Remote execution allows Bazel to execute actions on a separate platform, such as a datacenter. A gRPC protocol is currently in development. You can try remote execution with bazel-buildfarm, an open-source project that aims to provide a distributed remote execution platform. This document 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.
This document uses the following terminology when referring to different environment types or platforms:
When configuring a Bazel build for remote execution, you must follow the guidelines described in this document 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 document 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:
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 PATH, JAVA_HOME, or other local variables that may not be set to equivalent values (or at all) in the remote execution environment.
Toolchain rules currently exist for Bazel build and test rules for [Scala](https://github.com/bazelbuild/rules_scala/blob/master/scala/scala_toolch ain.bzl), 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.
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.
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.
configure-style WORKSPACE rulesBazel‘s 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 remote execution:
Building binaries. Executing compilation actions in WORKSPACE rules results in binaries that are incompatible with the remote execution platform if different from the host platform.
Installing pip packages. pip packages installed via WORKSPACE rules 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 WORKSPACE rules 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 runfiles tree. Do not use repository_ctx.symlink to symlink target files outside of the external repo directory.
Mutating the host platform. Avoid creating files outside of the Bazel runfiles tree, 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, we recommend splitting those operations between WORKSPACE and build rules as follows:
Platform inspection and dependency enumeration. These operations are safe to execute locally via WORKSPACE rules, 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_repository flag 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 cuda and python, the WORKSPACE rules produce the following BUILD files. 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 the repo. The BUILD files declare [genrules](https://github.com/tensorflow/tensorflow/blob/master/third_party/py/python_configure.bzl#L84) that can run both locally and remotely, and perform the necessary processing that was previously done via repository_ctx.symlink as shown here.