Skip to content
Branch: master
Find file History
Type Name Latest commit message Commit time
Failed to load latest commit information.
conformance kubetest: implement "KubectlCommand" Nov 30, 2018
e2e Refactor conformance & dind to use runner May 9, 2018
kind build kind with modules May 9, 2019
kubeadmdind Updated OWNERS files to include link to docs Jan 30, 2019
process Make kubetest pass lint Jun 12, 2018
util Fix 48 typos. Nov 12, 2018
BUILD.bazel Fix azblob usage in kubetest Mar 26, 2019
OWNERS bentheelder is no longer oncall Mar 18, 2019 clean up unused old federation code in kubetest Jan 29, 2019
anywhere.go kubetest/anywhere: write the NUM_NODES env variable Feb 24, 2019
anywhere_test.go remove multicluster bits from k/anywhere.go Jan 29, 2019
azure.go Merge pull request #11624 from adelina-t/agent_pool_count_update Apr 1, 2019
azure_helpers.go Skip resource group deletion errors because of bugs in go-autorest Mar 22, 2019
bash.go Support setting SSH_KUBE_BASTION env in kubetest. Apr 23, 2019
build.go kubetest: build windows and linux artifcats with bazel Apr 20, 2019
dump.go kubetest log dump: capture full systemd journal Oct 9, 2018
dump_test.go kubetest log dump: capture full systemd journal Oct 9, 2018
e2e.go Make it possible not to delete kubemark master. May 9, 2019
eks.go Merge pull request #10727 from shyamjvs/set-envs-for-eks-deployer Jan 15, 2019
extract_k8s.go Add method to fetch release version with http client Mar 8, 2019
extract_test.go Add method to fetch release version with http client Mar 8, 2019
gke.go Support setting SSH_KUBE_BASTION env in kubetest. Apr 23, 2019
kops.go Change kops overrides flag to string array Jan 28, 2019
kubernetes.go Make kubetest pass lint Jun 12, 2018
kubernetes_test.go kubetest: Add unit tests for node functions Oct 27, 2017
local.go Drop sudo from local-e2e DumpClusterLogs Jan 21, 2019
main.go Support setting SSH_KUBE_BASTION env in kubetest. Apr 23, 2019
main_test.go make kubetest write version to `revision` instead of `job-version` as Dec 5, 2018
node.go kubetest: implement "KubectlCommand" Nov 30, 2018
none.go kubetest: implement "KubectlCommand" Nov 30, 2018
stage.go clean up unused old federation code in kubetest Jan 29, 2019
util.go Support setting SSH_KUBE_BASTION env in kubetest. Apr 23, 2019
util_test.go Move kubetest utils to separate package Feb 24, 2018


Kubetest is the interface for launching and running e2e tests.

See the contributor documentation for information about e2e testing.

Kubetest sits between and various parts of the e2e lifecycle.

The library is a nominal/optional part of prow. This library is responsible for:

  • checking out each repository correctly
  • starting kubetest (or whatever the test binary is for other jobs)
  • uploading the test result (including artifacts) to gcs

The e2e lifecycle may:

  • --build kubernetes,
  • --stage this build to gcs,
  • --extract a staged build from gcs,
  • turn --up a new cluster using various --deployment strategies,
  • --test this cluster using ginkgo --test_args
  • --dump logs to a local folder, and finally
  • turn --down the cluster after completing testing,
  • --timeout after a particular duration (allowing extra time to clean up).

Note that developers frequently use kubetest by calling go run hack/e2e.go in the kubernetes/kubernetes repository. This hack/e2e.go program is a wrapper around updating kubetest (at most once a day) before calling it.

If you're making a change to kubetest, watch the canary jobs after it has merged; these run the latest images of kubekins-e2e, and thus don't have to wait for a version bump in the prow config. The canary jobs are used to give early signal on whether new kubetest features are working as intended, before bumping the image for general e2e tests.


Please run go get -u to install kubetest.

Common alternatives:

go run hack/e2e.go  # from kubernetes/kubernetes
go install  # if you check out test-infra
bazel run //kubetest  # use bazel to build and run


Right now kubetest is expected to run at head, regardless of the version of kubernetes being targeted.

Most e2e images, such as kubekins-e2e and kubekins-e2e-prow compile the latest version of kubetest whenever the image is updated (most updates to these images are done in order to update kubetest).


If PWD is in the kubernetes/kubernetes directory --build will build whatever changes you have made into a quick release.

Control the details of the --build=bazel by appending one of the build modes (see help for current list).

Stage a build

It is inefficient for every job to rebuild the same version. Instead our CI system defines build jobs which --stage the build somewhere on GCS. Some providers such as GKE require a staged build, whereas others like GCE allow you to scp over the binaries directly to each node.

