The Chromium Test Executable API

bit.ly/chromium-test-runner-api (*)

Introduction

This document defines the API that test executables must implement in order to be run on the Chromium continuous integration infrastructure (the LUCI system using the chromium and chromium_trybot recipes).

NOTE: This document specifies the existing isolated_scripts API in the Chromium recipe. Currently we also support other APIs (e.g., for GTests), but we should migrate them to use the isolated_scripts API. That work is not currently scheduled.

This spec applies only to functional tests and does not attempt to specify how performance tests should work, though in principle they could probably work the same way and possibly just produce different output.

This document is specifically targeted at Chromium and assumes you are using GN and Ninja for your build system. It should be possible to adapt these APIs to other projects and build recipes, but this is not an immediate goal. Similarly, if a project adapts this API and the related specifications it should be able to reuse the functionality and tooling we‘ve built out for Chromium’s CI system more easily in other LUCI deployments.


NOTE: It bears repeating that this describes the current state of affairs, and not the desired end state. A companion doc, Cleaning up the Chromium Testing Environment, discusses a possible path forward and end state.


Building and Invoking a Test Executable

There are lots of different kinds of tests, but we want to be able to build and invoke them uniformly, regardless of how they are implemented.

We will call the thing being executed to run the tests a test executable (or executable for short). This is not an ideal name, as this doesn't necessarily refer to a GN executable target type; it may be a wrapper script that invokes other binaries or scripts to run the tests.

We expect the test executable to run one or more tests. A test must be an atomically addressable thing with a name that is unique to that invocation of the executable, i.e., we expect that we can pass a list of test names to the test executable and only run just those tests. Test names must not contain a “::” (which is used as a separator between test names) and must not contain a “*” (which could be confused with a glob character) or start with a “-” (which would be confused with an indicator that you should skip the test). Test names should generally only contain ASCII code points, as the infrastructure does not currently guarantee that non-ASCII code points will work correctly everywhere. We do not specify test naming conventions beyond these requirements, and it is fully permissible for a test to contain multiple assertions which may pass or fail; this design does not specify a way to interpret or handle those “sub-atomic” assertions; their existence is opaque to this design. In particular, this spec does not provide a particular way to identify and handle parameterized tests, or to do anything with test suites beyond a supporting a limited form of globbing for specifying sets of test names.

To configure a new test, you need to modify one to three files:

  • The test must be listed in one or more test suites in //testing/buildbot/test_suites.pyl. Most commonly the test will be defined as a single string (e.g., “base_unittests”), which keys into an entry in //testing/buildbot/gn_isolate_map.pyl. In some cases, tests will reference a target and add additional command line arguments. These entries (along with //testing/buildbot/test_suite_exceptions.pyl and //testing/buildbot/waterfalls.pyl) determine where the tests will be run. For more information on how these files work, see //testing/buildbot/README.md
  • Tests entries must ultimately reference an entry in //testing/buildbot/gn_isolate_map.pyl. This file contains the mapping of ninja compile targets to GN targets (specifying the GN label for the latter); we need this mapping in order to be able to run gn analyze against a patch to see which targets are affected by a patch. This file also tells MB what kind of test an entry is (so we can form the correct command line) and may specify additional command line flags. If you are creating a test that is only a variant of an existing test, this may be the only file you need to modify. (Technically, you could define a new test solely in test_suites.pyl and reference existing gn_isolate_map.pyl entries, but this is considered bad practice).
  • Add the GN target itself to the appropriate build files. Make sure this GN target contains all of the data and data_deps entries needed to ensure the test isolate has all the files the test needs to run. If your test doesn‘t depend on new build targets or add additional data file dependencies, you likely don’t need this. However, this is increasingly uncommon.

Command Line Arguments

The executable must support the following command line arguments (aka flags):

--isolated-outdir=[PATH]

This argument is required, and should be set to the directory created by the swarming task for the task to write outputs into.

--out-dir=[PATH]

This argument mirrors --isolated-outdir, but may appear in addition to it depending on the bot configuration (e.g. IOS bots that specify the out_dir_arg mixin in //testing/buildbot/waterfalls.pyl). It only needs to be handled in these cases.

