Changes/InvokingTests

Standard Discovery, Packaging, Invocation of Integration Tests

This is a proposal
Feedback is more than welcome. There's a discussion tab above.

Summary

Lets define a clear delineation of between a testsuite (including framework) and the CI system that is running the test suite. This is the standard interface.

What follows is a standard way to discover, package and invoke integration tests for a package stored in a Fedora dist-git repo.

Many Fedora packages have unit tests. These tests are typically run during a %check RPM build step and run in a build root. On the other hand, integration testing should happen against a composed system. Upstream projects have integration tests, both Fedora QA and the Atomic Host team would like to create more integration tests, Red Hat would like to bring integration tests upstream.

Owner

• Name: Stef Walter
• Email: stefw@fedoraproject.org
• Name: Martin Pitt
• Email: martin@piware.de

Terminology

• Test Subject: The items that are to be tested.
  • Examples: RPMs, OCI image, ISO, QCow2, Module repository ...
• Test: A callable/runnable piece of code and corresponding test data and mocks which exercises and evaluates a test subject.
• Test Suite: The collection of all tests that apply to a test subject.
• Test Framework: A library or component that the test suite and tests use to accomplish their job.
  • Examples: Avocado, GNOME Installed Tests, Modularity Testing Framework, Ansible tests in Atomic Host, Tunir tests, docker test images, ...
• Test Result: A boolean pass/fail output of a test suite.
  • Test results are for consumption by automated aspects of a testing systems.
• Test Artifact: Any additional output of the test suite such as the stdout/stderr output, log files, screenshots, core dumps, or TAP/Junit/subunit streams.
  • Test artifacts are for consumption by humans, archival or big data analysis.
• Testing System: A CI or other testing system that would like to discover, stage and invoke tests for a test subject.
  • Examples: Jenkins, Taskotron, ZUUL, CentOS CI, Red Hat CI, Travis, Semaphore, Openshift CI/CD, Ubuntu CI, ...

Responsibilities

The testing system is responsible to:

• Build or otherwise acquire the test subject, such as package, container image, tree …
• Decide which test suite to run, often by using the standard interface to discover appropriate tests for the dist-git repo that a test subject originated in.
• Schedule, provision or orchestrate a job to run the test suite on appropriate compute, storage, ...
• Stage the test suite as described by the standard interface.
• Invoke the test suite as described by the standard interface.
• Gather the test results and test artifacts as described by the standard interface.
• Announce and relay the test results and test artifacts for gating, archival ...

The standard interface describes how to:

• Discover a test suite for a given dist-git repo.
• Uniquely identify a test suite.
• Stage a test suite and its dependencies such as test frameworks.
• Provide the test subject to the test suite.
• Invoke a test suite in a consistent way.
• Gather test results and test artifacts from the invoked test suite.

The test suite is responsible to:

• Declare its dependencies such as a test framework via the standard interface.
• Execute the test framework as necessary.
• Provision (usually locally) any containers or virtual machines necessary for testing the test subject.
• Provide test results and test subjects back according to the standard

The format of the textual logs and test artifacts that come out of a test suite is not prescribed by this document. Nor is it envisioned to be standardized across all possible test suites.

Requirements

• The test suite and test framework SHOULD NOT leak its implementation details into the testing system, other than via the standard interface.
• The test suite and test framework SHOULD NOT rely on behavior of the testing system other than the standard interface.
• The standard interface MUST enable a dist-git packager to run a test suite locally.
  • Test suites or test frameworks MAY call out to the network for certain tasks.
• It MUST be possible to stage an upstream test suite using the standard interface.
• Both in-situ tests, and more rigorous outside-in tests MUST be possible with the standard interface.
  • For in-situ tests the test suite is in the same file system tree and process space as the test subject.
  • For outside-in tests the test suite is outside of the file system tree and process space of the test subject.
• The test suite and test framework SHOULD be able to provision containers and virtual machines necessary for its testing without requesting them from the testing system.
• The standard interface SHOULD describe how to uniquely identify a test suite,

Detailed Description

This standard interface describes how to discover, stage and invoke tests. It is important to cleanly separate implementation details of the testing system from the test suite and its framework. It is also important to allow packagers to locally and manually invoke a test suite.

Packaging

The integration tests are packaged and delivered through Fedora as packages.

