Everyone is speaking about Big Data and Data Lakes these days. Many IT professionals see Apache Spark as the solution to every problem. At the same time, Apache Hadoop has been around for more than 10 years and won’t go away anytime soon. In this blog post I want to give a brief introduction to Big Data, demystify some of the main concepts such as Map Reduce, and highlight the similarities and differences between Hadoop and Spark.
You hear about Big Data everywhere. But what does it actually mean and what precisely can we do with it? When is a Big Data system required? We will try to answer these 2 questions in this section.
What is Big Data
What is Big Data? Have you ever heard of the popular definition of Big Data with the 3 Vs? This definition is very common and can be found in many text books and Wikipedia. It suggests that your data is Big data when one (or all - depending on the definition) of the following criteria are fulfilled:
I find this definition very concise and understandable but a bit imprecise which is probably intentional. Here is a more practical definition of what the 3 Vs stand for, based on my own experiences.
Volume describes a large amount of data you want to store, process or analyze. If we are speaking in terms of 100s of GBs to TBs to PBs then we are speaking about Big Data. An important aspect to consider, the data growth. As a rule of thumb: If your data is growing by multiple GBs per day, you are probably dealing with Big Data.
Velocity means a high data throughput to be stored, processed and/or analyzed; often a large amount of data over a short period of time. When we are processing thousands to millions of records per second, then we are most likely speaking of Big Data.
Variety stands for the large amount of different data types and formats that can be stored, processed or analyzed. This means one aims to process any kind of data, be it binary, text, structured, unstructured, compressed, uncompressed, nested, flat, etc. However, variety is rather a consequence of Big Data as all data is eventually stored on a distributed file system and so one has to care about different optimized file formats for different use-cases.
Big Data systems are built to handle data of high volume, velocity and variety. Apache Hadoop and Apache Spark are popular Big Data frameworks for large-scale distributed processing. We will learn the similarities and differences in the following sections.
Please note that other definitions vary slightly and you will find 4 or even more Vs, such as Veracity for example. Veracity refers to the trustworthiness of the data and hence describes how useful your data actually is. While these extended definitions are relevant for Big Data, they don’t necessarily apply only to Big Data systems or require Big Data systems in my opinion.
What can we do with Big Data
Big Data systems allow you to load, process and store data of high volume, velocity and variety. However, I prefer to see it the other way around. Once, you want to load, process and store data of high volume, velocity and variety, you need a Big Data system.
A great example is Google’s search index. In a very simplified way, one has to count occurrences of keywords and phrases in websites and store each keyword and its corresponding list of occurring websites as an inverted index. As you might imagine, both the amount of content and the inverted index probably don’t fit into the memory of a single machine.
Another classic example is Twitter’s trending hashtags (the word count example). One has to count the occurrences of a keyword in all tweets, maybe even weighted over time and aggregated per geographic region. As you might imagine, doing this on billion of messages per second probably exceeds the capabilities of a single machine.
Hence, both use-cases require scalable distributed systems to handle the load and process the data efficiently. There is one more thing that both use-cases have in common: they aggregate/group data by a key, e.g. the occurrences per keyword and the counts per hashtag. This is one of the key requirements for a Big Data system.
Hence most of the Big Data workloads can be categorized into 2 topics: Batch Processing or Stream Processing.
- Batch Processing
- Transformation, Join and Aggregation
- Analytics: (historical) Analytics, Prediction and Modeling
- Stream Processing
- Transformation, Join and (temporal) Aggregation
- Analytics: (real-time) Analytics, Inferencing Prediction models
What is Big Data Analytics
Big Data systems are used to store and process massive amounts of data, mostly batch and stream processing. They are often used for analytics. For clarity about which data is processed, I like to make the distinction between 3 different use-cases of increasing difficulty in Big Data analytics. The use cases are:
- (Historical) Analysis
In classical Analytics, we analyze historical/observed data. In Big Data, we analyze massive amounts of such data. A typical question to answer with analytics could be how to compute the number of visitors of the previous season based on all bookings of said season?.
