Fetching contributors…
Cannot retrieve contributors at this time
443 lines (304 sloc) 20.1 KB


Finagle is a library for building asynchronous RPC servers and clients in Java, Scala, or any JVM language. Built atop Netty, Finagle provides a rich set of tools that are protocol independent.

Finagle is flexible enough to support a variety of RPC styles, including request-response, streaming, and pipelining (e.g., HTTP pipelining and Redis pipelining). It also makes it easy to work with stateful RPC styles (e.g., those requiring authentication and those that support transactions).

Client Features

  • Connection Pooling
  • Load Balancing
  • Failure Detection
  • Failover/Retry
  • Distributed Tracing (a la Dapper)
  • Service Discovery (e.g., via Zookeeper)
  • Rich Statistics
  • Native OpenSSL bindings

Server Features

  • Backpressure (to defend against abusive clients)
  • Service Registration (e.g., via Zookeeper)
  • Distributed Tracing
  • Native OpenSSL bindings

Supported Protocols

  • HTTP
  • HTTP streaming (Comet)
  • Thrift
  • Memcached/Kestrel
  • More to come!

How do I start?

API Documentation

Links to the Scaladoc is available here


I think the best way to start is to look at examples:

  • Echo - A simple echo client and server using a newline-delimited protocol. Illustrates the basics of asynchronous control-flow.
  • Http - An advanced HTTP client and server that illustrates the use of Filters to compositionally organize your code. Filters are used here to isolate authentication and error handling concerns.
  • Memcached Proxy - A simple proxy supporting the Memcached protocol.
  • Stream - An illustration of Channels, the abstraction for Streaming protocols.
  • Spritzer 2 Kestrel - An illustration of Channels, the abstraction for Streaming protocols. Here the Twitter Firehose is "piped" into a Kestrel message queue, illustrating some of the compositionality of Channels.
  • Stress - A high-throughput HTTP client for driving stressful traffic to an HTTP server. Illustrates more advanced asynchronous control-flow.
  • Thrift - A simple client and server for a Thrift protocol.

But a great way to start might be to read the following conceptual overview:

Conceptual Overview

Here is a simple HTTP server and client. The server returns a simple HTTP 200 response:


val service: Service[HttpRequest, HttpResponse] = new Service[HttpRequest, HttpResponse] {
  def apply(request: HttpRequest) = Future(new DefaultHttpResponse(HTTP_1_1, OK))

val address: SocketAddress = new InetSocketAddress(10000)

val server: Server[HttpRequest, HttpResponse] = ServerBuilder()

The client connects to the server, and issues a simple HTTP GET request:

val client: Service[HttpRequest, HttpResponse] = ClientBuilder()

// Issue a request, get a response:
val request: HttpRequest = new DefaultHttpRequest(HTTP_1_1, GET, "/")
val responseFuture: Future[HttpResponse] = client(request)

// all done!

Note that the variable responseFuture in this example is of type Future[HttpResponse], which represents an asynchronous HTTP response (i.e., a response that will arrive sometime later). With a Future object, you can express your program in either a synchronous or asynchronous style: the program can either 1) block, awaiting the response, or 2) provide a callback to be invoked when a response is available. For example,


 // (1) Synchronously await the response for up to 1 second, then print:

 // (2) Alternatively, when the response arrives, invoke the callback, then print:
 responseFuture onSuccess { response =>

Futures allow the programmer to easily express a number of powerful idioms such as pipelining, scatter-gather, timeouts, and error handling. See the section "Using Futures" for more information.


In Finagle, RPC Servers are built out of Services and Filters. A Service is a simply a function that receives a request and returns a Future of a response. For example, here is a service that increments a number by one.


val plusOneService = new Service[Int, Int] {
  def apply(request: Int) = Future { request + 1 }                                  // (1)

Note that plusOneService acts as if it were asynchronous despite that it is not doing any asynchronous work. It adheres to the Service[Req, Rep] contract by wrapping the synchronously computed response in a "constant" Future of type Future[Int].

