AU2001265048A1 - System and method for automatically generating database queries - Google Patents

Description

SYSTEM AND METHOD FOR AUTOMATICALLY GENERATING DATABASE QUERIES

REFERENCE TO RELATED APPLICATION

The present application is based on Provisional Application 60/207,379 filed

May 26, 2000, the entire contents of which are herein incorporated by reference.

Field

The present specification relates to databases and, more specifically, to a system and method

for automatically generating database queries.

Background

Structured Query Language (SQL) is the ANSI standard language for interacting with

databases. SQL allows users to combine information from many different sets of records (tables)

in a database in order to retrieve data from the database. However, the power of SQL has some

drawbacks. For example, a user can combine different tables in an SQL query, which may create

queries that can be impossible for the database engine to resolve or that may utilize a significant

amount of system resources.

Database systems often include performance tuning systems or methods. For example,

Oracle 7 provides various scripts (e.g., UTILBSTAT and UTLESTAT) that can be used to assist in

performance tuning by summarizing the operating state of a database in one or more reports. These

scripts are a set of SQL scripts useful for capturing a snapshot of system- wide database performance

statistics and generating a report which can help an operator optimize the performance of the database. The database operator can use the reports for fine-tuning the database performance and for preventive maintenance of the database. A report may include, for example, information about database memory objects, including library cache, dictionary cache, latch usage, file I/O and rollback statistics.

A problem with existing systems is that they often limit the amount of time it takes to produce an optimal execution plan, h addition, a knowledgeable user might have more insight into the data than represented by the object statistics, if object statistics are present at all. Existing systems often do not take advantage of the knowledge of the user. In addition, although statistics may be useful when properly provided and analyzed, statistics might not exist at all if the objects have not yet been analyzed, or the statistics may be out-of-date. A specific statistic that would provide insight to the distribution of the data may not be present in a particular instance. On the other hand, the user may have first hand knowledge that may be useful in tuning database queries that my not otherwise be available. For example, the user may know that the query will be executed at a time when more resources, such as multiple CPUs, will be available, even though the system is normally fully utilized. Such information could prove very helpful during a database query tuning process.

SUMMARY

A method of tuning a database query comprises selecting a database query, parsing the selected database query to determine relationships between portions of the selected database query, selecting an optimization mode from a plurality of available optimization modes, tuning the selected database query by modifying at least one portion of the selected database query based on the determined relationships and the selected optimization mode and displaying the modified database query.

The parsing may determine tokens within the database query, tokens being words separated by delimiters. The plurality of available optimization modes may include Cost based and Rule based modes. The Cost based modes may include First_Rows mode and All_Rows mode. The method may further comprise determining a cost associated with using the tuned database query. The method may further comprise comparing a cost associated with using the selected database query to the cost associated with using the tuned database query. The method may further comprise parsing the selected database query to determine whether the database query includes at least one subquery joined by at least one of a NOT EXISTS, a NOT IN and an ALL clause. The method may further comprise prompting a user to select preferences to be used during tuning based on whether the database query includes at least one of a NOT EXISTS, a NOT IN and an ALL clause.

The preferences may include rewrite preferences for enabling a user to select at least one of a conversion of NOT EXISTS operators to a NOT IN operator and conversion of the selected database query to an outer join. The preferences may include rewrite preferences for enabling a user to select to convert subqueries joined by an ALL operator to a join or outer-join. The preferences may include rewrite preferences for enabling a user to select whether to use at least one of a NOT EXISTS operator and an outer-join to convert subqueries joined by a NOT In operator.

A computer storage medium includes computer executable code for tuning a database query, comprising computer executable code for allowing a user to select a database query, computer executable code for parsing the selected database query to determine relationships between portions of the selected database query, computer executable code for allowing a user to select an optimization mode from a plurality of available optimization modes, computer executable code for tuning the selected database query by modifying at least one portion of the selected database query based on the detennined relationsliips and the selected optimization mode and computer executable code for displaying the modified database query.

