Hadoop meets SQL

Hadoop meets SQL

Big data technologies like hadoop are providing enterprises a cost effective way to store and analyze data. Enterprises are looking at using hadoop  to augment their traditional data warehouse.  Compared to traditional data warehouse solutions, hadoop can scale using commodity hardware and can be used to store both structured as well as unstructured data. Traditional data warehouses based on relational database technologies have been around for a long time and have mature sets of tools for querying and analysis. Business Users use SQL as the query language to perform ad-hoc queries against these warehouses. Also reporting tools like Cognos , Business Objects, MicroStrategy rely on SQL heavily.  The real value of hadoop is realized when users can access and perform ad-hoc queries data directly on hadoop using tools that support SQL.

Querying a hadoop data store means knowing Map-Reduce programming or writing Pig and Hive scripts.  Hadoop at its core consists of HDFS storage and Map-Reduce Engine. Map-reduce programs being typically written in Java are difficult to write. They are not at all easy to use from a business users perspective.  The ability to use SQL to analyze data stored in hadoop will help in making hadoop go main stream.  This will also enable the business user to reuse their existing SQL knowledge to analyze the data store in hadoop. Various initiatives are underway both in the open source as well as in various companies to solve the problem of enabling SQL on Hadoop.  The following are some of the most common ones.

Hive: Facebook developed Hive as a way of bringing a SQL Like interface for querying hadoop. A Hive Warehouse needs to be created first that provides a schema on top of the data stored in HDFS. Using Hive QL ( Hive Query language) we can use SQL like syntax to query data stored in Hadoop HDFS.  The Sql Support for Hive is very limited at this point.  It does not offer support for full ANSI Sql.  In the case of joins Hive supports  ANSI Join syntax only. It only supports equi-joins at this point. Hive also does not support correlated sub queries which are commonly used in most traditional warehouse queries . Hive is not designed for low latency queries. Hive actually launches map-reduce jobs in the background. So even for small Hive tables, the query will take several minutes. It is really designed to run queries against massive amounts of data, where the query will return results in a few hours. Hive is not suited for real time querying and analysis

Impala:  Cloudera’s Impala provide a fast real time query capability for your data stored in Hadoop using Sql. Impala is based on Google’s Dremel paper. Currently Impala supports a subset of Ansi-92 Sql. There are still some issues on the join table sizes with Impala. If the join results do not fit into the amount of memory available in the hadoop cluster, the join would fail.  Your queries are limited by the amount of memory you have. Impala currently supports hash Joins. Cloudera does provide some recommendations on memory size for data nodes based on their beta customer experience Impala also provides connection using JDBC and ODBC as well as a command line tool.  Cloudera provides some interesting performance data on their site for Impala(http://blog.cloudera.com/blog/2013/05/cloudera-impala-1-0-its-here-its-real-its-already-the-standard-for-sql-on-hadoop/).

BigSql:  BigSql is an enterprise class SQL query engine from IBM and is available on IBM’s hadoop distribution BigInsights version 2.1.  BigSql provides full ANSI Sql support as well as support for correlated subqueries. There is no memory size limitation for join tables. BigSql runs on top of a Hive Warehouse or Hbase . BigSql also provides an option for using adaptive Map-reduce to improve the performance of Map-reduce Jobs.  Adaptive Map-reduce comes from IBM’s experience in high performance computing clusters. BigSql supports standard as well as ANSI Join, cross join and non-equi join syntax. It also provides a much wider support for data types compared to Hive. Any BI and visualization tool that uses JDBC/ODBC drivers can use BigSql to connect to a BigInsights Hadoop cluster. BigSql also comes with a command line query tool called JSqsh, which is similar to Oracle Sql*Plus or MySql command line tools. IBM has announced their free download for BigInsights Quickstart VM that comes bundled with BigSql (http://www-01.ibm.com/software/data/infosphere/biginsights/quick-start/)

Google BigQuery: Google launched BigQuery based on their Dremel tool to enable real time querying of their data using Sql Queries. BigQuery provides both synchronous and asynchronous running of queries. However BigQuery is available only if data is loaded into Google’s cloud storage. Google provides a set of RESTFUL API’s to access the queries as well. It supports joins, but there is a table size limitation for one the joining tables. BigQuery is a powerful querying tool if you are using Google Cloud to store your data.

HAWQ: Greenplum announced their HAWQ query engine that runs on top of their  Pivotal HD ( Greenplum’s hadoop Distribution) and can run execute SQL queries against Hadoop. With HAWQ users can query data stored in Hbase, Hive or HDFS. HAWQ  uses the same query optimizer that is used by Greenplum DB.  HAWQ uses dynamic pipelining which is the combination of a bunch of different Greenplum technologies that have been built for the parallel relational database. The dynamic pipelining is a job scheduler for queries (different than job tracker and Name node used by hadoop). This would be a good option for customers who already use Greenplum DB as their warehouse.  They can run the same queries against the Greenplum warehouse as well as the Pivotal HD hadoop cluster.  HAWQ is a proprietary solution from Greenplum.

What is coming next?

Other open source initiatives are in place to address the issue of providing real time query features on Hadoop.

Apache Drill: This is a new open source initiative based on Googles Dremel paper.  The aim is to provide near real time query capabilities on hadoop similar to Google BigQuery.

Stinger : The Stinger Initiative is a Apache project managed by HortonWorks and Microsoft. The aim is to leverage hadoop 2.0 and Yarn to help improve the performance and sql capabilties of Hive. With Stinger, Hive queries will be 100X faster than the current queries. Hive will also support  sub queries and better alignment with ANSI Sql.


