Coursera Architecture

Coursera is a venture backed for-profit educational technology company that offers massive open online courses (MOOCs).

It works with top universities and organizations to make some of their courses available online, and offers courses in physics, engineering, humanities, medicine, biology, social sciences, mathematics, business, computer science, digital marketing, data science and other subjects.  Itz an online educational startup with over 14 million learners across the globe to offer more than 1000 courses from over 120 top universities.

At Coursera, Amazon Redshift is used as primary data warehouse because it provides a standard SQL interface and has fast and reliable performance. AWS Data Pipeline is used to extract, transform, and load (ETL) data into the warehouse. Data Pipeline provides fault tolerance, scheduling, resource management and an easy-to-extend API for ETL processing.

Dataduct is a Python-based framework built on top of Data Pipeline that lets users create custom reusable components and patterns to be shared across multiple pipelines. This boosts developer productivity and simplifies ETL management.

At Coursera, 150+ pipelines were executed to pull the data from 15 data sources such as Amazon RDS, Cassandra, log streams, and third-party APIs. 300+ tables are loaded every day into Amazon Redshift, processing several terabytes of data. Subsequent pipelines push data back into Cassandra to power our recommendations, search, and other data products.

The attached image below illustrates the data flow at Coursera.

ScyllaDB

ScyllaDB is the world's fastest NoSQL column store database, which is written in C++. Itz fully compatible with Apache Cassandra at 10x throughput and jaw dropping low latency.

Scylla will work with existing Cassandra command line CQL clients. However, mixed clusters of Scylla and Cassandra nodes are not supported. A Scylla node cannot join a Cassandra cluster, and a Cassandra node cannot join a Scylla cluster.

To share the benchmark between Scylla and Cassandra, both throughput on a single multi core server is evaluated with the Hardward specification:
2x Xeon E5-2695v3: 2.3GHz base, 35M cache,
14 core -> 28 core with HT
64GB RAM
2x 400GB Intel NVMe P3700 SSD
Intel Ethernet CNA XL710-QDA1

In terms of software Scylla 0.8 & Cassandra 3.0, is enabled as TestBed.

In the attached image, average throughput for the test is presented as lines, latency as bars.

Scylla’s measured latency of less than 1 ms for the 99th percentile is significantly lower than Cassandra’s, while providing significantly higher throughput (the single client machine could not fully load the server).  The lack of garbage collection means that there are no externally imposed latency sources, so Scylla latency can be brought even lower.


Scylla’s order of magnitude improvement in performance opens a wide range of possibilities.  Instead of designing a complex data model to achieve adequate performance, use a straightforward data model, eliminate complexity, and finish your NoSQL project in less time with fewer bugs.

MapReduce Authors

We know that MapReduce is the ice breaker for the traditional computing model by introducing Scale Out technology in the easy way. Google has a separate research web page on Google's MapReduce at http://research.google.com/archive/mapreduce.html . Authors are Sanjay Ghemawat & Jeff Dean from Google Inc.

As the research scholar, I liked the motivation of their research paper - Large Scale Data Processing. It was achievable with super computing on earlier days. But key difference is parallel execution of hundreds or thousands of CPUs, with commodity box and easy mode.  More over, MapReduce provides:

1. Automatic parallelization and distribution
2. Fault-tolerance
3. I/O scheduling
4. Status and monitoring

Fault-tolerance is handled via re-execution.  On worker failure:

• Detect failure via periodic heartbeats
• Re-execute completed and in-progress map tasks
• Re-execute in progress reduce tasks
• Task completion committed through master

Data Locality Optimization, Skipping Bad Records and Compression of intermediate data are their few refinement technique to boost the performance on large scale data.

In their research paper, the use case was listed in August 2004 with the below metric:

• Number of jobs 29,423
• Average job completion time 634 secs
• Machine days used 79,186 days
• Input data read 3,288 TB
• Intermediate data produced 758 TB
• Output data written 193 TB
• Average worker machines per job 157
• Average worker deaths per job 1.2
• Average map tasks per job 3,351
• Average reduce tasks per job 55
• Unique map implementations 395
• Unique reduce implementations 269
• Unique map/reduce combinations 

Amazing and game changing methodology with easiness, as the result of great minds research from Google.  Herez an opportunity for me to highlight the authors of MapReduce - Sanjay Ghemawat & Jeff Dean

MapReduce Execution

On submission of the job by the User, Hadoop initiates the Job Tracker process at Master Node.  Internally, the execution undergoes 3 major tasks/steps as below:

1. Map 

A map task is created that runs the user-supplied map function on each record. The map function takes a key-value pair as input and produces zero or more intermediate key-value pairs. Map tasks are executed in parallel by various machines across the cluster.  