A programmed computer system for tuning a database query comprises a display for displaying at least one database queries to a user, a user input allowing the user to select a database query from among the displayed database queries and an optimization mode from a plurality of available optimization modes and a processor for parsing the selected database query to determine relationships between portions of the selected database query, and for tuning the selected database query by modifying at least one portion of the selected database query based on the determined relationships and the selected optimization mode, the modified database query being displayed to the user via the display.

Brief Description of the Drawings

The following drawings depict illustrative examples only and are provided for an even more complete appreciation and understanding of the following description.

Fig. 1 is a block diagram for describing a database query tuning process according to an embodiment;

Fig. 2 shows a dialog display including an SQL statement;

Fig. 3 shows a Select Optimization Mode dialog display;

Fig. 4 shows a Preferences window display;

Fig. 5 shows a Check SQL dialog display; Fig. 6 shows a Result window display; Fig. 7 shows an edit or tuning window display; Fig. 8 shows a Detailed Result window display.

Fig. 9 is a block diagram of a computer system to which the present system and method may be applied; and

Figs. 10 - 13 are charts showing hierarchies for tables in various SQL statements.

Detailed Description

In describing preferred embodiments of the present disclosure illustrated in the drawings, specific teraiinology is employed for sake of clarity. However, the present disclosure is not intended to be limited to the specific terminology so selected and it is to be understood that each specific element includes all technical equivalents which operate in a similar manner.

The present disclosure relates to a system and method that automatically tunes database queries for increasing performance. For example, the present system and method may be in the form of a database management system (DBMS) capable of tuning Structured Query Language (SQL) statements to produce an optimal execution plan. In one aspect, the present disclosure relates to systems and methods for analyzing a database query in the context of the database structure and modifying and/or restructuring the query so that it can be executed witliin the context of the DBMS structure. In addition, the system and method of the present disclosure can utilize the database relationship structure to minimize the DBMS resources used by the query and to ensure that the query is completed. One way to achieve these results is to parse the SQL statement into a custom data structure that permits the modification and/or restructuring of the SQL statement. As an example, each table in, e.g., a FROM clause, may be checked for referential integrity and compared against, e.g., a users WHERE condition, and the WHERE condition may be checked for correct outer join specifications.

The present system parses out the SQL statement and keeps track of each token, how tables are joined, and the relationships between portions of the SQL statement. A token may be any word within an SQL statement separated by delimiters. Tokens that stand for table column names may be used to identify joins and other relationships existing witliin the SQL statement. The system and method may then change, for example, the join order and may move subqueries to tune the SQL statement for top performance. The system and method may be implemented in the form of a software application running on a computer system such as a mainframe, personal computer (PC), handheld computer, etc. The computer system may be linked to a database. The link may be, for example, via a direct link such as a direct hard wire or wireless connection, via a network connection such as a local area network, or via the Internet.

An example of a computer system capable of implementing the present system and method is shown in Fig. 9. The computer system referred to generally as system 200 may include a central processing unit (CPU) 202, memory 204, a printer interface 206, a display unit 208, a LAN (local area network) data transmission controller 210, a LAN interface 212, a network controller 214, an intemal bus 216 and one or more input devices 218 such as, for example, a keyboard, mouse, etc. As shown, the system 200 may be connected to a database 220 via a link 222.

Fig. 1 depicts a process according to an embodiment of the present disclosure for tuning SQL statements for increased performance. A user first selects an SQL statement they desire to have tuned (Step S2). For example, from an SQL edit window displayed on display unit 208, a user can highlight an SQL statement they desire to have tuned by placing the cursor at the beginning of the SQL statement, left-clicking the mouse and scrolling down to the end of the SQL statement and then releasing the mouse button. This will highlight the SQL statement. The user can then select an SQL Tuner option from a Tools menu (not shown) by, for example, clicking on an appropriate button.

