Kubetest is the interface for launching and running e2e tests.
See the contributor documentation for information about e2e testing.
Kubetest sits between bootstrap.py and various parts of the e2e lifecycle.
The bootstrap.py 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:
--buildkubernetes,--stagethis build to gcs,--extracta staged build from gcs,- turn
--upa new cluster using various--deploymentstrategies, --testthis cluster using ginkgo--test_args--dumplogs to a local folder, and finally- turn
--downthe cluster after completing testing, --timeoutafter a particular duration (allowing extra time to clean up).
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 clone this repo and then run GO111MODULE=on go install ./kubetest from the clone.
Common alternatives:
go install k8s.io/test-infra/kubetest # 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).
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.
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.
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.
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.
The --boskos-wait-duration flag defines how long Kubetest waits on an Boskos resource
becomes available before quitting the job, default value is 5 minutes.
See the boskos docs for more details.
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.
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" --gcp-project=google.com:some-project
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_skew is the directory
corresponding to release indicated by first --extract flag, while kubernetes
corresponds to second flag.
The first --extract downloads and extracts the specified release into the kubernetes
directory. The second --extract renames the kubernetes directory as kubernetes_skew
and then downloads the second release into a new kubernetes directory.
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" --gcp-project=google.com:some-project
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" --gcp-project=google.com:some-project