A new release of VM Runtime on Google Distributed Cloud is available as part of the Google Distributed Cloud for bare metal 1.30.0-gke.1930 release, which is now available for download. For a more information about what's new in this release, see the Google Distributed Cloud for bare metal 1.30.0-gke.1930 release notes.
Added support for the virtctl diagnose command to diagnose VM issues and the virtctl snapshot command to collect VMRuntime logs for debugging.
Added support for configuring live migration maxDowntime through the kubevirt migration policy. This specifies the maximum time in milliseconds that a VM is allowed to be
paused at the end of a live migration. The default value is 300 (0.3 seconds). For VMs with workloads that are dirtying RAM faster than it can be copied across the network for live migration, increasing this setting can speed up live migration.
Updated the live migration flow to remove the old virt-launcher pod after a successful live migration.
Updated VM Runtime to mark a VMI as unready when the VM hits IO/sync error.
Increased the nodeMonitorGracePeriod default value for node high availability to 55 seconds to tolerate cluster API unavailability due to etcd failover.
Moved the configuration for node high availability to spec.haPolicy in the vmruntime resource.
Changed the KubeVirt restore short name to virtvmrestore to avoid conflict with vmrestore that's used by VM Runtime on GDC.
Refactored the high-availability (HA) logic to fix a corner case issue in HA.
Fixed a networking issue on live migration.
Disabled the VM generation check for the autoRestartOnConfigurationChange feature, because the check isn't accurate.
Removed the erroneous "No disk capacity" error from the virt-launcher log.
A new release of VM Runtime on Google Distributed Cloud is available as part of the GKE on Bare Metal 1.29.0-gke.1449 release, which is now available for download. For a more information about what's new in this release, see the GKE on Bare Metal 1.29.0-gke.1449 release notes.
Added support for applying real-time label patching to the vm Pod by adding labels to GVM resources.
Added support for configuring the running state (Running or Stopped) of a GVM through the runningState field in the GVM resource spec.
Reduced the CPU and memory resource request settings for the macvtap DaemonSet.
Added ResponseReadTimeout for use when fetching images from an S3 repository. Now, the importer Pod disconnects and restarts when the object read can't get new data for more than 10 minutes.
Added vmruntime preflight check result to the output of kubectl get vmruntime.
Enabled the ability to override the pod network configuration through the virtSpec ConfigMap. Overrides are needed when the cluster CIDR conflicts with the default mock CIDR (10.0.2.0/24) used by VMs that connect to the default pod network.
Changed the default VM high availability grace period from 90 seconds to 35 seconds. With this change, VMs fail over on node failure sooner than with previous releases. Included in this change is the addition of continuous retries on heartbeat failures.
Fixed the virt-launcher log container termination issues that sometimes caused VM statuses of not ready.
Fixed issues on the persistent non-volatile random access memory NVRAM feature, which caused the created PVC to get stuck in a pending state.
Fixed the guest agent installation script.
Fixed an issue with the virtctl create vm command that caused failures when os-type was set to windows.
Anthos clusters on bare metal 1.28.0 is now available for download. For a more information about what's new in this release, see the Anthos clusters on bare metal 1.28 release notes.
Added support for local image caching within the same namespace.
Increased the memory resource limit for CDI pods when importing or cloning images.
Added an error to catch attempts to create disks using a storage class (other than local-shared) with a storage profile that's missing access modes. In earlier versions of Anthos VM Runtime, the access mode defaults to ReadWriteOnce (RWO) for this case.
Added a check to see if the network controller is ready as part of overall VMRuntime readiness.
Added an Enabled column for the VMRuntime custom resource so that you can see whether VMRuntime is enabled or not without printing the whole custom resource.
Removed the node patching access from the virt-handler DaemonSet to reduce security risk.
Added ErrorConfiguration and Pending to the disk phase status.
Upgraded CDI version to v1.57.0. With this upgrade, support for the qcow2 VM format is obsolete. This image format configuration is ignored silently.
Increased the live migration Pod creation timeout to resolve cluster upgrade failures.
Enabled live migration auto-converge, by default. This fixes an issue where live migration wouldn't finish for some VMs.
FIxed an issue to prevent VM annotations from triggering VMs to restart when the autoRestartOnConfigurationChange flag is enabled.
