Big Data – part 4


Welcome, dear reader, to another post in our series on Big Data. If the reader has not read the previous posts in this series, the links can be found at the end of this post, or in the menus, “Big Data” section. In this post, we will cover a technology that has gained quite popularity in the world of Big Data: The Apache Spark.


Created in Berkeley University by AMPLab, the goal of Spark is to provide a computer model, according to the official website, up to 100x faster than a conventional mapreduce Hadoop job. But how it hopes to achieve this performance improvement?


Such gain is based on this one point in Hadoop mapreduce model. During the execution of a Hadoop job, we have 3 times when the data is “stored” in the processing:

  • At initial processing, before the map step;
  • In the midst of processing, when the data filtered by the map phase is being stored for later stages of sort and reduce;
  • At the end of the processing, when the final result is delivered;

In Hadoop, on these three aforementioned moments, we have an IO disk consumption, because the data is stored on disk, rather than kept in memory, including the intermediate step between the steps of map and reduce. In a production environment of Big Data, it is common to have iterative jobs, running several times on a given body of data, using the result of the previous run as input for the next run. It is precisely in this scenario that the Spark has its biggest gain: keeping the data in memory, the access / write of the data becomes much faster, thus ensuring the announced earnings. From this seemingly simple change, the Spark project, which allows constructing jobs following the BSP model (Bulk Synchronous Parallel), was born keeping as much as possible of the data in memory within a run, thus ensuring a fast and scalable computational model. In the picture below we can see the architecture of the Spark and its subprojects, which we will discuss below

Complementary modules

From the Spark initial project, 4 subprojects were born, that complement his use. All these modules are already part of the default installation of Spark and they are:

Spark SQL: Similar to what is the Hive for Hadoop, Spark SQL brings a language similar to SQL for data query on a Spark installation;

Spark Streaming: Spark streaming allows the build of streaming style applications, where the data can be read / written during the processing, instead of the traditional model, where results from a process can only be delivered at the end of a execution;

MLlib: Equivalent to Apache Mahout, allows the construction of machine learning processes. Machine learning is a field within computer science, where using of statistical and logical rules, programs can “learn” and draw your own conclusions from a mass of data provided as input, simulating a human reasoning;

GraphX: The Spark GraphX allows processing to be built in the Graph format, allowing the resolution of problems through algorithms like Pert, BFS and DFS.

Spark & Hadoop

The reader may be wondering at this point: may I use Spark or Hadoop in my Big Data project? Like everything in the world of technology, this is no simple answer. Several factors may influence this decision, not only technical, but also business, such as the absence, to date, of major players that provide distributions with commercial support, unlike Hadoop that already has commercial distributions of weight as Cloudera and Hortonworks. Due to his complementary nature – Spark integrates with most of the components that make up Hadoop – however, it is possible that Spark could go for a complementary technology over than a competing platform. An example of this is the distribution of Cloudera itself, which provides a Hadoop distribution that also has a Spark distribution. Thus, we have as an increasingly scenario, the combination of the two technologies, rather than using only one of the two. After all, why should we use only 1, if we can enjoy the best that each has to offer us?


And so we come to the conclusion of another chapter of our series. In the next and last post in our series, we will examine some cases of the use of Big Data in the world, in order that we see in practice all the benefits that the Big Data can offer us. Until next time.
Continue reading

Big Data – part 2


This is the second part of a series of posts on Big Data.On this post, let’s talk about the two most popular distributed  processing models of Big Data, the mapreduce, and the BSP (Bulk Synchronous Parallel). A process model is a kind of algorithm upon which to develop software.

Mapreduce model

Modelo map reduce

In the figure above, we can see the mapreduce model. This model is widely used in the market today, especially in companies that use Hadoop as her main Big Data technology. The model consists of two well-defined steps, called map and reduce:

  • In the step known as Map, hundreds – or even thousands – of parallel processes, called “threads”, perform a type of task called mapping, where a large mass of data is divided into pieces, and each performs a filtering process within a respective piece, creating a mass of values in the key-value format. At the end of this phase, there is a group phase, where the values for the same key are grouped to form data in the format key: {value1, value2, value3 …. valueN};
  • In the step known as reduce, the data generated by the map phase is again divided into pieces and passed to hundreds or even thousands of processes that perform processing on the received data bits and generate as a key-value output, which is the final output of the processing that is finally grouped into a mass of results;

In a future post, we’ll take a hands-on hadoop, where we can see an example of this processing model in practice with the WordCount.

BSP Model (Bulk Synchronous Parallel)


Although widespread, the mapreduce model is not without its drawbacks. When we talk about the model being applied in the context of Hadoop, for example, all of the cluster steps and mounting of the final mass with the results is done through files on the file system of Hadoop, HDFS, which generates an overhead in performance when it has to perform the same processing in a iterative manner.Another problem is that for graph algorithms such as DFS, BFS or Pert, MapReduce model is not satisfactory. For these scenarios, there is the BSP.

In the BSP algorithm, we have the concept of supersteps. A superstep consists of a unit of generic programming, which through a global communication component, makes thousands of parallel processing on a mass of data and sends it to a “meeting” called synchronization barrier. At this point, the data are grouped, and passed on to the next superstep chain. In this model, it is simpler to construct iterative workloads, since the same logic can be re-executed in a flow of supersteps. Another advantage pointed out by proponents of this model is that it has a simpler learning curve for developers coming from the procedural world.

Speaking in terms of platforms, Hadoop has the Apache Hama as implementation of this model. The main competitor of Hadoop, Spark, come with this feature natively.


And so we conclude another part of our series on Big Data. To date, these are the main models used by the Big Data platforms. As a technology booming, it is natural that in the future we could have more models emerging and gaining their adoption shares. In the next parts of our series, we’ll talk about the two most known implementations of Big Data to date: Hadoop and Spark. U.

Continue reading