images/

README

Kubernetes test images

Overview

All the images found here are used in Kubernetes tests that ensure its features and functionality. The images are built and published as manifest lists, allowing multiarch and cross platform support.

This guide will provide information on how to: make changes to images, bump their version, build the new images, test the changes made, promote the newly built staging images.

Prerequisites

In order to build the docker test images, a Linux node is required. The node will require make, docker (version 19.03.0 or newer), and docker buildx, which will be used to build multiarch images, as well as Windows images. In order to properly build multiarch and Windows images, some initialization is required:

docker run --rm --privileged multiarch/qemu-user-static --reset -p yes
docker buildx create --name img-builder --use
docker buildx inspect --bootstrap

The node must be able to push the images to the desired container registry, make sure you are authenticated with the registry you're pushing to.

Making changes to images

There are several thousands of tests in Kubernetes E2E testing. Not all of them are being run on new PRs, and thus, not all images are used, especially those that are not used by Conformance tests.

So, in order to prevent regressions in the images and failing jobs, any changes made to the image itself or its binaries will require the image's version to be bumped. In the case of a regression which cannot be immediately resolved, the image version used in E2E tests will be reverted to the last known stable version.

The version can easily be bumped by modifying the file test/images/${IMAGE_NAME}/VERSION, which will be used when building the image. Additionally, for the agnhost image, also bump the Version in test/images/agnhost/agnhost.go.

The typical image used in E2E testing is the agnhost image. It contains several subcommands with different functionalities, used to validate different Kubernetes behaviours. If a new functionality needs testing, consider adding an agnhost subcommand for it first, before creating an entirely separate test image.

Some test images (agnhost) are used as bases for other images (kitten, nautilus). If the parent image's VERSION has been bumped, also bump the version in the children's BASEIMAGE files in order for base image changes to be reflected in the child images as well.

Keep in mind that the Kubernetes CI will not run with the image changes you've made. It is a good idea to build the image and push it to your own registry first, and run some tests that are using that image. For these steps, see the sections below.

After the desired changes have been made, the affected images will have to be built and published, and then tested. After the pull request with those changes has been approved and merged, a postsubmit job will then be triggered which will build the images that were changed (for example, if a change was made in test/images/agnhost, then the job post-kubernetes-push-e2e-agnhost-test-images will be triggered). The image will then be pushed to the gcr.io/k8s-staging-e2e-test-images registry, from which it can then promoted to the k8s.gcr.io/e2e-test-images registry by adding a line in kubernetes/k8s.io. For this, you will need the image manifest list's digest, which can be obtained by running:

manifest-tool inspect --raw gcr.io/k8s-staging-e2e-test-images/${IMAGE_NAME}:${VERSION} | jq '.[0].Digest'

All the postsubmit jobs and their logs for all the images can be seen in testgrid.

The images are built through make. Since some images (e.g.: busybox) are used as a base for other images, it is recommended to build them first, if needed.

Creating and promoting new images

If you intend to add an entirely different image and have it automatically built by the Image Builder and used in E2E tests, you will also have to define the postsubmit prow job for it. This can easily be done by running this script in kubernetes/test-infra.

Windows test images considerations

Ideally, the same Dockerfile can be used to build both Windows and Linux images. However, that isn't always possible. If a different Dockerfile is needed for an image, it should be named Dockerfile_windows. When building, image-util.sh will first check for this file name when building Windows images.

The building process uses docker buildx to build both Windows and Linux images, but there are a few limitations when it comes to the Windows images:

