vCluster v0.38 release notes
This page lists the customer-facing changes in each vCluster v0.38 release. For an overview of the release, see What's new.
Some changes also shipped in a v0.37 patch release. Those items name the patch, so if you already run that patch or later, they don't change anything for you.
0.38.0​
Released November 18, 2026.
vCluster v0.38 is a limited availability release, recommended only for new deployments. Keep existing clusters on vCluster v0.37 and Platform v4.12, the current LTS releases. For upgrade guidance, contact vCluster Labs.
If vCluster Platform manages your clusters, v0.38 pairs with Platform v5.0.
Deprecations​
- The Istio integration,
integrations.istio, is deprecated, including sleep mode for Istio ingress. It remains available, and no removal date is announced. See Istio. - The
--docker-loginflag onvcluster loginandvcluster platform loginis deprecated and does nothing. It wrote registry credentials from shared secrets to the local Docker config, and Platform v5.0 removes shared secrets. The CLI still accepts the flag, so existing scripts keep working, and prints a deprecation warning.
New features​
- Platform node pools. With Platform v5.0, a cluster reads node pools
defined in the platform and runs them through Auto Nodes alongside the pools
in
vcluster.yaml. Give platform pools names that differ from the pools invcluster.yaml. There are no newvcluster.yamlkeys. See Node pools. - Embedded etcd defragmentation timeout.
controlPlane.backingStore.etcd.embedded.defragTimeoutlimits startup defragmentation, with a default of5m. Set0sto skip it. See Embedded etcd. Also in v0.37.7. - Docker driver on Podman. The Docker driver works with rootful Podman
through its Docker-compatible API. You still need the
dockerCLI. On macOS, LoadBalancer services aren't available on Podman, because Podman can't map privileged ports to a specific host IP. On Windows, the CLI doesn't find Podman's socket for you, so setDOCKER_HOSTyourself. See Docker container basics. - Argo CD cluster metadata.
integrations.argoCD.cluster.metadataadds labels and annotations to the cluster's Argo CD registration. Also in v0.37.4. See Label registered clusters. - Observability integration.
integrations.observabilitydelivers the Platform metrics-writer Secret to the cluster. See Observability. - ServiceMonitor scrape interval.
controlPlane.serviceMonitor.intervalsets how often Prometheus scrapes the control plane. Also in v0.37.3. See ServiceMonitor.
Changes​
vcluster platform startselects Platform v5.0.0 or a later compatible stable release when--versionisn't set. To install a specific version, set--version, for example--version 5.0.0.- The
ghcr.io/loft-sh/vcluster-cliimage no longer includes DevSpace. If your automation runs DevSpace in that image, install it separately. - The resource proxy enforces the cluster's validating admission policies and webhooks on proxied writes, and refuses server-side apply to a proxied resource. Also in v0.37.2.
- Istio 1.27 and 1.28 are no longer tested. The tested versions are 1.29.0, 1.30.0, and 1.31.1. See Istio.
- In clusters with private nodes, Karpenter resources become read-only once
vCluster starts its own Karpenter, for example to run a Platform node pool,
even when
autoNodesisn't configured. A Karpenter you installed yourself keeps working until then. - The platform CLI drops shared secrets, which Platform v5.0 removes.
vcluster platform get secretandset secretmanage only project secrets and fail without--projector a default project.vcluster platform list secretsfails without--project,--all, or--all-projects, and--allno longer includes shared secrets. On these three commands,-nand--namespaceare ignored with a warning. To manage shared secrets on Platform v4.12, use a CLI compatible with v4.12.vcluster platform backup managementstill backs up shared secrets from a platform that serves them, and--skip sharedsecretsstill works. - The default Kubernetes version is 1.36.6. The Kubernetes 1.35 image moves to
1.35.10, and the default etcd image is
3.6.16-0. Also in v0.37.5. The rebuilt Kubernetes images include upstream security fixes. - The image lists published with the release include only Kubernetes 1.35 and 1.36 images. Clusters pinned to an older minor keep working, but mirror those images yourself for air-gapped installations. Also in v0.37.5.
- Embedded etcd defragments on startup only when it can reclaim enough space. Before, it defragmented on every start and failed after 30 seconds. Also in v0.37.7.
- Auto Nodes removes unscheduled node claims first when it scales down a static pool. It replaces node claims that haven't joined when a pool's configuration changes, and with Platform v5.0 it removes nodes whose machine is gone. See Auto Nodes.
- A platform restored from a management backup taken with the CLI keeps its identity. Also in v0.37.4.
Fixes​
- The kubelet proxy rejects paths containing
.or..segments and container log paths with extra segments, so requests can't bypass the validated path when forwarded to the control plane cluster. vcluster registry pushhandles Helm chart filenames that start with-without interpreting them as Helm flags.- The control plane always elects a leader, so extra control plane pods, for example after scaling the StatefulSet or Deployment directly, can't run a second syncer and delete data. On embedded SQLite, only pod-0 can lead. Also in v0.37.7.
vcluster platform destroyworks against vCluster Platform v5.0. Earlier CLI versions stopped before deleting anything. See Uninstall the platform.- Restoring a snapshot of an embedded database keeps the store revision moving forward, and the restore refuses to run over a live database. Also in v0.37.5.
vcluster connecthandles kubeconfig servers without an explicit port, and port forwarding drops a path in the kubeconfig server URL.- Platform node claims work in clusters without Auto Nodes, and a pending node claim is deleted when its node pool is removed.
- When a sync fails part way, changes that weren't synced yet are kept instead of being overwritten.
policies.networkPolicyallows DNS traffic to the CoreDNS that vCluster deploys, for example on vCluster Standalone. Before, it blocked DNS there.- Dual-stack networking works on vCluster Standalone and private nodes.
Standalone accepts a dual-stack
networking.podCIDRandnetworking.serviceCIDR, the built-in flannel splits a dual-stack pod CIDR by address family, and private nodes with two node IPs initialize. Also in v0.37.7. - Containerd registry credentials that contain
$are preserved on private nodes. Also in v0.37.4. - Private nodes detect dead konnectivity agent connections. Also in v0.37.4.
- A stopping standalone control plane is removed from the Kubernetes endpoints. Also in v0.37.4.
- The
ghcr.io/loft-sh/vcluster-cliimage downloads kubectl fromdl.k8s.io, so it ships the stable kubectl release current at build time. Earlier images used a Google Cloud Storage bucket that no longer receives new releases.