Adv Topics: MapReduce and Hadoop

Hadoop allows for data processing through MapReduce and it also allows for data storage (Lublinsky et al., 2014). MapReduce is an analytical engine and pattern that takes advantage of distributed systems while keeping the processes and data in one machine (Sadalage & Fowler, 2012). MapReduce thus contains two functions that work in parallel on distributed systems (Hortonworks, 2013; Sadalage & Fowler, 2012; Sakr, 2014; Sathupadi, 2010):

    1. Mappers functions create and process transactions on the system by mapping and aggregating data by key values. Mappers can read only one data record at a time.
    2. Reducers functions know what that key values are and will take all those values stored in a map to reduce the data to what is relevant. Reducers help summarize the data into a single output. This helps deal with the amount of data moving between multiple computational nodes.

Lublinsky, Smith, and Yakubovich, (2014), stated that an intermediate component of MapReduce is known as the shuffle and sort, where the data from the mapping function outputs are moved and presented to the reducer function.

Thus, MapReduce is a framework that uses parallel sequential algorithms that capitalize on cloud architecture, which became popular under the open source Hadoop project, as its main executable analytic engine (Lublinsky et al., 2014; Sadalage & Fowler, 2012; Sakr, 2014). Essentially, a sequential algorithm is a computer program that runs on a sequence of commands, and a parallel algorithm runs a set of sequential commands over separate computational cores (Brookshear & Brylow, 2014; Sakr, 2014). Thus, a parallel sequential algorithm runs a full sequential program over multiple but separate cores (Sakr, 2014). Another feature of MapReduce is that a reduced output can become another’s map function (Sadalage & Fowler, 2012). Subsequently, the advantages and disadvantages of using MapReduce are (Lusblinksy et al., 2014; Sakr, 2014):

+ aggregation techniques under the mapper function can exploit multiple different techniques

+ no read or write of intermediate data, thus preserving the input data

+ no need to serialize or de-serialize code in either memory or processing

+ it is scalable based on the size of data and resources needed for processing the data

+ isolation of the sequential program from data distribution, scheduling, and fault tolerance

– too many mapper functions can create an infrastructure overhead, which increases resources and thus cost

– too few mapper functions can create huge workloads for certain types of computational nodes

– too many reducers can provide too many outputs, and too little reducers can provide too little outputs

 – it’s a different programming paradigm that most programmers are not familiar with

 – the use of available parallelism will be underutilized for smaller data sets

Given that Hadoop is predominately known for popularizing MapReduce tasks, it is also known for its Hadoop Distributed File System (HDFS) where the data is distributed across multiple systems (Rathbone, 2013). Hadoop’s service is part of the cloud (as Platform as a Service = PaaS).  For PaaS, the end users manage the applications and data, whereas the provider (Hadoop), administers the runtime, middleware, O/S, virtualization, servers, storage, and networking (Lau, 2001). Data is broken up into small blocks, like Legos, such that they are distributed across a distributed database system and across multiple servers and can be processed across all these servers, e.g. Hadoop Cluster (IBM, n.d.).

A common example of a parallel sequential program is dynamical weather forecasting models. In dynamical weather forecasting models, there is a set of defined geodynamic, thermodynamic, and physical sequential algorithms define and evolve the main seven variables of weathers across time. For each time step, the forecasting models run these sequential algorithms over each grid point, which can represent a finite geospatial region. Each of these geospatial regions is split amongst multiple computational scores. This example expands in complexity when data has to travel between different finite geospatial regions through the boundaries, which is an example of data parallelism (Sakr, 2014). MapReduce uses the concept of data parallelism to help map and reduce data. Therefore, weather models could be considered as a loose form of MapReduce algorithm.

Resources:

Advertisements

Data Tools: XML & Hadoop

Hadoop is predominately known for its Hadoop Distributed File System (HDFS) where the data is distributed across multiple systems and its code for running MapReduce tasks (Rathbone, 2013). MapReduce has two queries, one that maps the input data into a final format and split across a group of computer nodes, while the second query reduces the data in each node so that when combining all the nodes it can provide the answer sought (Eini, 2010). In other words, data is partitioned, sorted and grouped to provide a key and value as an output (Rathbone, 2013). As more data gets added in real time, data in motion, MapReduce can do the recalculations cheaper than before, and the data scientist doesn’t have to touch the data (Eini, 2010; Roy, 2014). Roy (2014) had suggested an example of using Intensive Care Unit (ICU) sensor data, which comes into a database multiple times per second to help avoid false positive alarms that could lead to overwork hospital staffers.  However, Hadoop is best used for non-realtime tasks with a huge demand for processing power (Rathbone, 2013). The issue for Hadoop is to identify the correct instance that an actionable item is needed and acting on that item (Roy, 2014).

