What is a Data Warehouse? And Why Are So Many Schools Setting Them Up? Richard Goerwitz What Is a Data Warehouse? Nobody can agree So I’m not actually going to define a DW Don’t feel cheated, though By the end of this talk, you’ll • Understand key concepts that underlie all warehouse implementations (“talk the talk”) • Understand the various components out of which DW architects construct real-world data warehouses • Understand what a data warehouse project looks like Why Are Schools Setting Up Data Warehouses? A data warehouse makes it easier to: • Optimize classroom, computer lab usage • Refine admissions ratings systems • Forecast future demand for courses, majors • Tie private spreadsheet data into central repositories • Correlate admissions and IR data with outcomes such as: GPAs Placement rates Happiness, as measured by alumni surveys • Notify advisors when extra help may be needed based on Admissions data (student vitals; SAT, etc.) Special events: A-student suddenly gets a C in his/her major Slower trends: Student’s GPA falls for > 2 semesters/terms • (Many other examples could be given!) Better information = better decisions • Better admission decisions • Better retention rates • More effective fund raising, etc. Talking The Talk To think and communicate usefully about data warehouses you’ll need to understand a set of common terms and concepts: • OLTP • ODS • OLAP, ROLAP, MOLAP • ETL • Star schema • Conformed dimension • Data mart • Cube • Metadata Even if you’re not an IT person, pay heed: • You’ll have to communicate with IT people • More importantly: Evidence shows that IT will only build a successful warehouse if you are intimately involved! OLTP OLTP = online transaction processing The process of moving data around to handle day-to-day affairs • Scheduling classes • Registering students • Tracking benefits • Recording payments, etc. Systems supporting this kind of activity are called transactional systems Transactional Systems Transactional systems are optimized primarily for the here and now • Can support many simultaneous users • Can support heavy read/write access • Allow for constant change • Are big, ugly, and often don’t give people the data they want As a result a lot of data ends up in shadow databases Some ends up locked away in private spreadsheets Transactional systems don’t record all previous data states Lots of data gets thrown away or archived, e.g.: • Admissions data • Enrollment data • Asset tracking data (“How many computers did we support each year, from 1996 to 2006, and where do we expect to be in 2010?”) Simple Transactional Database Map of Microsoft Windows Update Service (WUS) back-end database • Diagrammed using Sybase PowerDesigner Each green box is a database “table” Arrows are “joins” or foreign keys This is simple for an OLTP back end More Complex Example Recruitment Plus back-end database Used by many admissions offices Note again: • Green boxes are tables • Lines are foreign key relationships • Purple boxes are views Considerable expertise is required to report off this database! Imagine what it’s like for even more complex systems • Colleague • SCT Banner (over 4,000 tables) The “Reporting Problem” Often we require OLTP data as a snapshot, in a spreadsheet or report Reports require querying back-end OLTP support databases But OLTP databases are often very complex, and typically • Contain many, often obscure, tables • Utilize cryptic, unintuitive field/column names • Don’t store all necessary historical data As a result, reporting becomes a problem – • Requires special expertise • May require modifications to production OLTP systems • Becomes harder and harder for staff to keep up! Workarounds Ways of working around the reporting problem include: 1. Have OLTP system vendors do the work • Provide canned reports • Write reporting GUIs for their products 2. Hire more specialists • To create simplified views of OLTP data • To write reports, create snapshots 3. Periodically copy data from OLTP systems to a place where • The data is easier to understand • The data is optimized for reporting • Easily pluggable into reporting tools ODS ODS = operational data store ODSs were an early workaround to the “reporting problem” To create an ODS you • Build a separate/simplified version of an OLTP system • Periodically copy data into it from the live OLTP system • Hook it to operational reporting tools An ODS can be an integration point or real-time “reporting database” for an operational system It’s not enough for full enterprise-level, cross- database analytical processing OLAP OLAP = online analytical processing OLAP is the process of creating and summarizing historical, multidimensional data • To help users understand the data better • Provide a basis for informed decisions • Allow users to manipulate and explore data themselves, easily and intuitively More than just “reporting” Reporting is just one (static) product of OLAP OLAP Support Databases OLAP systems require support databases These databases typically • Support fewer simultaneous users than OLTP back ends • Are structured simply; i.e., denormalized • Can grow large Hold