The Bazel Platform Last updated Mar 24, 2025

Purpose: high-level overview of Bazel’s capabilities and how they are used by OBazl to support OCaml.

Overview

Bazel is not usually thought of as a platform. It is primarily a build system, but it is much more than that. It is also:

Furthermore, Bazel has powerful extension mechanisms:

These facilities make it possible to seamlessly integrate tools of all sorts into development workflows of all kinds, with Bazel acting as the master-of-ceremonies.

See also:

Bazel provides a great deal of information about your build program, and it provides fine-grained control over build actions. You can build any target in your project, and only it and its dependencies will be built. You can parameterize your builds at any level of granularity. If you include appropriate test targets, you can modularize your development even if your source code is not organized in a modular manner.

Furthermore, Bazel includes powerful query facilities that make it possible to explore the dependency structures of your code. You can easily list the dependency chain between two targets, or list the targets that have a certain parameter, etc. You can generate SVG graphs showing dependency graphs of your code.

Bazel also makes it easy to develop multiple, mutually-dependent projects simultaneously. If your main project depends on an external repository, you can easily configure your build to use a local copy of the dependency without altering your build code. One benefit of this is that it makes it easy to eliminate embedded git submodules.

No installation

Installation? Installation? We don’t need no stinkin’ installation!

Bazel’s package management is very different from opam’s. Installation, deployment, packaging etc. are not directly handled by Bazel. Those capabilities must be supplied by rules, for example rules_pkg.

Tool runners

bazel run can be used to build and run tools.

Bazel

OBazl recommends using Bazelisk (Installing Bazel using Bazelisk).

Official Bazel User Guide

A great deal of Bazel documentation is available, but it is not always easy to find what you need. If you do find it, you probably want to bookmark it.

If you use Bazelisk, you can pin the Bazel version by putting file .bazelversion in the root directory of your project, containing the required version string, e.g. latest or 5.1.1.

To work effectively with OBazl you must master the following material at minimum:

If you are just getting started with Bazel, you should work through one of the Tutorials

Setup

To get the most out of OBazl and Bazel, you need to decide on some conventions and do a little configuration. See OBazl Conventions for a list.

Aspects

An aspect may be thought of as an extension of the normal build program. The core of an aspect is one or more build actions that are executed in addition to or instead of the normal build actions. Typical examples are aspects that run formatters or linters.

There are two ways to implement an aspect in this sense: external and inline.

External aspects

External aspects use Bazel’s aspect mechanism, which is designed to support the use of external tools (tools not included in the toolchain) to extend build actions throughout a dependency graph defined by the aspect. External aspects specify a set of rule attributes that determine a “shadow” dependency graph, and Bazel will apply the aspect to all the targets in the dependency graph.

Inline aspects

The other method, which is not generally recognized, is to use Bazel’s --output_groups facility to enable inline aspects. An inline aspect is a build action that is directly supported by a build rule but only executed when the user requests a specific output group. Inline aspects are local; they do not propagate to additional build targets in the way that

For example, the OCaml SDK includes a tool, ocamlobjinfo, that dumps textual (human-readable) metadata about compiled artifacts to stdout. To use ocamlobjinfo, the developer must first compile the object of interest - say, //foo/bar:A, producing foo/bar/a.cmo - and then run ocamlobjinfo against the output.

Without an inline aspect, this would be impractical for Bazel users. Building a target puts the outputs in a filesystem area controlled by Bazel, on a path constructed by Bazel. So to apply ocamlobjinfo to the output of $ bazel build foo/bar:A, one would have to obtain from Bazel the path of the output object. This is possible, but very cumbersome.

But an inline aspect makes this almost trivial. The ocaml_module rule includes in its implementation function a build action to compile its input. It also includes an inline aspect that runs ocamlobjinfo against output of the compile action. The inline aspect is just a build action that takes the output of the normal compile action as its input, and uses ctx.actions.run_shell to run ocamlobjinfo. (Use of run_shell is required in this case because ocamlobjinfo writes to stdout and does not have a parameter like -o to redirect output to a file.)