Does XML have any impact on MapReduce application design?

XML is a machine and human readable data format (Smith, 2012). With a goal of using XML for MapReduce, we need to assume that we need to map and reduce huge files (Eini, 2010; Smith 2012). Unfortunately, XML doesn’t include sync markers in the data format and therefore MapReduce doesn’t support XML (Smith, 2012). There are posts out there by coders use workarounds to allow for XML processing in Hadoop (Atom, 2010; Krishna, 2014; Rohit, 2013; Smith, 2012).  Smith (2012) and Rohit (2013) used the XmlInputFormat class from mahout to work with XML input data into HBase.  So, depending on the path the data scientist chooses will mean how much work is needed to be able to use MapReduce: code a new version of reading, mapping and reducing XML data from scratch; or use libraries from other code that is compatible with Hadoop.  Smith (2012), stated that the Mahout’s code needs to know the exact sequence of XML start and end tags that will be searched for and Elements with attributes are hard for Mahout’s XML library to detect and parse. Depending on the complexity of the XML document, Smith’s (2012) statement may mean the more complex use of XML input codes may be needed.  Therefore, a well designed XML document could make this process a bit easier, but the complexity of the data stored in it will make the task of creating code for using MapReduce on XML data harder.  Finally, Smith (2012) recommended a preprocessing step to convert XML data and treat it as a line of a record into other libraries native for MapReduce.

References

Data Tools: Hadoop Vs Spark

 

Apache Spark

Apache Spark started from a working group inside and outside of UC Berkley, in search of an open-sourced, multi-pass algorithm batch processing model of MapReduce (Zaharia et al., 2012). Spark can have applications written in Java, Scala, Python, R, and interfaces with SQL, which increases ease of use (Spark, n.d.; Zaharia et al., 2012).

Essentially, Spark is a high-performance computing cluster framework, but it doesn’t have its distributed file system and thus uses Hadoop Distributed File System (HDFS, HBase) as in input and output (Gu & Li, 2013).  Not only can it access data from HDFS, HBase, it can also access data from Cassandra, Hive, Tachyon, and any other Hadoop data source (Spark, n.d.).  However, Spark uses its data structure called Resilient Distribution Datasets (RDD) which cache’s data and is a read-only operation to improve its processing time as long as there is enough memory for it in all the nodes of a cluster (Gu & Li, 2013; Zaharia et al., 2012). Spark tries to avoid data reloading from the disk that is why it stores its data in the node’s cache system, for initial and intermediate results (Gu & Li, 2013).

Machines in the cluster can be rebuilt if lost, thus making the RDDs are fault-tolerant without requiring replication (Gu &LI, 2013; Zaharia et al., 2012).  Each RDD is tracked in a lineage graph, and reruns the operations if data becomes lost, therefore reconstructing data, even if all the nodes running spark were to fail (Zaharia et al., 2012).

Hadoop

Hadoop is Java-based system that allows for manipulation and calculations to be done by calling on MapReduce function on its HDFS system (Hortonworks, 2013; IBM, n.d.).

HFDS big data is broken up into smaller blocks across different locations, no matter the type or amount of data, each of these blocs can be still located, which can be aggregated like a set of Legos throughout a distributed database system (IBM, n.d.; Minelli, Chambers, & Dhiraj, 2013). Data blocks are distributed across multiple servers.  This block system provides an easy way to scale up or down the data needs of the company and allows for MapReduce to do it tasks on the smaller sets of the data for faster processing (IBM, n.d). IBM (n.d.) boasts that the data blocks in the HFDS are small enough that they can be easily duplicated (for disaster recovery purposes) in two different servers (or more, depending on your data needs), offering fault tolerance as well. Therefore, IBM’s (n.d.) MapReduce functions use the HFDS to run its procedures on the server in which the data is stored, where data is stored in a memory, not in cache and allow for continuous service.

