Release Notes

Review release notes with the Kubernetes Support Matrix, Version and Lifecycle, and feature-specific documentation to understand release changes and exact support boundaries.

4.5.0

Issued: 2026-11-17

Features and Enhancements

Kubernetes 1.36

4.5 creates new platform-managed clusters on Kubernetes 1.36. Before you upgrade the global cluster to 4.5, workload clusters created by must run Kubernetes 1.36, 1.35, 1.34, or 1.33. The Kubernetes version of imported and registered third-party clusters does not block the global cluster upgrade, but the components on those clusters must be within the direct upgrade window.

For the version relationships and the Kubernetes, containerd, and etcd versions in each Alauda OS image, see Kubernetes Support Matrix and Alauda OS Support Matrix.

Alauda OS Cluster Upgrades in One Request

On Huawei DCS, Huawei Cloud Stack, VMware vSphere, and Bare Metal clusters that run Alauda OS, one upgrade request now upgrades Core, the installed Aligned Extensions, Kubernetes, and Alauda OS. You no longer upgrade Kubernetes on these clusters in a separate manual procedure.

  • Node replacement through every intermediate version — The platform replaces nodes one at a time and moves the control plane and workers through each intermediate Kubernetes minor version. On Bare Metal, workers are reinstalled only at the target version. Progress is reported in an InfraClusterVersion resource, and provider-specific preflight checks run before any node is replaced.
  • Infrastructure providers upgraded with global — The infrastructure providers for Huawei DCS, Huawei Cloud Stack, VMware vSphere, and Bare Metal are Aligned Extensions on the global cluster and are upgraded together with it.
  • Kube-OVN and CoreDNS managed by the platform — The platform installs and upgrades Kube-OVN and CoreDNS on Alauda OS clusters; do not upgrade them by hand. On a cluster upgraded from a version earlier than 4.5, the platform takes them over once the cluster and global both run 4.5. The takeover can finish after the upgrade request completes.

Upgrade boundaries for Alauda OS clusters:

  • The documented upgrade path starts at v4.3.2. For an Alauda OS cluster on an earlier version, contact technical support.
  • A cluster with a single control plane node is for testing and cannot be upgraded.

For details, see Upgrade Overview.

Alauda OS Cluster Installation with ACP CLI (Technology Preview)

You describe an Alauda OS global cluster or workload cluster in one install-config.yaml file and create it with the CLI: ac adm alpha global create for the global cluster, and ac adm alpha cluster create for a workload cluster on Huawei DCS, Huawei Cloud Stack, VMware vSphere, or Bare Metal. The CLI validates the configuration, the registry, and the infrastructure before it creates the cluster, replacing the previous step-by-step manifest procedure. The ac adm alpha commands are a Technology Preview.

For details, see Install on Alauda OS and Create a Cluster on Alauda OS.

Alauda OS Node Configuration

  • Machine Configuration included — Machine Configuration is delivered with every Alauda OS cluster and no longer needs a separate installation. For details, see Machine Configuration Overview.
  • Automatic node resource reservation — The new KubeletConfig API reserves CPU, memory, and ephemeral storage for the operating system and Kubernetes components on Alauda OS nodes. On supported clusters, the platform creates the acp-default-node-reservation object, which calculates the reservation for each node. You can also set static values for a node pool. For details, see Configure Node Resource Reservation.
  • Bare Metal system disk selection — A Bare Metal host now reports its disks when it registers and waits until you select its system disk. This replaces the fixed installation disk setting. For details, see Select the System Disk.

Platform Image Registry

A new 4.5 installation requires an external platform image registry, on Alauda OS and on other operating systems. The platform built-in Registry is no longer deployed on new installations. A global cluster upgraded from an earlier version can continue to use it.

  • mirror-registry — For environments without an image registry, provides mirror-registry, a single-instance registry available on the Core Package download page. It is not highly available and is not recommended for production environments. For details, see Create an Image Registry with mirror-registry.
  • Publishing Extensions — When you publish Extensions with violet push, set --dest-repo to the registry that the target cluster uses. For details, see violet CLI.

For registry requirements, see Prepare an External Platform Image Registry.

Upgrade Preparation

  • Package download based on installed Extensions — Before an upgrade, you export the Extensions installed on the global cluster. The Customer Portal checks the list and downloads the matching target-version packages, including the infrastructure providers.
  • Aligned and Agnostic download scenarios — For a new installation, the Customer Portal provides the ACP Common Aligned Extensions and ACP Common Agnostic Extensions scenarios. The Upgrade from ACP 4.3 or earlier scenario is needed only when the source version is 4.3 or earlier. For details, see Download Packages.
  • Direct upgrade window — A cluster can be upgraded directly to 4.5 from 4.2, 4.3, or 4.4. The window applies to every cluster registered on the global cluster, including the components on imported and registered third-party clusters.

For the preparation steps, see Prepare for the Upgrade.

