All of this equipment is here to help you get your work done more efficiently. However, learning how task-graphs are generated is probably not the work you are interested in doing. This section should help you accomplish some of the more common changes to the task graph with minimal fuss.
If you cannot accomplish what you need with the information provided here, please consider whether you can achieve your goal in a different way. Perhaps something simpler would cost a bit more in compute time, but save the much more expensive resource of developers' mental bandwidth. Task-graph generation is already complex enough!
If you want to proceed, you may need to delve into the implementation of task-graph generation. The documentation and code are designed to help, as are the authors -
hg blame may help track down helpful people.
As you write your new transform or add a new kind, please consider the next developer. Where possible, make your change data-driven and general, so that others can make a much smaller change. Document the semantics of what you are changing clearly, especially if it involves modifying a transform schema. And if you are adding complexity temporarily while making a gradual transition, please open a new bug to remind yourself to remove the complexity when the transition is complete.
The recommended process for changing task graphs is this:
Run one of the
mach taskgraph subcommands (see
mach) to generate a baseline against which to measure your changes.
./mach taskgraph tasks --json > old-tasks.json
Make your modifications under
Run the same
mach taskgraph command, sending the output to a new file, and use
diff to compare the old and new files. Make sure your changes have the desired effect and no undesirable side-effects. A plain unified diff should be useful for most changes, but in some cases it may be helpful to post-process the JSON to strip distracting changes.
When you are satisfied with the changes, push them to try to ensure that the modified tasks work as expected.
If you are working on an action task and wish to test it out locally, use the
./mach taskgraph test-action-callback command:
./mach taskgraph test-action-callback \ --task-id I4gu9KDmSZWu3KHx6ba6tw --task-group-id sMO4ybV9Qb2tmcI1sDHClQ \ --input input.yml hello_world_action
This invocation will run the hello world callback with the given inputs and print any created tasks to stdout, rather than actually creating them.
First, find the test description. This will be in
taskcluster/ci/*/tests.yml, for the appropriate kind (consult
kinds). You will find a YAML stanza for each test suite, and each stanza defines the test's characteristics. For example, the
chunks property gives the number of chunks to run. This can be specified as a simple integer if all platforms have the same chunk count, or it can be keyed by test platform. For example:
The full set of available properties is in
taskcluster/taskgraph/transforms/tests.py. Some other commonly-modified properties are
max-run-time (useful if tests are being killed for exceeding maxRunTime) and
Android tests are also chunked at the mozharness level, so you will need to modify the relevant mozharness config, as well.
To add a new test suite, you will need to know the proper mozharness invocation for that suite, and which kind it fits into (consult
Add a new stanza to
taskcluster/ci/<kind>/tests.yml, copying from the other stanzas in that file. The meanings should be clear, but authoritative documentation is in
taskcluster/taskgraph/transforms/tests.py should you need it. The stanza name is the name by which the test will be referenced in try syntax.
Add your new test to a test set in
test-sets.yml in the same directory. If the test should only run on a limited set of platforms, you may need to define a new test set and reference that from the appropriate platforms in
test-platforms.yml. If you do so, include some helpful comments in
test-sets.yml for the next person.
When a test is not yet reliably green, configuration for that test should not be landed on integration branches. Of course, you can control where the configuration is landed! For many cases, it is easiest to green up a test in try: push the configuration to run the test to try along with your work to fix the remaining test failures.
When working with a group, check out a "twig" repository to share among your group, and land the test configuration in that repository. Once the test is green, merge to an integration branch and the test will begin running there as well.
If you are adding a new task that is not a test suite, there are a number of options. A few questions to consider:
- Is this a new build platform or variant that will produce an artifact to be run through the usual test suites?
- Does this task depend on other tasks? Do other tasks depend on it?
- Is this one of a few related tasks, or will you need to generate a large set of tasks using some programmatic means (for example, chunking)?
- How is the task actually executed? Mozharness? Mach?
- What kind of environment does the task require?
Armed with that information, you can choose among a few options for implementing this new task. Try to choose the simplest solution that will satisfy your near-term needs. Since this is all implemented in-tree, it is not difficult to refactor later when you need more generality.
The simplest option is to add your task to an existing kind. This is most practical when the task "makes sense" as part of that kind -- for example, if your task is building an installer for a new platform using mozharness scripts similar to the existing build tasks, it makes most sense to add your task to the
build kind. If you need some additional functionality in the kind, it's OK to modify the implementation as necessary, as long as the modification is complete and useful to the next developer to come along.
Tasks in the
build kind generate Firefox installers, and the
test kind will add a full set of Firefox tests for each
The next option to consider is adding a new kind. A distinct kind gives you some isolation from other task types, which can be nice if you are adding an experimental kind of task.
Kinds can range in complexity. The simplest sort of kind uses the transform loader to read a list of jobs from the
jobs key, and applies the standard
Job descriptions are defined and documented in
If your task depends on other tasks, then the decision of which tasks to create may require some code. For example, the
test kind iterates over the builds in the graph, generating a full set of test tasks for each one. This specific post-build behavior is implemented as a loader defined in
A custom loader is useful when the set of tasks you want to create is not static but based on something else (such as the available builds) or when the dependency relationships for your tasks are complex.
Most loaders apply a series of "
transforms <transforms>" that start with an initial human-friendly description of a task and end with a task definition suitable for insertion into a Taskcluster queue.
Custom transforms can be useful to apply defaults, simplifying the YAML files in your kind. They can also apply business logic that is more easily expressed in code than in YAML.
Transforms need not be one-to-one: a transform can produce zero or more outputs for each input. For example, the test transforms perform chunking by producing an output for each chunk of a given input.
Ideally those transforms will produce job descriptions, so you can use the existing
Try to keep transforms simple, single-purpose and well-documented!
If the way your task is executed is unique (so, not a mach command or mozharness invocation), you can add a new implementation of the job description's "run" section. Before you do this, consider that it might be a better investment to modify your task to support invocation via mozharness or mach, instead. If this is not possible, then adding a new file in
taskcluster/taskgraph/transforms/jobs with a structure similar to its peers will make the new run-using option available for job descriptions.
If you make another change not described here that turns out to be simple or common, please include an update to this file in your patch.