Fixed an issue that caused nil pointer crashes when the KubeVirt custom resource is missing.
Enabled support for running a VM as a dynamic host configuration protocol (DHCP) server.
]]>Anthos clusters on bare metal 1.16.0 is now available for download. For a more information about what's new in this release, see the Anthos clusters on bare metal 1.16 release notes.
Added GPU support for Rocky Linux guest operating systems (OSes). Support is similar to what's provided for Ubuntu, but you must use the dnf package manager, instead of the Ubuntu apt.
Added support to persist states for Unified Extensible Firmware Interface (UEFI) and trusted platform module (TPM) across virtual machine (VM) restarts. This feature is enabled in the VMRuntime custom resource by setting
spec.featureGates.VMPersistentState to Enabled.
Added support for inserting or ejecting CD-ROMs for live virtual machines (VMs) without interruptions. You can eject currently attached CD-ROMs and you can insert a CD-ROM into any open slot. You can also start with open slots and insert CD-ROMs at any time.
Added READY column in VMRuntime custom resource to tell you if VMRuntime is ready or not without printing the whole resource manifest.
Upgraded Kubevirt version from v0.58.0 to v0.59.0.
Changed the default VM format from qCOW2 to RAW for file system storage. This change improves the underlying storage performance and is transparent to users.
Fixed the issue that prevented proper detection of GPU devices when the VM is booted with UEFI.
Fixed a network disconnect issue that occurred during VM live migration.
Fixed the issue where a daemon-reload for systemd on a worker node causes an existing GPU workload to fail.
Fixed the issue that blocked VMs from starting when secure boot and Windows guest agent are both enabled.
Fixed the issue where cluster upgrades or credential rotation prevented new GPU workloads from running.
]]>Make MIG controller resilient to K8s API server failure This change makes the GPU MIG component more resilient to short-time controller plane failures. For example, if there is a glitch on the API server, the MIG controller will retry connecting to the server multiple times till the API server returns to be responsive or it times out.
Change eviction behavior during cluster upgrade VM live migration/restart is triggered after a node is put into maintenance mode instead of before as in earlier versions, and the operation is subject to the common enteringMaintenanceWindow timeout set by Anthos on Bare Metal. The VM will try to migrate or move to other nodes that are already upgraded in a best-effort way. Also changed the Kubevirt migration taints to soft taints to fix the single-control-node upgrade issue.
Base image upgrade to address vulnerabilities
Kubevirt version upgrade from v0.57.1 to v0.58.0 This most notably removes the root-user requirement on virt-launcher pods.
Support ABM cluster proxy
Qemu version upgrade from 5.2 to 7.2 This upgrade changse the default Qemu machine type for VM instances. We will restart all existing VMs during upgrade to switch to the new machine type.
HA enhancement for Robin storage class When the Robin storage class is used, because Robin supports storage fencing, trigger Robin fencing to ensure no access to the storage from the failed node during HA failover. This allows HA to work with a block storage class with Robin.
Support configuration of the virtio-nontransitional driver for SLES11 Expose the Kubevirt knob on the virtio-nontransitional driver to the KubeVM layer.
Windows guest HyperV optimization For windows guest OS on a physical server environment, automatically enable some advanced configurations for better performance.
Fix the OOM error for VM with large memory This fix adds memory overhead for VMs with large memory.
Prevent multiple VMs from attaching to the same disk in non-read-only mode
VNC access improvement Use tablet mode to enhance VNC experience.
Support for virtual machine disk resize When customers modify the size of a VM disk, the underlying persistent volume claim (PVC) and disk used by the VM are expanded if the storage provider supports it.
Guest agent V2 Previously, the guest agent connected to the control plane through the Kubernetes API server watching mechanism. In Anthos for Bare Metal 1.15, we changed the default communication mode for guest agents on Linux VMs to the push model through the vsock.
Detects generation difference between Kubevirt VM and Kubevirt VMI This ensures that when the autoResetartVMOnConfigurationChange flag is enabled, VM instances will always be in sync with VM configurations irrespective of whether the changes were made before the flag enablement or not.
Support Rocky 8 guest OS, SLES11, and Ubuntu 22.04
]]>