Copy page

Releases

Release 10: Themisto

Released: 2026-06-05

---

Release 10 focuses on instance lifecycle management and observability. New capabilities include runtime user management, instance annotations, hotpluggable volumes for scale-to-zero instances, indefinite instance retention with post-mortem metrics, and the ability to wake VMs from TAP device traffic. On the tooling side, the CLI gains a new unikraft api passthrough and EROFS rootfs by default, while Kraftlet picks up full CSI integration and ROM-based ephemeral volumes.

---

Platform

Image tag response format change

The GET /v1/images and GET /v1/images/list endpoints now return only the actual OCI tag. For an image such as oci://unikraft.io/user/image:latest, the response used to contain user/image:latest and now contains latest. Update any clients that parse the tag field.

Retained instance state renamed

The lifecycle state for logically removed but still-retained instances changed from dead to deleted. Update any monitoring, automation, or filtering logic that matches on the old state name.

User management at runtime via userdb

A new user database (userdb) makes it possible to manage users at runtime rather than only at startup. This work also made internal database entry replacements atomic, improving robustness. This release removes the "no authorization" mode. Deployments must now run with authorization enabled.

Instance annotations

Instances can now carry user-defined annotations, which you can also include in the instance log output. This makes it easier to correlate instance behavior with deployment-specific metadata such as environment, tenant, or release identifiers.

Hotplug volumes for scale-to-zero and template-derived instances

You can now hotplug new volumes to instances during scale-to-zero cycles and add volumes to instances created from templates. This allows storage configuration to evolve over the lifetime of an instance without recreating it.

Indefinite retention of instances

You can now keep instances indefinitely instead of enforcing a time-based limit. This is useful for keeping post-mortem state available for longer investigation windows.

The instance DELETE endpoint now accepts a dont_retain flag that removes a retained instance immediately instead of keeping it around for the configured retention window. This gives operators a direct way to reclaim resources held by retained instances.

Metrics endpoint available for retained instances

You can now query the /v1/instances/metrics endpoint for retained instances. Operators can continue to export metrics for instances that the platform still retains after removal, which is useful for debugging and accounting.

Image pull policy forwarding

The image pull policy configured in the API is now forwarded to the host platform. This allows eliminating image pull latency when you use the if_not_present policy together with background image updates configured in the host platform.

Merge image default environment with instance environment

When starting an instance, the environment variables defined as defaults on the image are now merged with any instance-level environment variables. The instance-level variables no longer fully replace the defaults. This change also fixes a problem where environment variables weren't escaped correctly.

Image fetch errors propagated to the proxy

Image fetch failures now reach the proxy layer. Clients triggering an instance via the proxy receive a more meaningful response when the platform can't pull the underlying image, instead of a generic failure.

Wake VMs from TAP device traffic

Incoming traffic on a TAP device can now wake instances, broadening the set of wake-up sources beyond the existing trigger mechanisms. This allows private VM instances without a service group to have scale-to-zero configured and wake up when needed.

Configurable database storage location

A new --db-path option lets operators choose where the platform stores its databases. This makes it easier to place the databases on dedicated storage, separate from logs or other state.

---

Tooling

CLI

New unikraft api subcommand

unikraft api <endpoint> issues an authenticated HTTP request to the Unikraft Cloud API, reusing the current profile, default metro, and credentials. Use it for endpoints that don't yet have a dedicated subcommand, and for exploration and scripting.

Code
# List instances in the default metro
unikraft api /v1/instances

# Inspect a specific instance
unikraft api /v1/instances/abc123-...-def456

# Create a 256MB volume in fra
unikraft api /v1/volumes --metro fra -d '{"name":"data","size_mb":256}'

# POST a request body from a file or stdin
unikraft api /v1/volumes -d @volume.json
cat volume.json | unikraft api /v1/volumes -d @-

# Delete an instance
unikraft api /v1/instances/abc123-...-def456 -X DELETE

EROFS rootfs by default

unikraft build now produces EROFS root filesystems by default. EROFS images are smaller and faster to mount than the CPIO archives used before, which translates to lower cold-start latency and reduced per-image storage cost.

The runtime's advertised feature flags drive the selection: if every target platform reports EROFS support, the builder picks EROFS. Otherwise it falls back to CPIO. Existing Kraftfiles continue to work without changes.

Kraftfile schema updates

Two changes have landed in the v0.7 Kraftfile schema.

rootfs.format no longer defaults to cpio at parse time. Before this change, the parser rewrote a rootfs without an explicit format to format: cpio. The parser now leaves the field empty and lets the tooling pick a format based on the target runtime's advertised capabilities. Set format: explicitly if you need to pin one.

Code
spec: v0.7
rootfs:
  source: ./rootfs
  # format: cpio    # set explicitly to pin; previously implied

New optional dockerfile: field on rootfs.source. A Dockerfile-backed rootfs can now point at a Dockerfile independent of its build context, matching the behaviour of docker build -f.

Code
spec: v0.7
rootfs:
  source:
    type: dockerfile
    path: .
    dockerfile: ./custom/path/Dockerfile

ROM support

ROMs provide read-only image volumes that many instances can share at low overhead. The CLI can now build them, attach them at instance creation time, and edit them on existing instances.

Code
# Attach a ROM image at /rom on a new instance
unikraft instance create \
    --name demo \
    --metro fra \
    --image nginx:latest \
    --rom image=myuser/my-rom:latest,at=/rom,name=docs

# Attach an inline ROM built from a local directory
unikraft instance create \
    --name demo \
    --metro fra \
    --image nginx:latest \
    --rom dir=./mydata,at=/rom

# Add a ROM to an existing instance
unikraft instance edit demo --rom image=myuser/my-rom:latest,at=/rom2

ROMs are also surfaced as fields on unikraft instance get/list under roms.*.{name,image,dir,at}.

deleted Instance state

The CLI now understands the new deleted lifecycle state for instances that the platform has logically removed but still retains. You can see them in unikraft instance list and filter for them explicitly.

Code
unikraft instance list --filter 'state==deleted'

Volume access modes

unikraft volume create now accepts an --access-mode flag that controls how many instances can share a volume.

Code
# Read-write, single attachment (default)
unikraft volume create --name data --size 10GiB --access-mode rwo

# Read-only, may be attached to many instances
unikraft volume create --name assets --size 1GiB --access-mode rox

# Read-write, may be attached to many instances
unikraft volume create --name shared --size 1GiB --access-mode rwx

--rm Flag on unikraft instance create

unikraft instance create --rm marks the VM for automatic deletion once it exits. This matches docker run --rm and is the preferred form for one-shot and CI workloads.

Code
unikraft instance create \
    --name oneshot \
    --metro fra \
    --image my-job:latest \
    --rm

idle Scale-to-zero policy

--scale-to-zero now accepts policy=idle in addition to on and off. The platform scales idle instances down after cooldown-time of no activity, without requiring an explicit stateful/stateless distinction.

Code
unikraft instance create \
    --name web \
    --metro fra \
    --image nginx:latest \
    --scale-to-zero policy=idle,cooldown-time=30000

Insecure registries

unikraft build and unikraft images now accept --insecure to opt into HTTP or unverified-TLS connections to specific registries, for use with internal mirrors and self-signed deployments.

Code
unikraft build . --output my-org/my-app:latest --insecure registry.internal
unikraft images pull my-org/my-app:latest --insecure

KraftKit

Kraftfile schema v0.7

KraftKit picks up the v0.7 schema changes: rootfs.format is no longer defaulted at parse time, and a new optional dockerfile: field can point at a custom Dockerfile path.

Fetch images from per-metro image stores

kraft cloud image list and image-pull flows now also consult per-metro image stores alongside the global registry. When a metro has already cached an image locally, the pull resolves against it directly, shortening cold-start image pulls for clients colocated with the metro.

Code
# Lists images visible in the fra metro, including its local store
kraft cloud image list --metro fra

Kraftlet

Kraftlet is the Unikraft Cloud integration for Kubernetes, allowing you to run microVMs as Kubernetes pods.

CSI integration

Kraftlet now integrates with the Kubernetes CSI to provision and attach volumes from arbitrary storage backends to Unikraft Cloud instances. Kraftlet stages volumes by calling the appropriate CSI plugin based on the PVC's storage class, then publishes them to the VM via Unikraft Cloud managed volumes.

The Helm chart now exposes the CSI endpoint and supports setting the required node annotations.

Code
# values.yaml
csi:
  endpoint: unix:///var/lib/kubelet/plugins/csi.example.com/csi.sock
nodeAnnotations:
  volumes.kubernetes.io/controller-managed-attach-detach: "true"

Ephemeral volumes as ROMs

configMap, secret, and projected volumes are now mounted as inline ROMs instead of going through the previous import flow. Pod startup no longer has to wait for an import instance to populate the volume contents, which reduces start time for pods that mount ConfigMaps or Secrets.

Resource tagging

Kraftlet tags all Unikraft Cloud resources it creates with their origin, including the Kraftlet node name, originating Kubernetes namespace and resource name, and CSI plugin and storage class where applicable. This allows attributing Unikraft Cloud resources back to their Kubernetes owners without maintaining a separate mapping.

Custom node annotations

Kraftlet can advertise arbitrary annotations on its Kubernetes node object, configured through the Helm chart. This underpins the CSI controller-managed-attach annotation and allows operators to attach cluster-side metadata to the Kraftlet node.

---

Back to all releases