What makes this work as an aspect is that the output of the ocamlobinfo action is listed only in the OutputGroupInfo provider, as field modinfo. That means that the aspect action will only be executed if the user passes --output_groups=modinfo to the build command. So a normal $ bazel build foo/bar:A will ignore the inline aspect and proceed as usual. But $ bazel build foo/bar:A --output_groups=modinfo tells bazel that only the output specified as the modinfo field of the OutputGroupInfo provider need be built. Since that field contains the output of the ocamlobjinfo build action, which in turn depends on the output of the compile action, the result is that the target will be compile as normal, and then the ocamlobjinfo action will be run against the compilation output, producing the modinfo requested by the user.

Inline aspects only make sense for actions that use tools provided by the toolchain. An inline action that uses an external tool would make the ruleset dependent on that tool, whether the user wants to use it or not.

Observability

Inspecting the Bazel environment, logs, build actions, etc.

bazel info

The bazel info command will print a dictionary listing the parameters, file locations, etc. that Bazel uses internally. It supports a large number of options; run $ bazel help info to see them all; to see just the keys for the dictionary, run $ bazel help info-keys.

Most of entries in the dictionary, most of the time, can be safely ignored; but if you run into trouble, two of them can be helpful with debugging: command_log and output_base.

command_log

Bazel writes logs to a command_log file each time it executes a command; it overwrites the file. You can discover the location of the file by running $ bazel info command_log. Since the output of this command will overwrite the log file, you must use an alias or shell script to enable easy browsing. See the aliases recommendation in OBazl Conventions for an example.

output_base

The output_base directory contains a subdirectory, external, that contains the external repositories your project has configured. You can browse the BUILD.bazel files of an external repo, for example, to verify that you are using the correct target labels.

targets

Bazel can print a source text representation of each target :

$ bazel query '@ppx_tools//:*' --output build

This shows what the target looks like to Bazel after it has been processed (e.g. variables expanded, etc.).

actions

A single build target may generate multiple build actions. For example, if an ocaml_module rule is parameterized with a ppx argument, it will generate two actions: one to transform the source file with the PPX, and one to compile the result. Each action will have a command line string.

Normally there is no need to pay these actions any mind, but if something goes wrong with your build it may be useful to see exactly what a build rule is doing - what the actions are, what commands and arguments are used to run the actions, and what the inputs and outputs are. Fortunately this is easy to do. You can use the [action query]() facility to print all the actions generated by a rule without actually running the rule (so it does not trigger any compilation). For example, the following will print all the actions (and much additional information) generated by the //foo/bar:baz target:

$ bazel aquery //foo/bar:baz

See Transparency for more information.

packages

To list all the files in the directory containing associated with a package:

$ bazel query 'kind("source file", @<repo>//<pkg>:*)' --output label_kind

To include all subpackages (subdirectories containing BUILD.bazel file), use @<repo>//<pkg>/...:* instead of @<repo>//<pkg>:*.

E.g.

$ bazel query 'kind("source file", @ppx_tools//:*)' --output label_kind

$ bazel query 'kind("source file", @ppx_tools//metaquot:*)' --output label_kind

todo…​

Useful tips

  • The clean command "[r]emoves bazel-created output, including all object files, and bazel metadata." It will not refresh repository dependencies. Adding the --expunge option will delete everything; it will also stop the server, so that then next build command will start from scratch. You almost never need to do this.

  • You should rarely need to run $ bazel clean. Bazel caches a complete description of the build, so it always knows what needs to be rebuilt. However, if you change the build structure - especially if you remove build targets - you may need this command to rebuild the cache.

  • Do spend some time learning to use the query facilities. On a project of any size you’ll be glad you did.

  • To experiment with build rules etc. you can avoid cluttering the source tree by creating a BUILD.bazel in a work directory like dev and putting the rules there. Since dependencies are expressed as target labels, you can reach into the tree anywhere you like, although you may need to adjust the visibility attribute of targets.