MapReduce contains two job types that work in parallel on distributed systems: (1) Mappers which creates & processes transactions on the system by mapping/aggregating data by key values, and (2) Reducers which know what that key value is, will take all those values stored in a map and reduce the data to what is relevant (Hortonworks, 2013; Sathupadi, 2010). Reducers can work on different keys, and when huge amounts of data are entered into MapReduce, then the Mapper maps the data, where the data is then shuffled and sorted before it is reduced (Hortonworks, 2013).  Once the data is reduced, the researcher gets the output that they sought.

Significant Differences between Hadoop and Apache Spark              

Spark is faster than Hadoop in iterative operations by 25x-40x for really small datasets, 3x-5x for relatively large datasets, but Spark is more memory intensive, and speed advantage disappears when available memory goes down to zero with really large datasets (Gu & Li, 2013).  Apache Spark, on their website, boasts that they can run programs 100X faster than Hadoop’s MapReduce in Memory (Spark, n.d.). Spark outperforms Hadoop by 10x on iterative machine learning jobs (Gu & Li, 2013). Also, Spark runs 10x faster than Hadoop on disk memory (Spark, n.d.).

Gu and Li (2013), recommend that if speed to the solution is not an issue, but memory is, then Spark shouldn’t be prioritized over Hadoop; however, if speed to the solution is critical and the job is iterative Spark should be prioritized.

References

  • Gu, L., & Li, H. (2013). Memory or time: Performance evaluation for iterative operation on hadoop and spark. InHigh Performance Computing and Communications & 2013 IEEE International Conference on Embedded and Ubiquitous Computing (HPCC_EUC), 2013 IEEE 10th International Conference on (pp. 721-727). IEEE.

Data Tools: Hadoop and how to install it

What is Hadoop

Hadoop’s Distributed File System (HFDS) is where big data is broken up into smaller blocks (IBM, n.d.), which can be aggregated like a set of Legos throughout a distributed database system. Data blocks are distributed across multiple servers.  This block system provides an easy way to scale up or down the data needs of the company and allows for MapReduce to do it tasks on the smaller sets of the data for faster processing (IBM, n.d). Blocks are small enough that they can be easily duplicated (for disaster recovery purposes) in two different servers (or more, depending on the data needs).

HFDS can support many different data types, even those that are unknown or yet to be classified and it can store a bunch of data.  Thus, Hadoop’s technology to manage big data allows for parallel processing, which can allow for parallel searching, metadata management, parallel analysis (with MapReduce), the establishment of workflow system analysis, etc. (Gary et al., 2005, Hortonworks, 2013, & IBM, n.d.).

Given the massive amounts of data in Big Data that needs to get processed, manipulated, and calculated upon, parallel processing and programming are there to use the benefits of distributed systems to get the job done (Minelli et al., 2013).  Hadoop, which is Java based allows for manipulation and calculations to be done by calling on MapReduce, which pulls on the data which is distributed on its servers, to map key items/objects, and reduces the data to the query at hand (Hortonworks, 2013 & Sathupadi, 2010).

Parallel processing allows making quick work on a big data set, because rather than having one processor doing all the work, Hadoop splits up the task amongst many processors. This is the largest benefit of Hadoop, which allows for parallel processing.  Another advantage of parallel processing is when one processor/node goes out; another node can pick up from where that task last saved safe object task (which can slow down the calculation but by just a bit).  Hadoop knows that this happens all the time with their nodes, so the processor/node create backups of their data as part of their fail safe (IBM, n.d).  This is done so that another processor/node can continue its work on the copied data, which enhances data availability, which in the end gets the task you need to be done now.

Minelli et al. (2013) stated that traditional relational database systems could depend on hardware architecture.  However, Hadoop’s service is part of cloud (as Platform as a Service = PaaS).  For PaaS, we manage the applications, and data, whereas the provider (Hadoop), administers the runtime, middleware, O/S, virtualization, servers, storage, and networking (Lau, 2001).  The next section discusses how to install Hadoop and how to set up Eclipse to access map/reduce servers.

