Data Warehousing/Mining Comp 150 Data Warehousing Design (not in book)

1. Data Warehousing/MiningComp 150 Data Warehousing Design(not in book) Instructor: Dan Hebert

2. Warehouse Design • What to materialize in the warehouse • Which source data? • Which summary tables? • Which indices? • Influenced by both querying and maintenance • Trade storage space and update time for query speed

3. Designing a Data Warehouse • Data models designed to support DW require optimization strategies for DSS • Design option • Relational model in DW - ROLAP Servers for analysis • Special-purpose multi-dimensional data model in DW (MDDB) - MOLAP Servers for analysis

4. Why is DW Design Different? • DSS: few transactions, each accessing a large number of records • Typical ER designs tend to be complex and difficult to navigate

5. Multi-Dimensional Data • Measures - numerical data being tracked • Dimensions - business parameters that define a transaction • Example: Analyst may want to view sales data (measure) by geography, by time, and by product (dimensions) • Dimensional modeling is a technique for structuring data around the business concepts • ER models describe “entities” and “relationships” • Dimensional models describe “measures” and “dimensions”

6. Dimensional Modeling Using Relational DBMS • Special schema design: star, snowflake • Special indexes: bitmap, multi-table join • Special tuning: maximize query throughput • Proven technology (relational model, DBMS), tend to outperform specialized MDDB especially on large data sets • Products • IBM DB2, Oracle, Sybase IQ, RedBrick, Informix

7. Dimensional Modeling Using Special-Purpose Model (MDDB) • Facts stored in multi-dimensional arrays • Dimensions used to index array • Sometimes on top of relational DB • Products • Pilot, Arbor Essbase, Gentia

8. “Sales by product line over the past six months” “Sales by account between 1990 and 1995” Example Account Info Key columns joining fact table to dimension tables Numerical Measures Prod Code Time Code Acct Code Sales Qty Fact table for measures Product Info Dimension tables Time Info . . .

9. Dimensions are organized into hierarchies E.g., Time dimension: days  weeks  quarters E.g., Product dimension: product  product line  brand Dimensions have attributes Physical architecture describe by Star Schema Dimensional Modeling

10. Example Cont’d Geography Time Geography Code Region Code Region Mgr City Code City Name Time Code Quarter Code Quarter Name Week Code Day Code Day name Sales Geography Code Time Code Account Code Product Code Dollar Amount Units Product Account Product Code Product Name Brand Mgr Brand Code Prod. Line Code Prod. Line Name Prod. Name ... Account Code Key Account Code Account Name Account Type Account Market

11. Dimensional Modeling Cont’d • Fact tables are fully normalized • Dimension tables are denormalized • Repetitively stored for sake of simplicity and performance

12. Extending Dimensional Modeling • Some instances when star schema is not ideal • Denormalized schema may require too much storage • Very large dimension tables are affecting performance negatively • “Snowflake schema” • Normalized dimensions

13. Advantages of Dimensional Modeling • Define complex, multi-dimensional data with simple model • Reduces the number of phycial joins a query has to process • Allows the data warehouse to evolve with rel. low maintenance • HOWEVER! Star schema and rel. DBMS are not the magic solution • Query optimization is still problematic

14. Index Structures • Traditional access methods • B-trees, hash tables, grid files, etc. • Popular in warehouses • Inverted indexes (lists) • Bit map indexes • Join indexes

15. Inverted Index • Index for every keyword • Query: • “Get people with age =20 and name =‘Fred’” • (1) Use age index and retrieve ids: r4,r18,r34,r35 • (2) Use name index and retrieve ids: r18,r52 • (3) Answer is intersection: r18

16. Bit Map Index • Developed for Model 204 DBMS in 1987 1 1 18 0 19 1 1 18 0 20 0 23 20 0 21 0 22 1 0 0 23 0 25 0 26

17. Using Bit Maps • Query: • “Get people with age=20 and name =‘Fred’” • (1) Bit map for age =20: 1101100 • (2) Bit map for name=‘Fred’: 0100000 • (3) Answer is intersection: 0100000 • Good if domain cardinality is small • Bit vectors can be compressed

18. Join Index • Index on one table for a quantity that involves a column value of a different table

19. Aggregation • Process by which low-level data is summarized in advanced and placed into intermediate tables • Speeds up query processing, less ad-hoc • “Show me total US sales for 1990” • How much to aggregate? • Data cube data model • All possible aggregations along all dimensions • Cells contain aggregated values • How much of the cells in cube should be pre-computed?

20. Aggregation Cont’d • Special operators to navigate the hierarchies • Roll-up: remove a dimension element • e.g., Roll-up products to brands • Drill-down (opposite of roll-up), • Slice (defines a subcube) • Various visualization ops (e.g., pivot)

21. Example roll-up to region Dimensions: Time, Product, Geography Attributes: Product (upc, price, …) Geography … … Hierarchies: Product  Brand  … Day  Week  Quarter City  Region  Country Geography NY SF roll-up to brand LA 10 34 56 32 12 56 Juice Milk Coke Cream Soap Bread Product roll-up to week M T W Th F S S Time 56 units of bread sold in LA on M

22. Warehouse DBMS—Buzzwords • Used primarily for decision support(DSS) • A.K.A. On-Line Analytical Processing (OLAP) • Complex queries, substantial aggregation • TPC-D benchmark • Multidimensional data model • Can be implemented either using rel. model or proprietary data model • Multi-dimensional database (MDDB) • Aggregation: Data Cube • All possible groupings and aggregations

