Big Data is a hot topic these days, and one aspect of that problem space is processing streams of data in near-real time. One of the applications that can help you do this is Spark, which is produced at UC Berkeley’s AMP (Algorithms, Machines and People) Lab.

The first thing you need when you’re looking at data stream analysis techniques is a stream of data to analyse. I’m using a JSON/WebSocket representation of SIX Financial Information’s real time market data feed.

The next thing I need is a problem to solve using my data stream. Now, I’m no great financial wizard, but I’m going to suggest that there might be something useful to be gained from knowing what sectors are currently “trending” - where trending means showing an overall trend in a positive direction, through lots of price changes.

Spark - a quick introduction

For those of you that haven’t heard of Spark before, it’s a project written by the folks over at Berkeley, and is a key component of their Berkeley Data Analytics Stack. It is written mostly in Scala, and provides APIs for Scala, Java and Python. It is fully compatible with Hadoop Distributed File System, but extends on Hadoop’s core functionality by providing in-memory cluster computation, and, most importantly for this blog post, a stream handling framework.

If you’re interested in finding out more about Spark, they offer a free online introductory course (that will set you back about $10 in Amazon EC2 fees). However, at the time of writing, the streaming exercise doesn’t work as the EC2 image is based on an old version of Spark that uses a decommissioned Twitter API.

Spark Streaming

To consume a stream of data in Spark you need to have a StreamingContext in which you register an InputDStream that in turn can produce a Receiver object. Spark provides a number of default implementations of these (e.g. Twitter, Akka Actor, ZeroMQ, etc.) that are accessible from the context. As there is no default implementation for a WebSocket, so we’re going to have to define our own.

The real work is done by the Receiver implementation, so we’ll start there.

I planned to use the scalawebsocket library to access the WebSocket, but unfortunately it’s only available for Scala 2.10, and Spark is only available for Scala 2.9 - it’s when this happens that you have to swallow down your annoyance at the lack of binary compatibility between Scala versions - fortunately all I had to do was strip out the logging statements from scalawebsocket (they used 2.10 macros), and then I could recompile it for 2.9.

With that done, we can now implement a simple trait for using our WebSocket (it uses the Listings object, which just produces a list of all the available stock listings for which the sector is known, and a map of sector id to name for later):

This is useful because to start off with, we just want to check that we’re receiving messages from the WebSocket correctly. We can write a simple extension of this trait:

So once we’ve got our Scala application hooked up to the WebSocket correctly, we can implement a Receiver to consume messages using Spark. Seeing as we’re receiving our stream over a generic network protocol, we’re going to extend the NetworkReceiver. All we need to do then is create a block generator and append our messages onto it:

That was pretty simple, but all we’ve got here is a text string containing JSON data - we can extract the salient bits of data into a case class that is then going to be easier to manipulate. Let’s create a PriceUpdate case class:

Unfortunately I couldn’t find the financial listing attribute to give me the previous price for a listing. We’ll just close our eyes and pretend there isn’t a threading problem using a central map to hold onto the previous values of the price - obviously if we were writing a production application we wouldn’t be able to use this hack.

Now our receiver can look like this:

Much better. Now we need a corresponding InputDStream. Seeing as we’re only ever going to be returning new PriceReceiver objects whenever the getReceiver function is called, we can just create our stream as an object:

Right, let’s plug it into a Spark Streaming application and fire it up. If we follow the Spark Quick Start instructions and then the guide for using Spark Streaming, we need to wrap up the following basic outline into an application:

This code is initialising the streaming context, providing the cluster details (I’m just using a local single node), the name of the application, how much time to gather data from the stream before processing it, the location of the installed Spark software, and the jar file to run the application from. The latter we can use the output from sbt, all we need to do is make sure we use sbt package run from the command line, and it will produce a jar file in the target directory, and Spark then uses that as a result of our passing it on creation of the StreamingContext.

