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
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
from its sources, then depend on that target.
Always building from source ensures that a build is not using a library that was built 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-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
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.
For project-specific options, use the configuration file
If you want to support per-user options for your project that you do not want to check
into source control, include the line
(or any other file name) in your
user.bazelrc to your
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.
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
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
When using a
new_ repository rule, prefer to specify
A repository rule should generally be responsible for:
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
repository_ctx.download() and then
write a BUILD file that builds it, instead of running
Recommended code organization:
foo.protowill be in a rule named
foo_proto, which is located in the same package.
[language]_proto_librarythat wraps a
foo_protoshould be called
foo_[language]_proto, and be located in the same package.