Global Cluster Disaster Recovery Upgrades

For a patch upgrade within the same minor version, starting from 4.5.0, you can upgrade the primary global cluster and selected workload clusters first, and the standby global cluster afterwards. The Global Cluster Disaster Recovery plugin stays installed; synchronization is paused during the upgrades and resumed after the standby is upgraded. There is no disaster recovery protection while synchronization is paused.

For details, see Upgrade the Standby global Cluster After the Workload Clusters.

Alauda OS Documentation

The Immutable Infrastructure documentation is now part of the documentation. Alauda OS installation, workload cluster creation, node management, Machine Configuration, upgrades, the Cluster API and infrastructure provider API references, and the new Alauda OS Support Matrix are published here. Starting with 4.5, infrastructure provider changes are published in these release notes.

For an overview of the node operating system options, see Cluster Management Models.

Networking

  • EnvoyProxy resource quotas — The EnvoyProxyResourceQuota resource limits the total CPU and memory requests that EnvoyProxy resources in a namespace can declare. Requests that exceed the quota are rejected. For details, see Configure EnvoyProxy Resource Quotas.
  • HSTS and custom error pages on Envoy Gateway — Set the Strict-Transport-Security header for a whole Gateway, and return custom error pages by status code. For details, see Configure HSTS and Configure Custom Error Pages.

Storage

  • Ceph OSDs on Local Storage volumes — In Internal Mode, Ceph can use block persistent volumes provisioned by Local Storage as OSDs instead of host devices. Local Storage can provision these volumes with LocalVolume or the new LocalVolumeSet. A Ceph cluster cannot mix host devices and Local Storage volumes. For details, see Deploying in Internal Mode with Local Storage and Provision with LocalVolumeSet.
  • Ceph Tentacle upgrade — Alauda Storage provides Ceph 20.2 Tentacle in a separate Operator channel. Existing Ceph clusters are not upgraded automatically; you choose when to upgrade from Ceph 19.2 Squid. The upgrade requires 4.5 or later and Linux kernel 5.11 or later on every node that runs Ceph daemons or mounts Ceph volumes, and it cannot be rolled back. For details, see Upgrade Ceph from Squid (v19) to Tentacle (v20).

Virtualization

  • Live migration during node drain — A new virtualization installation sets the cluster eviction strategy to LiveMigrateIfPossible. When a node is drained, for example during maintenance or a cluster upgrade, virtual machines that can be live migrated move to another node, and the others are restarted on another node. An existing installation keeps its current setting; to change it, follow KubeVirt Node Maintenance.

Deprecations and Removals

  • Alauda Build of Prometheus Operator removed — The Alauda Build of Prometheus Operator is no longer shipped in OperatorHub starting in 4.5. On a cluster upgraded from 4.4, an existing installation keeps running but receives no further upgrades. Use Alauda Container Platform Advanced Monitoring or Alauda Container Platform Basic Monitoring instead. Unless you have a specific technology stack requirement, use Alauda Container Platform Advanced Monitoring. For details, see Monitoring Installation.
  • Platform built-in Registry not deployed on new installations — A new installation of 4.5 requires an external platform image registry. A global cluster upgraded from an earlier version can continue to use the platform built-in Registry. For details, see Platform Image Registry.
  • Cluster upgrade from the Web Console removed — Upgrading clusters from the Web Console is removed. Use ac adm upgrade instead. For details, see Upgrade Overview.
  • Huawei DCS and Huawei Cloud Stack cluster creation from the Web Console removed — Creating Huawei DCS and Huawei Cloud Stack clusters from the Web Console is removed. Use ac adm alpha cluster create instead. For details, see Create a Cluster on Alauda OS.
  • Legacy monitoring and alerting APIs remain deprecated — The /v2/metrics, /v2/alerts, and /v1/alerthistories endpoints, deprecated in 4.4, remain available in 4.5 and are planned for removal in a future release. Migrate to the corresponding /platform endpoints. For details, see Monitoring APIs and Alert APIs.

Other Changes

  • Plugin names — Some plugins are listed under new names on the Customer Portal. If a plugin uploaded to shows a name that you cannot find on the Customer Portal, see Plugin Name Reference.