Installation steps

  • Go to the Hadoop Main Page < http://hadoop.apache.org/ > and scroll down to the getting started section, and click “Download Hadoop from the release page.” (Birajdar, 2015)
  • In the Apache Hadoop Releases < http://hadoop.apache.org/releases.html > Select the link for the “source” code for Hadoop 2.7.3, and then select the first mirror: “http://apache.mirrors.ionfish.org/hadoop/common/hadoop-2.7.3/hadoop-2.7.3-src.tar.gz” (Birajdar, 2015)
  • Open the Hadoop-2.7.3 tarball file with a compression file reader like WinRAR archiver < http://www.rarlab.com/download.htm >. Then drag the file into the Local Disk (C:). (Birajdar, 2015)
  • Once the file has been completely transferred to the Local Disk drive, close the tarball file, and open up the hadoop-2.7.3-src folder. (Birajdar, 2015)
  • Download Hadoop 0.18.0 tarball file < https://archive.apache.org/dist/hadoop/core/hadoop-0.18.0/ > and place the copy the “Hadoop-vm-appliance-0-18-0” folder into the Java “jdk1.8.0_101” folder. (Birajdar, 2015; Gnsaheb, 2013)
  • Download Hadoop VM file < http://ydn.zenfs.com/site/hadoop/hadoop-vm-appliance-0-18-0_v1.zip >, unzip it and place it inside the Hadoop src file. (Birajdar, 2015)
  • Open up VMware Workstation 12, and open a virtual machine “Hadoop-appliance-0.18.0.vmx” and select play virtual machine. (Birajdar, 2015)
  • Login: Hadoop-user and password: Hadoop. (Birajdar, 2015; Gnsaheb, 2013)
  • Once in the virtual machine, type “./start-hadoop” and hit enter. (Birajdar, 2015; Gnsaheb, 2013)
    1. To test MapReduce on the VM: bin/Hadoop jar Hadoop-0.18.0-examples.jar pi 10 100000000
      1. You should get a “job finished in X seconds.”
      2. You should get an “estimated value of PI is Y.”
  • To bind MapReduce plugin to eclipse (Gnsaheb, 2013)
    1. Go into the JDK folder, under Hadoop-0.18.0 > contrib> eclipse-plugin > “Hadoop-0.18.0-eclipse-plugin” and place it into the eclipse neon 1 plugin folder “eclipse\plugins”
    2. Open eclipse, then open perspective button> other> map/reduce.
    3. In Eclipse, click on Windows> Show View > other > MapReduce Tools > Map/Reduce location
    4. Adding a server. On the Map/Reduce Location window, click on the elephant
      1. Location name: your choice
      2. Map/Reduce master host: IP address achieved after you log in via the VM
  • Map/Reduce Master Port: 9001
  1. DFS Master Port: 9000
  2. Username: Hadoop-user
  1. Go to the advance parameter tab > mapred.system.dir > edit to /Hadoop/mapped/system

Issues experienced in the installation processes (Discussion of any challenges and explain how it was investigated and solved)

Not one source has the entire solution Birajdar, 2015; Gnsaheb, 2013; Korolev, 2008).  It took a combination of all three sources, to get the same output that each of them has described.  Once the solution was determined to be correct, and the correct versions of the files were located, they were expressed in the instruction set above.  Whenever a person runs into a problem with computer science, google.com is their friend.  The links above will become outdated with time, and methods will change.  Each person’s computer system is different than those from my personal computer system, which is reflected in this instruction manual.  This instruction manual should help others google the right terms and in the right order to get Hadoop installed correctly onto their system.  This process takes about 3-5 hours to install correctly, with the long time it takes to download and install the right files, and with the time to set up everything correctly.

Resources

Big Data Analytics: Compelling Topics

Big Data and Hadoop:

According to Gray et al. (2005), traditional data management relies on arrays and tables in order to analyze objects, which can range from financial data, galaxies, proteins, events, spectra data, 2D weather, etc., but when it comes to N-dimensional arrays there is an “impedance mismatch” between the data and the database.    Big data, can be N-dimensional, which can also vary across time, i.e. text data (Gray et al., 2005). Big data, by its name, is voluminous. Thus, given the massive amounts of data in Big Data that needs to get processed, manipulated, and calculated upon, parallel processing and programming are there to use the benefits of distributed systems to get the job done (Minelli, Chambers, & Dhiraj, 2013).  Parallel processing allows making quick work on a big data set, because rather than having one processor doing all the work, you split up the task amongst many processors.

Hadoop’s Distributed File System (HFDS), breaks up big data into smaller blocks (IBM, n.d.), which can be aggregated like a set of Legos throughout a distributed database system. Data blocks are distributed across multiple servers. Hadoop is Java-based and pulls on the data that is stored on their distributed servers, to map key items/objects, and reduces the data to the query at hand (MapReduce function). Hadoop is built to deal with big data stored in the cloud.