  • Use Bazelisk to make sure you’re always using the latest version of Bazel. You can pin the version you want by using a .bazelversion file.

  • You can enable command-line completion (also known as tab-completion) in Bash and Zsh. This lets you tab-complete command names, flags names and flag values, and target names. Caveat: tab-completion may be an issue for Bazelisk; see Support bash autocomplete #29.)

  • If you need to make some kind of global change, e.g. renaming a target or adding a dependencie to multiple rules, do not search-and-replace. Use buildozer instead. (See Batch Editing for more information.)

MacOS

versions- os, sdk, xcode

Because the OCaml compilers may call a compiler (linker/assember), setting these Bazel switches may not be enough to eliminate those pesky "object was built for newer version.." (see below). You may also need to pass some options to the linker using -ccopt (TODO: details)

Locking Xcode versions in Bazel blog post, March 2021; slightly outdated

"Bazel discovers your currently installed Xcode versions by running xcode_locator, and then generating a BUILD file that contains an entry for every version you currently have installed…​ In order to enforce developers use the same version, you can short circuit bazel’s Xcode discovery and instead reference a local target that you provide."

NB: xcode-locator is a Bazel tool, not a macos tool.

the blog article is based on pre-bzlmod versions; it lists this command: cat bazel-$(basename $PWD)/external/local_config_xcode/BUILD To use it with bzlmod replace external/local_config_xcode with external/bazel_tools~xcode_configure_extension\~local_config_xcode

To retrieve the build number from your current Xcode version:

xcodebuild -version

better to use --macos_sdk_version? see below
  • --xcode_version=<a string>

    • If specified, uses Xcode of the given version for relevant build actions. If unspecified, uses the executor default version of Xcode.

  • --xcode_version_config=<a build target label>

    • default: "@bazel_tools//tools/cpp:host_xcodes"

    • The label of the xcode_config rule to be used for selecting the Xcode version in the build configuration.

  • xcrun --sdk macosx --show-sdk-version

  • xcrun --sdk macosx --show-sdk-path

sdk versions

  • --macos_sdk_version=<version>

    • Specifies the version of the macOS SDK to use to build macOS applications. If unspecified, uses default macOS SDK version from 'xcode_version'.

env vars

When using the cc_toolchain (e.g. to build C, C++, objc, …​):

  • SDKROOT

  • DEVELOPER_DIR

  • XCODE_VERSION_OVERRIDE

  • APPLE_SDK_PLATFORM

  • APPLE_SDK_VERSION_OVERRIDE

macos versions

<version> == a dotted version (for example '2.3' or '3.3alpha2.4')> default: see description

  • --host_macos_minimum_os=<version>

    • Minimum compatible macOS version for host targets. If unspecified, uses 'macos_sdk_version'.

  • --macos_minimum_os=<version>

    • Minimum compatible macOS version for targets. If unspecified, uses 'macos_sdk_version'.

warnings

You may see warnings like the following if your code links to CC libraries:

ld: warning: object file (bazel-out/darwin-fastbuild/bin/interop/ffi/case110/cclibs/liblibalpha.a(alpha.o)) was built for newer macOS version (12.3) than being linked (12.0)

This happens because we built liblibalpha.a with a Bazel rule (e.g. cc_library), which automatically adds --mmacos_version_min=x.y.z to its command line. We then pass it to the OCaml linker (on the command line, or using -cclib), which proceeds to call the system linker. But OCaml does not pass --macos_version_min to the linker, so if it has different ideas about the minimum version, you get this warning message.

One way to address this is to add build --macos_version_min=x.y.z to .bazelrc. Another would be to pass --cclib --macos_version_min=x.y.z to OCaml.

The latest version of rules_ocaml should eliminate this problem. The rules interrogate the Bazel cc toolchain to obtain the --macos_version_min option it uses, and then pass it to the OCaml linker.

todo

clang wrapper etc