Let's see how data is stored, using Hadoop's File System (HDFS), and processed, using the MapReduce algorithm, in the Hadoop Cluster.
Last time around, I took you on a tour through Hadoop's ecosystem, or in other words, I showed you the main components of its mechanism. Now let's see what makes it tick. Note, however, that this is no easy task, and I won't explain everything in detail. I'll try to keep the explanations simple, but Hadoop is everything but simple.
Let's start with Hadoop's file system, the Hadoop Distributed File System (HDFS). As I mentioned in the previous article, the HDFS is based upon the Google File System (GFS) architecture, which means that, like the GFS, the HDFS is a very resilient piece of computer engineering. The HDFS provides a distributed architecture for extremely large-scale storage, which can easily be extended by scaling out the hardware that supports it. There's an important nuance in the previous sentence. Typically, supercomputers scale up: By adding more resources (more processors, disks, and so on) to the supercomputer itself, performance is improved. With Hadoop and HDFS, it's a bit different: Scaling out means adding more small servers to the cluster, as opposed to adding more resources to one massive supercomputer, when more power or capacity is needed.
In order to scale out, HDFS has some peculiarities. Let's start with the most important: how files are stored. As you'd expect, HDFS is a file system, so it's obvious that it stores files. However, there are a couple of things that you need to know in order to understand how and why the HDFS performs this task. Hadoop can handle structured and unstructured data. The latter usually comes in the form of huge log files (typically larger than 500MB), which a regular file system would have difficulties processing. It's also important to mention that Hadoop was designed to work on a cluster of machines, sharing the workload among them, because none of them by itself can handle the copious amounts of data that need processing. Finally, there are a lot more reads than writes in the HDFS than in a regular file system because Hadoop digests data more often than it ingests: Hadoop's purpose being processing big data sets means that several different analyses are necessarily run over the same chunk of data.
In order to be able to cater to Hadoop's peculiar needs, the HDFS stores files in a particular way. When you save a file in the HDFS, the system breaks it down into blocks and stores these blocks in various slave nodes all over the Hadoop cluster. These blocks don't follow the original file-record markings (for instance, a CSV file can be split midline). Instead, the blocks are created based on the size of the data. HDFS only wants to make sure that files are split into similarly sized blocks that match the predefined block size for the Hadoop instance. Regular file systems also do this. However, Hadoop's file blocks are usually 128MB or larger, while a typical Linux block has 4KB. This is important because the MapReduce (and similar algorithms) will process these blocks in parallel. To enable efficient processing, a balance needs to be struck between the block size and the processing resources available. On one hand, the block size needs to be large enough to warrant the resources dedicated to an individual unit of data processing (typically a cluster node). On the other hand, the block size can't be so large that the system is waiting a very long time for one last unit of data processing to finish its work.
So, a file (usually a large chunk of unstructured data) is split in equal-sized blocks and spread across the Hadoop cluster. Wait...What if one of the cluster nodes fails? After all, we're talking about cheap, off-the-shelf servers. Well, that's where one of the other peculiarities of HDFS kicks in: When the files are split into blocks, HDFS sends the same block to several nodes, thus providing the necessary redundancy that allows the system to keep running smoothly even if some nodes fail. Performance will be affected, but data integrity will be maintained. This block replication process typically sends three copies of each block to different nodes. If a node failure is detected, a fresh copy of the data it contained is replicated to another node in order to comply with the "three-block-copy" principle across the Hadoop cluster.
In short, HDFS's main features are its Write Once, Read Many architecture; file block splitting to enhance parallel processing; and redundancy via replication.
The MapReduce Algorithm (Like Having an Octopus Make Lemonade)
Now that you know how the data is stored, let's see how it can be queried efficiently. In a conventional programming language, like RPG, data is usually processed sequentially. For instance, in order to determine how many records of a table contain a certain value, you'd follow a sequence of actions:
- Open the table.
- Read the first record (either sequentially from the top or using a key).
- Check whether it matches the predefined conditions.
- Increment the counter.
- Move to the next record, repeating the process until the end of the file is reached.
Naturally, this is a simplistic view of the process, because typically indices are used to speed up the search, and SQL might also come into play. However, the process is still sequential. MapReduce and similar algorithms introduce parallel processing into this logic. Imagine that you're making lemonade with your bare hands; even if you're using two hands, it'll take a while. Now imagine an (intelligent) octopus making lemonade: with its multiple arms, the octopus can perform several tasks simultaneously, or in parallel, and get the job done faster. Hadoop's MapReduce implementation is just that - a way to get things done faster by performing tasks in parallel.
Let's consider an example of finding which "things" match a certain set of conditions. Suppose I want to count how many files in my gigantic dataset contain the word "lemonade." Now remember that the files are scattered over several nodes of the Hadoop cluster. Sequential processing would take forever because I'd have to retrieve each file (just like I'd retrieve each record in the previous example) and analyze it. The MapReduce algorithm solves this problem by splitting the task into two subtasks: mapping (finding where the files that contain pertinent data are) and reducing (applying whichever operation was requested - in this case, it'd be a simple count) over that subset of data. Still, by itself, this wouldn't solve the problem, because there's a lot of data, all over the cluster. That's the beauty of MapReduce: these tasks are executed in parallel in the slave nodes, and the result is then sent to the primary node. Instead of bringing data to the primary node for processing, the code itself is sent to the slave nodes for execution. Only the partial results of each node are sent back, as opposed to the "raw" files being sent for processing. This makes any operation much faster than it would be if it were being performed sequentially...much like an octopus would make lemonade much faster than you would.
The map subtask consists in finding the relevant data by using several mathematical tricks, such as sorting, searching, indexing, and combining data into smaller, more manageable chunks of data. It can turn a whole file into a map-type object. In other words, everything is baked into key-value pairs. For instance, a text file containing the sentence "I really really really love cold cold lemonade" would be mapped into the following key-value pairs: (I, 1), (really, 3), (love, 1), (cold, 2), (lemonade, 1). The key is the mapped word and the value is the number of occurrences. This would then be the input for the reduce task, which would apply the search conditions and decide if this file is relevant for the query. In our example, it would be because we're looking for files containing the word "lemonade."
I won't go into great detail, but you can (and should) write your own Java classes (Hadoop is Java-based, even though you can use other programming languages for parts of the framework) to perform these "map" and "reduce" tasks. There's an optional "combine" task, which takes the output of the "map" task and further processes it to facilitate the work of the "reduce" task. If you are familiar with Java and want to learn more about this algorithm's implementation in Hadoop, Apache offers a great tutorial about the topic here.
So Much More
This is a bird's-eye view of the Hadoop framework, one of the main tools for processing Big Data. But there are more things you can do, more tools to explore, and more ways to use big datasets! Things like Machine Learning, Artificial Intelligence, and so on are becoming more and more mainstream and making their way from the academic to the business world. It's a brand-new field that you should explore!
LATEST COMMENTS
MC Press Online