protoc-gen-star (PG*)

!!! THIS PROJECT IS A WORK-IN-PROGRESS | THE API SHOULD BE CONSIDERED UNSTABLE !!!

PG* is a protoc plugin library for efficient proto-based code generation

package main

import "github.com/lyft/protoc-gen-star"

func main() {
  pgs.Init(pgs.DebugEnv("DEBUG")).
    RegisterModule(&myPGSModule{}).
    RegisterPostProcessor(&myPostProcessor{}).
    Render()
}

Features

Documentation

While this README seeks to describe many of the nuances of protoc plugin development and using PG*, the true documentation source is the code itself. The Go language is self-documenting and provides tools for easily reading through it and viewing examples. The docs can be viewed on GoDoc or locally by running make docs, which will start a godoc server and open them in the default browser.

Roadmap

• Interface-based and fully-linked dependency graph with access to raw descriptors
• Built-in context-aware debugging capabilities
• Exhaustive, near 100% unit test coverage
• End-to-end testable via overrideable IO & Interface based API
• Visitor pattern and helpers for efficiently walking the dependency graph
• BuildContext to facilitate complex generation
• Parsed, typed command-line Parameters access
• Extensible ModuleBase for quickly creating Modules and facilitating code generation
• Configurable post-processing (eg, gofmt) of generated files
• Support processing proto files from multiple packages
• Load comments (via SourceCodeInfo) from proto files into gathered AST for easy access
• Language-specific helper subpackages for handling common, nuanced generation tasks
• Load plugins/modules at runtime using Go shared libraries

Examples

protoc-gen-example, can be found in the testdata directory. It includes two Module implementations using a variety of the features available. It's protoc execution is included in the testdata/generated Makefile target. Examples are also accessible via the documentation by running make docs.

How It Works

The protoc Flow

Because the process is somewhat confusing, this section will cover the entire flow of how proto files are converted to generated code, using a hypothetical PG* plugin: protoc-gen-myplugin. A typical execution looks like this:

protoc \
  -I . \
  --myplugin_out="foo=bar:../generated" \
  ./pkg/*.proto

protoc, the PB compiler, is configured using a set of flags (documented under protoc -h) and handed a set of files as arguments. In this case, the I flag can be specified multiple times and is the lookup path it uses for imported dependencies in a proto file. By default, the official descriptor protos are already included.

myplugin_out tells protoc to use the protoc-gen-myplugin protoc-plugin. These plugins are automatically resolved from the system's PATH environment variable, or can be explicitly specified with another flag. The official protoc-plugins (eg, protoc-gen-python) are already registered with protoc. The flag's value is specific to the particular plugin, with the exception of the :../generated suffix. This suffix indicates the root directory in which protoc will place the generated files from that package (relative to the current working directory). This generated output directory is not propagated to protoc-gen-myplugin, however, so it needs to be duplicated in the left-hand side of the flag. PG* supports this via an output_path parameter.

protoc parses the passed in proto files, ensures they are syntactically correct, and loads any imported dependencies. It converts these files and the dependencies into descriptors (which are themselves PB messages) and creates a CodeGeneratorRequest (yet another PB). protoc serializes this request and then executes each configured protoc-plugin, sending the payload via stdin.

protoc-gen-myplugin starts up, receiving the request payload, which it unmarshals. There are two phases to a PG*-based protoc-plugin. First, PG* unmarshals the CodeGeneratorRequest received from protoc, and creates a fully connected abstract syntax tree (AST) of each file and all its contained entities. Any parameters specified for this plugin are also parsed for later consumption.

When this step is complete, PG* then executes any registered Modules, handing it the constructed AST. Modules can be written to generate artifacts (eg, files) or just performing some form of validation over the provided graph without any other side effects. Modules provide the great flexibility in terms of operating against the PBs.

Once all Modules are run, PG* writes any custom artifacts to the file system or serializes generator-specific ones into a CodeGeneratorResponse and sends the data to its stdout. protoc receives this payload, unmarshals it, and persists any requested files to disk after all its plugins have returned. This whole flow looks something like this:

foo.proto → protoc → CodeGeneratorRequest → protoc-gen-myplugin → CodeGeneratorResponse → protoc → foo.pb.go

The PG* library hides away nearly all of this complexity required to implement a protoc-plugin!

Modules

PG* Modules are handed a complete AST for those files that are targeted for generation as well as all dependencies. A Module can then add files to the protoc CodeGeneratorResponse or write files directly to disk as Artifacts.

