Travis Brown


Posted February 6, 2015

Finagle services made simple with Serial

Most internal services at Twitter speak the Thrift protocol, which provides many benefits—once you’ve defined your data types and service interfaces, for example, it’s possible to create bindings for a wide range of programming languages, and Twitter’s Scrooge in particular makes it easy to create high-performance Finagle servers and clients for your Thrift interfaces.

In some cases, though, it would be more convenient not to have to worry about things like interface description files, the build system plugins necessary to generate bindings from them, etc. In particular, being able to define Finagle services that take arbitrary types as inputs and outputs in a Scala REPL would make writing tutorials and quickstart projects much more straightforward, and would enable easier experimentation with other parts of the Finagle API.

With this in mind, Vladimir Kostyukov and I decided to spend part of Twitter’s January 2015 Hack Week putting together a library that would allow users to create Finagle clients and servers that use Scala (or Java) libraries for serialization instead of IDL-based systems like Thrift. The result was Finagle Serial, a new project in the Finagle organization. Serial is built on Mux, a generic session-layer RPC protocol that supports multiplexing, and currently provides support for using Scodec for object serialization. Other serialization libraries may be supported in the future, but we found that Scodec worked well for the initial proof-of-concept implementation.

Starting a server

With Serial, you don’t need to describe your service interfaces using an external IDL—instead you simply provide codecs for your input and output types. If you clone the Serial repository and install SBT, you can launch a REPL with the following command:

sbt "project scodec" console

And then define and start a Finagle server with just a few lines of code:

import com.twitter.finagle.Service
import com.twitter.util.Future
import io.github.finagle.serial.scodec.ScodecSerial
import scodec.Codec
import scodec.codecs._

val protocol = ScodecSerial(int32, double)

val reciprocalServer =
  protocol.serveFunction(new InetSocketAddress(8123))(1.0 / _)

And that’s all—we now have a Finagle server running on port 8123 that will return the floating-point reciprocal of any integer we pass to it.

Note that all you need to be able to serve a Finagle Service[Req, Rep] (or an ordinary function from Req to Rep) is a pair of codecs for the Req and Rep types.

In this example, int32 and double are the Scodec codecs that we’ve decided to use for the Int and Double input and output types. These are fairly simple, but Scodec makes it possible to define more complex codecs with very little boilerplate (see for example the ones for the User and Greeting case classes below).

We can try out our server by defining and calling a client:

val reciprocal = protocol.newService("localhost:8123")

reciprocal(1001).onSuccess(result => println(f"$result%.12f"))

Which should print 0.000999000999, exactly as we’d expect.

Handling exceptions

Now for a slightly more interesting service:

val protocol = ScodecSerial(variableSizeBits(uint8, utf8), int32)

val intParsingServer =
  protocol.serveFunction(new InetSocketAddress(8124))(_.toInt)

val intParser = protocol.newService("localhost:8124")


This should print 2015. But the function that we’re serving is of course not very well-behaved—it’ll throw an exception on strings that can’t be parsed into integers. So what happens if we call our client with invalid input?

intParser("I'm not a number").onFailure(result => println(s"Too bad: $result"))

You should see the following:

Too bad: java.lang.NumberFormatException: For input string: "I'm not a number"

Note that we’re getting a NumberFormatException, not a ServerApplicationError or some other kind of Finagle-explicit exception type. This is because the default Scodec protocol implementation knows how to serialize a few common exception types so that they can be sent over the wire.

Using your own input, output, and error types

Using types you’ve defined as your service’s input and output is as easy as providing codecs for those types, and you also can extend the functionality in the default implementation with information about how to serialize your own exception types. Custom error types requires a little more work than custom input and output types, but it’s still fairly concise (and is likely to improve in future iterations of the API).

Suppose for example that we’ve got a simple user greeting service with a couple of case classes and a custom error type:

case class User(name: String)
case class Greeting(u: User) {
  override def toString = s"Hello, ${}!"
case class GreetingError(message: String) extends Exception(message)

object GreetUser extends Service[User, Greeting] {
  def apply(u: User) = u match {
    case User("Mary") => Future.value(Greeting(u))
    case User(other) => Future.exception(GreetingError(s"Unknown user: $other"))

The easy part is defining our codecs for our input and output case classes:

val userCodec: Codec[User] = variableSizeBits(uint24, utf8).as[User]
val greetingCodec: Codec[Greeting] =[Greeting]

We have to create a new ScodecSerial instance to add support for our custom error type, but then we’re done:

import io.github.finagle.serial.scodec.ApplicationErrorCodec

object UserSerial extends ScodecSerial {
  override lazy val applicationErrorCodec =
    ApplicationErrorCodec.basic.add(GreetingError(_)) {
      case GreetingError(message) => message

val protocol = UserSerial(userCodec, greetingCodec)

val server = protocol.serve(new InetSocketAddress(8125), GreetUser)
val client = protocol.newService("localhost:8125")

Now Mary will get a greeting and everyone else will get a GreetingError:

scala> client(User("Mary")).onSuccess(println)
res33: com.twitter.util.Future[Greeting] = ...
Hello, Mary!

scala> client(User("Travis")).onFailure(println)
res34: com.twitter.util.Future[Greeting] = ...
GreetingError: Unknown user: Travis

Any exceptions that the implementation doesn’t know how to serialize will result in the exception’s message being wrapped in a Serial ApplicationError.

Using Serial

Serial is a very young project, but Finagle and Mux are a good foundation to build on, and we’d love to have other people give it a try.

The API documentation is available on the project’s GitHub Pages site, and the project’s README presents some initial benchmarks and goes into more detail on topics like exception serialization. If you’re interested in writing an implementation for another serialization library, we’ve put together some ScalaCheck-based integration testing tools that can help you verify that it’s working correctly.

Please let us know if you have any questions, either by filing a GitHub issue or getting in touch through @finagle or the Finaglers mailing list!