snapshots of data in OLTP systems Provide history/time depth to our analyses • Are optimized for read (not write) access • Updated via periodic batch (e.g., nightly) ETL processes ETL Processes ETL = extract, transform, load • Extract data from various sources • Transform and clean the data from those sources • Load the data into databases used for analysis and reporting ETL processes are coded in various ways • By hand in SQL, UniBASIC, etc. • Using more general programming languages • In semi-automated fashion using specialized ETL tools like Cognos Decision Stream Most institutions do hand ETL; but note well: • Hand ETL is slow • Requires specialized knowledge • Becomes extremely difficult to maintain as code accumulates and databases/personnel change! Where Does the Data Go? What sort of a database do the ETL processes dump data into? Typically, into very simple table structures These table structures are: • Denormalized • Minimally branched/hierarchized • Structured into star schemas So What Are Star Schemas? Star schemas are collections of data arranged into star-like patterns • They have fact tables in the middle, which contain amounts, measures (like counts, dollar amounts, GPAs) • Dimension tables around the outside, which contain labels and classifications (like names, geocodes, majors) • For faster processing, aggregate fact tables are sometimes also used (e.g., counts pre-averaged for an entire term) Star schemas should • Have descriptive column/field labels • Be easy for users to understand • Perform well on queries A Very Simple Star Schema Data Center UPS Power Output Dimensions: Phase Time Date Facts: Volts Amps Etc. A More Complex Star Schema Freshman survey data (HERI/CIRP) Dimensions: • Questions • Survey years • Data about test takers Facts: • Answer (text) • Answer (raw) • Count (1) Oops • Not a star • Snowflaked! Oops, answers should have been placed in their own dimension (creating a “factless fact table”). I’ll demo a better version of this star later! Data Marts One definition: • One or more star schemas that present data on a single or related set of business processes Data marts should not be built in isolation They need to be connected via dimensional tables that are • The same or subsets of each other • Hierarchized the same way internally So, e.g., if I construct data marts for… • GPA trends, student major trends, enrollments • Freshman survey data, senior survey data, etc. …I connect these marts via a conformed student dimension • Makes correlation of data across star schemas intuitive • Makes it easier for OLAP tools to use the data • Allows nonspecialists to do much of the work Simple Data Mart Example UPS Battery star By battery Run-time % charged Current Input star By phase Voltage Current Output star By phase Voltage Current Sensor star By sensor Temp Humidity Note conformed date, time dimensions! CIRP Star/Data Mart CIRP Freshman survey data Corrected from a previous slide Note the CirpAnswer dimension Note student dimension (ties in with other marts) CIRP Mart in Cognos BI 8 ROLAP, MOLAP ROLAP = OLAP via direct relational query • E.g., against a (materialized) view • Against star schemas in a warehouse MOLAP = OLAP via multidimensional database (MDB) • MDB is a special kind of database • Treats data kind of like a big, fast spreadsheet • MDBs typically draw data in from a data warehouse Built to work best with star schemas Data Cubes The term data cube means different things to different people Various definitions: 1. A star schema 2. Any DB view used for reporting 3. A three-dimensional array in a MDB 4. Any multidimensional MDB array (really a hypercube) Which definition do you suppose is technically correct? Metadata Metadata = data about data In a data warehousing context it can mean many things • Information on data in source OLTP systems • Information on ETL jobs and what they do to the data • Information on data in marts/star schemas • Documentation in OLAP tools on the data they manipulate Many institutions make metadata available via data malls or warehouse portals, e.g.: • University of New Mexico • UC Davis • Rensselear Polytechnic Institute • University of Illinois Good ETL tools automate the setup of malls/portals! The Data Warehouse OK now we’re experts in terms like OLTP, OLAP, star schema, metadata, etc. Let’s use some of these terms to describe how a DW works: • Provides ample metadata – data about the data • Utilizes easy-to-understand column/field names • Feeds multidimensional databases (MDBs) • Is updated via periodic (mainly nightly) ETL jobs • Presents data in a simplified, denormalized form • Utilizes star-like fact/dimension table schemas • Encompasses multiple, smaller data “marts” • Supports OLAP tools (Access/Excel, Safari, Cognos BI) • Derives data from (multiple) back-end OLTP systems • Houses historical data, and can grow very big A Data Warehouse is Not… Vendor and consultant proclamations aside, a data warehouse is not: • A project With a specific end date • A product you buy from a vendor Like an ODS (such as SCT’s) A canned “warehouse” supplied by iStrategy Cognos ReportNet • A database schema or