Skip to main content
Version: 2.1 (deprecated)

Add codegen

How to add a new code generator to Pants.

Pants supports code generators that convert a protocol language like Protobuf into other languages, such as Python or Java. The same protocol source may be used to generate multiple different distinct languages.

Pants will not actually write the generated files to disk, except when running ./pants export-codegen. Instead, any targets that depend on the protocol targets will cause their code to be generated, and those generated files will be copied over into the "chroot" (temporary directory) where Pants executes.

Example: Protobuf -> Python

This guide walks through each step of adding Protobuf to generate Python sources. See here for the final result.

This guide assumes that you are running a code generator that already exists outside of Pants as a stand-alone binary, such as running Protoc or Thrift.

If you are instead writing your own code generation logic inline, you can skip Step 2. In Step 4, rather than running a Process, use CreateDigest.

1. Create a target type for the protocol​

You will need a define a new target type to allow users to provide metadata for their protocol files, e.g. their .proto files. See Creating new targets for a guide on how to do this.

from pants.engine.target import COMMON_TARGET_FIELDS, Dependencies, Sources, Target

class ProtobufSources(Sources):
    default = ("*.proto",)
    expected_file_extensions = (".proto",)

class ProtobufLibrary(Target):
    """Protobuf files used to generate various languages."""

    alias = "protobuf_library"
    core_fields = (*COMMON_TARGET_FIELDS, Dependencies, ProtobufSources)

You should define a subclass of Sources, like ProtobufSources or ThriftSources. This is important for Step 3.

Typically, you will want to register the Dependencies field.

Target type already exists?​

If Pants already has a target type for your protocol—such as Pants already having a ProtobufLibrary defined—you should not create a new target type.

Instead, you can optionally add any additional fields that you would like through plugin fields. See Extending pre-existing targets.

Add dependency injection (Optional)​

Often, generated files will depend on a runtime library to work. For example, Python files generated from Protobuf depend on the protobuf library.

Instead of users having to explicitly add this dependency every time, you can dynamically inject this dependency for them.

To inject dependencies:

  1. Subclass the Dependencies field. Register this subclass on your protocol target type.
  2. Define a subclass of InjectDependenciesRequest and set the class property inject_for to the Dependencies subclass defined in the previous step. Register this new class with a UnionRule for InjectDependenciesRequest.
  3. Create a new rule that takes your new InjectDependenciesRequest subclass as a parameter and returns InjectedDependencies.
from pants.engine.addresses import Address
from pants.engine.target import Dependencies, InjectDependenciesRequest, InjectedDependencies
from pants.engine.rules import collect_rules, rule
from pants.engine.unions import UnionRule

class ProtobufDependencies(Dependencies):
    pass

...

class ProtobufLibrary(Target):
    alias = "protobuf_library"
    core_fields = (*COMMON_TARGET_FIELDS, ProtobufDependencies, ProtobufSources)

class InjectProtobufDependencies(InjectDependenciesRequest):
    inject_for = ProtobufDependencies

@rule
async def inject_dependencies(_: InjectProtobufDependencies) -> InjectedDependencies:
    address = Address("3rdparty/python", target_name="protobuf")
    return InjectedDependencies([address])

def rules():
    return [
        *collect_rules(),
        UnionRule(InjectDependenciesRequest, InjectProtobufDependencies),
    ]

This example hardcodes the injected address. You can instead add logic to your rule to make this dynamic. For example, in Pants's Protobuf implementation, the python-protobuf subsystem has a --runtime-dependencies option. The rule takes that subsystem as a parameter, and then converts the option's string value to an address with Get(Addresses, UnparsedAddressInputs). See protobuf/python/python_protobuf_subsystem.py.