Key things to consider when looking for Sql Capabilities n Hadoop:

  • Deploying a hadoop distribution that is open source or has open source support will be important.  Hadoop has been a open source initiative and there are lot of contributors who are adding and building capabilities into the platform.  Companies like IBM and EMC Greenplum  are now adding enterprise class features into it and enabling integration with other enterprise data stores.
  • These query tools are designed to be used by small group of users. All the Sql type query tools , run map reduce jobs in the background. Map-reduce inherently scale up very well but does not scale down. Enable a limited set of business users  to run SQL queries against Hadoop using these tools.
  • Enterprise class SQL features are available thru IBM BigSql and GreenPlum HAWQ.  These two query engines are both relatively new.  They are also proprietary solutions tied to their own hadoop distributions.  Both of them do aim to support full ANSI SQL and an enable enterprises  to port and reuse existing queries. If you have an existing warehouse running on DB2, Teradata etc that you want to augment and reuse queries quickly, a BigSql running on a BigInsights hadoop cluster would be a logical choice.
  • All the Query engines have limitations and are not as robust and mature as the standard SQL query tools that are available on traditional warehouses.  To expect business users to quickly use these tools is still a stretch and will need help from technical experts who understand hadoop.

Will Hadoop replace or augment your Enterprise Data Warehouse?

Will Hadoop replace or augment your Enterprise Data Warehouse?

There is all the buzz about BigData and Hadoop these days and its potential for replacing Enterprise Data Warehouse (EDW). The promise of Hadoop has been the ability to store and process massive amounts of data using commodity hardware that scales extremely well and at very low cost. Hadoop is good for batch oriented work and not really good at OLTP workloads.

The logical question then is do enterprises still need the EDW. Why not simply get rid of the expensive warehouse and deploy a Hadoop cluster with Hbase and Hive. After all you never hear about Google or Facebook using data warehouse systems from Oracle or Teradata or  Greenplum.

Before we get into that a little bit of overview on how Hadoop stores data. Hadoop comprise of two components. The Hadoop Distributed File System (HDFS) and the Map-Reduce Engine.  HDFS enables you to store all kinds of data (structured as well as unstructured) on commodity servers. Data is divided into blocks and  distributed across data nodes. The data itself is processed by using Map-Reduce programs which are typically  written in Java. NoSql Databases like HBase and Hive provide a layer on top of  HDFS storage that enables end users to use Sql language. In addition BI reporting ,visualization and analytical tools like Cognos, Business Objects, Tableau, SPSS, R etc can now connect to Hadoop/Hive.

A traditional EDW stores structured data from OLTP and back office ERP systems into a relational database using expensive storage arrays with RAID disks. Examples of this structured data may be your customer orders, data from your financial systems, sales orders, invoices etc.  Reporting tools like Cognos, Business Objects, SPSS etc are used to run reports and perform analysis on the data.

So are we ready to dump the EDW and move to Hadoop for all our Warehouse needs. There are some things the EDW does very well that Hadoop is still not very good at:

  • Hadoop and HBase/Hive are all still very IT  focused. They need people with lot of expertise in writing Map reduce Programs in Java, Pig etc. Business Users who actually need the data are not in a position to run ad-hoc queries and analytics easily without involving IT. Hadoop is still maturing and needs lot of IT Hand holding to make it work.
  • EDW is well suited for many common business processes, such as monitoring sales by geography, product or channel; extract insight from customer surveys;  cost and profitability analyses. The data is loaded into pre-defined schemas/data marts and business users can use familiar tools to perform analysis and run ad-hoc Sql Queries.
  • Most EDW come with pre built adaptors for various ERP Systems and databases. Companies have built complex ETL, data marts , analytics, reports etc on top of these warehouse. It will be extremely expensive, time consuming and risky  to recode that into a new Hadoop environment. People with Hadoop/Map-reduce expertise are not readily available and are in short supply.

Augment your EDW with Hadoop to add new capabilities and Insight. For the next couple of years, as the Hadoop/BigData landscape evolves augment and enhance your EDW with a Hadoop/BigData cluster as follows:

  • Continue to store summary structured data from your OLTP and back office systems into the EDW.
  • Store unstructured data into Hadoop that does not fit nicely into “Tables”.  This means all the communication with your customers from phone logs, customer feedbacks, GPS locations, photos, tweets, emails, text messages etc can be stored in Hadoop. You can store this lot more cost effectively in Hadoop.
  • Co-relate data  in your EDW with the data in your Hadoop cluster to get better insight about your customers, products, equipments etc. You can now use this data for analytics that are computation intensive like clustering and targeting. Run ad-hoc analytics and models against your data in Hadoop, while you are still transforming and loading your EDW.
  • Do not build Hadoop capabilities within your enterprise in a silo. Big Data/Hadoop technologies should work in tandem with and extend the value of your existing data warehouse and analytics technologies.
  • Data Warehouse vendors are adding capabilities of Hadoop and Map-reduce into their offerings. When adding Hadoop capabilities, I would recommend going with a  vendor that supports and enhances the open source Hadoop distribution.

In a few years as newer and better analytical and reporting capabilities develop on top of Hadoop, it may eventually be a good platform for all your warehousing needs. Solutions like IBM‘s BigSql and Cloudera’s Impala will make it easier for business users to move more of their warehousing needs to Hadoop by improving query performance and Sql capabilities.