Futures and Promises

By: Philipp Haller, Aleksandar Prokopec, Heather Miller, Viktor Klang, Roland Kuhn, and Vojin Jovanovic

Introduction

Futures provide a nice way to reason about performing many operations in parallel– in an efficient and non-blocking way. The idea is simple, a Future is a sort of a placeholder object that you can create for a result that does not yet exist. Generally, the result of the Future is computed concurrently and can be later collected. Composing concurrent tasks in this way tends to result in faster, asynchronous, non-blocking parallel code.

By default, futures and promises are non-blocking, making use of callbacks instead of typical blocking operations. To simplify the use of callbacks both syntactically and conceptually, Scala provides combinators such as flatMap, foreach, and filter used to compose futures in a non-blocking way. Blocking is still possible - for cases where it is absolutely necessary, futures can be blocked on (although this is discouraged).

Futures

A Future is an object holding a value which may become available at some point. This value is usually the result of some other computation:

1. If the computation has not yet completed, we say that the Future is not completed.
2. If the computation has completed with a value or with an exception, we say that the Future iscompleted.

Completion can take one of two forms:

1. When a Future is completed with a value, we say that the future was successfully completed with that value.
2. When a Future is completed with an exception thrown by the computation, we say that theFuture was failed with that exception.

A Future has an important property that it may only be assigned once. Once a Future object is given a value or an exception, it becomes in effect immutable– it can never be overwritten.

The simplest way to create a future object is to invoke the future method which starts an asynchronous computation and returns a future holding the result of that computation. The result becomes available once the future completes.

Note that Future[T] is a type which denotes future objects, whereas future is a method which creates and schedules an asynchronous computation, and then returns a future object which will be completed with the result of that computation.

This is best shown through an example.

Let’s assume that we want to use a hypothetical API of some popular social network to obtain a list of friends for a given user. We will open a new session and then send a request to obtain a list of friends of a particular user:

import scala.concurrent._
import ExecutionContext.Implicits.global
val session = socialNetwork.createSessionFor("user", credentials)
val f: Future[List[Friend]] = future {
  session.getFriends()
}

Above, we first import the contents of the scala.concurrent package to make the type Futureand the construct future visible. We will explain the second import shortly.

We then initialize a session variable which we will use to send requests to the server, using a hypothetical createSessionFor method. To obtain the list of friends of a user, a request has to be sent over a network, which can take a long time. This is illustrated with the call to the methodgetFriends that returns List[Friend]. To better utilize the CPU until the response arrives, we should not block the rest of the program– this computation should be scheduled asynchronously. The future method does exactly that– it performs the specified computation block concurrently, in this case sending a request to the server and waiting for a response.

The list of friends becomes available in the future f once the server responds.

An unsuccessful attempt may result in an exception. In the following example, the session value is incorrectly initialized, so the computation in the future block will throw aNullPointerException. This future f is then failed with this exception instead of being completed successfully:

val session = null
val f: Future[List[Friend]] = future {
  session.getFriends
}

The line import ExecutionContext.Implicits.global above imports the default global execution context. Execution contexts execute tasks submitted to them, and you can think of execution contexts as thread pools. They are essential for the future method because they handle how and when the asynchronous computation is executed. You can define your own execution contexts and use them with future, but for now it is sufficient to know that you can import the default execution context as shown above.

Our example was based on a hypothetical social network API where the computation consists of sending a network request and waiting for a response. It is fair to offer an example involving an asynchronous computation which you can try out of the box. Assume you have a text file and you want to find the position of the first occurrence of a particular keyword. This computation may involve blocking while the file contents are being retrieved from the disk, so it makes sense to perform it concurrently with the rest of the computation.

val firstOccurrence: Future[Int] = future {
  val source = scala.io.Source.fromFile("myText.txt")
  source.toSeq.indexOfSlice("myKeyword")
}

Callbacks

We now know how to start an asynchronous computation to create a new future value, but we have not shown how to use the result once it becomes available, so that we can do something useful with it. We are often interested in the result of the computation, not just its side-effects.

In many future implementations, once the client of the future becomes interested in its result, it has to block its own computation and wait until the future is completed– only then can it use the value of the future to continue its own computation. Although this is allowed by the Scala Future API as we will show later, from a performance point of view a better way to do it is in a completely non-blocking way, by registering a callback on the future. This callback is called asynchronously once the future is completed. If the future has already been completed when registering the callback, then the callback may either be executed asynchronously, or sequentially on the same thread.

