Configurations

Starlark configuration is Bazel’s API for customizing how your project builds.

This makes it possible to:

  • define custom flags for your project, obsoleting the need for --define
  • write transitions to configure deps in different configurations than their parents (e.g. --compilation_mode=opt or --cpu=arm)

  • bake better defaults into rules (e.g. automatically build //my:android_app with a specified SDK)

and more, all completely from .bzl files (no Bazel release required).

Current status

Ongoing bug/feature work can be found in the Bazel configurability roadmap.

This feature may have memory and performance impacts and we are still working on ways to measure and mitigate those impacts.

User-defined build settings

A build setting is a single piece of configuration information. Think of a configuration as a key/value map. Setting --cpu=ppc and --copt="-DFoo" produces a configuration that looks like {cpu: ppc, copt: "-DFoo"}. Each entry is a build setting.

Traditional flags like cpu and copt are native settings i.e. their keys are defined and their values are set inside native bazel java code. Bazel users can only read and write them via the command line and other APIs maintained natively. Changing native flags, and the APIs that expose them, requires a bazel release. User-defined build settings are defined in .bzl files (and thus, don’t need a bazel release to register changes). They also can be set via the command line (if they’re designated as flags, see more below), but can also be set via user-defined transitions.

Defining build settings

The build_setting rule() parameter

Build settings are rules like any other rule and are differentiated using the Starlark rule() function’s build_setting attribute.

# example/buildsettings/build_settings.bzl
string_flag = rule(
    implementation = _impl,
    build_setting = config.string(flag = True)
)

The build_setting attribute takes a function that designates the type of the build setting. The type is limited to a set of basic Starlark types like bool and string. See the config module documentation for details. More complicated typing can be done in the rule’s implementation function. More on this below.

The config function also takes an optional boolean parameter, flag, which is set to false by default. if flag is set to true, the build setting can be set on the command line by users as well as internally by rule writers via default values and transitions. Not all settings should be settable by users. For example if you as a rule writer have some debug mode that you’d like to turn on inside test rules, you don’t want to give users the ability to indiscriminately turn on that feature inside other non-test rules.

Using ctx.build_setting_value

Like all rules, build setting rules have implementation functions. The basic Starlark-type value of the build settings can be accessed via the ctx.build_setting_value method. This method is only available to ctx objects of build setting rules. These implementation methods can directly forward the build settings value or do additional work on it, like type checking or more complex struct creation. Here’s how you would implement an enum-typed build setting:

# example/buildsettings/build_settings.bzl
TemperatureProvider = provider(fields = ['type'])

temperatures = ["HOT", "LUKEWARM", "ICED"]

def _impl(ctx):
    raw_temperature = ctx.build_setting_value
    if raw_temperature not in temperatures:
        fail(str(ctx.label) + " build setting allowed to take values {"
             + ", ".join(temperatures) + "} but was set to unallowed value "
             + raw_temperature)
    return TemperatureProvider(type = raw_temperature)

temperature = rule(
    implementation = _impl,
    build_setting = config.string(flag = True)
)

Note: if a rule depends on a build setting, it will receive whatever providers the build setting implementation function returns, like any other dependency. But all other references to the value of the build setting (e.g. in transitions) will see its basic Starlark-typed value, not this post implementation function value.

Instantiating Build Settings

Rules defined with the build_setting parameter have an implicit mandatory build_setting_default attribute. This attribute takes on the same type as declared by the build_setting param.

# example/buildsettings/build_settings.bzl
FlavorProvider = provider(fields = ['type'])

def _impl(ctx):
    return FlavorProvider(type = ctx.build_setting_value)

flavor = rule(
    implementation = _impl,
    build_setting = config.string(flag = True)
)
# example/buildsettings/BUILD
load("//example/buildsettings:build_settings.bzl", "flavor")
flavor(
    name = "favorite_flavor",
    build_setting_default = "APPLE"
)

Predefined settings

The Skylib library includes a set of predefined settings you can instantiate without having to write custom Starlark.

For example, to define a setting that accepts a limited set of string values:

# example/BUILD
load("@bazel_skylib//rules:common_settings.bzl", "string_flag")
string_flag(
    name = "myflag",
    values = ["a", "b", "c"],
    build_setting_default = "a",
)

For a complete list, see Common build setting rules.

Using build settings

Depending on build settings

If a target would like to read a piece of configuration information, it can directly depend on the build setting via a regular attribute dependency.

# example/rules.bzl
load("//example/buildsettings:build_settings.bzl", "FlavorProvider")
def _rule_impl(ctx):
    if ctx.attr.flavor[FlavorProvider].type == "ORANGE":
        ...

drink_rule = rule(
    implementation = _rule_impl,
    attrs = {
        "flavor": attr.label()
    }
)
# example/BUILD
load("//example:rules.bzl", "drink_rule")
load("//example/buildsettings:build_settings.bzl", "flavor")
flavor(
    name = "favorite_flavor",
    build_setting_default = "APPLE"
)
drink_rule(
    name = "my_drink",
    flavor = ":favorite_flavor",
)

Languages may wish to create a canonical set of build settings which all rules for that language depend on. Though the native concept of fragments no longer exists as a hardcoded object in Starlark configuration world, one way to translate this concept would be to use sets of common implicit attributes. For example:

# kotlin/rules.bzl
_KOTLIN_CONFIG = {
    "_compiler": attr.label(default = "//kotlin/config:compiler-flag"),
    "_mode": attr.label(default = "//kotlin/config:mode-flag"),
    ...
}

...

kotlin_library = rule(
    implementation = _rule_impl,
    attrs = dicts.add({
        "library-attr": attr.string()
    }, _KOTLIN_CONFIG)
)

kotlin_binary = rule(
    implementation = _binary_impl,
    attrs = dicts.add({
        "binary-attr": attr.label()
    }, _KOTLIN_CONFIG)

Settings Build Settings on the command line

Build settings are set on the command line like any other flag. Boolean build settings understand no-prefixes and both equals and space syntaxes are supported. The name of build settings is their full target path:

$ bazel build //my/target --//example:favorite_flavor="PAMPLEMOUSSE"

There are plans to implement shorthand mapping of flag labels so users don’t need to use their entire target path each time i.e.:

$ bazel build //my/target --cpu=k8 --noboolean_flag

instead of

$ bazel build //my/target --//third_party/bazel/src/main:cpu=k8 --no//my/project:boolean_flag

Label-typed build settings

Unlike other build settings, label-typed settings cannot be defined using the build_setting rule parameter. Instead, bazel has two built-in rules: label_flag and label_setting. These rules forward the providers of the actual target to which the build setting is set. label_flag and label_setting can be read/written by transitions and label_flag can be set by the user like other build_setting rules can. Their only difference is they can’t customely defined.

Label-typed settings will eventually replace the functionality of late-bound defaults. Late-bound default attributes are Label-typed attributes whose final values can be affected by configuration. In Starlark, this will replace the configuration_field API.

# example/rules.bzl
MyProvider = provider(fields = ["my_field"])

def _dep_impl(ctx):
    return MyProvider(my_field = "yeehaw")

dep_rule = rule(
    implementation = _dep_impl
)

def _parent_impl(ctx):
    if ctx.attr.my_field_provider[MyProvider].my_field == "cowabunga":
        ...

parent_rule = rule(
    implementation = _parent_impl,
    attrs = { "my_field_provider": attr.label() }
)

# example/BUILD
load("//example:rules.bzl", "dep_rule", "parent_rule")

dep_rule(name = "dep")

parent_rule(name = "parent", my_field_provider = ":my_field_provider")

label_flag(
    name = "my_field_provider",
    build_setting_default = ":dep"
)

TODO(bazel-team): Expand supported build setting types.

Build settings and select()

Users can configure attributes on build settings by using select(). Build setting targets can be passed to the flag_values attribute of config_setting. The value to match to the configuration is passed as a String then parsed to the type of the build setting for matching.

config_setting(
    name = "my_config",
    flag_values = {
        "//example:favorite_flavor": "MANGO"
    }
)

User-defined transitions

A configuration transition is how we change configuration of configured targets in the build graph.

IMPORTANT: In order to use Starlark transitions, you need to attach a special attribute to the rule to which the transition is attached:

"_whitelist_function_transition": attr.label(
     default = "@bazel_tools//tools/whitelists/function_transition_whitelist"
 )

By adding transitions you can pretty easily explode the size of your build graph. This sets a whitelist on the packages in which you can create targets of this rule. The default value in the codeblock above whitelists everything. But if you’d like to restrict who is using your rule, you can set that attribute to point to your own custom whitelist. Contact bazel-discuss@googlegroups.com if you’d like advice or assistance understanding how transitions can affect on your build performance.

Defining

Transitions define configuration changes between rules. For example, a request like “compile my dependency for a different CPU than its parent” is handled by a transition.

Formally, a transition is a function from an input configuration to one or more output configurations. Most transitions are 1:1 e.g. “override the input configuration with --cpu=ppc”. 1:2+ transitions can also exist but come with special restrictions.

In Starlark, transitions are defined much like rules, with a defining transition() function and an implementation function.

# example/transitions/transitions.bzl
def _impl(settings, attr):
    _ignore = (settings, attr)
    return {"//example:favorite_flavor" : "MINT"}

hot_chocolate_transition = transition(
    implementation = _impl,
    inputs = [],
    outputs = ["//example:favorite_flavor"]
)

The transition() function takes in an implementation function, a set of build settings to read(inputs), and a set of build settings to write (outputs). The implementation function has two parameters, settings and attr. settings is a dictionary {String:Object} of all settings declared in the inputs parameter to transition().

attr is a dictionary of attributes and values of the rule to which the transition is attached. When attached as an outgoing edge transition, the values of these attributes are all configured i.e. post-select() resolution. When attached as an incoming edge transition, attr does not include any attributes that use a selector to resolve their value. If an incoming edge transition on --foo reads attribute bar and then also selects on --foo to set attribute bar, then there’s a chance for the incoming edge transition to read the wrong value of bar in the transition.

Note: since transitions are attached to rule definitions and select()s are attached to rule instantiations (i.e. targets), errors related to select()s on read attributes will pop up when users create targets rather than when rules are written. It may be worth taking extra care to communicate to rule users which attributes they should be wary of selecting on or taking other precautions.

The implementation function must return a dictionary (or list of dictionaries, in the case of transitions with multiple output configurations) of new build settings values to apply. The returned dictionary keyset(s) must contain exactly the set of build settings passed to the outputs parameter of the transition function. This is true even if a build setting is not actually changed over the course of the transition - its original value must be explicitly passed through in the returned dictionary.

Defining 1:2+ transitions

Outgoing edge transition can map a single input configuration to two or more output configurations. These are defined in Starlark by returning a list of dictionaries in the transition implementation function.

# example/transitions/transitions.bzl
def _impl(settings, attr):
    _ignore = (settings, attr)
    return [
        {"//example:favorite_flavor" : "LATTE"},
        {"//example:favorite_flavor" : "MOCHA"},
    ]

coffee_transition = transition(
    implementation = _impl,
    inputs = [],
    outputs = ["//example:favorite_flavor"]
)

Attaching transitions

Transitions can be attached in two places: incoming edges and outgoing edges. Effectively this means rules can transition their own configuration (incoming edge transition) and transition their dependencies’ configurations (outgoing edge transition).

NOTE: There is currently no way to attach Starlark transitions to native rules. If you need to do this, contact bazel-discuss@googlegroups.com and we can help you try to figure out a workaround.

Incoming edge transitions

Incoming edge transitions are activated by attaching a transition object (created by transition()) to rule()’s cfg parameter:

# example/rules.bzl
load("example/transitions:transitions.bzl", "hot_chocolate_transition")
drink_rule = rule(
    implementation = _impl,
    cfg = hot_chocolate_transition,
    ...

Incoming edge transitions must be 1:1 transitions.

Outgoing edge transitions

Outgoing edge transitions are activated by attaching a transition object (created by transition()) to an attribute’s cfg parameter:

# example/rules.bzl
load("example/transitions:transitions.bzl", "coffee_transition")
drink_rule = rule(
    implementation = _impl,
    attrs = { "dep": attr.label(cfg = coffee_transition)}
    ...

Outgoing edge transitions can be 1:1 or 1:2+.