• The Dockerfile can have multiple stages, including Windows and Linux stages for the same image, but the Windows stage cannot have any RUN commands (see the agnhost's Dockerfile_windows as an example).
• The Windows stage cannot have any WORKDIR commands due to a bug (https://github.com/docker/buildx/issues/378)
• When copying Windows symlink files to a Windows image, docker buildx changes the symlink target, prepending Files\ to them (https://github.com/docker/buildx/issues/373) (for example, the symlink target C:\bin\busybox.exe becomes Files\C:\bin\busybox.exe). This can be avoided by having symlink targets with relative paths and having the target duplicated (for example, the symlink target busybox.exe becomes Files\busybox.exe when copied, so the binary C:\bin\Files\busybox.exe should exist in order for the symlink to be used correctly). See the busybox's Dockerfile_windows as an example.
• docker buildx overwrites the image's PATH environment variable to a Linux PATH environment variable, which won't work properly on Windows. See https://github.com/moby/buildkit/issues/1560
• The base image for all the Windows images is nanoserver, which is ~10 times smaller than Windows Servercore. Most binaries added to the image will work out of the box, but some will not due to missing dependencies (atention: the image will still build successfully, even if the added binaries will not work). For example, coredns.exe requires netapi32.dll, which cannot be found on a nanoserver image, but we can copy it from a servercore image (see the agnhost image's Dockerfile_windows file as an example). A good rule of thumb is to use 64-bit applications instead of 32-bit as they have fewer dependencies. You can determine what dependencies are missing by running procmon.exe on the container's host (make sure that process isolation is used, not Hyper-V isolation). This is a useful guide on how to use procmon.exe.

Because of the docker buildx limitations regarding the RUN commands mentioned above, we need to use a Windows helper image in order to properly build the Windows images without requiring any Windows nodes in the regular image building process. This helper image can be found in at e2eteam/powershell-helper:6.2.7. It can be used by anyone, but if you need to build your own, you can read more about it here.

For Windows, in order to spawn process-isolated containers, the container OS version should closely match the host OS version. For this reason, we build test images for different Windows OS Versions: 1809, 1903, 1909, 2004, 20H2. In order add support for a new Windows OS version, a new entry for that OS version will have to be first added to the windows-servercore-cache and busybox images, followed by the rest of the images. These images are then used by the rest of the E2E test images as a cache / base image.

Building images

The images are built through make. Since some images (agnhost) are used as a base for other images, it is recommended to build them first, if needed.

An image can be built by simply running the command:

make all WHAT=agnhost

To build AND push an image, the following command can be used:

make all-push WHAT=agnhost

By default, the images will be tagged and pushed under the gcr.io/kubernetes-e2e-test-images registry. That can changed by running this command instead:

REGISTRY=foo_registry make all-push WHAT=agnhost

NOTE (for test gcr.io image publishers): Some tests (e.g.: should serve a basic image on each replica with a private image) require the agnhost image to be published in an authenticated repo as well:

REGISTRY=gcr.io/kubernetes-e2e-test-images make all-push WHAT=agnhost
REGISTRY=gcr.io/k8s-authenticated-test make all-push WHAT=agnhost

Additionally, WHAT=all-conformance can be used to build / push the images most commonly used in E2E Conformance tests.

Testing the new image

Once the image has been built and pushed to an accesible registry, you can run the tests using that image by having the environment variable KUBE_TEST_REPO_LIST set before running the tests that are using the image:

export KUBE_TEST_REPO_LIST=/path/to/repo_list.yaml

repo_list.yaml is a configuration file used by the E2E tests, in which you can set alternative registries to pull the images from. Sample file:

e2eRegistry: your-awesome-registry
promoterE2eRegistry: your-awesome-registry
gcRegistry: your-awesome-registry
sampleRegistry: your-awesome-registry

Keep in mind that some tests are using multiple images, so it is a good idea to also build and push those images.

Finally, make sure to bump the image version used in E2E testing by modifying the file test/utils/image/manifest.go, and recompile afterwards:

./build/run.sh make WHAT=test/e2e/e2e.test

After all the above has been done, run the desired tests.

Known issues and workarounds

docker manifest create fails due to permission denied on /etc/docker/certs.d/gcr.io (https://github.com/docker/for-linux/issues/396). This issue can be resolved by running:

sudo chmod o+x /etc/docker

A few images have been mirrored from dockerhub into the gcr.io/k8s-staging-e2e-test-images registry (busybox, glusterdynamic-provisioner, httpd, httpd-new, nginx, nginx-new, perl), and they only have a noop Dockerfile. However, due to an issue, the same SHA cannot be pushed twice. A small change to them is required in order to generate a new SHA, which can then be pushed and promoted.