23. Warehouse DBMS — Buzzwords (2) • ROLAP vs. MOLAP • Special purpose OLAP servers that directly implement multidimensional data and operations • Roll-up = aggregate on some dimension • Drill-down = deaggregate on some dimension • ROLAP: Oracle, Sybase IQ, RedBrick • MOLAP: Pilot, Essbase, Gentia

24. Warehouse DBMS - Buzzwords (3) • Clients: • Query and reporting tools • Analysis tools • Data mining: discovering patterns of various forms • Poses many new research issues in: • Query processing and optimization • Database design • View management

25. Data Warehouse Physical Design

26. Common Design Activities – OLTP • Schema design (base tables) • Normalization (3NF, BCNF, …) • Schema design (views) • Mostly for convenience, security • Usually NOT for performance • Exception: View indexing [Roussopolous 1982] • Materialize pointers to tuples instead of tuples themselves • Index selection • In practice, use rules of thumb • Tool: DBDSGN [IBM Almaden], RDT for System R

27. Relational Views • Part of the ANSI/SPARC architecture • Derived, virtual table • View definition is an SQL query statement • View update problem • Good for logical data independence, security • How to implement a view for querying • Query modification: modify view query into a query on the underlying base tables • View materialization: physically implementing view as table

28. View Indexing ... • In general, no need for materialized views in OLTP systems • Increase in performance through indexing • Secondary storage space used to be expensive • New idea (N. Roussopolous 1982) - view index • Store index whose elements point to tuples which comprise view • View selection problem: Find a subset of views, which, when indexed, minimizes the total cost of answering all queries as well as cost of maintaining the view structures

29. … View Indexing … • Assume N views to consider, 2N subsets • Can’t do simple enumeration (cost to answer all queries in a given subset) • NP-complete problem • Solution uses search algorithm to approximate the optimal view selection • Potential exponential worst case • Only subset of views needs to be considered • Cost function which computes for each state (set of views + remaining storage) (1) Cost to compute queries, maintenance of current index set + (2) Estimate of incremental cost that must be incurred in extending view set (upper bound on actual cost)

30. … View Indexing • But ... • Algorithm does not consider index selection on views (view indexes) • Indexes have impact on which view indexes to choose • Very simple cost model (maintenance cost ~ size of view) • Problem: Cost of maintaining view is a complex query optimization problem • Cannot be estimated without knowing which subview indexes are chosen • Good first treatment of subject

31. Indexing ... • Which type of index structure, which attribute(s) to index on • “Access path selection” -> DBA • Many choices, depend on many factors • Space-time trade-off • Index selection problem: Which ordering rule for stored records and which non-clustered indices • Database practitioners use rules/guidelines (e.g., SYBASE manual) • Design tools available • Support dba during creation and maintenance of database, i.e., solve the index-selection problem

32. Factors that Influence Index Selection • Maintenance • Storage cost • Global solution depends on index selection of all tables combined

33. Example ORDERS (OrderNo, SuppNo, PartNo, Date, Qty) PARTS (PartNo, Descrip, SuppNo, QtyOnhand, Color, …) Query: SELECT O.SuppNo FROM PARTS P, ORDERS O WHERE O.PartNo = P.PartNo AND O.SuppNo = 15 AND P.QtyOnHand BETWEEN 100 AND 150 • Situation 1: Assume PARTS clustered on Descrip and non-clustered index on PartNo Then: Best clustered index for ORDERS SuppNo • Situation 2: Assume PARTS clustered on PartNo Then: Best clustered index for ORDERS PartNo

34. Data Warehouse Design • Schema design (base tables) • Star schema (dimensions, measures) • Schema design (view/index selection) • Mostly for performance enhancement • Physical warehouse design. Balance three costs: (1) The cost of answering queries using warehouse relations and additional structures (2) Cost of maintaining additional structures (3) Cost of secondary storage

35. WH Schema Design • Tables must map efficiently to the operational requests • OLTP: maximize concurrency, optimize insert/update/delete performance • OLAP: Queries large, complex, ad-hoc, data-intensive, no updates • Query centric view -> Star schema (facts, dimensions) • Widely accepted, intuitive, easy to navigate (query formulation) • Problem: Poor performance on OLTP db engines • Join processing (pair-wise join problem) • Number of pair-wise joins for N tables = N! • e.g., 7 tables -> 5,040 combinations, 5 different join algorithms -> 25,200 combinations

36. Star Schema Join Problem • Heuristic: “pick directly related tables” doesn’t work in star schema • Options: • Join unrelated tables (Cartesian product) • Parallelism (speed-up, scale-up) • New join techniques (e.g., bit vector star joins) in combination with new indexing schemes (e.g., bit maps, variant indexes)

37. Warehouse Access Path (Physical) Design Problem • Materialize user queries as views (reduces cost 1) • How to reduce cost 2 and 3? • “View Index Selection Problem VIS” • Choose a set of supporting views and a set of indexes to materialize such that the total maintenance cost for the warehouse is minimized (cost 2 & 3)

38. Solutions - Relational DB Design Practices • Rel. DB design algorithms must be adapted • View index approach has no index selection, simple cost model (cannot achieve global solution by locally optimizing each materialized subview) • Index selection approach can be extended - but trouble ahead • Algorithms require queries and frequencies as input

39. Solutions - Rule Condition Maintenance • Work on rule condition evaluation • How to evaluate trigger conditions for rules efficiently ( ~ view maintenance problem: rule is triggered whenever view that satisfies its condition becomes non-empty) • Discrimination networks for each rule (view) • RETE model materializes selection and join nodes • TREAT materializes only selection nodes • Incremental evaluation techniques • Recommendations not generally applicable