GP

Beyond Conventional Data Warehousing Florian Waas Greenplum Inc. Takeaways • The basics • Who is Greenplum? What is Greenplum Database? • The problem • Data growth and other recent trends in DWH • A look at different customers and their requirements • The solution • Teaching an old dog new tricks: using an RDBMS for massively parallel data processing • Conclusion Greenplum Company Snapshot Greenplum Inc. • What: High-performance database software for Business Intelligence and Data Warehousing • Where: Based in San Mateo, CA • When: Founded in June 2003; product GA in February 2006 • Who: Technical pioneers in data warehousing (Teradata, Microsoft, Oracle, Informix, Tandem, PostgreSQL, …) • Strategic Partner: Powers the Sun Data Warehouse Appliance. Greenplum Database • DBMS • highly scalable • fault-tolerant • high-performance • Based on Postgres • Shared-nothing architecture • Commodity hardware • Currently supported on Solaris, Linux Architecture The Sun Data Warehouse Appliance • 2 dual core AMD processors • 48 Hitachi Deskstar SATA II 7200 rpm 500GB drives • 6 Marvell 8-port Serial- ATA 2.0 Storage Controllers • Leverages Hyper- transport architecture to achieve high- performance I/O capabilities Sample Hardware Configuration – Sun Fire X4500 Data Server Trends in DWH: Data Growth • Growth of customer base • E.g. phone carriers in Asia • Additional data sources • E.g. click-stream and ad-impression data • Data processing – “Data Bunnies” • E.g. intermediate results of analysis, aggregated/expanded • Data will continue to outpace Moore’s Law Trends in DWH: Customers • New Customers • Not your typical DB customers • No pre-existing DB infrastructure • Atypical data: logs, click-stream, etc. • “Weight” of data less significant • E.g. CDR – call detail records • Click-stream vs. sales/transaction records • Often: Reflects behavior, not deliberate purchase decision • “bankers vs. teens” • Analysis as service Trends in DWH: Analysis • Turn-around on reporting • Similar/same requirements despite increased data volume • Automated/on-line decision making process • E.g. ad placement in social network applications • Advanced data analysis processes over massive amounts of data • E.g. Bayesian classification Requirements • Petabyte-scale storage • High-performance query processing • Fault-tolerance/high-availability • Constant loading activity • “Richer processing capabilities” • Leverage parallelism automatically • Cannot move data (size, privacy concerns) • Integrate with existing programming environments • Not strictly a DWH requirement Leveraging Greenplum Database • GPDB designed for • Scalability • High-performance query processing • Fault-tolerance • How to address processing needs? How to use GPDB for Data Processing • Typical installation 10s to 100s of CPU cores • 100s GB memory • 100s TB disk space • Often largest individual system in data center • Slack resources during off-peak times Example: ETL (1) • Customer’s System • 40 nodes • 160 CPU cores • 1 TB main memory • 3.6 TB/h load rate • ETL jobs • 18 hours to process 1 day’s worth of data • 5 serial streams • Load time < 1hr Example: ETL (2) • ETL crucial in daily processing • Mainly data cleaning: string manipulations, conversions, etc. • Hard to parallelize effectively, load-balance • Hard to recover if falling behind • E.g. glitches in ETL logic, data contamination • Desired run time < 4hrs Example: ETL (3) • Load “dirty” raw data directly into GPDB • Trade-off: raw data bulkier • Rewrite ETL logic in SQL • Cleaner program • Run SQL statements on GPDB • Automatic parallelization, fully transparent • Max degree of parallelism • Run time < 3 hrs • Leverage existing query processing infrastructure • Rewrite procedural logic in SQL • Enjoy benefits of SQL • Automatic parallelization • Add UDFs and UDAggs in other languages as needed, e.g. Java, C#, etc. Solution • Query Processing does not mean read-only • Database Technology suffers image problem • SQL is difficult • Declarative programming perceived as non-intuitive • SQL dialects (portability issues) • Too powerful – overwhelming • Requires special skillset/expertise Challenges Summary • Database technology for DWH addresses scalability, fault-tolerance and performance needs • Users are looking for additional mileage from large-scale DWH installations • ELT, and tools like UDFs, UDAggs become more attractive • Existing database technology to be revamped into massively parallel processing engine Beyond Conventional Data Warehousing Florian Waas Greenplum Inc.