Big data is primarily defined by the volume of a data set. Big data sets are generally huge – measuring tens of terabytes – and sometimes crossing the threshold of petabytes. It is surprising to know how much data is generated every minute. As estimated by DOMO:
Over 2.5 quintillion bytes of data are created every single day, and it’s only going to grow from there. By 2020, it’s estimated that 1.7MB of data will be created every second for every person on earth.
You can read DOMO's full report, including industry-specific breakdowns, here.
To store and process even only a fraction of this amount of data, we need Big Data frameworks as the traditional Databases would not be able to store so much of data nor traditional processing systems would be able to process this data quickly. Here comes the frameworks like Apache Spark and MapReduce to our rescue and help us to get deep insights into this huge amount of structured, unstructured and semi-structured data and make more sense out of it.
Apache Spark was originally developed in 2009 at UC Berkeley by the team who later founded Databricks. Since its launch Spark has seen rapid adoption and growth. Most of the cutting-edge technology organizations like Netflix, Apple, Facebook, Uber have massive Spark clusters for data processing and analytics. The demand for Spark is increasing at a very fast pace. According to marketanalysis.com report forecast, the global Apache Spark market will grow at a CAGR of 67% between 2019 and 2022. The global Spark market revenue is rapidly expanding and may grow up $4.2 billion by 2022, with a cumulative market valued at $9.2 billion (2019 – 2022).
MapReduce has been there for a little longer after being developed in 2006 and gained industry acceptance during the initial years. But at last, 5 years or so with Apache Spark gaining more ground, demand for MapReduce as the processing engine has reduced. But, it cannot be said in black and white that MapReduce will be completely replaced by Apache Spark in the coming years. Both the technologies have their own pros and cons as we will see them below. One solution cannot fit at all the places, so MapReduce will have its own takers depending on the problem to be solved.
Also, Spark and MapReduce do complement each other on many occasions.
Both these technologies have made inroads in all walks of common man’s life. You name the industry and its there. Be it telecommunication, e-commerce, banking, insurance, healthcare, medicine, agriculture, biotechnology, etc.
As per Apache, “Apache Spark is a unified analytics engine for large-scale data processing”. Apache Spark is a fast and general-purpose cluster computing system. It provides high-level APIs in Java, Scala, Python and R, and an optimized engine that supports general execution graphs. It also supports a rich set of higher-level tools including Spark SQL for SQL and structured data processing, MLlib for machine learning, GraphX for graph processing, and Spark Streaming.
Spark, instead of just “map” and “reduce” functions, defines a large set of operations called transformations and actions for the developers and which are ultimately transformed to map/reduce by the spark execution engine and these operations are arbitrarily combined for highly optimized performance.
Spark is developed in Scala language and it can run on Hadoop in standalone mode using its own default resource manager as well as in Cluster mode using YARN or Mesos resource manager. It is not mandatory to use Hadoop for Spark, it can be used with S3 or Cassandra also. But, in the majority of the cases, Hadoop is the best fit as Spark’s data storage layer.
Speed: Spark enables applications running on Hadoop to run up to 100x faster in memory and up to 10x faster on disk. Spark achieves this by minimising disk read/write operations for intermediate results and storing in memory and perform disk operations only when essential. Spark achieves this using DAG, query optimizer and highly optimized physical execution engine.
Fault Tolerance: Apache Spark achieves fault tolerance using spark abstraction layer called RDD (Resilient Distributed Datasets), which are designed to handle worker node failure.
Lazy Evaluation: All the processing(transformations) on Spark RDD/Datasets are lazily evaluated, i.e. the output RDD/datasets are not available right away after transformation but will be available only when an action is performed.
Dynamic nature: Spark offers over 80 high-level operators that make it easy to build parallel apps.
Multiple Language Support: Spark provides multiple programming language support and you can use it interactively from the Scala, Python, R, and SQL shells.
Reusability: Spark code can be used for batch-processing, joining streaming data against historical data as well as run ad-hoc queries on streaming state.
Machine Learning: Apache Spark comes with out of the box support for machine learning called MLib which can be used for complex, predictive data analytics.
Graph Processing: GraphX is Apache Spark's API for graphs and graph-parallel computation. You can view the same data as both graphs and collections, transform and join graphs with RDDs efficiently, and write custom iterative graph algorithms using the Pregel API.
Real-Time Stream Processing: Spark Streaming brings Apache Spark's language-integrated API to stream processing, letting you write streaming jobs the same way you write batch jobs.
Spark is used by 1000+ organizations in Production. Many of these organizations are known to run Spark clusters of 1000+ nodes. In terms of data size, Spark has been shown to work well up to petabytes. It has been used to sort 100 TB of data 3X faster than Hadoop MapReduce (which sorted 100 TB of data in 23 min, using 2100 machines) using 10X fever machines, winning the 2014 Daytona GraySort Benchmark, as well as to sort 1 PB. Several production workloads use Spark to do ETL and data analysis on PBs of data. Below are some examples where Spark is used across industries:
AsiaInfo: Uses Spark Core, Streaming, MLlib and Graphx and Hadoop to build cost-effective data centre solution for our customers in the telecom industry as well as other industrial sectors.