PG* provides a ModuleBase struct to simplify developing modules. Out of the box, it satisfies the interface for a Module, only requiring the creation of Name and Execute methods. ModuleBase is best used as an anonyomous embedded field of a wrapping Module implementation. A minimal module would look like the following:

// ReportModule creates a report of all the target messages generated by the
// protoc run, writing the file into the /tmp directory.
type reportModule struct {
  *pgs.ModuleBase
}

// New configures the module with an instance of ModuleBase
func New() pgs.Module { return &reportModule{&pgs.ModuleBase{}} }

// Name is the identifier used to identify the module. This value is
// automatically attached to the BuildContext associated with the ModuleBase.
func (m *reportModule) Name() string { return "reporter" }

// Execute is passed the target files as well as its dependencies in the pkgs
// map. The implementation should return a slice of Artifacts that represent
// the files to be generated. In this case, "/tmp/report.txt" will be created
// outside of the normal protoc flow.
func (m *reportModule) Execute(targets map[string]pgs.File, pkgs map[string]Package) []pgs.Artifact {
  buf := &bytes.Buffer{}

  for _, f := range targets {
    m.Push(f.Name().String()).Debug("reporting")

    fmt.Fprintf(buf, "--- %v ---", f.Name())

    for i, msg := range f.AllMessages() {
      fmt.Fprintf(buf, "%03d. %v\n", i, msg.Name())
    }

    m.Pop()
  }

  m.OverwriteCustomFile(
    "/tmp/report.txt",
    buf.String(),
    0644,
  )

  return m.Artifacts()
}

ModuleBase exposes a PG* BuildContext instance, already prefixed with the module's name. Calling Push and Pop allows adding further information to error and debugging messages. Above, each file from the target package is pushed onto the context before logging the "reporting" debug message.

The base also provides helper methods for adding or overwriting both protoc-generated and custom files. The above execute method creates a custom file at /tmp/report.txt specifying that it should overwrite an existing file with that name. If it instead called AddCustomFile and the file existed, no file would have been generated (though a debug message would be logged out). Similar methods exist for adding generator files, appends, and injections. Likewise, methods such as AddCustomTemplateFile allows for Templates to be rendered instead.

After all modules have been executed, the returned Artifacts are either placed into the CodeGenerationResponse payload for protoc or written out to the file system. For testing purposes, the file system has been abstracted such that a custom one (such as an in-memory FS) can be provided to the PG* generator with the FileSystem InitOption.

Post Processing

Artifacts generated by Modules sometimes require some mutations prior to writing to disk or sending in the response to protoc. This could range from running gofmt against Go source or adding copyright headers to all generated source files. To simplify this task in PG*, a PostProcessor can be utilized. A minimal looking PostProcessor implementation might look like this:

// New returns a PostProcessor that adds a copyright comment to the top
// of all generated files.
func New(owner string) pgs.PostProcessor { return copyrightPostProcessor{owner} }

type copyrightPostProcessor struct {
  owner string
}

// Match returns true only for Custom and Generated files (including templates).
func (cpp copyrightPostProcessor) Match(a pgs.Artifact) bool {
  switch a := a.(type) {
  case pgs.GeneratorFile, pgs.GeneratorTemplateFile,
    pgs.CustomFile, pgs.CustomTemplateFile:
      return true
  default:
      return false
  }
}

// Process attaches the copyright header to the top of the input bytes
func (cpp copyrightPostProcessor) Process(in []byte) (out []byte, err error) {
  cmt := fmt.Sprintf("// Copyright © %d %s. All rights reserved\n",
    time.Now().Year(),
    cpp.owner)

  return append([]byte(cmt), in...), nil
}

The copyrightPostProcessor struct satisfies the PostProcessor interface by implementing the Match and Process methods. After PG* recieves all Artifacts, each is handed in turn to each registered processor's Match method. In the above case, we return true if the file is a part of the targeted Artifact types. If true is returned, Process is immediately called with the rendered contents of the file. This method mutates the input, returning the modified value to out or an error if something goes wrong. Above, the notice is prepended to the input.

PostProcessors are registered with PG* similar to Modules:

g := pgs.Init(pgs.IncludeGo())
g.RegisterModule(some.NewModule())
g.RegisterPostProcessor(copyright.New("PG* Authors"))

Protocol Buffer AST

While protoc ensures that all the dependencies required to generate a proto file are loaded in as descriptors, it's up to the protoc-plugins to recognize the relationships between them. To get around this, PG* uses constructs an abstract syntax tree (AST) of all the Entities loaded into the plugin. This AST is provided to every Module to facilitate code generation.

