1. Database Systems I Data Warehousing

2. 2222 I NTRODUCTION Increasingly, organizations are analyzing current and historical data to identify useful patterns and support business strategies (Decision Support). Emphasis is on complex, interactive, exploratory analysis of very large datasets created by integrating data from across all parts of an enterprise; data is fairly static. Contrast such On-Line Analytic Processing (OLAP) with traditional On-line Transaction Processing (OLTP) : mostly long queries, instead of short update transactions.

3. 3333 D ATA W AREHOUSING A Data Warehouse is a collection of data for the purpose of decision support that is: subject oriented - Data that gives information about a particular subject instead of about a company's ongoing operations. Integrated - Data that is gathered into the data warehouse from a variety of sources and merged into a coherent whole. time variant - All data in the data warehouse is identified with a particular time period. non volatile - Data is stable in a data warehouse. More data is added but data is never removed. This enables management to gain a consistent picture of the business. (Source: "What is a Data Warehouse?" W.H. Inmon, Prism, Volume 1, Number 1, 1995).

4. 4444 D ATA W AREHOUSING Integrates data from several operational (OLTP) databases. Keeps (relevant part of the) history of the data. Views data at a more abstract level than OLTP systems (aggregate over many detail records).

5. 5555 DBS FOR D ECISION S UPPORT Database: an application that manages data and allows fast storage and retrieval of that data A transaction makes small changes Data Warehouse: a centralized database Consolidate data from many sources in one large repository Loading, periodic synchronization of replicas Semantic integration OLAP: Complex SQL queries and views Queries based on “multidimensional” view of data and spreadsheet-style operations. Interactive and “online” ( manual ) analysis. Data Mining: Automatic discovery of interesting trends and other patterns.

6. 6666 OLTP VS DW

7. 7777 Data warehouses and OLTP systems have very different requirements: Workload DW designed to accommodate ad hoc queries might not know the workload of your data warehouse in advance should be optimized to perform well for a wide variety of queries OLTP support only predefined operations (ordering books, searching, etc) Data modifications DW Updated on a regular basis by the ETL process (run nightly or weekly) Uses bulk data modification techniques The end users of a data warehouse do not directly update the data warehouse OLTP end users routinely issue individual data modification statements always up to date reflects the current state of each business transaction

8. 8888 OLTP VS DW Data warehouses and OLTP systems have very different requirements: Historical data DW usually store many months or years of data to support historical analysis OLTP usually store data from only a few weeks or months only historical data as needed to successfully meet the current transaction Typical operations DW A typical query scans thousands or millions of rows For example, "Find the total sales for all customers last month." OLTP accesses only a handful of records For example, "Retrieve the current order for this customer."

9. 9999 OLTP VS DW Data warehouses and OLTP systems have very different requirements: Schema design DW denormalized or partially denormalized schemas Optimized for query performance Called a Star schema OLTP fully normalized schemas to optimize update/insert/delete performance To guarantee data consistency

10. 10 S TAR VS DB S CHEMA consists of one or more fact tables referencing any number of dimension tables more effective for handling simpler queries Fact Table consists of the measurements, metrics or facts of a business process often located at the centre of a star schema surrounded by dimension tables Ex: SALES fact table might contain data uniquely defined by a day, product and store. Dimension tables provides the means to "slice and dice" data provide structured labeling information to otherwise unordered numeric measures Ex: "Customer", "Date", and "Product"

11. 11

12. 12 S TAR S CHEMA Consists of a collection of dimensions (independent variables) and (numeric) facts (dependent variables). Each entry (cell) aggregates the value(s) of the measure(s) for all records that fall into that cell i.e. for all records that in each dimension have attribute values corresponding to the value of the cell in this dimension. Example: dimensions Product (pid), Location (locid), and Time (timeid) and measure Sales.