The highlighted SQL statement to be tuned is then displayed in an SQL Tuner window 1 (Step S4), as shown in Fig. 2. The user can then review the displayed SQL statement 2 to confirm that it is the SQL statement that they desire to tune. The user then has the option of clicking Back button 3, Next button 4, Finish button 5 or Cancel button 6. Back button 3 returns to the previous window, allowing the user to select a different SQL statement. Cancel button 6 exits the SQL tuner window 1 without any changes being made to the displayed SQL statement 2. Next button 4 moves to the next window (shown in Fig. 3), allowing the user to select optimization mode settings as will be described later below. Finish button 5 runs the SQL tuning process using the current optimization mode settings.

The present system is capable of operating utilizing several optimization modes, and using several different rewrite preferences. After Next button 4 is clicked on by the user, the present system parses the SQL statement (Step S6) and displays a Select Optimization Mode window 11 as shown in Fig. 3. The SQL statement is checked to determine, for example, what operators, functions, etc. exist in the statement. The system also checks to determine whether any subqueries within the SQL statement are joined by NOT EXISTS, NOT IN, or ALL clauses. The present system also checks table statistics. For example, table statistics generated by a system such as Oracle when executing an ANALYZE command on a table may be checked. The table statistics may include number of rows (NUM_ROWS), number of data blocks below the high water mark (e.g., the number of data blocks that have been formatted to receive data, regardless of whether they currently contain data or are empty) (BLOCKS), the number of data blocks allocated to the table that have never been used (EMPTY_BLOCKS), the average available free space in each data block in bytes (AVG_SPACE), the number of chained rows (CHATN_COUNT) and the average row length, including the row's overhead, m bytes (AVG_ROW_LEN).

If sufficient information exists (e.g., statistics exist for the relevant tables), the system automatically selects an optimization mode. The selected optimization mode is indicated by radio buttons 7-9. For example, a cost-based optimization (FIRST ROW or ALL ROWS) mode may be automatically selected since it will yield good results if statistics exist on the tables concerned. In Step S8, if there is insufficient information in the table statistics for the system to automatically select an optimization mode or if the user desires to change an optimization mode setting made automatically by the system, the user can manually select the optimization mode by clicking one of the radio buttons 7-9 in the Select Optimization Mode window 11 as shown in Fig. 3. The user can also modify rewrite preferences settings from window 11, as will be described in more detail below by clicking Preferences button 10.

One cost-based optimization mode is referred to as ALLJ OWS mode. The ALL_ROWS mode is automatically selected by default by the system if statistics exist on at least one of the referenced tables in the SQL statement. ALL_ROWS mode provides fast throughput for batch processing. The user then has the option of selecting a FIRST_ROWS mode by clicking radio button 8, unless the SQL statement contains a set operator (e.g., UNION, INTERSECT, MINUS, UNION ALL), a GROUP BY clause, a FOR UPDATE clause, aggregate functions, or a DISTINCT operator. If the SQL statement contains a set operator (e.g.., UNION, INTERSECT, MINUS, UNION ALL), a GROUP BY clause, a FOR UPDATE clause, aggregate functions, or a DISTINCT operator, the user is not given the option to select the FIRST ROWS mode. The ALL_ROWS mode can also be manually selected by the user by clicking radio button 7.

The FIRST_ROWS mode may be automatically selected by the system when the system cannot otherwise determine an optimization mode to use for the session. The FIRST_ROW mode may be manually selected if a fast response time for line processing is desired.

Another mode is referred to as Rule-based mode. The Rule-based mode is automatically selected by the system if no statistics exist on any of the referenced tables. Rule-based mode may be manually selected by clicking radio button 9.

If it was determined during the parsing process (Step S6) that the SQL statement included subqueries joined by NOT EXISTS, NOT IN or ALL clauses, Preferences button 10 is made active in window 11 (Fig. 3). Clicking on Preferences button 10 allows the user to specify rewrite preferences to be used to rewrite subqueries joined by the NOT IN operator, NOT EXISTS operator and ALL operator.

