) - [USER] Gardener does now wait until all needed controllers are active before cleaning the Kubernetes resources when deleting a Shoot. (#386, @rfranzke)
- [USER] Resources deployed into Shoot clusters that are not needed any longer are no correctly pruned. (#384, @rfranzke)
-
[USER] The version of calico has been upgraded to
v3.2.1
. (#380, @databus23) - [OPERATOR] An issue where the Shoot-specific Prometheus instance did use invalid certificates for scraping etcd and the kubelet has been resolved. (#394, @mvladev)
- [OPERATOR] Horizontal pod autoscaling for the kube-apiserver’s is now enabled for all Shoots. (#388, @zanetworker)
- [OPERATOR] Fixed an issue that prevented a correct error code mapping for errors returned by the machine-controller-manager. (#375, @Adracus)
-
[OPERATOR] Fixed an issue that caused unnecessary restarts of kube-/cloud-controller-manager for Shoot deletions. (
73365eb
)
Docker Images
gardener-apiserver: [eu.gcr.io/gardener-project/gardener/apiserver:0.12.0](http://eu.gcr.io/gardener-project/gardener/apiserver:0.12.0)
gardener-controller-manager: [eu.gcr.io/gardener-project/gardener/controller-manager:0.12.0](http://eu.gcr.io/gardener-project/gardener/controller-manager:0.12.0)