13. 13 S TAR S CHEMA In order to see the total sales for a particular month for a particular category, a SQL query would look something like this: SELECT Sum(SalesFact.SalesDollars) AS SumOfSalesDollars FROM TimeDimension INNER JOIN (ProductDimension INNER JOIN SalesFact ON ProductDimension.ProductID = SalesFact.ProductID) ON TimeDimension.TimeID = SalesFact.TimeID WHERE ProductDimension.Category='Brass Goods' AND TimeDimension.Month=3 AND TimeDimension.Year=1999

14. 14 More dimensions? Notice fact table contains IDs and very little data

15. 15 M ULTIDIMENSIONAL D ATA M ODEL pid timeid locid sales 8 10 10 30 20 50 25 8 15 1 2 3 timeid pid 11 12 13 locid Slice locid=1 is shown Tabular representation Multidimensional representation

16. 16 D ATA W AREHOUSING I SSUES Semantic integration: When getting data from multiple sources, must eliminate mismatches, e.g., different currencies, DB schemas. Heterogeneous sources: Must access data from a variety of source formats and repositories. Replication capabilities can be exploited here. Load, refresh, purge: Must load data, periodically refresh it, and purge too-old data. Metadata management: Must keep track of source, loading time, and other information for all data in the warehouse.

17. D ETAILS

18. 18 M ULTIDIMENSIONAL D ATA M ODEL For each dimension, the set of values can be organized in a concept hierarchy (subset relationship), e.g. PRODUCTTIMELOCATION category week month state pname date city year quarter country

19. 19 M ULTIDIMENSIONAL D ATA M ODEL Multidimensional data can be stored physically in a (disk-resident, persistent) array; called MOLAP (multi-dimensional OLAP) systems. Alternatively, can store as a relation; called ROLAP (relational OLAP) systems. The main relation, which relates dimensions to a measure, is called the fact table. Each dimension can have additional attributes and an associated dimension table. E.g., fact table Transactions(pid, locid, timeid, sales) and (one of the) dimension table Products(pid, pname, category, price) Fact tables are much larger than dimensional tables.

20. 20 OLAP Q UERIES Influenced by SQL and by spreadsheets. A common operation is to aggregate a measure over one or more dimensions. Find total sales. Find total sales for each city, or for each state. Find top five products ranked by total sales. We can aggregate at different levels of a dimension hierarchy. A roll-up operation aggregates along the next higher level of the dimension hierarchy. E.g., given total sales by city, we can roll-up to get sales by state.

21. 21 OLAP Q UERIES Drill-down : The inverse of roll-up. E.g., given total sales by state, can drill-down to get total sales by city. E.g., can also drill-down on different dimension to get total sales by product for each state. Pivoting : Aggregation on selected dimensions. E.g., pivoting on Location and Time yields this cross-tabulation: Slicing and Dicing : Equality and range selections on one or more dimensions. 63 81 144 38 107 145 75 35 110 WI CA Total 1995 1996 1997 176 223 339 Total

22. 22 D EMO http://www.youtube.com/watch?v=- j5J7lXav7Y#t=12m30s From 11:55 to 18:44

23. 23 C OMPARISON WITH SQL Q UERIES The cross-tabulation obtained by pivoting can also be computed using a collection of SQL queries, e.g. SELECT SUM (S.sales) FROM Sales S, Times T, Locations L WHERE S.timeid=T.timeid AND S.timeid=L.timeid GROUP BY T.year, L.state SELECT SUM (S.sales) FROM Sales S, Times T WHERE S.timeid=T.timeid GROUP BY T.year SELECT SUM (S.sales) FROM Sales S, Location L WHERE S.timeid=L.timeid GROUP BY L.state SELECT SUM (S.sales) FROM Sales S

24. 24 T HE C UBE O PERATOR Generalizing the previous example, if there are d dimensions, we have 2 d possible SQL GROUP BY queries that can be generated through pivoting on a subset of dimensions (without considering selections of specific values for certain dimensions). A Data Cube is a multi-dimensional model of a datawarehouse where the domain of each dimension is extended by the special value „ALL“ with the semantics of aggregating over all values of the corresponding dimension.

25. 25 T HE C UBE O PERATOR The Cube Operator computes the measures for all cells (evaluates all possible GROUP BY queries) at the same time. It can be much more efficiently processed than the set of all corresponding (independent) SQL GROUP BY queries. Observation: The results of more generalized queries (with fewer GROUP BY attributes) can be derived from more specialized queries (with more GROUP BY attributes) by aggregating over the irrelevant GROUP BY attributes.