The most general form of registering a callback is by using the onComplete method, which takes a callback function of type Try[T] => U. The callback is applied to the value of type Success[T] if the future completes successfully, or to a value of type Failure[T] otherwise.

The Try[T] is similar to Option[T] or Either[T, S], in that it is a monad potentially holding a value of some type. However, it has been specifically designed to either hold a value or some throwable object. Where an Option[T] could either be a value (i.e. Some[T]) or no value at all (i.e.None), Try[T] is a Success[T] when it holds a value and otherwise Failure[T], which holds an exception. Failure[T] holds more information that just a plain None by saying why the value is not there. In the same time, you can think of Try[T] as a special version ofEither[Throwable, T], specialized for the case when the left value is a Throwable.

Coming back to our social network example, let’s assume we want to fetch a list of our own recent posts and render them to the screen. We do so by calling a method getRecentPosts which returns a List[String]– a list of recent textual posts:

import scala.util.{Success, Failure}
val f: Future[List[String]] = future {
  session.getRecentPosts
}
f onComplete {
  case Success(posts) => for (post <- posts) println(post)
  case Failure(t) => println("An error has occured: " + t.getMessage)
}

The onComplete method is general in the sense that it allows the client to handle the result of both failed and successful future computations. To handle only successful results, the onSuccesscallback is used (which takes a partial function):

val f: Future[List[String]] = future {
  session.getRecentPosts
}
f onSuccess {
  case posts => for (post <- posts) println(post)
}

To handle failed results, the onFailure callback is used:

val f: Future[List[String]] = future {
  session.getRecentPosts
}
f onFailure {
  case t => println("An error has occured: " + t.getMessage)
}
f onSuccess {
  case posts => for (post <- posts) println(post)
}

The onFailure callback is only executed if the future fails, that is, if it contains an exception.

Since partial functions have the isDefinedAt method, the onFailure method only triggers the callback if it is defined for a particular Throwable. In the following example the registeredonFailure callback is never triggered:

val f = future {
  2 / 0
}
f onFailure {
  case npe: NullPointerException =>
    println("I‘d be amazed if this printed out.")
}

Coming back to the previous example with searching for the first occurrence of a keyword, you might want to print the position of the keyword to the screen:

val firstOccurrence: Future[Int] = future {
  val source = scala.io.Source.fromFile("myText.txt")
  source.toSeq.indexOfSlice("myKeyword")
}
firstOccurrence onSuccess {
  case idx => println("The keyword first appears at position: " + idx)
}
firstOccurrence onFailure {
  case t => println("Could not process file: " + t.getMessage)
}

The onComplete, onSuccess, and onFailure methods have result type Unit, which means invocations of these methods cannot be chained. Note that this design is intentional, to avoid suggesting that chained invocations may imply an ordering on the execution of the registered callbacks (callbacks registered on the same future are unordered).

That said, we should now comment on when exactly the callback gets called. Since it requires the value in the future to be available, it can only be called after the future is completed. However, there is no guarantee it will be called by the thread that completed the future or the thread which created the callback. Instead, the callback is executed by some thread, at some time after the future object is completed. We say that the callback is executed eventually.

Furthermore, the order in which the callbacks are executed is not predefined, even between different runs of the same application. In fact, the callbacks may not be called sequentially one after the other, but may concurrently execute at the same time. This means that in the following example the variable totalA may not be set to the correct number of lower case and upper case a characters from the computed text.

@volatile var totalA = 0
val text = future {
  "na" * 16 + "BATMAN!!!"
}
text onSuccess {
  case txt => totalA += txt.count(_ == ‘a‘)
}
text onSuccess {
  case txt => totalA += txt.count(_ == ‘A‘)
}

Above, the two callbacks may execute one after the other, in which case the variable totalA holds the expected value 18. However, they could also execute concurrently, so totalA could end up being either 16 or 2, since += is not an atomic operation (i.e. it consists of a read and a write step which may interleave arbitrarily with other reads and writes).

For the sake of completeness the semantics of callbacks are listed here:

1. Registering an onComplete callback on the future ensures that the corresponding closure is invoked after the future is completed, eventually.

2. Registering an onSuccess or onFailure callback has the same semantics as onComplete, with the difference that the closure is only called if the future is completed successfully or fails, respectively.