Fixed Issues

  • In environments that had been upgraded through many versions step by step, a large number of operator version records (ArtifactVersion) accumulated in the global cluster. Newly uploaded operator versions could then take a long time to be recognized as available, and the more ArtifactVersions existed, the more CPU and memory the related controller used, so it might not recover on its own when resources ran short. This issue is fixed in v4.5.0. The platform now automatically removes old ArtifactVersions of system and platform operators, while keeping the versions in use and any pinned version.
  • If a cluster contained two ModuleInfo resources with the same cpaas.io/module-name, for example when an operator mapped several of its custom resource instances to ModuleInfo, the platform upgrade failed while building the upgrade plan and stopped. This issue is fixed in v4.5.0. A duplicate name on a module outside the upgrade plan no longer fails the whole plan build.
  • After the platform was integrated with an external identity provider (IdP), if single sign-on with platform accounts was configured for Harbor, GitLab, or Kiali as documented, users who signed in to these tools through single sign-on lost their platform user groups, together with the platform permissions granted through those groups. This issue is fixed in v4.5.0.
  • The ACP CLI (ac) could not update an existing login session. After a session token expired, running ac login again with the same session name reported that authentication succeeded but failed to write the kubeconfig with the error session already exists, leaving the CLI unauthenticated. This issue is fixed in v4.5.0.
  • In TopoLVM, after the device class configuration of a node was deleted and added back, the node capacity was displayed correctly, but creating a PVC with a storage class that used the default device class and mounting it to a Pod failed with an error that no storage space was available. This issue is fixed in v4.5.0.
  • In clusters using MetalLB with Kube-OVN underlay networking, after the OVS port number of the underlay physical NIC or a related OVS port changed, the ingress flows kept referencing the old port number, so the LoadBalancer Service became unreachable from outside the cluster and did not recover automatically. This issue is fixed in v4.5.0. The flows are now updated automatically when the port number changes.
  • On Alauda OS M55 nodes, the node firewalld masqueraded traffic forwarded to Pods through the Kube-OVN node interface, so NodePort Services with externalTrafficPolicy set to Local did not preserve the client source IP. Backend Pods saw a node address instead, which broke access control, auditing, or rate limiting based on client IP. This issue is fixed in v4.5.0.

Known Issues

  • On vSphere clusters that use Alauda OS virtual machines, a virtual machine newly created during a control plane rollout (for example, a Kubernetes version upgrade) occasionally fails to mount /opt on its first boot and enters emergency mode. The machine stays in Provisioning without getting an IP address, and the rollout and upgrade stall. Workaround: delete the stuck Machine so that the control plane creates a new virtual machine.
  • When you install a global cluster or create a workload cluster on vSphere, if a node group in install-config.yaml omits cpuCore or memoryGiB, the virtual machines are created with 2 vCPUs and 2 GiB of memory instead of the VM template's own settings. The ac message shown when either value is 0 or less, which says to leave the field out to keep the template's value, does not match this behavior. Workaround: always set cpuCore and memoryGiB explicitly in each node group.
  • After Alauda OS KU is installed on a physical server, it does not register its own boot entry in the UEFI firmware and boots only through the generic disk entry. If the server's UEFI boot order places the network (PXE) before the disk, the node can boot from the network when it restarts after installation or after any later reprovisioning (for example, cluster creation, scale-out, or upgrade), and be reinstalled with another operating system by an on-site PXE service. Workaround: place the disk before the network in the BIOS boot order, or, after the node is installed, use efibootmgr once to register a boot entry for \EFI\uos\grubx64.efi on the ESP and confirm it is first in BootOrder; later upgrades and reprovisioning do not need this step again.
  • When a non-system user who has permission to create project members (for example, a user account with the platform administrator role) creates a ProjectMember resource directly through the API, the request is always rejected with a 404 not found error. Adding project members in the Web Console is not affected. Workaround: manage project members in the Web Console.
  • In global clusters upgraded from v3.18 or v4.0 to v4.1 or later, a legacy AppRelease named region may remain in the cpaas-system namespace. It is periodically re-synchronized and reverts the Web Console frontend (the alauda-console image of frontend) to the v4.0 version, which causes an authorization code error when users log in to the platform. Workaround: set the interval-sync annotation of this AppRelease to "false" so that it is no longer synchronized automatically. If the frontend still runs the old version, contact technical support to re-synchronize the platform frontend components.
  • In clusters whose nodes use 10 Gbps or faster NICs, cross-node Pod traffic on the Kube-OVN overlay network achieves significantly lower bandwidth than the host network, possibly less than half of the host network bandwidth. No workaround is currently available.
  • When you upgrade a workload cluster that uses Calico networking to v4.5.0, MetalLB speaker and calico-node Pods are restarted in a rolling manner during the same time window. Workloads accessed through ALB or NodePort may experience intermittent connection timeouts or interruptions for about 2 to 3 minutes, and the impact is more noticeable when MetalLB provides VIPs in L2 mode. No workaround is currently available; perform the upgrade during off-peak hours or a maintenance window.
  • In clusters that use Kube-OVN, Pod access to a Service through its ClusterIP has two limitations. When a Service sets sessionAffinity: ClientIP, sessionAffinityConfig.clientIP.timeoutSeconds is ignored and the session affinity timeout is always 10800 seconds (3 hours). internalTrafficPolicy: Local is not honored: traffic is still forwarded to backend Pods on other nodes, and requests are not rejected when the local node has no backend. No workaround is currently available.
  • The local-storage device discovery component updates its resources at a high frequency, continuously generating a large volume of audit logs on clusters with many nodes. No workaround is currently available.

The following product sites provide component-specific documentation, compatibility information, upgrade guidance, and, where published, release notes. Their versions and release schedules are independent of 4.5, so product documentation and release notes can be published after the 4.5 platform release.