Anatomy of a Big Data Application

As an application developer building a Big Data application, you are primarily concerned with five areas:

  • Data Collection: A method of getting data into the system so that it can be processed. We distinguish these types of data collecting:

    • A system or application service may poll an external source for available data and then retrieve it ("pull"), or external clients may send data to a public endpoint of the platform ("push").

    • Data can come steadily, one event at a time ("real-time"), or many events at once in bulk ("batch").

    • Data can be acquired with a fixed schedule ("periodic"), or whenever new data is available ("on-demand").

  • Data Exploration: One of the most powerful paradigms of Big Data is the ability to collect and store data without knowing details about its structure. These details are only needed at processing time. An important step, between collecting the data and processing it, is exploration; that is, examining data with ad-hoc queries to learn about its structure and nature.

  • Data Processing: After data is collected, we need to process it in various ways. For example:

    • Raw events are filtered and transformed into a canonical form, to ensure quality of input data for down-stream processing.

    • Events (or certain dimensions of the events) are counted or aggregated.

    • Events are annotated and used by an iterative algorithm to train a machine-learned model.

    • Events from different sources are joined to find associations, correlations or other views across multiple sources.

    Processing can happen in real time, where a stream processor consumes events immediately after they are collected. Real time processing provides less expressive power than other processing paradigms, but it provides insights into the data in a very timely manner. CDAP offers Spark Streaming as the real-time processing framework.

    Processing can also happen in batch, where many events are processed at the same time to analyze an entire data corpus at once. Batch processing is more powerful than real-time processing, but due its very nature is always time-lagging and thus often performed over historical data. In CDAP, batch processing can be done via MapReduce or Spark, and it can also be scheduled on a periodic basis as part of a workflow.

  • Data Storage: The results of processing data must be stored in a persistent and durable way that allows other programs or applications to further process or analyze the data. In CDAP, data is stored in datasets.

  • Data Serving: The ultimate purpose of processing data is not to store the results, but to make these results available to other applications. For example, a web analytics application may find ways to optimize the traffic on a website. However, these insights are worthless without a way to feed them back to the actual web application. CDAP allows serving datasets to external clients through services.

A CDAP application consists of combinations of these components:

  • Programs (MapReduce, Spark) for data processing in real time or in batch

  • Datasets for data storage abstraction

  • Services for data serving to external clients

This diagram illustrates a typical Big Data application:

It illustrates the power of data abstraction in CDAP: a stream is not just a means to collect data. It can be consumed by real-time and batch processing at the same time. Similarly, datasets allow sharing of data between programs of different paradigms, real time or batch, without compromising the consistency of the data, because all data access happens under ACID (Atomicity, Consistency, Isolation, and Durability) guarantees.

Created in 2020 by Google Inc.