/rules_go

Go rules for Bazel

Primary LanguageGoApache License 2.0Apache-2.0

Go rules for Bazel

Travis Jenkins
travis jenkins
December 6, 2017
Release 0.8.0 is now available.
November 27, 2017
Release 0.7.1 is now available.
November 3, 2017
Release 0.7.0 is now available.

The rules are in the alpha stage of development. They support:

They currently do not support (in order of importance):

  • bazel-style auto generating BUILD (where the library name is other than go_default_library)
  • C/C++ interoperation except cgo (swig etc.)
  • coverage
  • test sharding
Note:The latest version of these rules (0.8.0) requires Bazel ≥ 0.8.0 to work.

The master branch is only guaranteed to work with the latest version of Bazel.

  • Create a file at the top of your repository named WORKSPACE and add one of the snippets below, verbatim. This will let Bazel fetch necessary dependencies from this repository and a few others.

    If you want to use the latest stable release, add the following:

    http_archive(
        name = "io_bazel_rules_go",
        url = "https://github.com/bazelbuild/rules_go/releases/download/0.8.0/rules_go-0.8.0.tar.gz",
        sha256 = "8eaf2e62811169d9cf511209153effcb132826cea708b2f75d4dd5f9e57ea2aa",
    )
    load("@io_bazel_rules_go//go:def.bzl", "go_rules_dependencies", "go_register_toolchains")
    go_rules_dependencies()
    go_register_toolchains()

    If you want to use a specific commit (for example, something close to master), add the following instead:

    git_repository(
        name = "io_bazel_rules_go",
        remote = "https://github.com/bazelbuild/rules_go.git",
        commit = "a390e7f7eac912f6e67dc54acf67aa974d05f9c3",
    )
    load("@io_bazel_rules_go//go:def.bzl", "go_rules_dependencies", "go_register_toolchains")
    go_rules_dependencies()
    go_register_toolchains()

    If you plan to use the proto rules (go_proto_library and go_grpc_library), add the following to WORKSPACE.

    load("@io_bazel_rules_go//proto:def.bzl", "proto_register_toolchains")
    proto_register_toolchains()

    You can add more external dependencies to this file later (see go_repository).

  • Add a file named BUILD.bazel in the root directory of your project. In general, you need one of these files in every directory with Go code, but you need one in the root directory even if your project doesn't have any Go code there.

  • If your project can be built with go build, you can generate your build files using Gazelle. If your project isn't compatible with go build or if you prefer not to use Gazelle, you can write build files by hand.

If your project can be built with go build, you can generate and update your build files automatically using gazelle, a tool included in this repository.

  • Add the code below to the BUILD.bazel file in your repository's root directory. Replace the prefix string with the prefix you chose for your project earlier.

    load("@io_bazel_rules_go//go:def.bzl", "gazelle")
    
    gazelle(
        name = "gazelle",
        prefix = "github.com/example/project",
    )
  • If your project uses vendoring, add external = "vendored", below the prefix line.

  • After adding the gazelle rule, run the command below:

    bazel run //:gazelle
    

    This will generate a BUILD.bazel file for each Go package in your repository. You can run the same command in the future to update existing build files with new source files, dependencies, and options.

If your project doesn't follow go build conventions or you prefer not to use gazelle, you can write build files by hand.

  • In each directory that contains Go code, create a file named BUILD.bazel

  • Add a load statement at the top of the file for the rules you use.

    load("@io_bazel_rules_go//go:def.bzl", "go_binary", "go_library", "go_test")
  • For each library, add a go_library rule like the one below. Source files are listed in srcs. Other packages you import are listed in deps using Bazel labels that refer to other go_library rules. The library's import path should be specified with importpath.

    go_library(
        name = "go_default_library",
        srcs = [
            "foo.go",
            "bar.go",
        ],
        deps = [
            "//tools:go_default_library",
            "@org_golang_x_utils//stuff:go_default_library",
        ],
        importpath = "github.com/example/project/foo",
        visibility = ["//visibility:public"],
    )
  • For each test, add a go_test rule like either of the ones below. You'll need separate go_test rules for internal and external tests.

