IMPORTANT: Those upgrading from version 1.4.2 to 2.X should read the UPGRADING.md notes

Protobuf is an implementation of Google's protocol buffers in ruby. We currently support version 2.4.1 with support for the new 2.5 coming shortly after it becomes final.

You will likely need to install protobuf from your favorite package manager or from source. This gem currently supports protobuf 2.4.1. You may alternatively specify a PROTOC_SRC when installing with rubygems (see below).

$ brew install protobuf

$ sudo apt-get install -y protobuf

Once the protobuf package is installed, go ahead and install with rubygems. If you'd like to skip installing the protobuf package (above) and specify an alternate location for the protobuf package, for instance if you have a custom protoc package, then specify it with PROTOC_SRC=/path/to/src. Please note that this will void your warranty as it were. If you compiled with a custom protobuf package and are having issues it may be difficult to troubleshoot.

$ gem install protobuf

# Provide an alternative protoc source directory to build from.
$ PROTOC_SRC=/path/to/protobuf/src gem install protobuf

Protocol Buffers are great because they allow you to clearly define data storage or data transfer packets. Google officially supports Java, C++, and Python for compilation and usage. Let's make it ruby aware!

Let's say you have a defs.proto file that defines a User message.

package foo;
message User {
  required string first_name = 1;
  required string last_name = 2;
}

Now let's compile that definition to ruby:

$ rprotoc defs.proto --ruby_out ./lib

The previous line will take whatever is defined in defs.proto and output ruby classes to the ./lib directory, obeying the package directive. Assuming that's all defs.proto had defined, ./lib should now look like this:

- lib
  |- foo
      |- defs.pb.rb

The generated file defs.pb.rb should look something like this:

module Foo
  class User < ::Protobuf::Message; end

  class User
    required :string, :first_name, 1
    required :string, :last_name, 2
  end
end

Note: The generator will pre-define all the classes empty and then re-open to apply the defined fields. This is an optimization to prevent recursive field errors.

The generated class is now just a plain old ruby object. You can use it however you wish.

require 'lib/foo/user.pb'

# dot notation reading/writing fields
user = Foo::User.new
user.first_name = "Lloyd"
user.last_name = "Christmas"
user.first_name     # => "Lloyd"

# or pass in the fields as a hash to the initializer
user = Foo::User.new :first_name => "Lloyd", :last_name => "Christmas"
user.first_name     # => Lloyd
user.last_name     # => Christmas

Every message object comes ready for serialization or deserialization. Use serialize_to_string to write out the byte-string for the message. Use parse_from_string on a new message instance to inflate the byte-string back to a message in ruby.

user = Foo::User.new(:first_name => 'Bob')
bytes = user.serialize_to_string
puts bytes #=> binary representation of this message object

inflated_user = Foo::User.new.parse_from_string(bytes)
inflated_user == user #=> true

RPC is one of many technologies that tries to solve the problem of getting smaller pieces of data from one place to another. Many will argue for or against RPC and its usefulness, but I'm not going to do that here. Google's Protocol Buffers provides support for Services with RPC and that's why you're here.

Any discussion about RPC leads to a discussion about clients and servers and the remote procedures themselves. For our purposes, we'll talk about a Client (process that is calling the server/service), a Service (the remote procedure), and a Server (the process that manages one or more services). We'll start with the Service first.

Services are simply classes that have endpoint methods defined. Here's what one looks like in protobuf:

package foo;
message UserRequest {
  optional string email = 1;
}
message UserList {
  repeated User users = 1;
}
service UserService {
  rpc Find (UserRequest) returns (UserList);
}

And the equivalent ruby stub for the service (generated with rprotoc):

# lib/foo/user.pb.rb
module Foo
  # UserRequest and UserList Class definitions not shown (see above for generated output of classes).

  class UserService < ::Protobuf::Rpc::Service
    rpc :find, UserRequest, UserList
  end
end

Important Note: The UserService class here is a stub. You should not provide your implementation in this generated file as subsequent generations will wipe out your implmentation. Read on to learn how to use this stub.