--isolated-script-test-output=[FILENAME]

This argument is optional. If this argument is provided, the executable must write the results of the test run in the JSON Test Results Format into that file. If this argument is not given to the executable, the executable must not write the output anywhere. The executable should only write a valid version of the file, and generally should only do this at the end of the test run. This means that if the run is interrupted, you may not get the results of what did run, but that is acceptable.

--isolated-script-test-filter=[STRING]

This argument is optional. If this argument is provided, it must be a double-colon-separated list of strings, where each string either uniquely identifies a full test name or is a prefix plus a “*” on the end (to form a glob). The executable must run only the test matching those names or globs. “*” is only supported at the end, i.e., ‘Foo.*’ is legal, but ‘*.bar’ is not. If the string has a “-” at the front, the test (or glob of tests) must be skipped, not run. This matches how test names are specified in the simple form of the Chromium Test List Format. We use the double colon as a separator because most other common punctuation characters can occur in test names (some test suites use URLs as test names, for example). This argument may be provided multiple times; how to treat multiple occurrences (and how this arg interacts with --isolated-script-test-filter-file) is described below.

--isolated-script-test-filter-file=[FILENAME]

If provided, the executable must read the given filename to determine which tests to run and what to expect their results to be. The file must be in the Chromium Test List Format (either the simple or tagged formats are fine). This argument may be provided multiple times; how to treat multiple occurrences (and how this arg interacts with --isolated-script-test-filter) is described below.

--isolated-script-test-launcher-retry-limit=N

By default, tests are run only once if they succeed. If they fail, we will retry the test up to N times (so, for N+1 total invocations of the test) looking for a success (and stop retrying once the test has succeed). By default, the value of N is 3. To turn off retries, pass --isolated-script-test-launcher-retry-limit=0. If this flag is provided, it is an error to also pass --isolated-script-test-repeat (since -repeat specifies an explicit number of times to run the test, it makes no sense to also pass --retry-limit).

--isolated-script-test-repeat=N

If provided, the executable must run a given test N times (total), regardless of whether the test passes or fails. By default, tests are only run once (N=1) if the test matches an expected result or passes, otherwise it may be retried until it succeeds, as governed by --isolated-script-test-launcher-retry-limit, above. If this flag is provided, it is an error to also pass --isolated-script-test-launcher-retry-limit (since -repeat specifies an explicit number of times to run the test, it makes no sense to also pass -retry-limit).

--xcode-build-version [VERSION]

This flag is passed to scripts on IOS bots only, due to the xcode_14_main mixin in //testing/builtbot/waterfalls.pyl.

--xctest

This flag is passed to scripts on IOS bots only, due to the xctest mixin in //testing/builtbot/waterfalls.pyl.

If “--” is passed as an argument:

  • If the executable is a wrapper that invokes another underlying executable, then the wrapper must handle arguments passed before the “--” on the command line (and must error out if it doesn‘t know how to do that), and must pass through any arguments following the “--” unmodified to the underlying executable (and otherwise ignore them rather than erroring out if it doesn’t know how to interpret them).
  • If the executable is not a wrapper, but rather invokes the tests directly, it should handle all of the arguments and otherwise ignore the “--”. The executable should error out if it gets arguments it can't handle, but it is not required to do so.

If “--” is not passed, the executable should error out if it gets arguments it doesn't know how to handle, but it is not required to do so.

If the test executable produces artifacts, they should be written to the location specified by the dirname of the --isolated-script-test-output argument). If the --isolated-script-test-output-argument is not specified, the executable should store the tests somewhere under the root_build_dir, but there is no standard for how to do this currently (most tests do not produce artifacts).

The flag names are purposely chosen to be long in order to not conflict with other flags the executable might support.

Environment variables

The executable must check for and honor the following environment variables:

GTEST_SHARD_INDEX=[N]

This environment variable is optional, but if it is provided, it partially determines (along with GTEST_TOTAL_SHARDS) which fixed subset of tests (or “shard”) to run. GTEST_TOTAL_SHARDS must also be set, and GTEST_SHARD_INDEX must be set to an integer between 0 and GTEST_TOTAL_SHARDS. Determining which tests to run is described below.

GTEST_TOTAL_SHARDS=[N]

This environment variable is optional, but if it is provided, it partially determines (along with GTEST_TOTAL_SHARDS) which fixed subset of tests (or “shard”) to run. It must be set to a non-zero integer. Determining which tests to run is described below.

Exit codes (aka return codes or return values)

The executable must return 0 for a completely successful run, and a non-zero result if something failed. The following codes are recommended (2 and 130 coming from UNIX conventions):

ValueMeaning
0 (zero)The executable ran to completion and all tests either ran as expected or passed unexpectedly.
1The executable ran to completion but some tests produced unexpectedly failing results.
2The executable failed to start, most likely due to unrecognized or unsupported command line arguments.
130The executable run was aborted the user (or caller) in a semi-orderly manner (aka SIGKILL or Ctrl-C).

Filtering which tests to run

By default, the executable must run every test it knows about. However, as noted above, the --isolated-script-test-filter and --isolated-script-test-filter-file flags can be used to customize which tests to run. Either or both flags may be used, and either may be specified multiple times.

The interaction is as follows:

  • A test should be run only if it would be run when every flag is evaluated individually.
  • A test should be skipped if it would be skipped if any flag was evaluated individually.

If multiple filters in a flag match a given test name, the longest match takes priority (longest match wins). I.e.,. if you had --isolated-script-test-filter='a*::-ab*', then ace.html would run but abd.html would not. The order of the filters should not matter. It is an error to have multiple expressions of the same length that conflict (e.g., a*::-a*).

Examples are given below.

It may not be obvious why we need to support these flags being used multiple times, or together. There are two main sets of reasons:

  • First, you may want to use multiple -filter-file arguments to specify multiple sets of test expectations (e.g., the base test expectations and then MSAN-specific expectations), or to specify expectations in one file and list which tests to run in a separate file.
  • Second, the way the Chromium recipes work, in order to retry a test step to confirm test failures, the recipe doesn't want to have to parse the existing command line, it just wants to append --isolated-script-test-filter and list the tests that fail, and this can cause the --isolated-script-test-filter argument to be listed multiple times (or in conjunction with --isolated-script-test-filter-file).

You cannot practically use these mechanisms to run equally sized subsets of the tests, so if you want to do the latter, use GTEST_SHARD_INDEX and GTEST_TOTAL_SHARDS instead, as described in the next section.

Running equally-sized subsets of tests (shards)

If the GTEST_SHARD_INDEX and GTEST_TOTAL_SHARDS environment variables are set, GTEST_TOTAL_SHARDS must be set to a non-zero integer N, and GTEST_SHARD_INDEX must be set to an integer M between 0 and N-1. Given those two values, the executable must run only every Nth test starting at test number M (i.e., every ith test where (i mod N) == M). dd

This mechanism produces roughly equally-sized sets of tests that will hopefully take roughly equal times to execute, but cannot guarantee the latter property to any degree of precision. If you need them to be as close to the same duration as possible, you will need a more complicated process. For example, you could run all of the tests once to determine their individual running times, and then build up lists of tests based on that, or do something even more complicated based on multiple test runs to smooth over variance in test execution times. Chromium does not currently attempt to do this for functional tests, but we do something similar for performance tests in order to better achieve equal running times and device affinity for consistent results.

You cannot practically use the sharding mechanism to run a stable named set of tests, so if you want to do the latter, use the --isolated-script-test-filter flags instead, as described in the previous section.

Which tests are in which shard must be determined after tests have been filtered out using the --isolated-script-test-filter(-file) flags.

The order that tests are run in is not otherwise specified, but tests are commonly run either in lexicographic order or in a semi-fixed random order; the latter is useful to help identify inter-test dependencies, i.e., tests that rely on the results of previous tests having run in order to pass (such tests are generally considered to be undesirable).