Each dist-git repo that has integration tests should package those tests in one or more subpackages like %{name}-tests. This is similar to the %{name}-debuginfo or %{name}-docs subpackages we have today.

The spec file for a dist-git repo may install upstream integration tests as files in its %{name}-tests package. The spec file may also include tests directly from files in tests/ subdirectory of the dist-git repo itself.

The tests package should use Requires: to require any other package, testing framework, or dependency necessary to run the tests. In in-situ testing cases, the tests package will directly Requires: the package of the test subject.

Invocation

To invoke the test suite, the test package that contains it is installed. Each test of the suite installs an executable in the path /usr/tests/sourcepackage/ (this will avoid name collisions between packages).

To invoke the test suite, one would:

1. Create a temporary directory, referred to as: $TESTDIR
2. Place the test subject(s) being tested in $TESTDIR/subjects/
3. Execute all executable files in /usr/tests/*/ directories one at a time.
   1. Each executable test is invoked with a working directory of $TESTDIR
   2. Each executable test is invoked as root, and may drop privileges as desired.
   3. Treat the stdout/stderr of the test process as the test log. This is a standard test artifact and written to $TESTDIR/artifacts/testname.log.
   4. Examine the exit code of each test process. Zero exit code is a successful test result, non-zero is failure.
4. Tests can put any additional test artifacts like screenshots into $TESTDIR/artifacts/.

This ensures that tests can be run on a production system without accidentally clobbering permanent directories, don't require root privileges (simplifies test development), and that CI systems have one unique place from where to collect artifacts. It also avoids collecting temporary files such as downloaded container or VM images as artifacts, as these would usually get stored for a longer time period.

These steps would usually be done through a standard test driver tool (particularly for sensible stdout/stderr teeing and log capturing), but its usage is not mandatory for developing and calling tests manually.

Staging

The %{name}-test package should Requires: all other packages that the testsuite executable needs in order to run. This includes libraries or frameworks, or subsystems like libvirt.

Some integration tests may choose to test in-situ, on the system on which the test suite is installed. In these cases the %{name}-tests package should directly depend on the package being tested.

More rigorous integration tests are outside-in. They test an integrated system without affecting its contents. It is the responsibility of the %{name}-tests packages to provision virtual machines or containers necessary to do such testing. In almost all cases this will happen by way of a provisioning framework such as Avocado, Ansible, Module Testing Framework, linch-pin, etc.

Multiple tests packages may be installed as long as their dependencies do not conflict.

Discovery

A testing system needs to be able to efficiently answer the question "does this subject have any tests packages, and if so, what are their names". This should be automatically discoverable to the extent possible.

For any RPM test subject this process requires no additional metadata and can be fully automatic:

• It is possible to map a RPM to its SRPM source package (<rpm:sourcerpm> in the package index *-primary.xml.gz).
• One can map an SRPM to all the RPMs that it builds (from the same index), and using the *-filelists.xml.gz index one can mechanically tell which of the RPMs are of this test package kind described here.

TODO: For other types of test subject cases such as docker images or distribution ISO files this discovery still needs to be discussed.

• E. g. a Dockerfile might grow a reference to a test package RPM, or at least initially there is a manually maintained map of subject to test package in the testing system.

Scope

This change requires no initial changes to Fedora infrastructure itself. The change only affects contents spec files in dist-git repos.

TODO: However certain key infrastructure changes could mitigate usability or side-effects of this change. In particular, once this grows beyond the experimental phase, these test packages need to be put into a separate archive, similar to -debuginfo.

• How much effort is that to set up?
• Does this require any additional tags, keywords, or other explicit declaration in the spec file, other than "this RPM ships something in /usr/tests/*"?

Benefit to Fedora

Developers benefit by having a consistent target for how to describe tests, while also being able to execute them locally while debugging issues or iterating on tests.

By packaging, staging and invoking tests consistently in Fedora we create an eco-system for the tests that allows varied test frameworks as well as CI system infrastructure to interoperate. The integration tests outlast the implementation details of either the frameworks they're written in or the CI systems running them.

User Experience

A standard way to package tests benefits Fedora stability, and makes Fedora better for users.

Users could also benefit by having tests that they can reproduce on their own systems. They could install the similar to how they consume %{name}-doc or %{name}-debuginfo subpackages today.

