20 Schema Modeling Techniques

Star Transformation with a Bitmap Index

A prerequisite of the star transformation is that there be a single-column bitmap index on every join column of the fact table. These join columns include all foreign key columns.

For example, the sales table of the sh sample schema has bitmap indexes on the time_id, channel_id, cust_id, prod_id, and promo_id columns.

Consider the following star query:

SELECT ch.channel_class, c.cust_city, t.calendar_quarter_desc,
   SUM(s.amount_sold) sales_amount
FROM sales s, times t, customers c, channels ch
WHERE s.time_id = t.time_id
AND   s.cust_id = c.cust_id
AND   s.channel_id = ch.channel_id
AND   c.cust_state_province = 'CA'
AND   ch.channel_desc in ('Internet','Catalog')
AND   t.calendar_quarter_desc IN ('1999-Q1','1999-Q2')
GROUP BY ch.channel_class, c.cust_city, t.calendar_quarter_desc;

This query is processed in two phases. In the first phase, Oracle Database uses the bitmap indexes on the foreign key columns of the fact table to identify and retrieve only the necessary rows from the fact table. That is, Oracle Database retrieves the result set from the fact table using essentially the following query:

SELECT ... FROM sales
WHERE time_id IN
  (SELECT time_id FROM times 
   WHERE calendar_quarter_desc IN('1999-Q1','1999-Q2'))
   AND cust_id IN
  (SELECT cust_id FROM customers WHERE cust_state_province='CA')
   AND channel_id IN
  (SELECT channel_id FROM channels WHERE channel_desc IN('Internet','Catalog'));

This is the transformation step of the algorithm, because the original star query has been transformed into this subquery representation. This method of accessing the fact table leverages the strengths of bitmap indexes. Intuitively, bitmap indexes provide a set-based processing scheme within a relational database. Oracle has implemented very fast methods for doing set operations such as AND (an intersection in standard set-based terminology), OR (a set-based union), MINUS, and COUNT.

In this star query, a bitmap index on time_id is used to identify the set of all rows in the fact table corresponding to sales in 1999-Q1. This set is represented as a bitmap (a string of 1's and 0's that indicates which rows of the fact table are members of the set).

A similar bitmap is retrieved for the fact table rows corresponding to the sale from 1999-Q2. The bitmap OR operation is used to combine this set of Q1 sales with the set of Q2 sales.

Additional set operations will be done for the customer dimension and the product dimension. At this point in the star query processing, there are three bitmaps. Each bitmap corresponds to a separate dimension table, and each bitmap represents the set of rows of the fact table that satisfy that individual dimension's constraints.

These three bitmaps are combined into a single bitmap using the bitmap AND operation. This final bitmap represents the set of rows in the fact table that satisfy all of the constraints on the dimension table. This is the result set, the exact set of rows from the fact table needed to evaluate the query. Note that none of the actual data in the fact table has been accessed. All of these operations rely solely on the bitmap indexes and the dimension tables. Because of the bitmap indexes' compressed data representations, the bitmap set-based operations are extremely efficient.

Once the result set is identified, the bitmap is used to access the actual data from the sales table. Only those rows that are required for the end user's query are retrieved from the fact table. At this point, Oracle Database has effectively joined all of the dimension tables to the fact table using bitmap indexes. This technique provides excellent performance because Oracle Database is joining all of the dimension tables to the fact table with one logical join operation, rather than joining each dimension table to the fact table independently.

The second phase of this query is to join these rows from the fact table (the result set) to the dimension tables. Oracle uses the most efficient method for accessing and joining the dimension tables. Many dimension are very small, and table scans are typically the most efficient access method for these dimension tables. For large dimension tables, table scans may not be the most efficient access method. In the previous example, a bitmap index on product.department can be used to quickly identify all of those products in the grocery department. Oracle Database's optimizer automatically determines which access method is most appropriate for a given dimension table, based upon the optimizer's knowledge about the sizes and data distributions of each dimension table.

The specific join method (as well as indexing method) for each dimension table will likewise be intelligently determined by the optimizer. A hash join is often the most efficient algorithm for joining the dimension tables. The final answer is returned to the user once all of the dimension tables have been joined. The query technique of retrieving only the matching rows from one table and then joining to another table is commonly known as a semijoin.

Execution Plan for a Star Transformation with a Bitmap Index

The following typical execution plan might result from "Star Transformation with a Bitmap Index":

SELECT STATEMENT
 SORT GROUP BY
  HASH JOIN
   TABLE ACCESS FULL                          CHANNELS
   HASH JOIN
    TABLE ACCESS FULL                         CUSTOMERS
    HASH JOIN
     TABLE ACCESS FULL                        TIMES
     PARTITION RANGE ITERATOR
      TABLE ACCESS BY LOCAL INDEX ROWID       SALES
       BITMAP CONVERSION TO ROWIDS
        BITMAP AND
         BITMAP MERGE
          BITMAP KEY ITERATION
           BUFFER SORT
            TABLE ACCESS FULL                 CUSTOMERS
           BITMAP INDEX RANGE SCAN            SALES_CUST_BIX
         BITMAP MERGE
          BITMAP KEY ITERATION
           BUFFER SORT
            TABLE ACCESS FULL                 CHANNELS
           BITMAP INDEX RANGE SCAN            SALES_CHANNEL_BIX
         BITMAP MERGE
          BITMAP KEY ITERATION
           BUFFER SORT
            TABLE ACCESS FULL                 TIMES
           BITMAP INDEX RANGE SCAN            SALES_TIME_BIX

In this plan, the fact table is accessed through a bitmap access path based on a bitmap AND, of three merged bitmaps. The three bitmaps are generated by the BITMAP MERGE row source being fed bitmaps from row source trees underneath it. Each such row source tree consists of a BITMAP KEY ITERATION row source which fetches values from the subquery row source tree, which in this example is a full table access. For each such value, the BITMAP KEY ITERATION row source retrieves the bitmap from the bitmap index. After the relevant fact table rows have been retrieved using this access path, they are joined with the dimension tables and temporary tables to produce the answer to the query.

Star Transformation with a Bitmap Join Index

In addition to bitmap indexes, you can use a bitmap join index during star transformations. Assume you have the following additional index structure:

CREATE BITMAP INDEX sales_c_state_bjix
ON sales(customers.cust_state_province)
FROM sales, customers
WHERE sales.cust_id = customers.cust_id
LOCAL NOLOGGING COMPUTE STATISTICS;

The processing of the same star query using the bitmap join index is similar to the previous example. The only difference is that Oracle utilizes the join index, instead of a single-table bitmap index, to access the customer data in the first phase of the star query.

Execution Plan for a Star Transformation with a Bitmap Join Index

The following typical execution plan might result from "Execution Plan for a Star Transformation with a Bitmap Join Index":

SELECT STATEMENT
 SORT GROUP BY
  HASH JOIN
   TABLE ACCESS FULL                          CHANNELS
   HASH JOIN
    TABLE ACCESS FULL                         CUSTOMERS
    HASH JOIN
     TABLE ACCESS FULL                        TIMES
     PARTITION RANGE ALL
      TABLE ACCESS BY LOCAL INDEX ROWID       SALES
       BITMAP CONVERSION TO ROWIDS
        BITMAP AND
         BITMAP INDEX SINGLE VALUE            SALES_C_STATE_BJIX
         BITMAP MERGE
          BITMAP KEY ITERATION
           BUFFER SORT
            TABLE ACCESS FULL                 CHANNELS
           BITMAP INDEX RANGE SCAN            SALES_CHANNEL_BIX
         BITMAP MERGE
          BITMAP KEY ITERATION
           BUFFER SORT
            TABLE ACCESS FULL                 TIMES
           BITMAP INDEX RANGE SCAN            SALES_TIME_BIX

The difference between this plan as compared to the previous one is that the inner part of the bitmap index scan for the customer dimension has no subselect. This is because the join predicate information on customer.cust_state_province can be satisfied with the bitmap join index sales_c_state_bjix.

How Oracle Chooses to Use Star Transformation

The optimizer generates and saves the best plan it can produce without the transformation. If the transformation is enabled, the optimizer then tries to apply it to the query and, if applicable, generates the best plan using the transformed query. Based on a comparison of the cost estimates between the best plans for the two versions of the query, the optimizer then decides whether to use the best plan for the transformed or untransformed version.

If the query requires accessing a large percentage of the rows in the fact table, it might be better to use a full table scan and not use the transformations. However, if the constraining predicates on the dimension tables are sufficiently selective that only a small portion of the fact table must be retrieved, the plan based on the transformation will probably be superior.

Note that the optimizer generates a subquery for a dimension table only if it decides that it is reasonable to do so based on a number of criteria. There is no guarantee that subqueries will be generated for all dimension tables. The optimizer may also decide, based on the properties of the tables and the query, that the transformation does not merit being applied to a particular query. In this case the best regular plan will be used.

Star Transformation Restrictions

Star transformation is not supported for tables with any of the following characteristics:

• Queries with a table hint that is incompatible with a bitmap access path

• Queries that contain bind variables

• Tables with too few bitmap indexes. There must be a bitmap index on a fact table column for the optimizer to generate a subquery for it.

• Remote fact tables. However, remote dimension tables are allowed in the subqueries that are generated.

• Anti-joined tables

• Tables that are already used as a dimension table in a subquery

• Tables that are really unmerged views, which are not view partitions

• Tables where the fact table is an unmerged view.

• Tables where a partitioned view is used as a fact table.

The star transformation may not be chosen by the optimizer for the following cases:

• Tables that have a good single-table access path

• Tables that are too small for the transformation to be worthwhile

In addition, temporary tables will not be used by star transformation under the following conditions:

• The database is in read-only mode

• The star query is part of a transaction that is in serializable mode