    # Internal test
    go_test(
        name = "go_default_test",
        srcs = ["foo_test.go"],
        importpath = "github.com/example/project/foo",
        library = ":go_default_library",
    )
    
    # External test
    go_test(
        name = "go_default_xtest",
        srcs = ["bar_test.go"],
        deps = [":go_default_library"],
        importpath = "github.com/example/project/foo",
    )
  • For each binary, add a go_binary rule like the one below.

    go_binary(
        name = "foo",
        srcs = ["main.go"],
        deps = [":go_default_library"],
        importpath = "github.com/example/project/foo",
    )
  • For instructions on how to depend on external libraries, see _vendoring

Yes, this setup was deliberately chosen to be compatible with go build. Make sure your project appears in GOPATH, and it should work.

Note that go build won't be aware of dependencies listed in WORKSPACE, so these will be downloaded into GOPATH. You may also need to check in generated files.

This was used to keep import paths consistent in libraries that can be built with go build before the importpath attribute was available.

In order to compile and link correctly, the Go rules need to be able to translate Bazel labels to Go import paths. Libraries that don't set the importpath attribute explicitly have an implicit dependency on //:go_prefix, a special rule that specifies an import path prefix. The import path is the prefix concatenated with the Bazel package and target name. For example, if your prefix was github.com/example/project, and your library was //foo/bar:bar, the Go rules would decide the import path was github.com/example/project/foo/bar/bar. The stutter at the end is incompatible with go build, so if the label name is go_default_library, the import path is just the prefix concatenated with the package name. So if your library is //foo/bar:go_default_library, the import path is github.com/example/project/foo/bar.

We are working on deprecating go_prefix and making importpath mandatory (see #721). When this work is complete, the go_default_library name won't be needed. We may decide to stop using this name in the future (see #265).

Bazel executes tests in a sandbox, which means tests don't automatically have access to files. You must include test files using the data attribute. For example, if you want to include everything in the testdata directory:

go_test(
    name = "go_default_test",
    srcs = ["foo_test.go"],
    data = glob(["testdata/**"]),
    importpath = "github.com/example/project/foo",
)

By default, tests are run in the directory of the build file that defined them. Note that this follows the Go testing convention, not the Bazel convention followed by other languages, which run in the repository root. This means that you can access test files using relative paths. You can change the test directory using the rundir attribute. See go_test.

Gazelle will automatically add a data attribute like the one above if you have a testdata directory unless it contains buildable .go files or build files, in which case, testdata is treated as a normal package.

References:

In order to run Bazel tests on Travis CI, you'll need to install Bazel in the before_install script. See our configuration file linked above.

You'll want to run Bazel with a number of flags to prevent it from consuming a huge amount of memory in the test environment.

  • --batch: Don't start the Bazel server.
  • --host_jvm_args=-Xmx500m --host_jvm_args=-Xms500m: Set the maximum and initial JVM heap size. Keeping the same means the JVM won't spend time growing the heap. The choice of heap size is somewhat arbitrary; other configuration files recommend limits as high as 2500m. Higher values mean a faster build, but higher risk of OOM kill.
  • --bazelrc=.test-bazelrc: Use a Bazel configuration file specific to Travis CI. You can put most of the remaining options in here.
  • build --spawn_strategy=standalone --genrule_strategy=standalone: Disable sandboxing for the build. Sandboxing may fail inside of Travis's containers because the mount system call is not permitted.
  • test --test_strategy=standalone: Disable sandboxing for tests as well.
  • --local_resources=1536,1.5,0.5: Set Bazel limits on available RAM in MB, available cores for compute, and available cores for I/O. Higher values mean a faster build, but higher contention and risk of OOM kill.
  • --noshow_progress: Suppress progress messages in output for cleaner logs.
  • --verbose_failures: Get more detailed failure messages.
  • --test_output=errors: Show test stderr in the Travis log. Normally, test output is written log files which Travis does not save or report.

Downloads on Travis are relatively slow (the network is heavily contended), so you'll want to minimize the amount of network I/O in your build. Downloading Bazel and a Go SDK is a huge part of that. To avoid downloading a Go SDK, you may request a container with a preinstalled version of Go in your .travis.yml file, then call go_rules_dependencies(go_version = "host") in a Travis-specific WORKSPACE file.

You may be tempted to put Bazel's cache in your Travis cache. Although this can speed up your build significantly, Travis stores its cache on Amazon, and it takes a very long time to transfer. Clean builds seem faster in practice.