26. 26 T HE C UBE O PERATOR Process more specialised queries first and, based on their results, determine the outcome of more generalised queries. Significant reduction of I/O cost, since intermediate results are much smaller than original (fact) table.

27. 27 T HE C UBE O PERATOR Lattice of GROUP-BY queries of a CUBE query w.r.t. derivability of the results Example {pid, locid, timeid} {pid, locid}{pid, timeid}{locid, timeid} {pid} {locid}{timeid} {} {A,B,...}: set of GROUP BY attributes, X Y: Y derivable from X

28. 28 T HE C UBE O PERATOR An entry of a data cube is called a cell. The number of cells of a datacube with d dimensions is Each SQL group corresponds to a datacube cell. A single GROUP BY of the 2 d different SQL GROUP BY queries can compute the measures for multiple datacube cells.

29. I MPLEMENTATION I SSUES

30. 30 I MPLEMENTATION I SSUES In the following, adopting a ROLAP implementation. Fact table normalized (redundancy free). Dimension tables un-normalized. Dimension tables are small; updates/inserts/deletes are rare. So, anomalies less important than query performance. This kind of schema is very common in OLAP applications, and is called a star schema ; computing the join of all these relations is called a star join.

31. 31 I MPLEMENTATION I SSUES Example star schema pricecategorypnamepidcountrystatecitylocid saleslocidtimeidpid holiday_flagweekdatetimeidmonthquarteryear SALES TIMES PRODUCTSLOCATIONS Fact table: Sales Dimension tables: Times, Products, Locations

32. 32 B ITMAP I NDEXES New indexing techniques: Bitmap indexes, Join indexes, array representations, compression, precomputation of aggregations, etc. Example Bitmap index: (maps the index to a bit) sex custid name sex rating rating M F Bit-vector: 1 bit for each possible value. One row per record.

33. 33 B ITMAP I NDEXES Selections can be processed using (efficient!) bit- vector operations. Example 1: Find all male customers Example 2: Find all male customer with a rating of 3  AND the relevant bit-vectors from the bitmap indexes for sex and rating sex custid name sex rating rating

34. 34 J OIN I NDEXES Consider the join of Sales, Products, Times, and Locations, possibly with additional selection conditions (e.g., country=“USA”). A join index can be constructed to speed up such joins (in a relatively static data warehouse). It basically materializes the result of a join. The index contains [s,p,t,l] if there are tuples with sid s in Sales, pid p in Products, timeid t in Times and locid l in Locations that satisfy the join (and selection) conditions.

35. 35 J OIN I NDEXES Problem: Number of join indexes can grow rapidly. In order to efficiently support all possible selections in a data cube, you need one join index for each subset of the set of dimensions. E.g, one join index each for [s,p,t,l], [s,p,t], [s,p,l], [s,t,l], [s,p], [s,t], [s,l]

36. 36 B ITMAPPED J OIN I NDEXES A variation of join indexes addresses this problem, using the concept of Bitmap indexes. For each attribute of each dimension table with an additional selection (e.g., country), build a Bitmap index. Index contains, e.g., entry [c,s] if a dimension table tuple with value c in the selection column joins with a Sales tuple with sid s. Note that s denotes the compound key of the fact table, e.g. [pid, timeid, locid]. The Bitmap index version is especially efficient ( Bitmapped Join Index ). Makes sense? Good!