instance Like Oracle SQL Server • A cut-down version of your live transactional database Kimball & Caserta’s Definition According to Ralph Kimball and Joe Caserta, a data warehouse is: A system that extracts, cleans, conforms, and delivers source data into a dimensional data store and then supports and implements querying and analysis for the purpose of decision making. Another def.: The union of all the enterprise’s data marts Aside: The Kimball model is not without some critics: • E.g., Bill Inmon Example Data Warehouse (1) This one is RPI’s 5 parts: • Sources • ETL stuff • DW proper • Cubes etc. • OLAP apps Example Data Warehouse (2) Caltech’s DW Five Parts: • Source systems • ETL processes • Data marts • FM/metadata • Reporting and analysis tools • Note: They’re also customers of Cognos! So Where is Colorado College? Phil Goldstein (Educause Center for Applied Research fellow) identifies the major deployment levels: • Level 1: Transactional systems only • Level 2a: ODS or single data mart; no ETL • Level 2: ODS or single data mart with ETL tools • Level 3a: Warehouse or multiple marts; no ETL; OLAP • Level 3b: Warehouse or multiple marts; ETL; OLAP • Level 3: Enterprise-wide warehouse or multiple marts; ETL tools; OLAP tools Goldstein’s study was just released in late 2005 It’s very good; based on real survey data Which level is Colorado College at? Implementing a Data Warehouse In many organizations IT people want to huddle and work out a warehousing plan, but in fact • The purpose of a DW is decision support • The primary audience of a DW is therefore College decision makers • It is College decision makers therefore who must determine Scope Priority Resources Decision makers can’t make these determinations without an understanding of data warehouses It is therefore imperative that key decision makers first be educated about data warehouses • Once this occurs, it is possible to Elicit requirements (a critical step that’s often skipped) Determine priorities/scope Formulate a budget Create a plan and timeline, with real milestones and deliverables! Is This Really a Good Plan? Sure, according to Phil Goldstein (Educause Center for Applied Research) He’s conducted extensive surveys on “academic analytics” (= business intelligence for higher ed) His four recommendations for improving analytics: 1. Key decisionmakers must lead the way 2. Technologists must collaborate • Must collect requirements • Must form strong partnerships with functional sponsors 3. IT must build the needed infrastructure • Carleton violated this rule with Cognos BI • As we discovered, without an ETL/warehouse infrastructure, success with OLAP is elusive 4. Staff must train and develop deep analysis skills Goldstein’s findings mirror closely the advice of industry heavyweights – Ralph Kimball, Laura Reeves, Margie Ross, Warren Thornthwaite, etc. Isn’t a DW a Huge Undertaking? Sure, it can be huge Don’t hold on too tightly to the big- sounding word, “warehouse” Luminaries like Ralph Kimball have shown that a data warehouse can be built incrementally • Can start with just a few data marts • Targeted consulting help will ensure proper, extensible architecture and tool selection What Takes Up the Most Time? You may be surprised 90 to learn what DW step 80 70 takes the most time 60 Hardware Try guessing which: 50 40 East Database ETL West Schemas • Hardware 30 North OLAP tools • Physical database setup 20 10 • Database design 0 1st Qtr 2nd Qtr 3rd Qtr 4th Qtr • ETL • OLAP setup Acc. to Kimball & Caserta, ETL will eat up 70% of the time. Other analysts give estimates ranging from 50% to 80%. The most often underestimated part of the warehouse project! Eight Month Initial Deployment Step Duration Step Duration Begin educating decision makers 21 days Secure, configure network 1 day Collect requirements 14 days Deploy physical “target” DB 4 days Decide general DW design 7 days Learn/deploy ETL tool 28 days Determine budget 3 days Choose/set up modeling tool 21 days Identify project roles 1 day Design initial data mart 7 days Eval/choose ETL tool 21 days Design ETL processes 28 days Eval/choose physical DB 14 days Hook up OLAP tools 7 days Spec/order, configure server 20 days Publicize, train, train 21 days Conclusion Information is held in transactional systems • But transactional systems are complex • They don’t talk to each other well; each is a silo • They require specially trained people to report off of For normal people to explore institutional data, data in transactional systems needs to be • Renormalized as star schemas • Moved to a system optimized for analysis • Merged into a unified whole in a data warehouse Note: This process must be led by “customers” • Yes, IT people must build the infrastructure • But IT people aren’t the main customers So who are the customers? • Admissions officers trying to make good admission decisions • Student counselors trying to find/help students at risk • Development offers raising funds that support the College • Alumni affairs people trying to manage volunteers • Faculty deans trying to right-size departments • IT people managing software/hardware assets, etc….