2. Install your code generator​

There are several ways for Pants to install your tool. See Installing tools. This example will use ExternalTool because there is already a pre-compiled binary for Protoc.

from pants.core.util_rules.external_tool import ExternalTool
from pants.engine.platform import Platform

class Protoc(ExternalTool):
    """The protocol buffer compiler (https://developers.google.com/protocol-buffers)."""

    options_scope = "protoc"
    default_version = "3.11.4"
    default_known_versions = [
        "3.11.4|darwin|8c6af11e1058efe953830ecb38324c0e0fd2fb67df3891896d138c535932e7db|2482119",
        "3.11.4|linux |6d0f18cd84b918c7b3edd0203e75569e0c8caecb1367bbbe409b45e28514f5be|1591191",
    ]

    def generate_url(self, plat: Platform) -> str:
        plat_str = "osx" if plat == Platform.darwin else "linux"
        return (
            f"https://github.com/protocolbuffers/protobuf/releases/download/"
            f"v{self.version}/protoc-{self.version}-{plat_str}-x86_64.zip"
        )

    def generate_exe(self, _: Platform) -> str:
        return "./bin/protoc"

3. Create a GenerateSourcesRequest​

GenerateSourcesRequest tells Pants the input and the output of your code generator, such as going from ProtobufSources -> PythonSources. Pants will use this to determine when to use your code generation implementation.

Subclass GenerateSourcesRequest:

from pants.engine.target import GenerateSourcesRequest

class GeneratePythonFromProtobufRequest(GenerateSourcesRequest):
    input = ProtobufSources
    output = PythonSources

The input should be the Sources class for your protocol target from Step 1.

The output should typically be the Sources class corresponding to the "language" you're generating for, such as JavaSources or PythonSources. The output type will understand subclasses of what you specify, so, generally, you should specify PythonSources instead of something more specific like PythonLibrarySources or PythonBinarySources.

Note that your rule will not actually return an instance of the output type, e.g. PythonSources. Codegen rules only return a Snapshot, rather than a whole Sources field. The output field is only used as a signal of intent.

Finally, register your new GenerateSourcesRequest with a UnionRule.

from pants.engine.rules import collect_rules
from pants.engine.unions import UnionRule

def rules():
    return [
        *collect_rules(),
        UnionRule(GenerateSourcesRequest, GeneratePythonFromProtobufRequest),
    ]

4. Create a rule for your codegen logic​

Your rule should take as a parameter the GenerateSourcesRequest from Step 3 and the Subsystem (or ExternalTool) from Step 2. It should return GeneratedSources.

from pants.engine.rules import rule
from pants.engine.target import GeneratedSources

...

@rule
async def generate_python_from_protobuf(
    request: GeneratePythonFromProtobufRequest, protoc: Protoc
) -> GeneratedSources:
    ...
    return GeneratedSources(result_snaphsot)

The GenerateSourcesRequest parameter will have two fields: protocol_sources: Snapshot and protocol_target: Target. Often, you will want to include protocol_sources in the input_digest to the Process you use to run the generator. You can use protocol_target to look up more information about the input target, such as finding its dependencies.

The rule should return GeneratedSources, which take a Snapshot as its only argument. This should be a Snapshot of the generated files for the input target.

If you used ExternalTool in step 1, you will use Get(DownloadedExternalTool, ExternalToolRequest) to install the tool. Be sure to merge this with the protocol_sources and any other relevant input digests via Get(Digest, MergeDigests).

For many code generators, you will need to get the input target's direct or transitive dependencies and include their sources in the input_digest. See Rules and the Target API.

Use DependenciesRequestLite and TransitiveTargetsRequestLite

For codegen, you will get a rule graph error if you use DependenciesRequest and TransitiveTargetsRequest like normal. Instead, use DependenciesRequestLite and TransitiveTargetsRequestLite, both from pants.engine.target.

These behave almost identically and take the same arguments in their constructors. However, these "lite" versions do not include dependency inference in the result. Unless you are generating code with PythonSources as the input, this should be fine.

This is fixed in Pants 2.2.