For example, rewrite preferences for NOT EXISTS operators can be specified enabling the user to select conversion of the NOT EXISTS operators to a NOT IN operator and/or conversion of the statement to an outer join. Rewrite preferences for the ALL operator can be specified, allowing the user to select conversion of the statement to a join or an outer-join. Rewrite preferences for the NOT IN operator can be specified enabling the user to select conversion of the NOT IN operator to NOT EXISTS or conversion of the statement to an outer-join.

Preferences can be further defined by the user for the rewrite by choosing options that limit the conversion to instances where the surrounding interface columns are defined as NOT NULL. The user can also specify that the present system always convert and add IS NOT NULL criterion to the statement if necessary.

When Preferences button 10 is clicked by the user, a preferences window 50 as shown in Fig. 4 is displayed. An SQL Tuner preferences page 52 can then be displayed by clicking on Tuner tab 51. The options in this window enable the user to set the rewrite preferences as described above for subqueries joined by the operators ALL, NOT IN and NOT EXISTS.

In the example shown in Fig. 4, to set rewrite preferences for the NOT EXISTS operator, the user selects the NOT EXISTS operator from the pulldown menu 53. This displays a list of options available for the NOT EXISTS operator. For example, the user can then specify whether to enable the present system to use NOT IN and/or outer-join to convert subqueries joined by the NOT EXISTS operator by selecting or de-selecting the appropriate check boxes for options, "Convert Subquery Joined Via NOT EXISTS Operator To NOT IN" (check box 55) and "Convert Subquery Joined Via NOT EXISTS Operator To An Outer- Join" (check box 56).

If the "Convert Subquery Joined Via NOT EXISTS Operator To NOT IN" (check box 55) option is selected, option boxes 57 and 58 become active. Option box 57 allows the user to select the "Check For NOT NULL Conditions On The Joined Columns In Outer Query" option and option box 58 allows the user to select the "Check For NOT NULL Conditions On The Joined Columns In The Subquery" option. The user may select one or both of the these options by selecting or deselecting check boxes 59 and 60, respectively. Conversion options then become active for each selected option box. For example, "Convert Only If The Columns Are Defined As NOT NULL" can be selected for the outer query and subqueiy by clicking radio buttons 61 and 63, respectively and "Always Convert And Add "IS NOT NULL" If Necessary" can be selected by clicking radio buttons 62 and 64, respectively. The user can then click on the "OK" button 54 to save the specified preferences and exit the preferences window. Selecting Cancel button 66 exits the window 52 without saving any changes made to the preferences. Selecting Help button 65 displays a help menu to the user.

In a similar manner, to set rewrite preferences for the ALL operator, the user selects "ALL operator" from the pulldown menu 53 to display a list of options. The user will then be presented with options to specify whether to enable the present system to convert subqueries joined by the ALL operator to a join or outer-join by selecting or de-selecting a "Convert subqueries joined via the ALL operator to a join or outer-join" option.

If the user selects this option, the user will then be presented with options to "Check For NOT NULL Conditions On The Joined Columns In The Outer Query" and "Check For NOT NULL Conditions On The Joined Columns In The Subquery". The user can then select one or both of the these options. For each of these options selected, the user is then presented with two conversion options "Convert Only If The Columns Are Defined As NOT NULL" and "Always Convert And Add "IS NOT NULL" If Necessary".

The user can then click on the "OK" button to save the specified preferences and exit the Preferences window.

To set rewrite preferences for the NOT IN operator, the user selects the NOT IN operator from the pulldown menu 53 to display a list of options. The user can then specify whether to enable the present system to use the NOT EXISTS operator and /or outer-join to convert subqueries joined by the NOT IN operator by selecting or de-selecting the options "Convert Subquery Joined Via NOT IN Operator To NOT EXISTS" and "Convert Subquery Joined Via NOT IN Operator To An Outer- Join".

