Querying data through SQL or Hive query language is possible through Spark SQL. Those familiar with RDBMS can easily relate to the syntax of Spark SQL. Locating tables and metadata couldn’t be easier than with Spark SQL. Spark SQL is also known for working with structured and semi-structured data. Structured data is something which has a schema that has a known set of fields. When the schema and the data have no separation, then the data is said to be semi-structured.
Spark SQL Definition: Putting it simply, for structured and semi structured data processing, Spark SQL is used which is nothing but a module of Spark.
Apache Hive was originally designed to run on top of Apache Spark. But, it had considerable limitations:
1) For running the ad-hoc queries, Hive internally launches MapReduce jobs. In the processing of medium-sized datasets, MapReduce lags in performance.
2) If during the execution of a workflow the processing suddenly fails, then Hive can’t resume from the point where it failed as the system returns back to normal.
3) If trash is enabled, it leads to an execution error when encrypted databases are dropped in cascade.
Spark SQL was incepted to over come these inefficiencies.
Architecture of Spark SQL
Spark SQL consists of three main layers such as:
Language API: Spark is compatible and even supported by the languages like Python, HiveQL, Scala, and Java.
SchemaRDD: RDD (resilient distributed dataset) is a special data structure which the Spark core is designed with. As Spark SQL works on schema, tables, and records, you can use SchemaRDD or data frame as a temporary table.
Data Sources: For Spark core, the data source is usually a text file, Avro file, etc. Data sources for Spark SQL are different like JSON document, Parquet file, HIVE tables, and Cassandra database.
Components of Spark SQL
Spark SQL DataFrames: There were some shortcomings on part of RDDs which the Spark DataFrame overcame in the version 1.3 of Spark. First of all, there was no provision to handle structured data and there was no optimization engine to work with it. On the basis of attributes, developers had to optimize each RDD. Spark DataFrame is a distributed collection of data ordered into named columns. You might be knowing what a table is in a relational database. Spark DataFrame is quite similar to that.
Spark SQL Datasets: In the version 1.6 of Spark, Spark dataset was the interface that was added. The catch with this interface is that it provides the benefits of RDDs along with the benefits of optimized execution engine of Apache Spark SQL. To achieve conversion between JVM objects and tabular representation, the concept of encoder is used. Using JVM objects, a dataset can be incepted, and functional transformations like map, filter, etc. have to be used to modify them. The dataset API is available both in Scala and Java, but it is not supported in Python.
Spark Catalyst Optimizer: Catalyst optimizer is the optimizer used in Spark SQL and all queries written by Spark SQL and DataFrame DSL is optimized by this tool. This optimizer is better than the RDD, and hence the performance of the system is increased.
Features of Spark SQL
Let’s take a stroll into the aspects which make Spark SQL so popular in data processing.
Integrated: One can mix SQL queries with Spark programs easily. Structured data can be queried inside Spark programs using either Spark SQL or a DataFrame API. Running SQL queries, alongside analytic algorithms, is easy because of this tight integration.
Hive compatibility: Hive queries can be run as they are as Spark SQL supports HiveQL, along with UDFs (user-defined functions) and Hive SerDes. This allows one to access the existing Hive warehouses.
Unified data access: Loading and querying data from variety of sources is possible. One only needs a single interface to work with structured data which the schema-RDDs provide.
Standard connectivity: Spark SQL includes a server mode with high-grade connectivity to JDBC or ODBC.
Performance and scalability: To make queries agile, alongside computing hundreds of nodes using the Spark engine, Spark SQL incorporates a code generator, cost-based optimizer, and columnar storage. This provides complete mid-query fault tolerance. Note that, as it is mentioned in Hive limitations section, this kind of tolerance was lacking in Hive. Spark has ample information regarding the structure of data, as well as the type of computation being performed which is provided by the interfaces of Spark SQL. This leads to extra optimization from Spark SQL, internally. Faster execution of Hive queries is possible as Spark SQL can directly read from multiple sources like HDFS, Hive, existing RDDs, etc.
There is a lot to learn about Spark SQL as how it is applied in the industry scenario, but the below three use cases can give an apt idea:
Twitter sentiment analysis: Initially, you used to get all data from Spark streaming. Later, Spark SQL came into the picture to analyze everything about a topic, say, Narendra Modi. Every tweet regarding Modi is gathered, and then Spark SQL does its magic by classifying tweets as neutral tweets, positive tweets, negative tweets, very positive tweets, and very negative tweets. This is just one of the ways how sentiment analysis is done. This is useful in target marketing, crisis management, and service adjusting.
Stock market analysis: As you are streaming data in the real time, you can also do the processing in the real time. Stock movements and market movements generate so much data and traders need an edge, an analytics framework, which will calculate all the data in the real time and provide the most rewarding stock or contract, all within the nick of time. As said earlier, if there is a need for a real-time analytics framework, then Spark, along with its components, is the technology to be considered.
Banking: Real-time processing is required in credit card fraud detection. Assume that a transaction happens in Bangalore where there is a purchase worth 4,000 rupees has been done swiping a credit card. Within 5 minutes, there is another purchase of 10,000 rupees in Kolkata swiping the same credit card. Banks can make use of real-time analytics provided by Spark SQL in detecting fraud in such cases.
Apache Software Foundation has given a carefully-thought-out component for real-time analytics. When the analytics world starts seeing the shortcomings of Hadoop in providing real-time analytics, then migrating to Spark will be the obvious outcome. Similarly, when the limitations of Hive become more and more apparent, then users will obviously shift to Spark SQL. It is to be noted that the processing which takes 10 minutes to perform via Hive can be achieved in less than a minute if one uses Spark SQL. On top of that the migration is also easy as Hive support is provided by Spark SQL. Here comes the great opportunity for those who want to learn Spark SQL and DataFrames. Currently, there aren’t many professionals who can work around in Hadoop. The demand is still higher for Spark, and those who learn it and have hands-on experience on it will be in great demand when the technology is used more and more in the future.
You can get ahead the rest of analytics professionals by learning Spark SQL right now. Intellipaat’s Spark SQL training is designed for you!