This document assumes that you are familiar with Bazel and provides advice on structuring your projects to take full advantage of Bazel's features.
The overall goals are:
These guidelines are not requirements: few projects will be able to adhere to all of them. As the man page for lint says, “A special reward will be presented to the first person to produce a real program that produces no errors with strict checking.” However, incorporating as many of these principles as possible should make a project more readable, less error-prone, and faster to build.
This document uses the requirement levels described in this RFC.
A project should always be able to run bazel build //... and bazel test //... successfully on its stable branch. Targets that are necessary but do not build under certain circumstances (e.g., require specific build flags, do not build on a certain platform, require license agreements) should be tagged as specifically as possible (e.g., “requires-osx”). This tagging allows targets to be filtered at a more fine-grained level than the “manual” tag and allows someone inspecting the BUILD file to understand what a target's restrictions are.
Prefer declaring third party dependencies as remote repositories in the WORKSPACE file. If it's necessary to check third party dependencies into your repository, put them in a directory called third_party/ under your workspace directory. Note that all BUILD files in third_party/ must include license declarations.
Everything should be built from source whenever possible. Generally this means that, instead of depending on a library some-library.so, you'd create a BUILD file and build some-library.so from its sources, then depend on that target.
Building from source prevents a build from using an library that was build with incompatible flags or a different architecture. There are also some features like coverage, static analysis, or dynamic analysis that will only work on the source.
Prefer building all code from head whenever possible. When versions must be used, avoid including the version in the target name (e.g., //guava, not //guava-20.0). This naming makes the library easier to update (only one target needs to be updated). It is also more resilient to diamond dependency issues: if one library depends on guava-19.0 and one depends on guava-20.0, you could end up with a library that tries to depend on two different versions. If you created a misleading alias to point both targets to one guava library, then the BUILD files are misleading.
.bazelrcFor project-specific options, use the configuration file _your-workspace_/tools/bazel.rc.
For options that you do not want to check into source control, create the configuration file _your-workspace_/.bazelrc and add .bazelrc to your .gitignore. Note that this file has a different name than the file above (bazel.rc vs .bazelrc).
Every directory that contains buildable files should be a package. If a BUILD file refers to files in subdirectories (e.g., srcs = ["a/b/C.java"]) it is a sign that a BUILD file should be added to that subdirectory. The longer this structure exists, the more likely circular dependencies will be inadvertently created, a target's scope will creep, and an increasing number of reverse dependencies will have to be updated.
See the BUILD file style guide.
Buildifier should be used to achieve the correct formatting for BUILD files. Editors should be configured to automatically format BUILD files on save. Humans should not try to format BUILD files themselves.
If there is a question as to what the correct formatting is, the answer is “how buildifier formats it.”
Files should be referred to by their paths relative to the package directory (without ever using up-references, such as ..). Generated files should be prefixed with “:” to indicate that they are not sources. Source files should not be prefixed with :. Rules should be prefixed with :. For example, assuming x.cc is a source file:
cc_library( name = "lib", srcs = ["x.cc"], hdrs = [":gen-header"], ) genrule( name = "gen-header", srcs = [], outs = ["x.h"], cmd = "echo 'int x();' > $@", )
Target names should be descriptive. If a target contains one source file, the target should generally be named after that source (e.g., a cc_library for chat.cc should be named “chat”).
The eponymous target for a package (the target with the same name as the containing directory) should provide the functionality described by the directory name. If there is no such target, do not create an eponymous target.
Prefer using the short name when referring to an eponymous target (//x instead of //x:x). If you are in the same package, prefer the local reference (:x instead of //x).
Do not set the default visibility of a package to //visibility:public. //visibility:public should be individually set for targets in the project‘s public API. These could be libraries which are designed to be depended on by external projects or binaries that could be used by an external project’s build process.
Otherwise, visibility should be scoped as tightly as possible, while still allowing access by tests and reverse dependencies. Prefer using __pkg__ to __subpackages__.
Dependencies should be restricted to direct dependencies (dependencies needed by the sources listed in the rule). Do not list transitive dependencies.
Package-local dependencies should be listed first and referred to in a way compatible with the References to targets in the current package section above (not by their absolute package name).
Do not use recursive globs (e.g., glob(["**/*.java"])). Recursive globs make BUILD files difficult to read, as they skip subdirectories containing BUILD files. Non-recursive globs are generally acceptable, see language-specific advice below for details.
Indicate “no targets” with []. Do not use a glob that matches nothing: it is more error-prone and less obvious than an empty list.
See the Style guide for .bzl files for Skylark rule guidelines.
See Packaging rules for advice on how to structure and where to put new Skylark rules.
When using a language for which Bazel has built-in rules (e.g., C++), prefer using these rules to writing your own in Skylark. These rules are documented in the build encyclopedia.
Prefer http_archive and new_http_archive to git_repository, new_git_repository, and maven_jar.
git_repository depends on jGit, which has several unpleasant bugs, and maven_jar uses Maven's internal API, which generally works but is less optimized for Bazel than http_archive's downloader logic. Track the following issues filed to remediate these problems:
Do not use bind(). See “Consider removing bind” for a long discussion of its issues and alternatives.
When using a new_ repository rule, prefer to specify build_file_content, not build_file.
A Skylark repository rule should generally be responsible for:
Avoid using repository_ctx.execute when possible. For example, when using a non-Bazel C++ library that has a build using Make, it is preferable to use respository_ctx.download() and then write a BUILD file that builds it, instead of running ctx.execute(["make"]).
Prefer Maven's standard directory layout (sources under src/main/java, tests under src/test/java).
Use one BUILD file per package containing Java sources. Every BUILD file should contain one java_library rule that looks like this:
java_library( name = "directory-name", srcs = glob(["*.java"]), deps = [...], )
The name of the library should be the name of the directory containing the BUILD file. The sources should be a non-recursive glob of all Java files in the directory.
Tests should be in a matching directory under src/test and depend on this library.
Each BUILD file should contain one cc_library rule target per compilation unit in the directory. C++ libraries should be as fine-grained as possible to provide as much incrementality as possible.
If there is a single source file in srcs, the library should be named based on that C++ file‘s name. This library should contain a C++ file(s), any matching header file(s), and the library’s direct dependencies. For example,
cc_library( name = "mylib", srcs = ["mylib.cc"], hdrs = ["mylib.h"], deps = [":lower-level-lib"] )
There should be one cc_test rule target per cc_library target in the file. The cc_test's source should be a file named [libname]_test.cc. For example, a test for the target above might look like:
cc_test(
name = "mylib_test",
srcs = ["mylib_test.cc"],
deps = [":mylib"]
)
All include paths should be relative to the workspace directory. Use includes only if a public header needs to be widely used at a non-workspace-relative path (for legacy or third_party code). Otherwise, prefer to use the copts attribute, not the includes attribute.
Using cc_inc_library is discouraged, prefer copts or includes. See the design document on C++ include directories for reasoning.
proto_library rule per .proto file.foo.proto will be in a rule named foo_proto, which is located in the same package.[language]_proto_library that wraps a proto_library named foo_proto should be called foo_[language]_proto, and be located in the same package.