3. Registering a callback on the future which is already completed will result in the callback being executed eventually (as implied by 1).

4. In the event that multiple callbacks are registered on the future, the order in which they are executed is not defined. In fact, the callbacks may be executed concurrently with one another. However, a particular ExecutionContext implementation may result in a well-defined order.

5. In the event that some of the callbacks throw an exception, the other callbacks are executed regardless.

6. In the event that some of the callbacks never complete (e.g. the callback contains an infinite loop), the other callbacks may not be executed at all. In these cases, a potentially blocking callback must use the blocking construct (see below).

7. Once executed, the callbacks are removed from the future object, thus being eligible for GC.

Functional Composition and For-Comprehensions

The callback mechanism we have shown is sufficient to chain future results with subsequent computations. However, it is sometimes inconvenient and results in bulky code. We show this with an example. Assume we have an API for interfacing with a currency trading service. Suppose we want to buy US dollars, but only when it’s profitable. We first show how this could be done using callbacks:

val rateQuote = future {
  connection.getCurrentValue(USD)
}
rateQuote onSuccess { case quote =>
  val purchase = future {
    if (isProfitable(quote)) connection.buy(amount, quote)
    else throw new Exception("not profitable")
  }
  
  purchase onSuccess {
    case _ => println("Purchased " + amount + " USD")
  }
}

We start by creating a future rateQuote which gets the current exchange rate. After this value is obtained from the server and the future successfully completed, the computation proceeds in theonSuccess callback and we are ready to decide whether to buy or not. We therefore create another future purchase which makes a decision to buy only if it’s profitable to do so, and then sends a request. Finally, once the purchase is completed, we print a notification message to the standard output.

This works, but is inconvenient for two reasons. First, we have to use onSuccess, and we have to nest the second purchase future within it. Imagine that after the purchase is completed we want to sell some other currency. We would have to repeat this pattern within the onSuccess callback, making the code overly indented, bulky and hard to reason about.

Second, the purchase future is not in the scope with the rest of the code– it can only be acted upon from within the onSuccess callback. This means that other parts of the application do not see the purchase future and cannot register another onSuccess callback to it, for example, to sell some other currency.

For these two reasons, futures provide combinators which allow a more straightforward composition. One of the basic combinators is map, which, given a future and a mapping function for the value of the future, produces a new future that is completed with the mapped value once the original future is successfully completed. You can reason about mapping futures in the same way you reason about mapping collections.

Let’s rewrite the previous example using the map combinator:

val rateQuote = future {
  connection.getCurrentValue(USD)
}
val purchase = rateQuote map { quote => 
  if (isProfitable(quote)) connection.buy(amount, quote)
  else throw new Exception("not profitable")
}
purchase onSuccess {
  case _ => println("Purchased " + amount + " USD")
}

By using map on rateQuote we have eliminated one onSuccess callback and, more importantly, the nesting. If we now decide to sell some other currency, it suffices to use map on purchaseagain.

But what happens if isProfitable returns false, hence causing an exception to be thrown? In that case purchase is failed with that exception. Furthermore, imagine that the connection was broken and that getCurrentValue threw an exception, failing rateQuote. In that case we’d have no value to map, so the purchase would automatically be failed with the same exception asrateQuote.

In conclusion, if the original future is completed successfully then the returned future is completed with a mapped value from the original future. If the mapping function throws an exception the future is completed with that exception. If the original future fails with an exception then the returned future also contains the same exception. This exception propagating semantics is present in the rest of the combinators, as well.

One of the design goals for futures was to enable their use in for-comprehensions. For this reason, futures also have the flatMap, filter and foreach combinators. The flatMap method takes a function that maps the value to a new future g, and then returns a future which is completed onceg is completed.

Lets assume that we want to exchange US dollars for Swiss francs (CHF). We have to fetch quotes for both currencies, and then decide on buying based on both quotes. Here is an example offlatMap and withFilter usage within for-comprehensions:

val usdQuote = future { connection.getCurrentValue(USD) }
val chfQuote = future { connection.getCurrentValue(CHF) }
val purchase = for {
  usd <- usdQuote
  chf <- chfQuote
  if isProfitable(usd, chf)
} yield connection.buy(amount, chf)
purchase onSuccess {
  case _ => println("Purchased " + amount + " CHF")
}

The purchase future is completed only once both usdQuote and chfQuote are completed– it depends on the values of both these futures so its own computation cannot begin earlier.