We may choose to avoid having such packages available in the standard repositories. We may choose to only have them in updates-testing or an arrangement similar to debuginfo. These choices will require some markup and/or change to infrastructure.

Upgrade/compatibility impact

Although there may already be packages that are named %{name}-tests this is merely a convention, and such packages will not affect the behavior of this proposal.

Comparison with Debian's autopkgtest

Debian/Ubuntu have used CI with packaged tests (called "autopkgtests") for many years, with over 7.000 tests. These are good candidates or at least bases for taking into Fedora packages. This compares the structure of autopkgtest with this proposal to learn from autopkgtest's experiences and take what works, and justifies the differences. See the format definition for details.

Packaging

Similarities: Both specifications use an existing test metatada format (RPM spec files with Requires: here, Debian RFC822 control files with Depends:in autopkgtest).

Differences:

• This specification requires packaging tests as binary RPM packages, whereas autopkgtest opted for keeping the test in the source package (equivalent of dist-git) only. The latter avoids the overhead of packaging the tests and having to create a separate archive for them. An important point is also that installing an RPM -test package requires root privileges, while invoking autopkgtest doesn't.
• As autopkgtest uses a separate control file (debian/tests/control instead of debian/control which describes the binary packages), it offers a much richer set of test metadata which cannot be expressed with debian/control or RPM spec files.

Invocation

Simimlarities: The test interface is very similar: In both specifications, a test is an executable (of any script or compiled language), the exit code is the primary indicator of pass/fail, the executable's stdout/err is a standard test artifact ("test log"), and tests can write additional artifacts into the $AUTOPKGTEST_ARTIFACTS dir (like ./artifacts/ here).

Differences:

• By default, autopkgtest considers a test as failed if it produces anything on stderr, for catching unexpected new warnings. This can be disabled with adding Restrictions: allow-stderr to the test metadata. However, this turned out to be not overly useful, and tests which want to intercept warnings should better do that themselves.
• autopkgtest has no concept of passing test subjects to the test. Tests expect that their subjects are already available/installed, i. e. they get called in a testbed of the desired kind and state. It is the responsibility of the autopkgtest command line tool (the "test driver/executor") to install proposed new package(s) into the testbed (due to its origin of being primarily focussed on testing packages). For testing desktop/cloud images, upgrades, or other non-package subjects, it is instead the testing system's responsibility to produce the desired testbed and call autopkgtest on it. As the scope of this specification puts the staging into the hand of the test instead of the testing system, passing the test subject is a necessary consequence.

Staging

Similarities: Both specifications use standard dpkg/rpm package dependencies (Depends: for dpkg, Requires: for rpm) to pull in test dependencies, and both can opt into doing their own provisioning of containers/VMs etc. for doing outside-in tests instead of in-situ. However, of all the ~ 7.000 autopkgtests, only a small handful is actually doing that (known cases are systemd and open-iscsi), as the vast majority of package/upgrade/image tests can (because it's sufficient) and should (because it's magnitudes faster) be run in-situ.

Differences: Here the test itself is responsible for installing the test subjects, while in autopkgtest it's the testing system's responsibility (see above).

Discovery

Similarities: The idea is the same in both specifications. Here, as soon as there is a binary package that ships /usr/tests/* it can be discovered through file lists. In autopkgtest, as soon as there is a debian/tests/control, the source package index entry will automatically get a Testsuite: autopkgtest tag. So in both cases the developer does not need to explicitly do anything other than adding the tests.

Differences: None concerning the interface, just technical implementation details due to how rpm/dpkg work.

Examples

What follows are examples of writing and/or packaging existing tests to this standard.

There is a mock test system' which is a simple shell script: run-installed-test. It runs all /usr/tests/*, can pass arbitrary subjects to them, and report/capture the results/logs. This is purely to study what a CI system would do and whether the standard interface works.