Did you read the note above? Go read it. I'll wait.

Ok, now that you have a generated service stub, you'll want to require it from lib and implement the methods. Create a service implementation file in your project. In rails I'd put this in app/services/user_service.rb.

# app/services/user_service.rb
require 'lib/foo/user.pb'

# Reopen the class and provide the implementation for each rpc method defined.
module Foo
  class UserService

    # request -> Foo::UserRequest
    # response -> Foo::UserResponse
    def find
      # request.email will be the unpacked string that was sent by the client request
      User.find_by_email(request.email).each do |user|
        # must only use a proto instance of Foo::User when appending to the `users` field
        response.users << user.to_proto
      end
    end

  end
end

Simply implement the instance method for the defined rpc. You can provide any other methods in this class as helpers, but only those defined in the proto file will be callable by remote clients. Every request made by a client will provide a non-empty request of the defined type. The server creates a new service instance based on the request, so you should not be constrained to just the endpoint method. This is similar to rails controllers where only methods defined by the routes file are hooked up to HTTP requests, but it's very common to implement private methods to aid in code quality and simpilicity.

Every instance has a request and response object used for fulfilling the call, again, similar to a rails controller action. You should never attempt to modify the request object. The response object however should be modified or replaced entirely. If you need to create your own response object (a valid case), simply use respond_with(new_response). The returned object should conform to one of three properties:

1. Response should be of same type as defined by the rpc definition (in this case, Foo::UserList), or
2. Response should be a hash. This hash will be used to construct an instance of the defined type and should therefore conform to the appropriate fields for that type.
3. Response should respond to the to_proto method. The object returned by to_proto should be an instance of the defined response type.

If at any time the implementation encounters an error, the client can be instructed of the error using rpc_failed:

#...
def find
  if request.email.blank?
    rpc_failed 'Unable to find user without an email'
  else
    # query/populate response
  end
end
#...

This means that the client's on_failure callback will be invoked instead of the on_success callback. Read more below on client callbacks. One drawback to the rpc_failed approach is that it does not short-circuit the rest of the method. This means that you must explicitly return from the method if you do not wish the remainder to be executed.

Service Filters provides ActionController-style filter support to service instances, specifically adding before_filter, after_filter, and around_filter.

require 'lib/foo/user.pb'

class Foo::UserService
  before_filter :start_request_timer
  after_filter :end_request_timer
  around_filter :benchmark_request

  # Provide a list of rpc methods to call (or exclude calling) for the given filter(s).
  # The following two filters are essentially equivalent.
  before_filter :verify_user_present, :only => [ :update, :delete ]
  before_filter :verify_user_present, :except => [ :find, :create ]

  # Using if/unless filters options to achieve the same goal, reporting a login after the login has been processed.
  # Note that you can provide a method name or lambda, but you must return a boolean value.
  after_filter :report_login, :only => :login, :if => :user_found?
  after_filter :report_login, :only => :login, :if => lambda { |service| service.response.user_guid.present? }
  after_filter :report_login, :only => :login, :unless => :user_missing?
  after_filter :report_login, :only => :login, :unless => lambda { |service| service.response.user_guid.empty? }

  #... rpc instance methods

  private

  def start_request_timer
    @time_start = Time.now
  end

  def end_request_timer
    @time_end = Time.now
    log_info { ... }
  end

  def benchmark_request
    Benchmark.benchmark do
      yield
    end
  end
end

Around Filters – Inside of an around filter, if you wish to halt request processing and return, simply do not yield the block. Since the filter is implemented as an instance method, you can use rpc_failed or respond_with just like you can in the endpoint methods.

Before Filters – Returning false from a before filter will cancel any other filter calls which would run afterwards, as well as canceling invocation of the service method. Note: You must actually return false, not just a "falsey" value such as nil.

After Filters – There is no request shortcutting since the after filter runs after the request. Duh.

The following options can be applied to any of the filters as the final argument in the filter configuration. (See example above).