The for-comprehension above is translated into:

val purchase = usdQuote flatMap {
  usd =>
  chfQuote
    .withFilter(chf => isProfitable(usd, chf))
    .map(chf => connection.buy(amount, chf))
}

which is a bit harder to grasp than the for-comprehension, but we analyze it to better understand theflatMap operation. The flatMap operation maps its own value into some other future. Once this different future is completed, the resulting future is completed with its value. In our example,flatMap uses the value of the usdQuote future to map the value of the chfQuote into a third future which sends a request to buy a certain amount of Swiss francs. The resulting futurepurchase is completed only once this third future returned from map completes.

This can be mind-boggling, but fortunately the flatMap operation is seldom used outside for-comprehensions, which are easier to use and understand.

The filter combinator creates a new future which contains the value of the original future only if it satisfies some predicate. Otherwise, the new future is failed with a NoSuchElementException. For futures calling filter has exactly the same effect as does calling withFilter.

The relationship between the collect and filter combinator is similar to the relationship of these methods in the collections API.

It is important to note that calling the foreach combinator does not block to traverse the value once it becomes available. Instead, the function for the foreach gets asynchronously executed only if the future is completed successfully. This means that the foreach has exactly the same semantics as the onSuccess callback.

Since the Future trait can conceptually contain two types of values (computation results and exceptions), there exists a need for combinators which handle exceptions.

Let’s assume that based on the rateQuote we decide to buy a certain amount. Theconnection.buy method takes an amount to buy and the expected quote. It returns the amount bought. If the quote has changed in the meanwhile, it will throw a QuoteChangedException and it will not buy anything. If we want our future to contain 0 instead of the exception, we use therecover combinator:

val purchase: Future[Int] = rateQuote map {
  quote => connection.buy(amount, quote)
} recover {
  case QuoteChangedException() => 0
}

The recover combinator creates a new future which holds the same result as the original future if it completed successfully. If it did not then the partial function argument is applied to theThrowable which failed the original future. If it maps the Throwable to some value, then the new future is successfully completed with that value. If the partial function is not defined on thatThrowable, then the resulting future is failed with the same Throwable.

The recoverWith combinator creates a new future which holds the same result as the original future if it completed successfully. Otherwise, the partial function is applied to the Throwablewhich failed the original future. If it maps the Throwable to some future, then this future is completed with the result of that future. Its relation to recover is similar to that of flatMap tomap.

Combinator fallbackTo creates a new future which holds the result of this future if it was completed successfully, or otherwise the successful result of the argument future. In the event that both this future and the argument future fail, the new future is completed with the exception from this future, as in the following example which tries to print US dollar value, but prints the Swiss franc value in the case it fails to obtain the dollar value:

val usdQuote = future {
  connection.getCurrentValue(USD)
} map {
  usd => "Value: " + usd + "$"
}
val chfQuote = future {
  connection.getCurrentValue(CHF)
} map {
  chf => "Value: " + chf + "CHF"
}
val anyQuote = usdQuote fallbackTo chfQuote
anyQuote onSuccess { println(_) }

The andThen combinator is used purely for side-effecting purposes. It returns a new future with exactly the same result as the current future, regardless of whether the current future failed or not. Once the current future is completed with the result, the closure corresponding to the andThen is invoked and then the new future is completed with the same result as this future. This ensures that multiple andThen calls are ordered, as in the following example which stores the recent posts from a social network to a mutable set and then renders all the posts to the screen:

val allposts = mutable.Set[String]()
future {
  session.getRecentPosts
} andThen {
  posts => allposts ++= posts
} andThen {
  posts =>
  clearAll()
  for (post <- allposts) render(post)
}

In summary, the combinators on futures are purely functional. Every combinator returns a new future which is related to the future it was derived from.

Projections

To enable for-comprehensions on a result returned as an exception, futures also have projections. If the original future fails, the failed projection returns a future containing a value of typeThrowable. If the original future succeeds, the failed projection fails with aNoSuchElementException. The following is an example which prints the exception to the screen:

val f = future {
  2 / 0
}
for (exc <- f.failed) println(exc)

The following example does not print anything to the screen:

val f = future {
  4 / 2
}
for (exc <- f.failed) println(exc)

Extending Futures

Support for extending the Futures API with additional utility methods is planned. This will allow external frameworks to provide more specialized utilities.