In Prediction, we analyze the past to build a model that can predict the future. In more general terms, one fits a model on a set of training data to use it for inferring any unknown/unseen observation. We often use statistical methods (such as Generalized Linear Models, Logistic Regression, etc.) as well as Machine Learning (SVM, Gradient Boosted Trees, Deep Learning, etc.) techniques to build these models. A typical question to answer with prediction could be how to forecast the number of visitors for the following season based on all bookings of previous seasons?.
Modeling builds on both analytics and prediction capabilities. In Modeling, the aims is to analyze the past and build a model to predict different possibilities of the future depending on the model parameters. These models are often more complicated than a simple statistical or Machine Learning model and take into account multiple state variables and parameters that can be modified. A typical question to answer with modeling could be how to forecast the number of visitors for the following season if the winter will be two weeks shorter based on all bookings of previous seasons plus additional data sources (weather data, etc.)?.
Hadoop: HDFS, Yarn and MapReduce
- HDFS: a distributed file system
- MapReduce: a framework for distributed processing
- Yarn: a cluster resource manager
HDFS (Hadoop Distributed File System) is a distributed file system that stores and replicates data in blobs across multiple nodes in a cluster. HDFS is the open-source implementation of the Google File System (GFS) paper published by Jeff Dean and Sanjay Ghemawat at Google in 2003.
HDFS consists of a name node and multiple data nodes. The name node holds the references to the data blobs and takes care of the file system and meta operations whereas the data nodes store the data in blobs on the local file system.
MapReduce is a high-level framework for distributed processing of large data sets that abstracts developers’ code into Map (transformation) and Reduce (aggregation) operations. By doing so, the code can automatically run in parallel on a distributed system. MapReduce is an open-source implementation of the MapReduce: Simplified Data Processing on Large Clusters paper published by Jeff Dean and Sanjay Ghemawat at Google in 2004.
Due to the same name of the paper and the open-source implementation, the term MapReduce can lead to confusion between the original concept and the framework.
Similar to both Google papers, HDFS and MapReduce were designed and developed to function as one single framework for distributed processing of large data sets. MapReduce takes advantage of data replication in HDFS by moving computations to the same physical machine were the data is stored. In Hadoop 1, the MapReduce services ran directly on the data nodes without any resource manager.
Apache Yarn (acronym for Yet Another Resource Negotiator) is a distributed resource manager and job scheduler for managing the cluster resources (CPUs, RAM, GPUs, etc.) and for scheduling and running distributed jobs on a Hadoop cluster. It was introduced in Hadoop 2 to decouple the MapReduce engine from the cluster resource management and allowed more services to run on-top of Hadoop. Hence, instead of starting services on each of the nodes individually one can submit a service to Yarn which takes care of the resource negotiation, distribution of the service to all requested nodes, execution of the service, log collection, etc.
Yarn consists of a resource manager service to negotiate cluster resources and multiple node manager services that manage the execution of processes on each of the nodes.
Although not managed in the same repository as Apache Hadoop, I often like to mention Apache Zookeeper as another integral building block of Hadoop. Apache Zookeeper is a distributed synchronized transaction-based in-memory key-value store. Many Hadoop services use Zookeeper for storing dynamic configuration (available nodes per partition, current master, etc.), leader election, synchronization, and much more.
Nowadays, there are many other services related to or included in the Hadoop stack. Here is a (small) list of distributed services that usually run on-top of Hadoop:
- Batch Processing
- Stream processing
- Spark Streaming
- Data Storage
- HDFS (File Store)
- HBase (NoSql)
- Cassandra (NoSQL)
- Accumulo (NoSQL)
- Kafka (Log Store)
- Solr (Inverted Document Index)
Most of these service run on top of Hadoop because they utilize one or more of its components. Typical examples of reused components are:
- HDFS as distributed storage (used in Hive, HBase, etc.)
- Yarn as resource manager (used in Hive, Spark, Storm, etc.)
- Zookeeper for synchronization and leader election (used in Hive, Kafka, Hive, etc. )
- Hive Metastore as a meta data storage (used in Spark, Impala, etc.)