Atp: Predictive models and learning algorithms to improve the relevance of programmatic marketing.
Credit Karma: Creates personalized experiences using Spark
eBay Inc: Using Spark core for log transaction aggregation and analytics
Kelkoo: Using Spark Core, SQL, and Streaming. Product recommendations, BI and analytics, real-time malicious activity filtering, and data mining.
More examples can be found on Apache’s Powered By page
// “sc” is a “Spark context” – this transforms the file into an RDD val textFile = sc.textFile("data.txt") // Return number of items (lines) in this RDD; count() is an action textFile.count() // Demo filtering. Filter is a transform. By itself this does no real work val linesWithSpark = textFile.filter(line => line.contains("Spark")) // Demo chaining – how many lines contain “Spark”? count() is an action. textFile.filter(line => line.contains("Spark")).count() // Length of line with most words. Reduce is an action. textFile.map(line => line.split(" ").size).reduce((a, b) => if (a > b) a else b) // Word count – traditional map-reduce. collect() is an action val word Counts = text File.flatMap(line => line.split(" ")).map(word => (word, 1)).reduceByKey((a, b) => a + b) word Counts.collect()
map(func): Return a new distributed dataset formed by passing each element of the source through a function func.
filter(func): Return a new dataset formed by selecting those elements of the source on which func returns true
union(other Dataset): Return a new dataset that contains the union of the elements in the source dataset and the argument.
reduce(func): Aggregate the elements of the dataset using a function func (which takes two arguments and returns one). The function should be commutative and associative so that it can be computed correctly in parallel.
collect(): Return all the elements of the dataset as an array at the driver program. This is usually useful after a filter or other operation that returns a sufficiently small subset of the data.
count(): Return the number of elements in the dataset.
The data is referred from the RDD Programming guide.
MapReduce is a programming model for processing and generating large data sets with a parallel, distributed algorithm on a cluster.
Programmers have been writing parallel programs for a long time in different languages like C++, Java, C#, Python. But, they have their own nuances and maintaining these, is the programmer's responsibility. There are chances of application crashing, performance hit, incorrect results. Also, such systems if grows very large is not very fault tolerant or difficult to maintain.
MapReduce has simplified all these. Fault tolerance, parallel execution, resources management is all responsibility of the Resource manager and the framework. Programmers have to only concentrate on business logic by writing only map and reduce functions.
A MapReduce application has broadly two functions called map and reduce.
Map: Mapper process takes input as key/value pair, processes them i.e. performs some computation and then produces intermediate results as key/value pairs
i.e. map(k1,v1) ---> list(k2,v2)
Reduce: Reducer process receives an intermediate key and a set of values in sorted order. It processes these and generates output key/value pairs by grouping values for each key.
i.e. reduce(k2, list(v2)) ---> list(v3)
Can also define an option function “Combiner” (to optimize bandwidth)
If defined, runs after Mapper & before Reducer on every node that has run a map task
Combiner receives as input all data emitted by the Mapper instances on a given node
Combiner output sent to the Reducers, instead of the output from the Mappers
Is a "mini-reduce" process which operates only on data generated by one machine
MapReduce is usually applied to huge datasets. A MapReduce job splits the input data into smaller independent chunks called partitions and then processes them independently using map tasks and reduce tasks. Below is an example.
MapReduce Word Count (Pseudocode)
map(String input_key, String input_value): // input_key: document name // input_value: document contents for each word w in input_value: EmitIntermediate(w, "1"); reduce(String output_key, Iterator intermediate_values): // output_key: a word // output_values: a list of counts int result = 0; for each v in intermediate_values: result += ParseInt(v); Emit(AsString(result));
Apply a function to all the elements of List
list1=[1,2,3,4,5]; square x = x * x list2=Map square(list1) print list2 -> [1,4,9,16,25]0
Combine all the elements of list for a summary
list1 = [1,2,3,4,5]; A = reduce (+) list1 Print A -> 15
After getting off hangover how Apache Spark and MapReduce works, we need to understand how these two technologies compare with each other, what are their pros and cons, so as to get a clear understanding which technology fits our use case.
As we can see, MapReduce involves at least 4 disk operations whereas Spark only involves 2 disk operations. This is one reason for Spark to be much faster than MapReduce. Spark also caches intermediate data which can be used in further iterations helping Spark improve its performance further. The more iterative the process the better is the Spark performance due to in-memory processing and caching. This is where MapReduce performance not as good as Spark due to disk read/write operations for every iteration.
Let’s see a comparison between Spark and MapReduce on different other parameters to understand where to use Spark and where to use MapReduce
MapReduce is designed for batch processing and is not as fast as Spark. It is used for gathering data from multiple sources and process it once and store in a distributed data store like HDFS. It is best suited where memory is limited and processing data size is so big that it would not fit in the available memory.