37. 37 B ITMAPPED J OIN I NDEXES Consider a query with conditions price=10 and country=“USA”. Suppose tuple (with sid) s in Sales joins with a tuple p with price=10 and a tuple l with country =“USA”. There are two (Bitmap) join indexes; one containing [10,s] and the other [USA,s]. Intersecting these indexes tells us which tuples in Sales are in the join and satisfy the given selection. pricecategorypnamepidcountrystatecitylocid saleslocidtimeidpid holiday_flagweekdatetimeidmonthquarteryear SALES TIMES PRODUCTSLOCATIONS 10 s USA 10,s,USA

38. D IFFERENCES TO STANDARD SQL

39. 39 S EQUENCES IN SQL SQL-92 supports only (unordered) sets of tuples. Trend analysis is difficult to do in SQL-92, e.g.: Find the % change in monthly sales Find the top 5 product by total sales Find the trailing n -day moving average of sales The first two queries can be expressed with difficulty, but the third cannot even be expressed in SQL-92 if n is a parameter of the query. The WINDOW clause in SQL:1999 allows us to formulate such queries over a table viewed as a sequence of tuples based on user-specified sort keys

40. 40 T HE WINDOW C LAUSE A window is an ordered group of tuples around each (reference) tuple of a table. The order within a window is determined based on an attribute specified by the SQL statement. The width of the window is also specified by the SQL statement. The tuples of the window can be aggregated using the standard (set-oriented) SQL aggregate functions (SUM, AVG, COUNT,...).

41. 41 T HE WINDOW C LAUSE Let the result of the FROM and WHERE clauses be “Temp”. Conceptually, Temp is partitioned according to the PARTITION BY clause. Similar to GROUP BY, but the answer has one tuple for each tuple in a partition, not one tuple per partition! Each partition is sorted according to the ORDER BY clause. SELECT L.state, T.month, AVG(S.sales) OVER W AS movavg FROM Sales S, Times T, Locations L WHERE S.timeid=T.timeid AND S.locid=L.locid WINDOW W AS (PARTITION BY L.state ORDER BY T.month RANGE BETWEEN INTERVAL `1’ MONTH PRECEDING AND INTERVAL `1’ MONTH FOLLOWING);

42. 42 T HE WINDOW C LAUSE For each tuple in a partition, the WINDOW clause creates a “window” of nearby (preceding or succeeding) tuples. Definition of window width can be value-based, as in example, using RANGE. Can also be based on number of tuples to include in the window, using ROWS clause. The aggregate function is evaluated for each tuple in the partition based on the corresponding window. SELECT L.state, T.month, AVG(S.sales) OVER W AS movavg FROM Sales S, Times T, Locations L WHERE S.timeid=T.timeid AND S.locid=L.locid WINDOW W AS (PARTITION BY L.state ORDER BY T.month RANGE BETWEEN INTERVAL `1’ MONTH PRECEDING AND INTERVAL `1’ MONTH FOLLOWING);

43. 43 T OP N Q UERIES Sometimes, want to find only the „best“ answers (e.g., web search engines). If you want to find only the 10 (or so) cheapest cars, the DBMS should avoid computing the costs of all cars before sorting to determine the 10 cheapest. Idea: Guess a cost c such that the 10 cheapest cars all cost less than c, and that not too many other cars cost less than c. Then add the selection cost<c and evaluate the query. If the guess is right: we avoid computation for cars that cost more than c. If the guess is wrong: need to reset the selection and recompute the original query.

44. 44 T OP N Q UERIES „Cut-off value“ c is chosen by query optimizer SELECT TOP 10 P.pid, P.pname, S.sales FROM Sales S, Products P WHERE S.pid=P.pid AND S.locid=1 AND S.timeid=3 ORDER BY S.sales DESC SELECT P.pid, P.pname, S.sales FROM Sales S, Products P WHERE S.pid=P.pid AND S.locid=1 AND S.timeid=3 AND S.sales > c ORDER BY S.sales DESC

45. 45 O NLINE A GGREGATION Consider an aggregate query, e.g., finding the average sales by state. If we do not have a corresponding (materialized) data cube, processing this query from scratch can be very expensive. In general, we have to scan the entire fact table. But the user expects interactive response time. An approximate result may be acceptable to the user.

46. 46 O NLINE A GGREGATION Can we provide the user with some approximate results before the exact average is computed for all states? Can show the current “running average” for each state as the computation proceeds. Even better, we can use statistical techniques and sample tuples to aggregate instead of simply scanning the aggregated table. E.g., we can provide bounds such as “the average for Wisconsin is 2000  102 with 95% probability“.