Blocking

As mentioned earlier, blocking on a future is strongly discouraged for the sake of performance and for the prevention of deadlocks. Callbacks and combinators on futures are a preferred way to use their results. However, blocking may be necessary in certain situations and is supported by the Futures and Promises API.

In the currency trading example above, one place to block is at the end of the application to make sure that all of the futures have been completed. Here is an example of how to block on the result of a future:

import scala.concurrent._
import scala.concurrent.duration._
def main(args: Array[String]) {
  val rateQuote = future {
    connection.getCurrentValue(USD)
  }
  
  val purchase = rateQuote map { quote =>
    if (isProfitable(quote)) connection.buy(amount, quote)
    else throw new Exception("not profitable")
  }
  
  Await.result(purchase, 0 nanos)
}

In the case that the future fails, the caller is forwarded the exception that the future is failed with. This includes the failed projection– blocking on it results in a NoSuchElementException being thrown if the original future is completed successfully.

Alternatively, calling Await.ready waits until the future becomes completed, but does not retrieve its result. In the same way, calling that method will not throw an exception if the future is failed.

The Future trait implements the Awaitable trait with methods method ready() and result(). These methods cannot be called directly by the clients– they can only be called by the execution context.

To allow clients to call 3rd party code which is potentially blocking and avoid implementing theAwaitable trait, the same blocking primitive can also be used in the following form:

blocking {
  potentiallyBlockingCall()
}

The blocking code may also throw an exception. In this case, the exception is forwarded to the caller.

Exceptions

When asynchronous computations throw unhandled exceptions, futures associated with those computations fail. Failed futures store an instance of Throwable instead of the result value.Futures provide the onFailure callback method, which accepts a PartialFunction to be applied to a Throwable. The following special exceptions are treated differently:

1. scala.runtime.NonLocalReturnControl[_] – this exception holds a value associated with the return. Typically, return constructs in method bodies are translated to throws with this exception. Instead of keeping this exception, the associated value is stored into the future or a promise.

2. ExecutionException - stored when the computation fails due to an unhandledInterruptedException, Error or a scala.util.control.ControlThrowable. In this case the ExecutionException has the unhandled exception as its cause. These exceptions are rethrown in the thread executing the failed asynchronous computation. The rationale behind this is to prevent propagation of critical and control-flow related exceptions normally not handled by the client code and at the same time inform the client in which future the computation failed.

See NonFatal for a more precise semantics description.

Promises

So far we have only considered Future objects created by asynchronous computations started using the future method. However, futures can also be created using promises.

While futures are defined as a type of read-only placeholder object created for a result which doesn’t yet exist, a promise can be thought of as a writable, single-assignment container, which completes a future. That is, a promise can be used to successfully complete a future with a value (by “completing” the promise) using the success method. Conversely, a promise can also be used to complete a future with an exception, by failing the promise, using the failure method.

A promise p completes the future returned by p.future. This future is specific to the promise p. Depending on the implementation, it may be the case that p.future eq p.

Consider the following producer-consumer example, in which one computation produces a value and hands it off to another computation which consumes that value. This passing of the value is done using a promise.

import scala.concurrent.{ future, promise }
import scala.concurrent.ExecutionContext.Implicits.global
val p = promise[T]
val f = p.future
val producer = future {
  val r = produceSomething()
  p success r
  continueDoingSomethingUnrelated()
}
val consumer = future {
  startDoingSomething()
  f onSuccess {
    case r => doSomethingWithResult()
  }
}

Here, we create a promise and use its future method to obtain the Future that it completes. Then, we begin two asynchronous computations. The first does some computation, resulting in a value r, which is then used to complete the future f, by fulfilling the promise p. The second does some computation, and then reads the result r of the completed future f. Note that the consumercan obtain the result before the producer task is finished executing thecontinueDoingSomethingUnrelated() method.

As mentioned before, promises have single-assignment semantics. As such, they can be completed only once. Calling success on a promise that has already been completed (or failed) will throw an IllegalStateException.