Then we just register our input stream, do some processing on it, and start the streaming context. We can start off by just using the print function, which will just print the first 10 items off the stream:

Which gives us output like this:

-------------------------------------------
Time: 1375194945000 ms
-------------------------------------------
Croda International PLC - 24.82 - 24.82
ASOS PLC - 47.485 - 47.485
Arian Silver Corp - 0.0435 - 0.0435
Medicx Fund Ltd - 0.7975 - 0.7975
Supergroup PLC - 10.73 - 10.73
Diageo PLC - 20.07 - 20.075
Barclays PLC - 2.891 - 2.8925
QinetiQ Group PLC - 1.874 - 1.874
CSR PLC - 5.7 - 5.7
United Utilities Group PLC - 7.23 - 7.23
...

Processing the data

Now we just need to process the blocks of data. First, we want to turn them into a list of sector, price change and change frequency. If we first turn each item into sector, change and a count of 1, we can do this as follows:

val sectorPriceChanges = stream.map(pu => (listingSectors(pu.id), (pu.price - pu.lastPrice, 1)))

The result of this line is that the stream has been transformed into a tuple of sector id combined with another tuple containing price change and the count of 1. Stream elements that are tuple pairs have an extra set of functions that we can use in the PairDStreamFunctions class, for which there is an implicit conversion function available in the StreamingContext, so by importing StreamingContext._ we can now use the reduceByKeyAndWindow function. This function allows us to use a moving frame to reduce over, using the first value of the pair as the key for the reduction. We supply a reduce function and an inverse reduce function - then for each iteration within the frame, Spark will reduce the new data and “un-reduce” the old. Here’s a picture to try and illustrate this:

Here we’re looking at a sliding window in its old state (red) and new state (blue), with the Spark iterations marked by the dashed lines. As each iteration passes by, the purple area is staying the same, so all Spark needs to do is undo the reduction of the red section that has fallen off the end, and add on the reduction of the new blue section.

So now we need a reduce and inverse reduce function to use - I want my reduction to sum all of the price changes (positive and negative), and then I want to know whether the number of changes were more in a positive direction than negative, so if the price change was positive, I’m going to increase my count, but if negative, I’m going to decrease it:

Now we’ve got a reduced stream of net price change and movement trend. We only want to display the biggest positive movers, so we can filter the stream for those values that have a positive movement trend, then we can switch the tuples around so that we have a key of something we can sort by. Now my statistics isn’t that great, but I want my ordering to be weighted by net price change and the positive movement trend, so I’m just going to multiply the two values together. Finally we can sort the data and print out the top 5. Put it all together and we’ve got our streaming application:

Note that we’ve had to add a location for Spark to checkpoint the DStream that is created when we use reduceByKeyAndWindow with an inverse function - it isn’t really explained why this is needed (it is just mentioned in passing in the Spark streaming guide), but my assumption is that Spark needs to store the data that is received in each interval so that it has it available when it comes to applying the inverse reduction function.

When we run this (you may run into PermGen space problems the first time you try running your application - I found these went away on a subsequent run), we then get the trending sectors being printed:

----------------------------------------------
Positive Trending (Time: 1375269240035 ms)
----------------------------------------------
Real estate
Telecommunication
Graphics, publishing & printing media
Environmental services & recycling
Agriculture & fishery
----------------------------------------------
Positive Trending (Time: 1375269255035 ms)
----------------------------------------------
Real estate
Graphics, publishing & printing media
Environmental services & recycling
Agriculture & fishery
Electrical appliances & components
----------------------------------------------
Positive Trending (Time: 1375269270034 ms)
----------------------------------------------
Environmental services & recycling
Agriculture & fishery
Electrical appliances & components
Vehicles
Precious metals & precious stones

It looks like Agriculture & fishery or Environmental services & recycling are worth investing in right now, but don’t take my word for it!

In this blog we’ve looked at how stream processing can be achieved using Spark - obviously if we were developing a real application we’d use much more solid statistical analysis, and we might use a smaller sliding interval to do our reduction over. Spark is a powerful application, and its future is definitely looking good - it has a sound footing at UC Berkeley, and has just been accepted as an Apache Incubator project, so expect to see more about it as it becomes a real alternative to Hadoop.

blog comments powered by Disqus