Mapper output (intermediate data) is stored on the Local file system (NOT HDFS) of each individual mapper nodes. This is typically a temporary directory location which can be setup in config by the Hadoop administrator. The intermediate data is cleaned up  after the Hadoop Job completes.

2. Shuffle and Sort

It is preformed by the reducers (reduce tasks). Each reducer is assigned one of the partitions on which it should work. This is a flurry of network copies between each reducer in the cluster so it can get the partition (intermediate key-value) data it was assigned to work on.

The output of the reduce is normally stored in HDFS for reliability. This step uses a lot of bandwidth between servers and benefits from very fast networking like 10G.

3. Reduce

After the partition data has been copied we can start performing a merge sort of the data. A merge sort takes a number of sorted items and merges them together to form a fully sorted list. 

   

Each reducer produces a separate output file, usually in HDFS. Each reducer output file usually named part-, where  is the number of the reduce task within the job. The output format of the file is specified by the author of the MapReduce job. The number of reducers is defined by the developer.

Attached image depicts the execution flow of Hadoop job.

MapReduce

Computers are always driven by TWO key factors

1. Storage
2. Process

So far, we covered SIX tips on Storage - HDFS model during last 2 months.  Let us start exploring the process layer - Map Reduce.

MapReduce is a programming model and an associated implementation for processing and generating large data sets with a parallel, distributed algorithm on a cluster. Conceptually similar approaches have been very well known since 1995 with the Message Passing Interface standard having reduce and scatter operations.

As depicted in the diagram, MapReduce program is composed of a Map procedure that performs filtering and sorting and a Reduce procedure that performs a summary operation. MapReduce Framework orchestrates the processing by marshaling the distributed servers, running the various tasks in parallel, managing all communications and data transfers between the various parts of the system, and providing for redundancy and fault tolerance.

MapReduce libraries have been written in many programming languages, with different levels of optimization. A popular open-source implementation that has support for distributed shuffles is part of Apache Hadoop. The name MapReduce originally referred to the proprietary Google technology, but has since been generalized.

HDFS ecosystem has TWO different versions on MapReduce, namely V1 and V2.

• V1 is the orginal MapReduce that uses TaskTracker and JobTrackerdaemons. 
• V2 is called YARN. YARN to splits up the two major functionalities of the JobTracker, resource management and job scheduling/monitoring, into separate daemons. Resource manager, application master, and node manager.

We'll see more details on Map Reduce V1 & V2 in the next blogs.

Back to School

Go after your Dream, No matter how unattainable other think it is!  Interesting quote, isn't it?

I always dream about the top-notch premium institute. God blessed me to admit into Indian Institute of Technology as Research Scholor, since August 2015.

At IIT Madras, research is a preoccupation of around 550 faculty members, 3000+ MS and PhD research scholars, more than 800 project staff, and a good number of undergraduates as well.  IIT is always world class institute and the alumni speaks the industry value across the globe.

My blessed journey is starting with the emerging technology course on Searching and Indexing in Large Datasets.  Wishing your prayer and support to execute this life filling opportunity.

Hadoop Command Line

All Hadoop commands are invoked by the bin/hadoop script. Running the hadoop script without any arguments prints the description for all commands.

hadoop [--config confdir] [COMMAND] [GENERIC_OPTIONS] [COMMAND_OPTIONS]

The File System (FS) shell includes various shell-like commands that directly interact with the Hadoop Distributed File System (HDFS) as well as other file systems that Hadoop supports, such as Local FS, HFTP FS, S3 FS, and others. The FS shell is invoked by:
bin/hadoop fs

Hadoop comes with a number of command-line tools that enable basic filesystem operations. HDFS commands are subcommands of the hadoop command-line utility. To display basic usage information, the command is:

hadoop fs  

Hadoop uses fs.default.name value in core-site.xml file if full url syntax is not used.  Command to list files in a dir.

hadoop fs -ls /user/myid
or
hadoop fs -ls hdfs://NameNode.blah.com:8020/home/myid

Command to upload file with -put or -copyFromLocal which copies file form local filesystem

hadoop fs -put /etc/mytest.conf /user/myid/

To download file from HDFS using -get or -copyToLocal.

hadoop fs -get /user/myid/mytest.conf ./

Process to set a replication factor for a file or dir of files with the -R

hadoop fs -setrep 5 -R /user/myid/rep5/

fsck is command to run HDFS filesystem checking utility. Run a fsck on the files we set the rep factor on and see if it looks correct

hadoop fsck /user/myid/rep5 -files -blocks -locations

These are basic frequently used HDFS commands using command line operations.