Execution groups

Execution groups allow for multiple execution platforms within a single target. Each execution group has its own toolchain dependencies and performs its own toolchain resolution.

Current status

To use Starlark rules that create execution groups, set --experimental_exec_groups=true on your command line or in a .bazelrc:

$ bazel build //target/with/exec/group --experimental_exec_groups

Background

Execution groups allow the rule author to define sets of actions, each with a potentially different execution platform. Multiple execution platforms can allow actions to execution differently, for example compiling an iOS app on a remote (linux) worker and then linking/code signing on a local mac worker.

Being able to define groups of actions also helps alleviate the usage of action mnemonics as a proxy for specifying actions. Mnemonics are not guaranteed to be unique and can only reference a single action. This is especially helpful in allocating extra resources to specific memory and processing intensive actions like linking in c++ builds without over-allocating to less demanding tasks.

Defining execution groups

During rule definition, rule authors can declare a set of execution groups. On each execution group, the rule author can specify everything needed to select an execution platform for that execution group, namely any constraints via exec_compatible_with and toolchain types via toolchain.

# foo.bzl
my_rule = rule(
    _impl,
    exec_groups = {
        link: exec_group(
            exec_compatible_with = [ "@platforms//os:linux" ]
            toolchains = ["//foo:toolchain_type"],
        ),
        test: exec_group(
            toolchains = ["//foo_tools:toolchain_type"],
        ),
    },
    attrs = {
        "_compiler": attr.label(cfg = config.exec("link"))
    },
)

In the code snippet above, you can see that tool dependencies can also specify transition for an exec group using the cfg attribute param and the config module. The module exposes an exec function which takes a single string parameter which is the name of the exec group for which the dependency should be built.

Accessing execution groups

In the rule implementation, you can declare that actions should be run on the execution platform of an execution group. You can do this by using the exec_group param of action generating methods, specifically [ctx.actions.run] (https://docs.bazel.build/versions/master/skylark/lib/actions.html#run) and ctx.actions.run_shell.

# foo.bzl
def _impl(ctx):
  ctx.actions.run(
     inputs = [ctx.attr._some_tool, ctx.srcs[0]]
     exec_group = "compile",
     # ...
  )

Rule authors will also be able to access the [resolved toolchains] (toolchains.html#toolchain-resolution) of execution groups, similarly to how you can access the resolved toolchain of a target:

# foo.bzl
def _impl(ctx):
  foo_info = ctx.exec_groups["link"].toolchains["//foo:toolchain_type"].fooinfo
  ctx.actions.run(
     inputs = [foo_info, ctx.srcs[0]]
     exec_group = "link",
     # ...
  )

Note: If an action uses a toolchain from an execution group, but doesn’t specify that execution group in the action declaration, that may potentially cause issues. A mismatch like this may not immediately cause failures, but is a latent problem.

Using execution groups to set execution properties

Execution groups are integrated with the exec_properties attribute that exists on every rule and allows the target writer to specify a string dict of properties that is then passed to the execution machinery. For example, if you wanted to set some property, say memory, for the target and give certain actions a higher memory allocation, you would write an exec_properties entry with an execution-group-augmented key, e.g.:

# BUILD
my_rule(
    name = 'my_target',
    exec_properties = {
        'mem': '12g',
        'link.mem': '16g'
    }
    
)

All actions with exec_group = "link" would see the exec properties dictionary as {"mem": "16g"}. As you see here, execution-group-level settings override target-level settings.

Execution groups for native rules

The following execution groups are available for actions defined by native rules:

  • test: Test runner actions.
  • cpp_link: C++ linking actions.

Creating exec groups to set exec properties

Sometimes you want to use an exec group to give specific actions different exec properties but don’t actually want different toolchains or constraints than the rule. For these situations, you can create exec groups using the copy_from_rule parameter:

# foo.bzl

# Creating an exec group with `copy_from_rule=True` is the same as explicitly
# setting the exec group's toolchains and constraints to the same values as the
# rule's respective parameters.
my_rule = rule(
    _impl,
    exec_compatible_with = [ "@platforms//os:linux" ],
    toolchains = ["//foo:toolchain_type"],
    exec_groups = {
        # The following two groups have the same toolchains and constraints:
        foo: exec_group(copy_from_rule = True),
        "bar": exec_group(
            exec_compatible_with = [ "@platforms//os:linux" ],
            toolchains = ["//foo:toolchain_type"],
        ),
    },
)

#

Execution groups and platform execution properties

It is possible to define exec_properties for arbitrary execution groups on platform targets (unlike exec_properties set directly on a target, where properties for unknown execution groups are rejected). Targets then inherit the execution platform’s exec_properties that affect the default execution group and any other relevant execution groups.

For example, suppose running a C++ test requires some resource to be available, but it isn’t required for compiling and linking; this can be modelled as follows:

constraint_setting(name = "resource")
constraint_value(name = "has_resource", constraint_setting = ":resource")

platform(
    name = "platform_with_resource",
    constraint_values = [":has_resource"],
    exec_properties = {
        "test.resource": "...",
    },
)

cc_test(
    name = "my_test",
    srcs = ["my_test.cc"],
    exec_compatible_with = [":has_resource"],
)

exec_properties defined directly on targets take precedence over those that are inherited from the execution platform.