The following example shows how to fail a promise.

val p = promise[T]
val f = p.future
val producer = future {
  val r = someComputation
  if (isInvalid(r))
    p failure (new IllegalStateException)
  else {
    val q = doSomeMoreComputation(r)
    p success q
  }
}

Here, the producer computes an intermediate result r, and checks whether it’s valid. In the case that it’s invalid, it fails the promise by completing the promise p with an exception. In this case, the associated future f is failed. Otherwise, the producer continues its computation, and finally completes the future f with a valid result, by completing promise p.

Promises can also be completed with a complete method which takes a potential value Try[T]– either a failed result of type Failure[Throwable] or a successful result of type Success[T].

Analogous to success, calling failure and complete on a promise that has already been completed will throw an IllegalStateException.

One nice property of programs written using promises with operations described so far and futures which are composed through monadic operations without side-effects is that these programs are deterministic. Deterministic here means that, given that no exception is thrown in the program, the result of the program (values observed in the futures) will always be the same, regardless of the execution schedule of the parallel program.

In some cases the client may want to complete the promise only if it has not been completed yet (e.g., there are several HTTP requests being executed from several different futures and the client is interested only in the first HTTP response - corresponding to the first future to complete the promise). For these reasons methods tryComplete, trySuccess and tryFailure exist on future. The client should be aware that using these methods results in programs which are not deterministic, but depend on the execution schedule.

The method completeWith completes the promise with another future. After the future is completed, the promise gets completed with the result of that future as well. The following program prints 1:

val f = future { 1 }
val p = promise[Int]
p completeWith f
p.future onSuccess {
  case x => println(x)
}

When failing a promise with an exception, three subtypes of Throwables are handled specially. If the Throwable used to break the promise is a scala.runtime.NonLocalReturnControl, then the promise is completed with the corresponding value. If the Throwable used to break the promise is an instance of Error, InterruptedException, orscala.util.control.ControlThrowable, the Throwable is wrapped as the cause of a newExecutionException which, in turn, is failing the promise.

Using promises, the onComplete method of the futures and the future construct you can implement any of the functional composition combinators described earlier. Let’s assume you want to implement a new combinator first which takes two futures f and g and produces a third future which is completed by either f or g (whichever comes first), but only given that it is successful.

Here is an example of how to do it:

def first[T](f: Future[T], g: Future[T]): Future[T] = {
  val p = promise[T]
  f onSuccess {
    case x => p.trySuccess(x)
  }
  g onSuccess {
    case x => p.trySuccess(x)
  }
  p.future
}

Note that in this implementation, if neither f nor g succeeds, then first(f, g) never completes (either with a value or with an exception).

Utilities

To simplify handling of time in concurrent applications scala.concurrent introduces a Durationabstraction. Duration is not supposed to be yet another general time abstraction. It is meant to be used with concurrency libraries and resides in scala.concurrent package.

Duration is the base class representing length of time. It can be either finite or infinite. Finite duration is represented with FiniteDuration class which is constructed from Long length andjava.util.concurrent.TimeUnit. Infinite durations, also extended from Duration, exist in only two instances , Duration.Inf and Duration.MinusInf. Library also provides several Durationsubclasses for implicit conversion purposes and those should not be used.

Abstract Duration contains methods that allow :

1. Conversion to different time units (toNanos, toMicros, toMillis, toSeconds, toMinutes,toHours, toDays and toUnit(unit: TimeUnit)).
2. Comparison of durations (<, <=, > and >=).
3. Arithmetic operations (+, -, *, / and unary_-).
4. Minimum and maximum between this duration and the one supplied in the argument (min,max).
5. Check if the duration is finite (isFinite).

Duration can be instantiated in the following ways:

1. Implicitly from types Int and Long. For example val d = 100 millis.
2. By passing a Long length and a java.util.concurrent.TimeUnit. For example val d = Duration(100, MILLISECONDS).
3. By parsing a string that represent a time period. For example val d = Duration("1.2 μs").

Duration also provides unapply methods so it can be used in pattern matching constructs. Examples:

import scala.concurrent.duration._
import java.util.concurrent.TimeUnit._
 
// instantiation
val d1 = Duration(100, MILLISECONDS) // from Long and TimeUnit
val d2 = Duration(100, "millis") // from Long and String
val d3 = 100 millis // implicitly from Long, Int or Double
val d4 = Duration("1.2 μs") // from String
// pattern matching
val Duration(length, unit) = 5 millis
Pasted from: <http://docs.scala-lang.org/overviews/core/futures.html>