Realtime Data Pipelines

In life there are really two major types of data analytics.  Firstly, we don’t know what we want to know – so we need analytics to tell us what is interesting.  This is broadly called discovery.  Secondly, we already know what we want to know – we just need analytics to tell us this information, often repeatedly and as quickly as possible.  This is called anything from reporting or dashboarding through more general data transformation and so on.

Typically we are using the same techniques to achieve this.  We shove lots of data into a repository of some from (SQL, MPP SQL, NoSQL, HDFS etc) then run queries/ jobs/ processes across that data to retrieve the information we care about.  

Now this makes sense for data discovery.  If we don’t know what we want to know, having lots of data in a big pile that we can slice and dice in interesting ways is good.   But when we already know what we want to know, continued batch based processing across mounds of data to produce “updated” results of data, that is often changing in constantly, can be highly inefficient.

Enter Realtime Data Pipelines.  Data is fed in one end, results are computed in real time as data flows down the pipeline and come out the other end whenever relevant changes we care about occur.  Data Pipelines / workflow / streams are becoming much more relevant for processing massive amounts of data with real time results.  Moving relevant forms of analytics out of large repositories into the actual data flow from producer to consumer, I believe, will be a fundamental step forward in big data management.

There are some emerging technologies looking to address this, more details to follow.