This repository contains references to tests that are run with every commit to Flutter to verify that no breaking changes have been introduced. The tests referenced by this repository are typically maintained by people outside of the Flutter team, as part of the development of their applications. They are intended to give the Flutter team visibility into how their changes affect real-world developers using Flutter.
Adding more tests
You are welcome to add a new batch of tests. To do so, copy the file
registry/template.test to create a new file in the
directory. Fill in the fields and delete all the comments. Then,
submit a PR with this new file.
Tests must fulfill the following criteria to be added:
All the code must be available publicly on GitHub under a license compatible with this effort.
Tests must be hermetic. For example, a test should not involve network activity, spawn processes, or access the local file system except to access files that are packaged with the test.
Tests must be resilient to being run concurrently with other tests, including concurrent runs of themselves.
Tests must be reliable. A test must not claim to pass if it is failing. Running a test multiple times in a row must always have the same result.
Tests must have no output when they are passing.
Tests must be as fast as possible given the hardware. For example, tests must not use real timers or depend on the wall clock.
The time taken by tests must be proportional to their value. A few thousands tests are expected to run within a few minutes. An upper limit of about five minutes will be applied to each contributed test suite (not including the time to download the tests), but it is expected that most suites will complete in seconds.
The tests must be compatible with any tools for automatically updating Flutter code (e.g. they cannot rely on custom code generation unless such code generation can hook into the automatic update mechanism).
The tests must represent good practices as described in Flutter's documentation. For example, using an object after calling its
disposemethod violates the contract described by that method. Accessing the fields of a private
Statesubclass from another package by casting it to dynamic is similarly sketchy and would not be supported behaviour.
The tests must pass at the time they are contributed.
Dependencies must be pinned. (Generally, checking in the
pubspec.lockfile is sufficient for this purpose.)
Running the tests locally
To run these tests locally, check out this directory in a directory
parallel to your
flutter repository checkout, then, from this
pushd ../flutter/dev/customer_testing && pub get && popd ../flutter/bin/cache/dart-sdk/bin/dart ../flutter/dev/customer_testing/run_tests.dart --skip-template --verbose registry/*.test
The first command retrieves the Dart packages used by
and can be omitted for subsequent executions.
If a test is broken
The point of these tests is to make sure we don't break existing code, while still being able to make improvements to Flutter APIs.
If you find that a PR you have created in flutter/flutter causes one these tests to fail, you have the following options:
Change your PR so that the test no longer fails. This is the preferred option, so long as the result is one we can be proud of. Is the resulting API something that you would plausibly come up with without the backwards-compatibility constraint? That's good. Is the resulting API something that, as soon as you see it, you think "why?" or "that's weird"? That's bad. Consider the advice in the Style guide: https://github.com/flutter/flutter/wiki/Style-guide-for-Flutter-repo
Go through the breaking change process, as documented here: https://github.com/flutter/flutter/wiki/Tree-hygiene#handling-breaking-changes If you're going to do this, you will need to contact the relevant people responsible for the breaking test(s) (see the relevant .test files), help them fix their code, and update this repository to use the new version of their tests, in addition to the steps described on the wiki. You will also need to land your change in two parts, so that people have time to migrate (a "soft-breaking" change).
Remove the test in question. This is by far the least ideal solution. To go down this path, we must first establish that one of the following is true:
the people listed as contacts for the test are not responsive (within 72 hours).
the test is poorly written (e.g. it contains a race condition or relies on assumptions that violate clearly documented API contracts), and the people listed as contacts are not willing to fix the test or accept fixes for the test.
we have gone through the breaking change process cited above, but are unable to update the test accordingly (e.g. the people listed as contacts are not willing to work with us to update their code).