For each option selected, the user is then presented with an option to check for NOT NULL conditions on the joined columns in outer query and to check for NOT NULL conditions on the joined columns in the subquery. The user can select one or both of these options. Conversion options then become active for each selected option. That is, the user is given the option to convert only if the columns are defined as NOT NULL or to always convert and add "IS NOT NULL" if necessary. The user can thus select a conversion option for joined columns on the outer query and/or the subquery. The user clicks on the "OK" button 54 to save the specified preferences and exit the Preferences window. Upon exiting the Preferences window, the display returns to the Select Optimization Mode Window 11 (Fig. 3). After Step S8 is complete and the user is satisfied with the optimization mode and preferences selections, the process can be continued by clicking Next button 8.

In Step S10 the present system checks the SQL statement for referential integrity, outer join, and/or NULL logic problems. If errors or inefficiencies are found, specific suggestions for altering the SQL statement to correct the problems are displayed. For example, suggestions may be displayed for invalid Outer Joins, which can cause invalid result sets. Suggestions may also be displayed for NULL logic problems causing invalid result sets, incorrect use of the HAVING clause, cartesian products , expressions in join criteria for better index utilization, expressions in non-join criteria for better index utilization, and unindexed foreign keys, etc.

In Step S12, a Check SQL window 70 as shown in Fig. 5 is then displayed to the user. Window portion 73 displays a diagram of each subquery in the statement. Each table in the SQL statement displays as a box with a header containing the table name and alias. The primary key, foreign keys, and unique keys are also displayed for each table.

A green line between tables represents a correct join. If a red line connects the tables, the join contains an error (such as a referential integrity breach or in improper outer join). If a blue line connects the tables, the join is inefficient due to lack of referential integrity.

Suggested coιτections to the SQL statement are displayed in window portion 71. Check SQL window 70 can be re-sized by the user as desired to view all of the information contained in any of the window portions.

Clicking on "Next SQL Module" button 72 and the "Previous SQL Module" button 74 highlights individual subqueries in the SQL statement in blue and displays only the highlighted subquery' s associated diagram.

The user can then decide (Step SI 4) which SQL statement to tune, the original statement or the fixed statement. For example, clicking on "Tune Original SQL" button 76 tunes the original SQL statement which is displayed in the field 75 at the bottom of the Check SQL window 70. Clicking on "Tune Fixed SQL" button 78 implements the suggested corrections displayed in the field 71 on the right of the Check SQL window 71 and then tunes the corrected SQL statement.

In Step SI 6, the selected SQL statement is tuned by performing one or more steps to optimize the statement. For example, the system may make transitive changes in the join criteria, may determine all subquery transformations, may determine transitivities in the non-join criteria if necessary and may determine join orders. The system may generate several transformed SQL statements. The system may then determine which of the transformed SQL statements has the best performance and sort the resulting SQL statements accordingly. When tuning or transforming a SQL statement, the present system may check to see if columns of a surrounding query belong to the same table and make transitive changes, if necessary. If they do not belong to the same table, the system checks to see if any of the join criteria reference a column interfacing the subquery and identifies the equivalent column. The present system does this for each column interfacing the subquery and then uses the transitive property to generate permutations of the equivalent columns in an attempt to find a set where all of the columns belong to the same table. The present system may make transitive changes to non-correlated subqueries joined via the NOT IN operator and correlated subqueries interfaced with NOT IN, NOT EXISTS, or ALL operators.

The present system may also determine all the subquery transformations that can be performed on the SQL syntax by evaluating each subquery and assigning a rank to each of the transformations. A transformation may be mandatory (always or never) or optional. This information is then used to generate the various alternative SQL versions. For example, if a SQL statement contains four subqueries, the present system may rank none of the subqueries "always transform", one "never transform", and three "optional" to generate possible alternate SQL statements.