You will likely need to add logic for handling source roots. For example, the code generator may not understand source roots so you may need to strip source roots before putting the sources in the input_digest. Likely, you will want to restore a source root after generation because most Pants code will assume that there is a source root present. In the below example, we restore the original source root, e.g. src/protobuf/f.proto becomes src/protobuf/f_pb2.py. See protobuf/python/rules.py for a more complex example that allows the user to specify what source root to use through a field on the protobuf_library.

@rule
async def generate_python_from_protobuf(
    request: GeneratePythonFromProtobufRequest, protoc: Protoc
) -> GeneratedSources:
    download_protoc_request = Get(
        DownloadedExternalTool, ExternalToolRequest, protoc.get_request(Platform.current)
    )

    # Protoc needs all transitive dependencies on `protobuf_libraries` to work properly. It won't
    # actually generate those dependencies; it only needs to look at their .proto files to work
    # with imports.
    transitive_targets = await Get(
        TransitiveTargets,
        TransitiveTargetsRequestLite([request.protocol_target.address]),
    )

    # NB: By stripping the source roots, we avoid having to set the value `--proto_path`
    # for Protobuf imports to be discoverable.
    all_stripped_sources_request = Get(
        StrippedSourceFiles,
        SourceFilesRequest(
            (tgt.get(Sources) for tgt in transitive_targets.closure),
            for_sources_types=(ProtobufSources,),
        ),
    )
    target_stripped_sources_request = Get(
        StrippedSourceFiles, SourceFilesRequest([request.protocol_target[ProtobufSources]]),
    )

    (
        downloaded_protoc_binary,
        all_sources_stripped,
        target_sources_stripped,
    ) = await MultiGet(
        download_protoc_request,
        all_stripped_sources_request,
        target_stripped_sources_request,
    )

    input_digest = await Get(
        Digest,
        MergeDigests(
            (all_sources_stripped.snapshot.digest, downloaded_protoc_binary.digest)
        ),
    )

    output_dir = "_generated_files"
    result = await Get(
        ProcessResult,
        Process(
            (
                downloaded_protoc_binary.exe,
                "--python_out",
                output_dir,
                *target_sources_stripped.snapshot.files,
            ),
            input_digest=input_digest,
            description=f"Generating Python sources from {request.protocol_target.address}.",
            output_directories=(output_dir,),
        ),
    )

    # We must do some path manipulation on the output digest for it to look like normal sources,
    # including adding back a source root.
    source_root_request = SourceRootRequest.for_target(request.protocol_target)
    normalized_digest, source_root = await MultiGet(
        Get(Digest, RemovePrefix(result.output_digest, output_dir)),
        Get(SourceRoot, SourceRootRequest, source_root_request),
    )
    source_root_restored = (
        await Get(Snapshot, AddPrefix(normalized_digest, source_root.path))
        if source_root.path != "."
        else await Get(Snapshot, Digest, normalized_digest)
    )
    return GeneratedSources(source_root_restored)

Finally, update your plugin's register.py to activate this file's rules.

pants-plugins/protobuf/register.py
from protobuf import python_support

def rules():
    return [*python_support.rules()]
Tip: use export-codegen to test it works

Run ./pants export-codegen path/to/file.ext to ensure Pants is correctly generating the file. This will write the generated file(s) under the dist/ directory, using the same path that will be used during Pants runs.

5. Audit call sites to ensure they've enabled codegen​

Call sites must opt into using codegen, and they must also specify what types of sources they're expecting. See Rules and the Target API about the Sources field.

For example, if you added a code generator that goes from ProtobufSources -> JavaSources, then Pants's Python backend would not use your new implementation because it ignores JavaSources.

You should check that everywhere you're expecting is using your new codegen implementation by manually testing it out. Create a new protocol target, add it to the dependencies field of a target, and then run goals like ./pants package and ./pants test to make sure that the generated file works correctly.

It may also be helpful to inspect the code for the rest of your plugin and/or Pants's own rules that you're using. Check that HydrateSourcesRequest and SourceFilesRequest have set enable_codegen=True and that they put the correct value for the argument for_sources_type.

6. Add tests (optional)​

Refer to Testing rules.