Extract a build

Aside from the build jobs, most of our CI systems --extract a prebuilt version. This saves a bunch of time compiling.

The most common options are either a specific version--extract=v1.7.0-beta.1, a release --extract=release/stable or --extract=ci/latest-1.8.

Note that you can extract 1 or 2 versions. Using 2 versions is useful for skew and upgrade testing.

See extract_k8s.go for further details.


There are various ways to deploy kubernetes. Choose a strategy with the --deployment flag (for example --deployment=kops or --deployment=kubernetes-anywhere). See kubetest --help for a full list of options.


The --up flag will tell kubetest to turn up a new cluster for you.

It will first attempt to tear down an old instance of the same cluster.

Currently requires a complicated set of flags and environment variables such as --gcp-project, etc.

We are in the process of converting all environment variables into flags. See the current set of flag options with kubetest -h.

Save/load credentials

The --save flag tells kubetest to upload your cluster credentials onto gcs somewhere. Later calling kubetest --save without an --up flag tells kubetest to load these credentials instead of turning up a new cluster.

Dynamic project selection

Most e2e jobs assume control of a GCP project (see leaks section below).

If kubetest is running inside a pod then it will attempt to talk to boskos to dynamically reserve a project.

This makes it easier for developers to add and remove jobs. With boskos they no longer need to worry about creating, provisioning, naming, etc a project for this new job.

See the boskos docs for more details.

Dump logs

The --dump flag tells kubetest to try and collect logs if there is a problem. Typically this means master and node logs.

Collecting these logs may take a long time. This typically involves sshing to each node, searching for and downloading any relevant logs.

There is also a --logexporter-gcs-path option which tells kubetest to run a container on each node which uploads logs directly to GCS. This dramatically reduces time required to dump logs, especially for scalability tests.


The --down flag tells kubetest to clear up the cluster after finishing.

Kubetest will try its best to tear down the cluster in spite of problems such as failing --up, --test, etc. The --timeout also includes some buffer to allow time for --down to clean up.


The --check-leaked-resources option tells kubetest to look for any extra GCP resources after tearing down the cluster.

The expectation is that any resources created by kubernetes will be cleaned up during cluster teardown.

This logic may be buggy so this options takes a snapshot of the resources at various points in time (start, after cluster up, after testing, after cluster down) and ensures that there are no resources present after down that weren't already present at the start.


Most testing uses ginkgo but there are other options available.


The --test flag tells kubetest to run the test.e2e binary built/extracted from the kubernetes/kubernetes repo.

Typically jobs also include a --test_args=--ginkgo.focus=FOO --ginkgo.skip=BAR flag to filter down to a particular set of interesting tests.

Upgrade, skew, kubemark

You can also run --kubemark tests instead of the standard tests.

Tests can use --skew and --upgrade_args if they provided multiple --extract flags (or manually created a kubernetes/kubernetes_skew directory as a sibling to kubernetes/kubernetes). This will cause tests to run from the skew directory, potentially to upgrade/downgrade kubernetes to another version.

A simple example is:
kubetest --up --check-version-skew=false --extract=v1.8.0-beta.1 --extract=v1.7.5 --upgrade_args=--ginkgo.focus="Feature:SomeUpgrade"

The command runs all (and only) upgrade tests tagged with Feature:SomeUpgrade label on GCE. The command downloads v1.7.5 and v1.8.0-beta.1 releases, unzips downloaded files, and runs the tests to upgrade the cluster from v1.7.5 to v1.8.0-beta.1. You will be able to find 2 new directories named kubernetes and kubernetes_skew at current directory. kubernetes is the directory corresponding to release indicated by first --extract flag, while kubernetes_skew corresponds to second flag.

Note that order of the 2 --extract flags matters: --extract=v2 --extract=v1 means upgrading from v1 to v2. The command does not run other e2e tests after completing the upgrade tests. If you want to run the e2e tests, specify also --test and --test_args flags.

Tips: CI upgrade tests listed at sig-cluster-lifecycle config show flags used in the real CI test environment, which is a good source to learn more about how the flags are used.


If you want to create a new release with your own changes, you have to upload built manifests to gcs. The command is very similar: kubetest --build --stage=gs://some/path/to/v1.8.0-beta.1 --check-version-skew=false --extract=gs://some/path/to/v1.8.0-beta.1 --extract=v1.7.5 --upgrade_args=--ginkgo.focus="Feature:SomeUpgrade"

If you already have release of a specific version, you do not need to fetch the release again. For instance, if you have v1.7.5 release and its directory is at the right path, the command below does the same as above: kubetest --build --stage=gs://some/path/to/v1.8.0-beta.1 --check-version-skew=false --extract=gs://some/path/to/v1.8.0-beta.1 --upgrade_args=--ginkgo.focus="Feature:SomeUpgrade"

You can’t perform that action at this time.