More sophisticated Services than plusOneService might make truly asynchronous calls (e.g., by making further RPC calls or by scheduling work in a queue), and apply having a return type of Future[Rep] is general enough to handle most cases.

Once you have defined your Service, it can be bound to a SocketAddress, thus becoming an RPC Server:



However, an RPC Server must speak a specific codec/protocol. One nice way to design an RPC Server is to decouple the protocol handling code from the implementation of the business rules. For example, we can have an adding service exposed via HTTP and Thrift. A Filter provides a easy way to do this.

Here is a Filter that adapts the HttpRequest => HttpResponse protocol to the Int => Int plusOneService:


val httpToIntFilter = new Filter[HttpRequest, HttpResponse, Int, Int] {              // (1)
  def apply(httpRequest: HttpRequest, intService: Service[Int, Int]) = {
    val intRequest = httpRequest.getContent.toString(CharsetUtil.UTF_8).toInt
    intService(intRequest) map { intResponse =>                                      // (2)
      val httpResponse = new DefaultHttpResponse(HTTP_1_1, OK)

val httpPlusOneService: Service[HttpRequest, HttpResponse] =
  httpToIntFilter.andThen(plusOneService)                                            // (3)

This example illustrates three important concepts:

  1. A Filter wraps a Service and (potentially) converts the input and output types of the service to other types. Here Int => Int is mapped to HttpRequest => HttpResponse.
  2. The result of a Future[A] computation can be converted to a Future[B] computation by calling the map function. This function is applied asynchronously and is analogous to the map function on sequences used in many programming languages. See the section "Using Futures" for more information.
  3. A Filter is wrapped around a Service by calling the andThen function. Any number of Filters can be composed using andThen, as we will see shortly.

Let's consider a more involved example. Often it is nice to isolate distinct phases of your application into a pipeline, and Filters provide a great way to accomplish this. In order to prevent Integer Overflow errors, our Service will check that the request is < 2**32 - 1:


val preventOverflowFilter = new SimpleFilter[Int, Int] {                             // (1)
  def apply(request: Int, continue: Service[Int, Int]) =
    if (request < Int.MaxValue)
      Future.exception(new OverflowException)                                        // (2)


  1. A SimpleFilter is a kind of Filter that does not convert the request and response types. It saves a little bit of typing.
  2. An exception can be returned asynchronously by calling Future.exception. See the section "Using Futures" for more information.

Another Filter typical of an RPC Service is authentication and authorization. Our Service wants to ensure that the user is authorized to perform addition:


val ensureAuthorizedFilter = new SimpleFilter[HttpRequest, HttpResponse] {
  def apply(request: HttpRequest, continue: Service[Int, Int]) =
    if (request.getHeader("Authorization") == "Basic ...")
      Future.exception(new UnauthorizedException)

Finally, all of the Filters can be composed with our Service in the following way:

val myService =
  ensureAuthorizedFilter andThen
  httpToIntFilter        andThen
  preventOverflowFilter  andThen

Building a robust RPC client

Finagle makes it easy to build RPC clients with connection pooling, load balancing, logging, and statistics reporting:


val client = ClientBuilder()
    .connectionTimeout(1.second)        // max time to spend establishing a TCP connection.
    .retries(2)                         // (1) per-request retries
    .reportTo(new OstrichStatsReceiver) // export host-level load data to ostrich

This creates a load balanced HTTP client that balances requests among 3 (local) endpoints. The balancing strategy is to pick the endpoint with the least number of outstanding requests (this is similar to "least connections" in other load balancers). The load-balancer deliberately introduces jitter to avoid synchronicity (and thundering herds) in a distributed system. In general a lot of thought has gone into making robust load-balancing and failover.