Cloud Computing:

Clouds come in three different privacy flavors: Public (all customers and companies share the all same resources), Private (only one group of clients or company can use a particular cloud resources), and Hybrid (some aspects of the cloud are public while others are private depending on the data sensitivity.  Cloud technology encompasses Infrastructure as a Service (IaaS), Platform as a Service (PaaS), and Software as a Service (SaaS).  These types of cloud differ in what the company managers on what is managed by the cloud provider (Lau, 2011).  Cloud differs from the conventional data centers where the company managed it all: application, data, O/S, virtualization, servers, storage, and networking.  Cloud is replacing the conventional data center because infrastructure costs are high.  For a company to be spending that much money on a conventional data center that will get outdated in 18 months (Moore’s law of technology), it’s just a constant sink in money.  Thus, outsourcing the data center infrastructure is the first step of company’s movement into the cloud.

Key Components to Success:

You need to have the buy-in of the leaders and employees when it comes to using big data analytics for predictive, prescriptive or descriptive purposes.  When it came to buy-in, Lt. Palmer had to nurture top-down support as well as buy-in from the bottom-up (ranks).  It was much harder to get buy-in from more experienced detectives, who feel that the introduction of tools like analytics, is a way to tell them to give up their long-standing practices and even replace them.  So, Lt. Palmer had sold Blue PALMS as “What’s worked best for us is proving [the value of Blue PALMS] one case at a time, and stressing that it’s a tool, that it’s a compliment to their skills and experience, not a substitute”.  Lt. Palmer got buy-in from a senior and well-respected officer, by helping him solve a case.  The senior officer had a suspect in mind, and after feeding in the data, the tool was able to predict 20 people that could have done it in an order of most likely.  The suspect was on the top five, and when apprehended, the suspect confessed.  Doing, this case by case has built the trust amongst veteran officers and thus eventually got their buy in.

Applications of Big Data Analytics:

A result of Big Data Analytics is online profiling.  Online profiling is using a person’s online identity to collect information about them, their behaviors, their interactions, their tastes, etc. to drive a targeted advertising (McNurlin et al., 2008).  Profiling has its roots in third party cookies and profiling has now evolved to include 40 different variables that are collected from the consumer (Pophal, 2014).  Online profiling allows for marketers to send personalized and “perfect” advertisements to the consumer, instantly.

Moving from online profiling to studying social media, He, Zha, and Li (2013) stated their theory, that with higher positive customer engagement, customers can become brand advocates, which increases their brand loyalty and push referrals to their friends, and approximately 1/3 people followed a friend’s referral if done through social media. This insight came through analyzing the social media data from Pizza Hut, Dominos and Papa Johns, as they aim to control more of the market share to increase their revenue.  But, is this aiding in protecting people’s privacy when we analyze their social media content when they interact with a company?

HIPAA described how we should conduct de-identification of 18 identifiers/variables that would help protect people from ethical issues that could arise from big data.   HIPAA legislation is not standardized for all big data applications/cases; it is good practice. However, HIPAA legislation is mostly concerned with the health care industry, listing those 18 identifiers that have to be de-identified: Names, Geographic data, Dates, Telephone Numbers, VIN, Fax, Device ID and serial numbers, emails addresses, URLs, SSN, IP address, Medical Record Numbers, Biometric ID (fingerprints, iris scans, voice prints, etc), full face photos, health plan beneficiary numbers, account numbers, any other unique ID number (characteristic, codes, etc), and certifications/license numbers (HHS, n.d.).  We must be aware that HIPAA compliance is more a feature of the data collector and data owner than the cloud provider.

HIPAA arose from the human genome project 25 years ago, where they were trying to sequence its first 3B base pair of the human genome over a 13 year period (Green, Watson, & Collins, 2015).  This 3B base pair is about 100 GB uncompressed and by 2011, 13 quadrillion bases were sequenced (O’Driscoll et al., 2013). Studying genomic data comes with a whole host of ethical issues.  Some of those were addressed by the HIPPA legislation while other issues are left unresolved today.

One of the ethical issues that arose were mentioned in McEwen et al. (2013), for people who have submitted their genomic data 25 years ago can that data be used today in other studies? What about if it was used to help the participants of 25 years ago to take preventative measures for adverse health conditions?  However, ethical issues extend beyond privacy and compliance.  McEwen et al. (2013) warn that data has been collected for 25 years, and what if data from 20 years ago provides data that a participant can suffer an adverse health condition that could be preventable.  What is the duty of the researchers today to that participant?

Resources: