This page covers basic style guidelines for Starlark and also includes information on macros and rules.
Starlark is a language that defines how software is built, and as such it is both a programming and a configuration language.
You will use Starlark to write
BUILD files, macros, and build rules. Macros and
rules are essentially meta-languages - they define how
BUILD files are written.
BUILD files are intended to be simple and repetitive.
All software is read more often than it is written. This is especially true for
Starlark, as engineers read
BUILD files to understand dependencies of their
targets and details of their builds. This reading will often happen in passing,
in a hurry, or in parallel to accomplishing some other task. Consequently,
simplicity and readability are very important so that users can parse and
BUILD files quickly.
When a user opens a
BUILD file, they quickly want to know the list of targets in
the file; or review the list of sources of that C++ library; or remove a
dependency from that Java binary. Each time you add a layer of abstraction, you
make it harder for a user to do these tasks.
BUILD files are also analyzed and updated by many different tools. Tools may not
be able to edit your
BUILD file if it uses abstractions. Keeping your
files simple will allow you to get better tooling. As a code base grows, it
becomes more and more frequent to do changes across many
BUILD files in order to
update a library or do a cleanup.
When in doubt, follow the PEP 8 style guide where possible. In particular, use four rather than two spaces for indentation to follow the Python convention.
Starlark is not Python,
some aspects of Python style do not apply. For example, PEP 8 advises that
comparisons to singletons be done with
is, which is not an operator in
Document files and functions using docstrings.
Use a docstring at the top of each
.bzl file, and a docstring for each public
Document rules and aspects
Rules and aspects, along with their attributes, as well as providers and their
fields, should be documented using the
- Variables and function names use lowercase with words separated by
[a-z][a-z0-9_]*), such as
- Top-level private values start with one underscore. Bazel enforces that private values cannot be used from other files. Local variables should not use the underscore prefix.
BUILD files, there is no strict line length limit as labels can be long.
When possible, try to use at most 79 characters per line (following Python's
style guide, PEP 8). This guideline
should not be enforced strictly: editors should display more than 80 columns,
automated changes will frequently introduce longer lines, and humans shouldn't
spend time splitting lines that are already readable.
In keyword arguments, spaces around the equal sign are preferred:
def fct(name, srcs): filtered_srcs = my_filter(source = srcs) native.cc_library( name = name, srcs = filtered_srcs, testonly = True, )
False (rather than of
0) for boolean values
(such as when using a boolean attribute in a rule).
Use print only for debugging
Do not use the
print() function in production code; it is only intended for
debugging, and will spam all direct and indirect users of your
.bzl file. The
only exception is that you may submit code that uses
print() if it is disabled
by default and can only be enabled by editing the source -- for example, if all
print() are guarded by
if DEBUG: where
DEBUG is hardcoded to
False. Be mindful of whether these statements are useful enough to justify
their impact on readability.
A macro is a function which instantiates one or more rules during the loading phase. In general, use rules whenever possible instead of macros. The build graph seen by the user is not the same as the one used by Bazel during the build - macros are expanded before Bazel does any build graph analysis.
Because of this, when something goes wrong, the user will need to understand
your macro's implementation to troubleshoot build problems. Additionally,
query results can be hard to interpret because targets shown in the results
come from macro expansion. Finally, aspects are not aware of macros, so tooling
depending on aspects (IDEs and others) might fail.
A safe use for macros is for defining additional targets intended to be referenced directly at the Bazel CLI or in BUILD files: In that case, only the end users of those targets need to know about them, and any build problems introduced by macros are never far from their usage.
For macros that define generated targets (implementation details of the macro which are not supposed to be referred to at the CLI or depended on by targets not instantiated by that macro), follow these best practices:
- A macro should take a
nameargument and define a target with that name. That target becomes that macro's main target.
- Generated targets, that is all other targets defined by a macro, should:
- Have their names prefixed by
_<name>. For example, using
name = '%s_bar' % (name).
- Have restricted visibility (
- Have a
manualtag to avoid expansion in wildcard targets (
- Have their names prefixed by
nameshould only be used to derive names of targets defined by the macro, and not for anything else. For example, don't use the name to derive a dependency or input file that is not generated by the macro itself.
- All the targets created in the macro should be coupled in some way to the main target.
- Keep the parameter names in the macro consistent. If a parameter is passed
as an attribute value to the main target, keep its name the same. If a macro
parameter serves the same purpose as a common rule attribute, such as
deps, name as you would the attribute (see below).
- When calling a macro, use only keyword arguments. This is consistent with rules, and greatly improves readability.
Engineers often write macros when the Starlark API of relevant rules is insufficient for their specific use case, regardless of whether the rule is defined within Bazel in native code, or in Starlark. If you're facing this problem, ask the rule author if they can extend the API to accomplish your goals.
As a rule of thumb, the more macros resemble the rules, the better.
See also macros.
- Rules, aspects, and their attributes should use lower_case names ("snake case").
- Rule names are nouns that describe the main kind of artifact produced by the
rule, from the point of view of its dependencies (or for leaf rules, the
user). This is not necessarily a file suffix. For instance, a rule that
produces C++ artifacts meant to be used as Python extensions might be called
py_extension. For most languages, typical rules include:
*_library- a compilation unit or "module".
*_binary- a target producing an executable or a deployment unit.
*_test- a test target. This can include multiple tests. Expect all tests in a
*_testtarget to be variations on the same theme, for example, testing a single library.
*_import: a target encapsulating a pre-compiled artifact, such as a
.jar, or a
.dllthat is used during compilation.
- Use consistent names and types for attributes. Some generally applicable
label_list, allowing files: source files, typically human-authored.
label_list, typically not allowing files: compilation dependencies.
label_list, allowing files: data files, such as test data etc.
label_list: runtime dependencies that are not needed for compilation.
- For any attributes with non-obvious behavior (for example, string templates
with special substitutions, or tools that are invoked with specific
requirements), provide documentation using the
dockeyword argument to the attribute's declaration (
- Rule implementation functions should almost always be private functions
(named with a leading underscore). A common style is to give the
implementation function for
- Pass information between your rules using a well-defined provider interface. Declare and document provider fields.
- Design your rule with extensibility in mind. Consider that other rules might want to interact with your rule, access your providers, and reuse the actions you create.
- Follow performance guidelines in your rules.