Dear Kettle users,

Most of you usually use a data integration engine to process data in a batch-oriented way. Pentaho Data Integration (Kettle) is typically deployed to run monthly, nightly, hourly workloads. Sometimes folks run micro-batches of work every minute or so. However, it’s lesser known that our beloved transformation engine can also be used to stream data indefinitely (never ending) from a source to a target. This sort of data integration is sometimes referred to as being “streaming“, “real-time“, “near real-time“, “continuous” and so on. Typical examples of situations where you have a never-ending supply of data that needs to be processed the instance it becomes available are JMS (Java Message Service), RDBMS log sniffing, on-line fraud analyses, web or application log sniffing or of-course … Twitter! Since Twitter is easily accessed it’s common for examples to pop up regarding it’s usage and in this blog post too we will use this service to demo the Pentaho Data Integration capabilities wrt to processing streaming data.

Here’s what we want to do:

  1. Continuously read all the tweets that are being sent on Twitter.
  2. Extract all the hash-tags used
  3. Count the number of hash-tags used in a one-minute time-window
  4. Report on all the tags that are being used more than once
  5. Put the output in a browser window, continuously update every minute.
This is a very generic example but the logic of this can be applied to different fields like JMS, HL7, log sniffing and so on. It differs from the excellent work that Vincent from Open-BI described earlier this week on his blog in the sense that his Talend job finishes where ours will never end and where ours will do time-based aggregation in contrast to aggregation over a finite data set.

Also note that in order for Kettle to fully support multiple streaming data sources we would have to implement support for “windowed” (time-based) joins and other nifty things. We’ve seen very little demand for this sort of requirement in the past, perhaps because people don’t know it’s possible with Kettle. In any case, if you currently are in need of full streaming data support, have a look at SQLStream, they can help you. SQLStream is co-founded by Pentaho’s Julian Hyde of Mondrian fame.

OK, let’s see how we can solve our little problem with Kettle instead…

1. Continuously read all the tweets that are being sent on Twitter.

For this we are going to use one of the public Twitter web services, one that delivers a never-ending stream of JSON messages:

Since the format of the output is never-ending and specific in nature I wrote a small “User Defined Java Class” script:

<blockquote>public boolean processRow(StepMetaInterface smi, StepDataInterface sdi) throws KettleException{HttpClient client = SlaveConnectionManager.getInstance().createHttpClient();client.setTimeout(10000);client.setConnectionTimeout(10000);Credentials creds = new UsernamePasswordCredentials(getParameter("USERNAME"), getParameter("PASSWORD"));client.getState().setCredentials(AuthScope.ANY, creds);client.getParams().setAuthenticationPreemptive(true);HttpMethod method = new PostMethod("");// Execute request//InputStream inputStream=null;BufferedInputStream bufferedInputStream=null;try {int result = client.executeMethod(method);// the response//inputStream = method.getResponseBodyAsStream();bufferedInputStream = new BufferedInputStream(inputStream, 1000);StringBuffer bodyBuffer = new StringBuffer();int opened=0;int c;while ( (!=-1 && !isStopped()) {char ch = (char)c;bodyBuffer.append(ch);if (ch=='{') opened++; else if (ch=='}') opened--;if (ch=='}' && opened==0) {// one JSON block, pass it on!//Object[] r = createOutputRow(new Object[0], data.outputRowMeta.size());String jsonString = bodyBuffer.toString();int startIndex = jsonString.indexOf("{");if (startIndex