Ready-to-run unikernel images for Datum Cloud compute. Reference one in a Workload the same way you would reference a container image and it boots as a unikernel on a Datum cell:
ghcr.io/datum-cloud/unikernel/<name>:<version>
<name> is the upstream project name (nginx, redis, python, ...) and
<version> is the upstream software version, never a distro release or a
date. Every image also publishes its minor line (1.30), its major line where
that is meaningful (8), and latest.
- Small. Each image ships only the application and the files it needs: single-binary images are a few megabytes, full runtimes under 130 MB.
- Fast. Instances boot in well under a second.
- Dual-stack. Every network image listens on IPv6 and IPv4 out of the box. Datum compute networks are IPv6-first, so this is what makes the image reachable the moment it starts.
- Tracks upstream. Tags follow the upstream version and are rebuilt weekly, so patch releases flow through without any change on your side.
Deploy one with datumctl. The image name is the only thing that changes
from deploying any other container; --runtime-class=unikernel puts the
Instances on the unikernel tier.
datumctl compute deploy web \
--image=ghcr.io/datum-cloud/unikernel/nginx:1.31 \
--city=DFW --http-port=8080 --runtime-class=unikernelA successful deploy prints the HTTPS URL the workload is published on. Omit
--http-port for internal services such as redis or memcached; they are
reachable on their port from other Instances on the same network.
Shell images open at a prompt, so they work as scratch machines:
datumctl compute deploy scratch \
--image=ghcr.io/datum-cloud/unikernel/debian:13 \
--city=DFW --runtime-class=unikernel
datumctl compute exec scratch-dfw-0 -it -- bashEverything else is ordinary workload management: datumctl compute scale,
restart, instances, and destroy all apply. Use -f with a manifest when
you need more than the flags express, for example a custom command or a
ConfigMap mounted into the python image.
Instances keep no state across restarts: the root filesystem lives in RAM. Tags are rebuilt in place; deploy by digest when you need a reproducible rollout.
Images are added one directory at a time under images/; each directory's
image.yaml records the upstream version, the published tags, the default
port, and whether the entrypoint is a static PIE or ships its own dynamic
loader. Browse the directory for the current catalogue.
Open an issue to request one, or send a pull request that adds
images/<name>/ together with its workflow. Each image is an ordinary
Dockerfile packaged with datumctl compute build, and each is its own build
pipeline: a change to that directory builds and verifies the image on the
pull request, and publishes it when the change lands on main.
CONTRIBUTING.md walks through the layout, and
docs/building.md explains the constraints a unikernel
image has to meet (position-independent entrypoints, shipping the exact
shared-library closure, the root filesystem size ceiling, and why the images
carry no kernel).
Apache-2.0; see LICENSE.