Hierarchy

The hierarchy generated by the PG* gatherer is fully linked, starting at a top-level Package down to each individual Field of a Message. The AST can be represented with the following digraph:

A Package describes a set of Files loaded within the same namespace. As would be expected, a File represents a single proto file, which contains any number of Message, Enum or Service entities. An Enum describes an integer-based enumeration type, containing each individual EnumValue. A Service describes a set of RPC Methods, which in turn refer to their input and output Messages.

A Message can contain other nested Messages and Enums as well as each of its Fields. For non-scalar types, a Field may also reference its Message or Enum type. As a mechanism for achieving union types, a Message can also contain OneOf entities that refer to some of its Fields.

Visitor Pattern

The structure of the AST can be fairly complex and unpredictable. Likewise, Module's are typically concerned with only a subset of the entities in the graph. To separate the Module's algorithm from understanding and traversing the structure of the AST, PG* implements the Visitor pattern to decouple the two. Implementing this interface is straightforward and can greatly simplify code generation.

Two base Visitor structs are provided by PG* to simplify developing implementations. First, the NilVisitor returns an instance that short-circuits execution for all Entity types. This is useful when certain branches of the AST are not interesting to code generation. For instance, if the Module is only concerned with Services, it can use a NilVisitor as an anonymous field and only implement the desired interface methods:

// ServiceVisitor logs out each Method's name
type serviceVisitor struct {
  pgs.Visitor
  pgs.DebuggerCommon
}

func New(d pgs.DebuggerCommon) pgs.Visitor {
  return serviceVistor{
    Visitor:        pgs.NilVisitor(),
    DebuggerCommon: d,
  }
}

// Passthrough Packages, Files, and Services. All other methods can be
// ignored since Services can only live in Files and Files can only live in a
// Package.
func (v serviceVisitor) VisitPackage(pgs.Package) (pgs.Visitor, error) { return v, nil }
func (v serviceVisitor) VisitFile(pgs.File) (pgs.Visitor, error)       { return v, nil }
func (v serviceVisitor) VisitService(pgs.Service) (pgs.Visitor, error) { return v, nil }

// VisitMethod logs out ServiceName#MethodName for m.
func (v serviceVisitor) VisitMethod(m pgs.Method) (pgs.Vistitor, error) {
  v.Logf("%v#%v", m.Service().Name(), m.Name())
  return nil, nil
}

If access to deeply nested Nodes is desired, a PassthroughVisitor can be used instead. Unlike NilVisitor and as the name suggests, this implementation passes through all nodes instead of short-circuiting on the first unimplemented interface method. Setup of this type as an anonymous field is a bit more complex but avoids implementing each method of the interface explicitly:

type fieldVisitor struct {
  pgs.Visitor
  pgs.DebuggerCommon
}

func New(d pgs.DebuggerCommon) pgs.Visitor {
  v := &fieldVisitor{DebuggerCommon: d}
  v.Visitor = pgs.PassThroughVisitor(v)
  return v
}

func (v *fieldVisitor) VisitField(f pgs.Field) (pgs.Visitor, error) {
  v.Logf("%v.%v", f.Message().Name(), f.Name())
  return nil, nil
}

Walking the AST with any Visitor is straightforward:

v := visitor.New(d)
err := pgs.Walk(v, pkg)

All Entity types and Package can be passed into Walk, allowing for starting a Visitor lower than the top-level Package if desired.

Build Context

Modules registered with the PG* Generator are initialized with an instance of BuildContext that encapsulates contextual paths, debugging, and parameter information.

Output Paths

The BuildContext's OutputPath method returns the output directory that the PG* plugin is targeting. This path is also initially . but refers to the directory in which protoc is executed. This default behavior can be overridden by providing an output_path in the flag.

The OutputPath can be used to create file names for Artifacts, using JoinPath(name ...string) which is essentially an alias for filepath.Join(ctx.OutputPath(), name...). Manually tracking directories relative to the OutputPath can be tedious, especially if the names are dynamic. Instead, a BuildContext can manage these, via PushDir and PopDir.

ctx.OutputPath()                // foo
ctx.JoinPath("fizz", "buzz.go") // foo/fizz/buzz.go

ctx = ctx.PushDir("bar/baz")
ctx.OutputPath()                // foo/bar/baz
ctx.JoinPath("quux.go")         // foo/bar/baz/quux.go

ctx = ctx.PopDir()
ctx.OutputPath()                // foo

ModuleBase wraps these methods to mutate their underlying BuildContexts. Those methods should be used instead of the ones on the contained BuildContext directly.