Examples

Assume that out/Default is a debug build (i.e., that the “Debug” tag will apply), and that you have tests named Foo.Bar.bar{1,2,3}, Foo.Bar.baz, and Foo.Quux.quux, and the following two filter files:

$ cat filter1
Foo.Bar.*
-Foo.Bar.bar3
$ cat filter2
# tags: [ Debug Release ]
[ Debug ] Foo.Bar.bar2 [ Skip ]
$

Filtering tests on the command line

$ out/Default/bin/run_foo_tests \
    --isolated_script-test-filter='Foo.Bar.*::-Foo.Bar.bar3'
[1/2] Foo.Bar.bar1 passed in 0.1s
[2/2] Foo.Bar.bar2 passed in 0.13s

2 tests passed in 0.23s, 0 skipped, 0 failures.
$

Using a filter file

$ out/Default/bin/run_foo_tests --isolated-script-test-filter-file=filter1
[1/2] Foo.Bar.bar1 passed in 0.1s
[2/2] Foo.Bar.bar2 passed in 0.13s

2 tests passed in 0.23s, 0 skipped, 0 failures.

Combining multiple filters

$ out/Default/bin/run_foo_tests --isolated-script-test-filter='Foo.Bar.*' \
    --isolated-script-test-filter='Foo.Bar.bar2'
[1/1] Foo.Bar.bar2 passed in 0.13s

All 2 tests completed successfully in 0.13s
$ out/Default/bin/run_foo_tests --isolated-script-test-filter='Foo.Bar.* \
    --isolated-script-test-filter='Foo.Baz.baz'
No tests to run.
$ out/Default/bin/run_foo_tests --isolated-script-test-filter-file=filter2 \
    --isolated-script-test-filter=-FooBaz.baz
[1/4] Foo.Bar.bar1 passed in 0.1s
[2/4] Foo.Bar.bar3 passed in 0.13s
[3/4] Foo.Baz.baz passed in 0.05s

3 tests passed in 0.28s, 2 skipped, 0 failures.
$

Running one shard of tests

$ GTEST_TOTAL_SHARDS=3 GTEST_SHARD_INDEX=1 out/Default/bin/run_foo_tests
Foo.Bar.bar2 passed in 0.13s
Foo.Quux.quux1 passed in 0.02s

2 tests passed in 0.15s, 0 skipped, 0 failures.
$

Related Work

This document only partially makes sense in isolation.

The JSON Test Results Format document specifies how the results of the test run should be reported.

The Chromium Test List Format specifies in more detail how we can specify which tests to run and which to skip, and whether the tests are expected to pass or fail.

Implementing everything in this document plus the preceding three documents should fully specify how tests are run in Chromium. And, if we do this, implementing tools to manage tests should be significantly easier.

On Naming Chromium Builders and Build Steps is a related proposal that has been partially implemented; it is complementary to this work, but not required.

Cleaning up the Chromium Testing Conventions describes a series of changes we might want to make to this API and the related infrastructure to simplify things.

Additional documents that may be of interest:

Document history

[ Significant changes only. ]

DateComment
2017-12-13Initial version. This tried to be a full-featured spec that defined common flags that devs might want with friendly names, as well the flags needed to run tests on the bots.
2019-05-24Second version. The spec was significantly revised to just specify the minimal subset needed to run tests consistently on bots given the current infrastructure.
2019-05-29All TODOs and discussion of future work was stripped out; now the spec only specifies how the isolated_scripts currently behave. Future work was moved to a new doc, Cleaning up the Chromium Testing Environment.
2019-09-16Add comment about ordering of filters and longest match winning for --isolated-script-test-filter.
2020-07-01Moved into the src repo and converted to Markdown. No content changes otherwise.

Notes

(*) The initial version of this document talked about test runners instead of test executables, so the bit.ly shortcut URL refers to the test-runner-api instead of the test-executable-api. The author attempted to create a test-executable-api link, but pointed it at the wrong document by accident. bit.ly URLs can't easily be updated :(.