This is an example Play application that shows how to use Play's Websocket API in Scala, by showing a series of stock tickers updated using WebSocket.
The Websocket API is built on Akka Streams, and so is async, non-blocking, and backpressure aware. Using Akka Streams also means that interacting with Akka Actors is simple.
This application uses a WebSocket to push data to the browser in real-time. To create a WebSocket connection in Play, first a route must be defined in the routes file. Here is the route which will be used to setup the WebSocket connection:
GET /ws controllers.HomeController.ws
The ws
method in the HomeController.java controller handles the request and does the protocol upgrade to the WebSocket connection. The UserActor
stores the input and output streams to the WebSocket connection, and can manipulate the streams in response to messages.
Once the UserActor
is created, the default stocks (defined in application.conf
) are added to the user's list of watched stocks. The flow of stock quotes is managed using MergeHub and BroadcastHub as a publish/subscribe method to dynamically add and remove streams to the Websocket. The StockHistory
and StockQuote
presentation objects are converted using Play-JSON using the implicit Reads
and Writes
defined on the companion objects.
On the client-side, a Reactive UI updates the stock charts every time a message is received. The index.scala.html file produces the web page at http://localhost:9000 and loads the JavaScript and CSS needed render the page and setup the UI.
The JavaScript for the page is compiled from the index.coffee file which is written in CoffeeScript (an elegant way to write JavaScript). Using jQuery, a page ready handler sets up the WebSocket connection and sets up functions which will be called when the server sends a message to the client through the WebSocket:
$ ->
ws = new WebSocket $("body").data("ws-url")
ws.onmessage = (event) ->
message = JSON.parse event.data
The message is parsed and depending on whether the message contains the stock history or a stock update, a stock chart is either created or updated. The charts are created using the Flot JavaScript charting library. Using CoffeeScript, jQuery, and Flot makes it easy to build Reactive UI in the browser that can receive WebSocket push events and update the UI in real-time.
When a web server gets a request, it allocates a thread to handle the request and produce a response. In a typical model the thread is allocated for the entire duration of the request and response, even if the web request is waiting for some other resource. A Reactive Request is a typical web request and response, but handled in an asynchronous and non-blocking way on the server. This means that when the thread for a web request is not actively being used, it can be released and reused for something else.
The route is defined in the routes file:
GET /sentiment/:symbol controllers.StockSentiment.get(symbol)
A GET
request to /sentiment/GOOG
will call get("GOOG")
on the StockSentiment.scala controller. That method begins with:
def get(symbol: String): Action[AnyContent] = Action.async {
The async
block indicates that the controller will return a Future[Result]
which is a handle to something that will produce a Result
in the future. The Future
provides a way to do asynchronous handling but doesn't necessarily have to be non-blocking. Often times web requests need to talk to other systems (databases, web services, etc). If a thread can't be deallocated while waiting for those other systems to respond, then it is blocking.
In this case a request is made to Twitter and then for each tweet, another request is made to a sentiment service. All of these requests, including the request from the browser, are all handled as Reactive Requests so that the entire pipeline is Reactive (asynchronous and non-blocking). This is called Reactive Composition.
Combining multiple Reactive Requests together is Reactive Composition. The StockSentiment controller does Reactive Composition since it receives a request, makes a request to Twitter for tweets about a stock, and then for each tweet it makes a request to a sentiment service. All of these requests are Reactive Requests. None use threads when they are waiting for a response. Scala's for comprehensions make it very easy and elegant to do Reactive Composition. The basic structure is:
for {
tweets <- tweetsFuture
sentiments <- Future.sequence(futuresForTweetSentiment(tweets))
} yield Ok(sentiments)
Because the web client library in Play, WS
, is asynchronous and non-blocking, all of the requests needed to get a stock's sentiments are Reactive Requests. Combined together these Reactive Requests are Reactive Composition.
The client-side of Reactive Requests and Reactive Composition is no different than the non-Reactive model. The browser makes an Ajax request to the server and then calls a JavaScript function when it receives a response. In the Reactive Stocks application, when a stock chart is flipped over it makes the request for the stock's sentiments. That is done using jQuery's ajax
method in the index.coffee file. When the request returns data the success
handler updates the UI.
The Reactive Stocks example combines Reactive Push, Reactive Requests, Reactive Composition, and a Reactive UI to create a Resilient, Interactive, Scalable, and Event-Driven application. For more information, please see the documentation for Websockets and Akka Streams: