2. Using Storm for Streaming Data

This module begins by defining a data stream as an unbounded sequence of event data, or in other words a continuous flow of data without an end. When processing streams, aggregation would most probably find place during a meaningful temporal window.

Storm is an open source Apache project that provides an event processor for data streams and employs the concept of topologies. You can implement a Storm topology by using Spouts and Bolts. A Spout consumes data streams and emits tuples of data to be consumed by one ore more Bolts. A Bolt contains the logic to be applied to the data tuples. A Storm topology can be used to process real-time data for different purposes such as real-time monitoring and reporting or simply in order to populate an HBase table.  

Within HDInsight you can create Storm clusters. Such a cluster consists of a master node running Nimbus and worker nodes running Supervisor. Nimbus assigns the processing tasks to the worker nodes, Supervisor manages the processing on the particular worker node, while Zookeeper takes responsibility of the overall management of the distributed processing. Zookeeper is again an Apache project that fits to the Hadoop ecosystem and is generic in nature, in the sense that it is not specific to Storm.

When a Storm topology is submitted for execution to the Storm cluster, you can benefit from parallelism. In the simple example shown below, where there is one worker process with two tasks (one spout and one bolt) you can set the parallelism hint to create more instances of any task. In this example, 3 instances of the bolt task have been created. As you can also see, each task is residing within an executor. 

By looking closely at the image above you will notice the shuffleGrouping class. By using grouping in Storm, you can determine the way that tuples are being distributed over the respective tasks. In the example above the tuples will be evenly distributed among the 3 bolt tasks. You can also choose for fieldGrouping that will distribute the tuples to executors according to the value of a certain field. Although there are more kinds of groupings you could use, shuffleGrouping en fieldGrouping are the most common ones.

A Storm topology has a set of specific tasks, called Ackers, by which you can track the processing status of a tuple. In order for this to work you will need to cache and uniquely stamp each tuple at the spout. By using the Acker after each bolt, you will either send a ‘success’ or ‘fail’ message to the Spout depending on whether the tuple has been successfully processed or not. The retry logic (which tuples will have to be re-processed by which bolts) will have to be manually created and thus requires explicit programming. Another different class of topologies is the transactional Storm topology, where re-sending of tuples in case of failure happens automatically and a commiter bolt has to be used at the end of the topology.

As it has already been mentioned a common operation on streams of data will be aggregation. The two most common aggregation types are tumbling window and sliding window aggregation. When talking about a tumbling window aggregation we refer to aggregating events in a fixed timespan, while a sliding window refers to aggregating events in overlapping timespans. 

In order to implement aggregation in Storm you will have to cache the incoming tuples at the aggregation bolt and act on those based on a Tick tuple. A Tick tuple is a special kind of tuple that you can configure within Storm and that whenever it is being sent from the spout it will initiate the aggregation logic. 

3. Using Spark for Interactive Analysis

Spark is designed as a fast, general purpose computation engine that supports in-memory operations. While not everything in Spark is necessarily in-memory, it is much faster than Pig and Hive, both of which do not support interactive data analysis. Since Spark is a general purpose engine, it can be used for different use -cases such as streaming, interactive and predictive analysis. As you might have expected, Spark is supported in HDInsight in the form of a different type of cluster that you can create.

The building block of Spark is the resilient distributed dataset (RDD) which is a collection of items. There are different APIs to work with RDDs. HDInsight supports Python and Scala, although you can also program in Java. In order to work with Spark you will need the Driver program that will use the Spark Context to connect to the master node of the cluster and one or more worker nodes to perform the processing. The Driver program initiates all Spark processing.

The image below shows how to create a Spark context and an RDD.

With RDDs you can call transformations that will create a new RDD or actions to output results. The most common are shown below.

In order to work with Spark you can of course create your own standalone application in Java or Python. If you want to work interactively with the data, you can use Notebooks at a web browser. The Notebooks used at HDInsight are based on Jupyter, which is a notebook format widely used in Data Science

A more recent addition to Spark is the DataFrames. By using DataFrames you can structure the data so that they can be queried with Spark SQL. The schema for a DataFrame can either be declared beforehand or be inferred at runtime, with explicit declaration being the preferred method so as to not negatively impact performance. DataFrames can be persisted as shared tables or temporary views and be queried along with Hive tables by using Spark SQL.

The final concept on Spark that this course talks about is Spark streaming. This is a concept similar to Storm streaming, that is specific to Spark. It operates on RDDs by handling the incoming stream as a sequence of RDDs, thus creating discretized streams or DStreams.

In Spark 2.0, when streaming data, the concept of RDD is replaced with DataFrames so that you will be using the same concepts when working with data at rest as with streaming data.

4. Introducing Kafka

Kafka is offered as a preview at this course. HDInsight offers Kafka clusters by which you can store messages in topics. Data stored in Kafka is produced by a producer and consumed by a consumer. Kafka will take care of storing these messages and properly prepare them for processing by the respective consumer. Kafka can be combined, for instance, with Spark streaming if both clusters are in the same virtual network in order to build complete Big Data solutions.