  1. If retries are specified (using retries(n: Int)), Finagle will retry the request in the event of an error, up to the number of times specified. Finagle does not assume your RPC service is Idempotent. Retries occur only in the event of TCP-related WriteExceptions, where we are certain the RPC has not been transmitted to the remote server.

Once you have constructed a client, a request is issued like this:


val request: HttpRequest = new DefaultHttpRequest(HTTP_1_1, Get, "/")
val futureResponse: Future[HttpResponse] = client(request)


HttpRequest request = new DefaultHttpRequest(HTTP_1_1, Get, "/")
Future<HttpResponse> futureResponse = client.apply(request)


A robust way to use RPC clients is to have an upper-bound on how long to wait for a response to arrive. With Futures, there are two ways to do this: synchronously and asynchrously. 1) The synchronous technique is to block, waiting for a response to arrive, and throw an exception if it does not arrive in time. 2) The asynchronous way is to register a callback to handle the result if it arrives in time, and invoke another callback if it fails.


// (1) synchronous timeouts
try {
  val response = futureResponse(1.second)
} catch {
  case e: TimeoutException => ...

// (2) asynchronous timeouts require an (implicit) Timer object
import com.twitter.finagle.util.Timer._

futureResponse.within(1.second) onSuccess { response =>
  println("yay it worked: " + response)
} onFailure {
  case e: TimeoutException => ...

Using Futures

Finagle uses com.twitter.util.Futures as the unifying abstraction for all asynchronous computation. A Future represents a computation that has not yet completed, and that can succeed or fail. The two most basic ways to use a Future is to 1) wait for the computation to return, or 2) register a callback to be invoked when the computation eventually succeeds or fails.

In the example below, we define a function f that takes an Int and returns a Future[Int]. It errors if given an odd number.

See the Scaladoc


def f(a: Int): Future[Int] =
  if (a % 2 == 0)
    Future.exception(new OddNumberException)

val myFuture: Future[Int] = f(2)

// an alternative way to define the function `f`:

def f(a: Int): Future[Int] = Future {
  if (a % 2 == 0) a
  else throw new OddNumberException

// 1) Wait 1 second the for computation to return
try {
} catch {
  case e: TimeoutException   => ...
  case e: OddNumberException => ...

// 2) Invoke a callback when the computation succeeds or fails
myFuture onSuccess { i =>
} onFailure { e =>
  println("uh oh!")
} ensure {

In addition to waiting for results to return, Futures can be transformed in interesting ways. For instance, it is possible to convert a Future[String] to a Future[Int] by using map:

val stringFuture: Future[String] = Future("1")
val intFuture: Future[Int] = stringFuture map (_.toInt)

Similar to map, there is flatMap. This allows you to easily "pipeline" a sequence of Futures:

val authenticateUser: Future[User] = User.authenticate(email, password)
val lookupTweets: Future[Seq[Tweet]] = authenticateUser flatMap { user =>

In this example, Tweet.findAllByUser(user) is a function of type User => Future[Seq[Tweet]].

As a final example of Futures let's consider the problem of scatter/gather patterns. The challenge is to issue a series of requests in parallel and wait for all of them to arrive. Suppose for example, we have a sequence of Futures. To wait for all to return, we do like so:

val myFutures: Seq[Future[Int]] = ...

val waitTillAllComplete: Future[Seq[Int]] = Future.collect(myFutures)

A more complex variation of scatter/gather is to perform a sequence of asynchronous operations and harvest only those that return within a certain time -- using a default value for those that don't return. A concrete example might be to issue a set of parallel requests to N partitions of a search index; those that don't return in time we consider to have returned the empty set.

import com.twitter.finagle.util.Timer._

val results: Seq[Future[Result]] = { partition =>
  partition.get(query).within(1.second) handle {
    case _: TimeoutException => EmptyResult
val allResults: Future[Seq[Result]] = Future.collect(timedResults)