Spark: The Evolution of MapReduce
Apache Spark got popular in 2014 as a fast general-purpose compute framework for distributed processing which claimed to be more than 100 times faster than the traditional MapReduce implementation. It provides high level operations for working with distributed data sets which are optimized and executed in-memory of the cluster nodes. Spark runs on top of multiple resource managers such as Yarn or Mesos.
Conceptually, Spark’s execution engine is similar to the other distributed processing frameworks:
- MapReduce: Simplified Data Processing on Large Clusters
- Tez: A Unifying Framework for Modeling and Building Data Processing Applications
- Impala: A Modern, Open-Source SQL Engine for Hadoop
- Spark: A Fault-Tolerant Abstraction for In-Memory Cluster Computing
Apache Tez (Tez is Hindi for speed) is a faster MapReduce engine based on Apache Yarn that optimizes complex execution graphs into single jobs to avoid intermediate writes to HDFS. It is the default execution engine powering Apache Pig and Apache Hive (a large scale data warehouse solution on top of Hadoop) on the Hortonworks Hadoop distribution.
Apache Impala is a Big Data SQL engine on top of HDFS, HBase and Hive (Metastore) with its own specialized distributed query engine. It is the default engine for Apache Hive on the Cloudera and MapR Hadoop distributions.
As we can see in the following figure, the main differences from the newer processing frameworks (Tez, Impala, and Spark) compared to the traditional MapReduce engine (left) is that they avoid writing intermediate results to HDFS and heavily optimize the execution graph. Another optimization strategy is processing/caching data in memory of the local nodes across the execution graph.
Traditional MapReduce (left) vs. Tez/Impala/Spark optimized engines (right) (Source: hortonworks.com)
What sets Apache Spark aside from the other frameworks, is the in-memory processing engine as well as the rich set of included libraries (GraphX for graph processing, MLib for Machine Learning, Spark Streaming for mini batch streaming, and Spark SQL) and SDKs (Scala, Python, Java, and R). Please note that these libraries are for distributed processing, so distributed graph processing, distributed machine learning, etc. out-of-the-box.
The amazing performance of Spark’s in-memory engine comes with a trade-off. Tuning and operating Spark pipelines with varying amounts of data requires a lot of manual tuning of configurations, digging through log files, and reading books, articles, and blog posts. And since the execution parallelism can be modified in a fine-grained way, one has to configure/set the number of tasks per JVM, the number of JVMs per worker, and the number of workers as well as all the memory settings (heap, shuffle, and storage) for these executors and the driver.
We speak about Big Data, when we speak about large volumes (> 10s GB), high velocity (> 10.000s records/second) or large variety (binary, text, unstructured, compressed, etc.) of data. We use Big Data systems to store and process massive data sets - either as batch (process partitions of data) or as stream (process single records). In Big Data analytics, we usually differentiate between historical analytics, prediction and modeling.
Apache Hadoop is a collection of services for large-scale distributed storage and processing, mainly HDFS (a distributed filesystem), MapReduce (a processing framework), Apache Yarn (a cluster resource manager), and Apache Zookeeper (a fast distributed key-value storage). Many other services such as Hive, HBase, etc. run on top of Hadoop.
Apache Spark is a fast (100 times faster than traditional MapReduce) distributed in-memory processing engine with high-level APIs, libraries for distributed graph processing and machine learning, and SDKs for Scala, Java, Python and R. It also has support for SQL and streaming.
- Big Data - Wikipedia
- Apache Hadoop - Wikipedia
- Apache Spark - Wikipedia
- Apache Spark - Website
- Apache Hadoop - Website
- Apache Hadoop - Github Repository
- Apache Impala - Website (Overview)
- The Google File System (2003)
- MapReduce: Simplified Data Processing on Large Clusters (2004)
- Introducing Tez - Hortonworks Blog
- Resilient Distributed Datasets: A Fault-Tolerant Abstraction for In-Memory Cluster Computing (2011)
- Apache Tez: A Unifying Framework for Modeling and Building Data Processing Applications (2015)
- Impala: A Modern, Open-Source SQL Engine for Hadoop (2015)