Example: Simple in-situ test

Add simple downstream integration test for gzip:

• Package: gzip-tests
• dist-git: https://github.com/martinpitt/fedora-gzip-test
• Reference: https://patches.ubuntu.com/g/gzip/ (taken verbatim from Ubuntu's autopkgtest)

With this you can install test RPM from above gzip repo:

 $ sudo rpm -i results_gzip/1.8/2.fc27/gzip-tests-1.8-2.fc25.x86_64.rpm

and run the gzip tests on the already installed package (as user) with

 $ ~/run-installed-test
 Subjects/artifacts directory: /tmp/test.vsR
 -----------------------------------------
 Running /usr/tests/gzip/test-simple
 -----------------------------------------
 ++ ls 'subjects/*.rpm'
 + echo Bla
 + cp bla.file bla.file.orig
 + gzip bla.file
 + gunzip bla.file.gz
 + cmp bla.file bla.file.orig
 + rm bla.file bla.file.orig
 PASS: /usr/tests/gzip/test-simple

 $ ls -l /tmp/test.vsR/artifacts/
 -rw-r--r-- 1 martin martin 156 Mar 28 16:49 test-simple.log

or run them as root (as officially specified) with a subject (locally built gzip RPM):

 $ sudo ~/run-installed-test results_gzip/1.8/2.fc27/gzip-1.8-2.fc25.x86_64.rpm
 Installing subject results_gzip/1.8/2.fc27/gzip-1.8-2.fc25.x86_64.rpm
 Subjects/artifacts directory: /tmp/test.Cck
 -----------------------------------------
 Running /usr/tests/gzip/test-simple
 -----------------------------------------
 ++ ls subjects/gzip-1.8-2.fc25.x86_64.rpm
 + '[' -w / ']'
 + rpm --verbose --force -U subjects/gzip-1.8-2.fc25.x86_64.rpm
 Preparing packages...
 gzip-1.8-2.fc25.x86_64
 + echo Bla
 + cp bla.file bla.file.orig
 + gzip bla.file
 + gunzip bla.file.gz
 + cmp bla.file bla.file.orig
 + rm bla.file bla.file.orig
 PASS: /usr/tests/gzip/test-simple

Example: GNOME style "Installed Tests"

Add downstream integration test running in gnome installed tests.

• Package: glib2-tests
• dist-git: https://github.com/martinpitt/fedora-glib2-test
• Reference: https://wiki.gnome.org/Initiatives/GnomeGoals/InstalledTests

Example: Tests run in Docker Container

Add integration test running glib2 installed tests in a docker container. This is also an example of having two different tests packages being created by the same dist-git repo.

• Package: glib2-container-tests
• dist-git: https://github.com/martinpitt/fedora-glib2-test
• Based on Debian/Ubuntu's autopkgtest glib2 related script

Example: Modularity testing Framework

TODO: Port an example

Example: Ansible with Atomic Host

TODO: Port an existing test

Example: Beakerlib based test

TODO: Port and shim a beakerlib test

Example: Cockpit upstream test

Run upstream integration test, which uses VMs through libvirt, in a docker container; the entire libvirt/bridge setup is confined to the container, so this can be run without interfering with the host system.

• Package: cockpit-tests
• dist-git: https://github.com/martinpitt/cockpit/commits/packaged-tests - The penultimate commit creates a new cockpit-integration-tests package containing the actual tests and libvirt/qemu etc. dependencies, the topmost commit provides the /usr/tests/ script that runs the former in docker.
• Reference: https://github.com/cockpit-project/cockpit/blob/master/test/README

Evaluation

Instructions: Copy the block below, sign your name and fill in each section with your evaluation of that aspect. Add additional bullet points with overall summary or notes.

Full Name -- SignAture

• Summary: ...
• Staging: ...
• Invocation: ...
• Discovery: ...

Stef Walter -- Stefw

• Summary:
  • Disclaimer: I am one of the owners above.
  • PRO: RPM is used for staging. RPM and YUM-style reposotiries are a standard part of Fedora. No other technology is involved in the standard.
  • PRO: Super simple Unix invocation mechanism: executable + stdin/stdout + environment variables.
  • CON: RPM has a learning curve. Although a dist-git maintainer is required to already know about this.
  • CON: /usr/tests is a new FHS directory, should probably be /usr/libexec/tests.
  • CON: Only a partial way to describe whether tests are compatible with or conflict with a specific NVR of test subjects.
  • CON: The *-tests packages may require special handling if the distro does not want to have users able to install/run tests.
• Staging:
  • Requires rpm and yum/dnf as well known staging dependencies.
  • The *-tests suffix is implied by the standard, not required. Is this confusing?
• Invocation:
  • The standard describes how multiple test suites can be staged together and executed in one shot.
• Discovery:
  • An NVR is the unique identifier for a test suite.
  • This uses capabilities of how YUM repositories work, but requires no additional technology.