// Process the results asynchronously.
// Note: this takes no longer than 1 second.
allResults onSuccess { results =>

Streaming Protocols

Finagle makes streaming and pubsub-like RPCs easy. Streams rely on a generalization of Futures called Channels. Channels represent a stream of events that can be listened to. To publish and subscribe to a Channel, do the following:

See the Scaladoc

// a ChannelSource is a readable-writable stream of messages, whereas a Channel is only readable.
val source = new ChannelSource[Int]
val sink: Channel = source

// start listening for messages:
val observer1 = sink respond { message =>
  Future { println("1: " + message) }                              // (1)
val observer2 = sink respond { message =>                          // (2)
  Future { println("2: " + message) }

// send some messages on the channel
// etc.

// stop listening for messages:


  1. Subscribers must return a Future indicating when processing of the received message is complete. This allows consumers to exhibit backpressure to producers.
  2. Just as with Futures there can be any number of responders to a Channel.

Channels have some of the usual sequence operations such as map and filter:

val channel = new ChannelSource[Int]
val evenChannel = channel filter (_ % 2 == 0)
val stringChannel = channel map (_.toString)

Most of the tricky issues concerning Channels involve data-loss and backpressure. In order to ensure that no data is lost, a typical pattern is not to broadcast on a Channel until there is at least one subscriber. This is easily accomplished:

channel.responds.first { _ =>                                       // (1) (2)
  // the first responder has arrived, start sending!
  // ...


  1. channel.responds is a Channel itself. In other words, every Channel has a sub-channel indicating when subscribers subscribe (and unsubscribe!).
  2. channel.first is a Future indicating when the first message arrives on that Channel.

However, a more robust producer stops producing when all consumers dispose(). This is easily accomplished.

channel.numObservers.respond { i =>
  i match {
    case 0 => start()
    case 1 => stop()
    case _ => // continue

Backpressure (that is, slowing down production if consumers get backed up) is also easily accomplished. The channel.send method returns a Seq[Future[Unit]] -- a sequence of Futures indicating when all consumers have processed the message. You can use Future.join to slow down production in the following way:

Future.join(channel.send(1)) onSuccess { _ =>
  Future.join(channel.send(2)) onSuccess { ... }

You can use recursion to send 10 messages, throttled by consumer performance. Additionally, Future provides a number of control flow constructs that are useful:

val i = new AtomicInteger(0)
Future.times(10) { channel.send(i.getAndIncrement()) } // (1)


  1. Future has methods like times(), parallel(), and whileDo() that implement advanced control-flow patterns.

With Channels, building a streaming RPC service is straightforward. You will need a codec that supports streaming (such as HTTP with chunked encoding). Then, build a Service that returns a Channel:

val channel = new ChannelSource[ChannelBuffer]
val myService = new Service[HttpRequest, Channel[ChannelBuffer]] {
  def apply(request: HttpRequest) = Future.value(channel)

Some incomplete API documentation is available: See scaladoc.


finagle-serversets is an implementation of the Finagle Cluster interface using ServerSets.

Here's how to use it:

Instantiate a ServerSet and a Cluster

val serverSet = new ServerSetImpl(zookeeperClient, "/twitter/services/silly")
val cluster = new ZookeeperServerSetCluster(serverSet)

If you're a Server, join the Cluster:

val serviceAddress = new InetSocketAddress(...)
val server = ServerBuilder()


If you're a Client,

val client = ClientBuilder()
  .cluster(cluster) // check this out!!
  .codec(new StringCodec)

That's it!


1.2.3 (2011-03-18)

  • A faster memcache codec (joint work with Arya from profiling rooster usage)
  • Fix a bug in memcache decoding of big responses (due to Arya)
  • More exported pool & connection statistics
  • Fix a bug in the CachingPool wherein connection accounting broke
  • Speed up leastqueued loadbalancer strategy
  • finagle-serversets is now included in the opensource distribution, instructions at