Transitions on native options

WARNING: Long term, we plan to reimplement all native options as build settings. When that happens, this syntax will be deprecated. Currently other issues are blocking that migration but be aware you may have to migrate your transitions at some point in the future.

Starlark transitions can also declare reads and writes on native options via a special prefix to the option name.

# example/transitions/transitions.bzl
def _impl(settings, attr):
    _ignore = (settings, attr)
    return {"//command_line_option:cpu": "k8"}

cpu_transition = transition(
    implementation = _impl,
    inputs = [],
    outputs = ["//command_line_option:cpu"]

NOTE: Transitioning on –define using “//command_line_option:define” is not supported - create a custom build setting to cover this functionality.

Accessing attributes with transitions

When attaching a transition to an outgoing edge (regardless of whether the transition is a 1:1 or 1:2+ transition) access to values of that attribute in the rule implementation changes. Access through ctx.attr is forced to be a list if it isn’t already. The order of elements in this list is unspecified.

# example/transitions/rules.bzl
def _transition_impl(settings, attr):
    return {"//example:favorite_flavor" : "LATTE"},

coffee_transition = transition(
    implementation = _transition_impl,
    inputs = [],
    outputs = ["//example:favorite_flavor"]
)

def _rule_impl(ctx):
    # Note: List access even though "dep" is not declared as list
    transitioned_dep = ctx.attr.dep[0]

    # Note: Access doesn't change, other_deps was already a list
    for other dep in ctx.attr.other_deps:
      # ...


coffee_rule = rule(
    implementation = _rule_impl,
    attrs = {
        "dep": attr.label(cfg = coffee_transition)
        "other_deps": attr.label_list(cfg = coffee_transition)
    })

Access to the value of a single branch of a 1:2+ has not been implemented yet.

Integration with platforms and toolchains

Many native flags today, like --cpu and --crosstool_top are related to toolchain resolution. In the future, explicit transitions on these types of flags will likely be replaced by transitioning on the target platform

Also see

Memory and performance considerations

Adding transitions, and therefore new configurations, to your build comes at a cost: larger build graphs, less comprehensible build graphs, and slower builds. It’s worth considering these costs when considering using transitions in your build rules. Below is an example of how a transition might create exponential growth of your build graph.

Badly behaved builds: a case study

Say you have the following target structure:

a graph showing a top level target, //pkg:app, which depends on two targets, //pkg:1_0 and //pkg:1_1. Both these targets depend on two targets, //pkg:2_0 and //pkg:2_1. Both these targets depend on two targets, //pkg:3_0 and //pkg:3_1. This continues on until //pkg:n_0 and //pkg:n_1, which both depend on a single target, //pkg:dep.

Building //pkg:app requires \(2n+2\) targets:

  • //pkg:app
  • //pkg:dep
  • //pkg:i_0 and //pkg:i_1 for \(i\) in \([1..n]\)

Imagine you implement) a flag --//foo:owner=<STRING> and //pkg:i_b applies

depConfig = myConfig + depConfig.owner="$(myConfig.owner)$(b)"

In other words, //pkg:i_b appends b to the old value of --owner for all its deps.

This produces the following configured targets:

//pkg:app                              //foo:owner=""
//pkg:1_0                              //foo:owner=""
//pkg:1_1                              //foo:owner=""
//pkg:2_0 (via //pkg:1_0)              //foo:owner="0"
//pkg:2_0 (via //pkg:1_1)              //foo:owner="1"
//pkg:2_1 (via //pkg:1_0)              //foo:owner="0"
//pkg:2_1 (via //pkg:1_1)              //foo:owner="1"
//pkg:3_0 (via //pkg:1_0 → //pkg:2_0)  //foo:owner="00"
//pkg:3_0 (via //pkg:1_0 → //pkg:2_1)  //foo:owner="01"
//pkg:3_0 (via //pkg:1_1 → //pkg:2_0)  //foo:owner="10"
//pkg:3_0 (via //pkg:1_1 → //pkg:2_1)  //foo:owner="11"
...

//pkg:dep produces \(2^n\) configured targets: config.owner= “\(b_0b_1…b_n\)” for all \(b_i\) in \({0,1}\).

This makes the build graph exponentially larger than the target graph, with corresponding memory and performance consequences.

TODO: Add strategies for measurement and mitigation of these issues.