Debugging

The BuildContext exposes a DebuggerCommon interface which provides utilities for logging, error checking, and assertions. Log and the formatted Logf print messages to os.Stderr, typically prefixed with the Module name. Debug and Debugf behave the same, but only print if enabled via the DebugMode or DebugEnv InitOptions.

Fail and Failf immediately stops execution of the protoc-plugin and causes protoc to fail generation with the provided message. CheckErr and Assert also fail with the provided messages if an error is passed in or if an expression evaluates to false, respectively.

Additional contextual prefixes can be provided by calling Push and Pop on the BuildContext. This behavior is similar to PushDir and PopDir but only impacts log messages. ModuleBase wraps these methods to mutate their underlying BuildContexts. Those methods should be used instead of the ones on the contained BuildContext directly.

Parameters

The BuildContext also provides access to the pre-processed Parameters from the specified protoc flag. The only PG*-specific key expected is "output_path", which is utilized by a module's BuildContext for its OutputPath.

PG* permits mutating the Parameters via the MutateParams InitOption. By passing in a ParamMutator function here, these KV pairs can be modified or verified prior to the PGG workflow begins.

Language-Specific Subpackages

While implemented in Go, PG* seeks to be language agnostic in what it can do. Therefore, beyond the pre-generated base descriptor types, PG* has no dependencies on the protoc-gen-go (PGG) package. However, there are many nuances that each language's protoc-plugin introduce that can be generalized. For instance, PGG package naming, import paths, and output paths are a complex interaction of the proto package name, the go_package file option, and parameters passed to protoc. While PG*'s core API should not be overloaded with many language-specific methods, subpackages can be provided that can operate on Parameters and Entities to derive the appropriate results.

PG* currently implements the pgsgo subpackage to provide these utilities to plugins targeting the Go language. Future subpackages are planned to support a variety of languages.

PG* Development & Make Targets

PG* seeks to provide all the tools necessary to rapidly and ergonomically extend and build on top of the Protocol Buffer IDL. Whether the goal is to modify the official protoc-gen-go output or create entirely new files and packages, this library should offer a user-friendly wrapper around the complexities of the PB descriptors and the protoc-plugin workflow.

Setup

For developing on PG*, you should install the package within the GOPATH. PG* uses glide for dependency management.

go get -u github.com/lyft/protoc-gen-star
cd $GOPATH/src/github.com/lyft/protoc-gen-star
make vendor

To upgrade dependencies, please make the necessary modifications in glide.yaml and run glide update.

Linting & Static Analysis

To avoid style nits and also to enforce some best practices for Go packages, PG* requires passing golint, go vet, and go fmt -s for all code changes.

make lint

Testing

PG* strives to have near 100% code coverage by unit tests. Most unit tests are run in parallel to catch potential race conditions. There are three ways of running unit tests, each taking longer than the next but providing more insight into test coverage:

# run code generation for the data used by the tests
make testdata

# run unit tests without race detection or code coverage reporting
make quick

# run unit tests with race detection and code coverage
make tests

# run unit tests with race detection and generates a code coverage report, opening in a browser
make cover

protoc-gen-debug

PG* comes with a specialized protoc-plugin, protoc-gen-debug. This plugin captures the CodeGeneratorRequest from a protoc execution and saves the serialized PB to disk. These files can be used as inputs to prevent calling protoc from tests.

Documentation

Go is a self-documenting language, and provides a built in utility to view locally: godoc. The following command starts a godoc server and opens a browser window to this package's documentation. If you see a 404 or unavailable page initially, just refresh.

make docs

Demo

PG* comes with a "kitchen sink" example: protoc-gen-example. This protoc plugin built on top of PG* prints out the target package's AST as a tree to stderr. This provides an end-to-end way of validating each of the nuanced types and nesting in PB descriptors:

# create the example PG*-based plugin
make bin/protoc-gen-example

# run protoc-gen-example against the demo protos
make testdata/generated

CI

PG* uses TravisCI to validate all code changes. Please view the configuration for what tests are involved in the validation.