:if – The object supplied to :if can either be a symbol/string indicating the instance method to call, or, an object that responds to call. The method or callable should return true or false indicating if the filter should be invoked or not. Akin to the if keyword.

:unless – The opposite of the :if option is :unless. Accepts the same object types. The method or callable should return true or false indicating if the filter should be invoked or not. Akin to the unless keyword.

:only – A string/symbol or Array of strings/symbols values that reference instance methods. The names of these methods should be the rpc method you wish to invoke the filter for. Methods not identified in this list would not have the filter applied.

:except – The opposite of the :only option. A string/symbol or Array of strings/symbols values that reference instance methods. The names of these methods should be the rpc method you wish to skip invokation of the given filter. Methods not identified in this list would have the filter applied.

A service is nothing without being hooked up to a socket. It's the nerdy kid waiting by the telephone for someone to call without knowing that the phone company disconnected their house. Sad and pathetic. So hook up the phone lines!

$ rpc_server -o myserver.com -p 9399 -l ./log/protobuf.log ./config/environment.rb

The previous call will start a Socket server running on the given host and port which will load your application into memory. You certainly don't have to run rails or any other framework, just make sure you have some kind of file that will load your services all into memory. The server doesn't know where you put your code, so tell it.

Be aware that the server needs to be able to translate the socket stream of bytes into an actual protobuf request object. If the definition for that request object aren't known to the server, you're going to have a long day getting this going. It's necessary to store all your definitions and their generated classes in a shared repository (read: gem) so that both client and server have access to the ruby classes in their respective load paths.

Once the server starts, you should see it as a running process with ps. Sending a KILL, QUIT, or TERM signal to the pid will result in shutting the server down gracefully.

$ ps aux | grep rpc_server
1234 ... rpc_server myservice.com:9399

$ kill -QUIT 1234
rpc_server shutdown

A lot of work has gone into making the client calls simple and easy to use yet still powerful. Clients have a DSL that feels very ajaxy.

# require the defs from the shared gem/repo
require 'sharedgem/foo/user.pb'

# Create a request object for the method we are invoking
req = Foo::UserRequest.new(:email => '[email protected]')

# Use the UserService class to generate a client, invoke the rpc method
# while passing the request object.
# We could also simply pass a hash to find.
Foo::UserService.client.find(req) do |c|
  # This block will be executed (registering the callbacks)
  # before the request actualy occurs.
  # the `c` param in this block is the `.client` object
  # that is generated from the call above

  # Register a block for execution when the response
  # is deemed successful from the service. Accepts
  # the unpacked response as its only parameter
  c.on_success do |response|
    response.users.each do |u|
      puts u.inspect
    end
  end

  # Register a block for execution when the response
  # is deemed a failure. This can be either a client-side
  # or server-side failure. The object passed to the
  # block has a `message` and a `code` attribute
  # to aid in logging/diagnosing the failure.
  c.on_failure do |err|
    puts 'It failed: ' + err.message
  end
end

Many different options can be passed to the .client call above (such as :timeout => 600). See the lib/protobuf/rpc/client.rb and lib/protobuf/rpc/service.rb files for more documentation.

The main reason I wrote this gem was to provide a ruby implementation to google's protobuf that worked on the RPC layer with a Java Service layer that was already running protobuf-socket-rpc, the supported socket rpc library for protobuf from Google. Other ruby protobuf implementations I've used did not provide this kind of support.

It must be noted that this gem was started originally as a fork of the ruby-protobuf gem. Its authors and I were unable to reach a communication point to be able to merge all of my RPC updates in with their master. Unfortunately I just simply couldn't use their RPC code and so I forked the code. Myself and others have significantly changed the internals of the gem, including the rpc implementation, the message/field implementation, and the compiler implementation. These changes were made to address glaring performance and quality issues in the code. The code was initially diverged at their 0.4.0 version.

It should also be noted that there are many more features I haven't really shown here, so please let me know if you have any questions on usage or support for various features. Happy protobufing.

-- BJ Neilsen, @localshred