Spark is 10-100 times faster because of in-memory processing and its caching mechanism. It can deliver near real-time analytics. It is used in Credit Card Processing, Fraud detection, Machine learning and data analytics, IoT sensors etc
As it is part of Apache Open Source there is no software cost.
Hardware cost is less in MapReduce as it works with smaller memory(RAM) as compared to Spark. Even commodity hardware is sufficient.
Spark also is Apache Open Source so no license cost.
Hardware cost is more than MapReduce as even though Spark can work on commodity hardware it needs a lot more memory(RAM) as compared to MapReduce since it should be able to fit all the data in Memory for optimal performance. Cluster needs little high-end commodity hardware with lots of RAM else performance gets hit
Ease of Use
MapReduce is a bit complex to write. MapReduce is written in Java and the APIs are a bit complex to code for new programmers, so there is a steep learning curve involved. The Pig has SQL like syntax and it is easier for SQL developers to get onboard easily. Also, there is no interactive mode available in MapReduce
Spark has APIs in Scala, Java, Python, R for all basic transformations and actions. It also has rich Spark SQL APIs for SQL savvy developers and it covers most of the SQL functions and is adding more functions with each new release. Also, Spark has scope for writing User Defined Analytical Functions and Functions (UDF/UDAF) for anyone who would like to have custom functions.
MapReduce is also compatible with all data sources and file formats Hadoop supports. But MapReduce needs another Scheduler like YARN or Mesos to run, it does not have any inbuilt Scheduler like Spark’s default/standalone scheduler.
Apache Spark can in standalone mode using default scheduler. It can also run on YARN or Mesos. It can run on-premise or on the cloud. Spark supports most of the data formats like parquet, Avro, ORC, JSON, etc. It also supports multiple languages and has APIs for Java, Scala, Python, R.
MapReduce can only be used for batch processing where throughput is more important and latency can be compromised.
Spark supports Batch as well as Stream processing, so fits both use cases and can be used for Lambda design where applications need both Speed layer and slower layer/data processing layer
MapReduce has more security features.MapReduce can enjoy all the Hadoop security benefits and integrate with Hadoop security projects, like Knox Gateway and Sentry.
Spark is a bit bare at the moment. Spark currently supports authentication via a shared secret. Spark can integrate with HDFS and it can use HDFS ACLs and file-level permissions. Spark can also run on YARN leveraging the capability of Kerberos.
MapReduce uses replication for fault tolerance. If any slave daemon fails, master daemons reschedule all pending and in-progress operations to another slave. This method is effective, but it can significantly increase the completion times for operations with a single failure also
In Spark, RDDs are the building blocks and Spark also uses it RDDs and DAG for fault tolerance. If an RDD is lost, it will automatically be recomputed by using the original transformations.
MapReduce has high latency
Spark provides low latency performance
MapReduce does not have any interactive mode of operation.
Spark can be used interactively also for data processing. It has out-of-the-box support for spark shell for scala/python/R
Machine Learning/Graph Processing
No support for these. A mahout has to be used for ML
Spark has dedicated modules for ML and Graph processing
MapReduce is best suited for Analysis of archived data where data size is huge and it is not going to fit in memory, and if the instant results and intermediate solutions are not required. MapReduce also scales very well and the cluster can be horizontally scaled with ease using commodity machines.
Offline Analytics is a good fit for MapReduce like Top Products per month, Unique clicks per banner.
MapReduce is also suited for Web Crawling as well as Crawling tweets at scale and NLP like Sentiment Analysis.
Another use case for MapReduce is de-duplicating data from social networking sites, job sites and other similar sites.
MapReduce is also heavily used in Data mining for Generating the model and then classifying.
Spark is fast and so can be used in Near Real Time data analysis.
A lot of organizations are moving to Spark as their ETL processing layer from legacy ETL systems like Informatica. Spark as very good and optimized SQL processing module which fits the ETL requirements as it can read from multiple sources and can also write to many kinds of data sources.
Spark can also handle Streaming data so its best suited for Lambda design.
Most graph processing algorithms like page rank perform multiple iterations over the same data and this requires a message passing mechanism. Spark has great support for Graph processing using GraphX module.
Almost all machine learning algorithms work iteratively. Spark has a built-in scalable machine learning library called MLlib which contains high-quality algorithms that leverage iterations and yields better results than one pass approximations sometimes used on MapReduce.
Hadoop MapReduce is more mature as it has been there for a longer time and its support is also better in the open source community. It can be beneficial for really big data use case where memory is limited and data will not fit the RAM. Most of the time, Spark use case will involve Hadoop and other tools like Hive, Pig, Impala and so when these technologies complement each other it will be a win for both Spark and MapReduce.
Hadoop Mapreduce is more mature as it has been there for a longer time and its support is also better in the open source community. It can be beneficial for really big data use case where memory is limited and data will not fit the RAM. Most of the time, Spark use case will involve Hadoop and other tools like Hive, Pig, Impala and so when these technologies complement each other it will be a win for both Spark and Mapreduce.