Create archive for the minikube images

[afbjorklund]

(breakout from issue #7737, as a released file)

As an alternative to the current preload tarball, where it cannot be used (for various reasons).
It would contain the same essentials and auxiliary images, but in a more portable format

Additionally it would work in all container runtimes, the preload requires one tarball for each.
As well as allow loading with all storage drivers, those also require different tarball variants.

Current:
preloaded-images-k8s-v2-v1.18.0-docker-overlay2-amd64.tar.lz4

preloaded-images-k8s-v2-v1.18.0-containerd-overlay2-amd64.tar.lz4
preloaded-images-k8s-v2-v1.18.0-crio-overlay-amd64.tar.lz4

Side note: the overlayfs storage format is just called "overlay" in github.com/containers/storage
(probably because there was no need for a "overlay2" to replace a previous "overlay" attempt*)

* see https://docs.docker.com/storage/storagedriver/overlayfs-driver/

Example:
images-k8s-v1.18.0-amd64.tgz

One could additionally use different format (oci-archive instead of docker-archive) as well as
different compression (xz instead of gz), but then it will not work with all container runtimes.

docker save $images | gzip -9 > images.tgz

podman save --format oci-archive $images > images.tar && xz < images.tar > images.txz

The preload tarball also contains the kubernetes programs, but those would be excluded here.
(e.g.kubeadm kubelet kubectl, as they are not container images but regular executables)

Related issues:

• Preload containerd images  #6933
• Preload CRI-O images #6934
• autoset --preload=false when storage driver is not overlay #7626
• Host preload image on a site accessible from China: Get Alibaba account for minikube #7695

For Docker, the current preload is still the fastest option (prepared format, faster compression)

So this is a portable alternative, with a more compatible format and more compatible compression:

• from ./lib/docker/overlay2/ (and friends) to manifest.json (and friends)

• from lz4 to gzip

[afbjorklund]

One small problem with this is that the feature is currently broken in podman:

• BUG: loading multiple images saved by docker fails containers/podman#2822 - loading multiple images saved by docker fails

$ docker images k8s.gcr.io/pause
REPOSITORY          TAG                 IMAGE ID            CREATED             SIZE
k8s.gcr.io/pause    3.2                 80d28bedfe5d        2 months ago        683kB
k8s.gcr.io/pause    3.1                 da86e6ba6ca1        2 years ago         742kB
$ docker save k8s.gcr.io/pause:3.1 k8s.gcr.io/pause:3.2 > images.tar
$ docker load -i images.tar 
Loaded image: k8s.gcr.io/pause:3.1
Loaded image: k8s.gcr.io/pause:3.2
$ podman load -i images.tar
  Unexpected tar manifest.json: expected 1 item, got 2
  open images.tar/manifest.json: not a directory
Error: error pulling "": unable to pull dir:images.tar: unable to pull image: Error determining manifest MIME type for dir:images.tar: open images.tar/manifest.json: not a directory

That is, load currently only supports one image per archive. And save is broken:

• podman save not saving multiple images containers/podman#2669 - only saves the first image, all tags pointing to it (!)

$ podman images k8s.gcr.io/pause
REPOSITORY         TAG   IMAGE ID       CREATED        SIZE
k8s.gcr.io/pause   3.2   80d28bedfe5d   2 months ago   688 kB
k8s.gcr.io/pause   3.1   da86e6ba6ca1   2 years ago    749 kB
$ podman save k8s.gcr.io/pause:3.1 k8s.gcr.io/pause:3.2 > images.tar
$ docker load -i images.tar
e17133b79956: Loading layer [==================================================>]  744.4kB/744.4kB
Loaded image: k8s.gcr.io/pause:3.1
Loaded image: k8s.gcr.io/pause:3.2
$ docker images k8s.gcr.io/pause
REPOSITORY          TAG                 IMAGE ID            CREATED             SIZE
k8s.gcr.io/pause    3.1                 da86e6ba6ca1        2 years ago         742kB
k8s.gcr.io/pause    3.2                 da86e6ba6ca1        2 years ago         742kB

So this feature is mostly for containerd and other storage drivers, until podman is fixed...
Actually the real issue is in the underlying github.com/containers/image library, but anyway.

[afbjorklund]

The workaround for cri-o (and others, as well) is just to use the current minikube cache...

Since it creates one cache file per container image, the problems above are not exposed.

The OCI format is not widely used at the moment, so might as well stay with the Docker format.

And since by default the layers are stored as .tar.gz, compressing them again don't yield much.

[afbjorklund]

With the amount of problems this still has, and the lack of benefit vs. the preload - lowered priority.
The current cache mechanism should be "good enough", and the preload is a bigger improvement.

[afbjorklund]

One simple option, if wanting a one-file download also for the others is to tar up the cache folder...
That is: to create an archive of archives, where the top archive contains one file for each image

gcr.io/
gcr.io/k8s-minikube/
gcr.io/k8s-minikube/storage-provisioner_v1.8.1
k8s.gcr.io/
k8s.gcr.io/pause_3.2
k8s.gcr.io/kube-apiserver_v1.18.0
k8s.gcr.io/kube-controller-manager_v1.18.0
k8s.gcr.io/kube-scheduler_v1.18.0
k8s.gcr.io/etcd_3.4.3-0
k8s.gcr.io/kube-proxy_v1.18.0
k8s.gcr.io/coredns_1.6.7
kubernetesui/
kubernetesui/metrics-scraper_v1.0.2
kubernetesui/dashboard_v2.0.0-rc6

386M images.tar
383M images.tgz

[fejta-bot]

Issues go stale after 90d of inactivity.
Mark the issue as fresh with /remove-lifecycle stale.
Stale issues rot after an additional 30d of inactivity and eventually close.

If this issue is safe to close now please do so with /close.

Send feedback to sig-testing, kubernetes/test-infra and/or fejta.
/lifecycle stale