This page is intended for Bazel users writing custom build and test rules who want to understand the requirements for Bazel rules in the context of remote execution.
Remote execution allows Bazel to execute actions on a separate platform, such as a datacenter. Bazel uses a gRPC protocol for its remote execution. You can try remote execution with bazel-buildfarm, an open-source project that aims to provide a distributed remote execution platform.
This page uses the following terminology when referring to different environment types or platforms:
- Host platform - where Bazel runs.
- Execution platform - where Bazel actions run.
- Target platform - where the build outputs (and some actions) run.
When configuring a Bazel build for remote execution, you must follow the guidelines described in this page to ensure the build executes remotely error-free. This is due to the nature of remote execution, namely:
Isolated build actions. Build tools do not retain state and dependencies cannot leak between them.
Diverse execution environments. Local build configuration is not always suitable for remote execution environments.
This page describes the issues that can arise when implementing custom Bazel build and test rules for remote execution and how to avoid them. It covers the following topics:
- Invoking build tools through toolchain rules
- Managing implicit dependencies
- Managing platform-dependent binaries
- Managing configure-style WORKSPACE rules
Invoking build tools through toolchain rules
A Bazel toolchain rule is a configuration provider that tells a build rule what
build tools, such as compilers and linkers, to use and how to configure them
using parameters defined by the rule's creator. A toolchain rule allows build
and test rules to invoke build tools in a predictable, preconfigured manner
that's compatible with remote execution. For example, use a toolchain rule
instead of invoking build tools via the
JAVA_HOME, or other local
variables that may not be set to equivalent values (or at all) in the remote
Toolchain rules currently exist for Bazel build and test rules for Scala, Rust, and Go, and new toolchain rules are under way for other languages and tools such as bash. If a toolchain rule does not exist for the tool your rule uses, consider creating a toolchain rule.
Managing implicit dependencies
If a build tool can access dependencies across build actions, those actions will fail when remotely executed because each remote build action is executed separately from others. Some build tools retain state across build actions and access dependencies that have not been explicitly included in the tool invocation, which will cause remotely executed build actions to fail.
For example, when Bazel instructs a stateful compiler to locally build foo, the compiler retains references to foo's build outputs. When Bazel then instructs the compiler to build bar, which depends on foo, without explicitly stating that dependency in the BUILD file for inclusion in the compiler invocation, the action executes successfully as long as the same compiler instance executes for both actions (as is typical for local execution). However, since in a remote execution scenario each build action executes a separate compiler instance, compiler state and bar's implicit dependency on foo will be lost and the build will fail.
To help detect and eliminate these dependency problems, Bazel 0.14.1 offers the local Docker sandbox, which has the same restrictions for dependencies as remote execution. Use the sandbox to prepare your build for remote execution by identifying and resolving dependency-related build errors. See Troubleshooting Bazel Remote Execution with Docker Sandbox for more information.
Managing platform-dependent binaries
Typically, a binary built on the host platform cannot safely execute on an arbitrary remote execution platform due to potentially mismatched dependencies. For example, the SingleJar binary supplied with Bazel targets the host platform. However, for remote execution, SingleJar must be compiled as part of the process of building your code so that it targets the remote execution platform. (See the target selection logic.)
Do not ship binaries of build tools required by your build with your source code unless you are sure they will safely run in your execution platform. Instead, do one of the following:
Ship or externally reference the source code for the tool so that it can be built for the remote execution platform.
Pre-install the tool into the remote execution environment (for example, a toolchain container) if it's stable enough and use toolchain rules to run it in your build.
Managing configure-style WORKSPACE rules
WORKSPACE rules can be used for probing the host platform for tools
and libraries required by the build, which, for local builds, is also Bazel's
execution platform. If the build explicitly depends on local build tools and
artifacts, it will fail during remote execution if the remote execution platform
is not identical to the host platform.
The following actions performed by
WORKSPACE rules are not compatible with
Building binaries. Executing compilation actions in
WORKSPACErules results in binaries that are incompatible with the remote execution platform if different from the host platform.
pippackages installed via
WORKSPACErules require that their dependencies be pre-installed on the host platform. Such packages, built specifically for the host platform, will be incompatible with the remote execution platform if different from the host platform.
Symlinking to local tools or artifacts. Symlinks to tools or libraries installed on the host platform created via
WORKSPACErules will cause the build to fail on the remote execution platform as Bazel will not be able to locate them. Instead, create symlinks using standard build actions so that the symlinked tools and libraries are accessible from Bazel's
runfilestree. Do not use
repository_ctx.symlinkto symlink target files outside of the external repo directory.
Mutating the host platform. Avoid creating files outside of the Bazel
runfilestree, creating environment variables, and similar actions, as they may behave unexpectedly on the remote execution platform.
To help find potential non-hermetic behavior you can use Workspace rules log.
If an external dependency executes specific operations dependent on the host
platform, you should split those operations between
WORKSPACE and build
rules as follows:
Platform inspection and dependency enumeration. These operations are safe to execute locally via
WORKSPACErules, which can check which libraries are installed, download packages that must be built, and prepare required artifacts for compilation. For remote execution, these rules must also support using pre-checked artifacts to provide the information that would normally be obtained during host platform inspection. Pre-checked artifacts allow Bazel to describe dependencies as if they were local. Use conditional statements or the
--override_repositoryflag for this.
Generating or compiling target-specific artifacts and platform mutation. These operations must be executed via regular build rules. Actions that produce target-specific artifacts for external dependencies must execute during the build.
To more easily generate pre-checked artifacts for remote execution, you can use
WORKSPACE rules to emit generated files. You can run those rules on each new
execution environment, such as inside each toolchain container, and check the
outputs of your remote execution build in to your source repo to reference.
For example, for Tensorflow's rules for
WORKSPACE rules produce the following
For local execution, files produced by checking the host environment are used.
For remote execution, a conditional statement
on an environment variable allows the rule to use files that are checked into