That is, the type of transformation determines whether the transformation should always be done, never be done or optionally be done. For example, some transformations may be desirable because the optimizer will use a better plan if the transformation were done. In this case, the transformation should always be done. Accordingly, the transformation would be assigned a highest ranlc. Some transfomiations may be undesirable since the optimizer will use a less efficient plan if the transformation were done. In this case, the transformation should not be done. Accordingly, the transformation would be assigned a lowest rank. Some transformations may or may not result in a

better optimization plan being selected, since it may depend on other factors as well. In this case,

the transformation may optionally be done. Accordingly, the transformation would be assigned a

middle ranlc.

After performing the subquery transformations, the present system identifies equivalent

columns in the non-join criteria that can be substituted for one another and makes the substitution.

This makes it possible for the present system to rewrite the SQL statement to allow a database (e.g.,

such as ORACLE) to take advantage of concatenated key indexes.

For example, in the following SQL statement the e.emp_seq criterion on the subquery of

DEPENDENTS can be interfaced to a.emp_seq or t.emp_seq.

SELECT * FROM employees e, assignments a, time_sheets t WHERE e.emp_seq = a.emp_seq AND a.emp_seq = t.emp_seq AND a.proj_seq = t.proj_seq

AND e.emp_seq IN (SELECT emp_seq FROM dependents WHERE relation = 'SPOUSE')

Switching e.emp_seq to a.emp_seq as shown below (Line 5) enables the database to use the

concatenated key index on ASSIGNMENTS:

SELECT * FROM employees e, assignments a, time_sheets t WHERE e.emp_seq = a.emp_seq AND a.emp_^seq = t.emp_seq AND a.proj_seq = t.proj_seq

AND a.emp_seq IN (SELECT emp_seq FROM dependents WHERE relation = 'SPOUSE') Various columns can replace another column when an equality join criterion exists equating the columns. However, the present system can also use transitivities for non-join criteria that is AND'ed to the join criteria. If a non-join criterion is OR'ed to the join criteria, the transitivities that result should not be used.

If a multiple column operand has transitivities, the present system adds the columns to the operand, and duplicates the corresponding column on the select list. All multi-column transitivities are used for each transformation except for the original statement. If the operand is equivalent to another column via an outer-joined criterion, the equivalent column cannot be used.

If a FROM clause has a nested select statement that does not have a GROUP BY syntax or aggregates and the main query has non-join criteria referencing one of the join columns between the main query and the nested select, the present system duplicates the non-join criterion in the nested select statement in all copies of the SQL statement, including the original.

The present system can use an ORDERED hint to determine join order. The ORDERED hint forces the optimizer to join tables in the order in which they appear in the FROM clause in the SQL statement. This allows SQL statements to be written that will join tables referenced in the FROM clause in a particular order. The system first tests the SQL statement with no ORDERED hint. The present system then creates additional pennutations using the ORDERED hint with different driving tables to force a specific join order, a driving table being, for example, a parent table in a nested- loop-join operation. Each row from a driving table can be used to find matching rows in a joined table to complete the join operation.

If the default join of the transformed SQL statement has a better cost/row value than the original SQL statement, then most likely the ORDERED hint will produce a more efficient plan but a higher cost/row than the initial transformation with a default join order. The present system can display the ORDERED transfonnation as a viable best solution.

To determine best performance from among each of the created alternative SQL statements, the present system generates a cost statistic for the original SQL and each of the SQL statements generated by the rewrite process. The cost is determined after creating the plan associated with the SQL. For example, Oracle's EXPLAIN PLAN statement can be used to generate a plan that the optimizer uses to execute the SQL statement. The EXPLAIN PLAN statement can be run on the SQL statements to get the execution plan which will describe the steps that the optimizer plans to use to execute the SQL statements. Oracle will also provide a number (e.g., cost) for each step in the plan. The cost is a relative figure of usage to execute the step in the plan. The costs can be summed up for all of the steps in the plan to arrive at a total cost for the plan. The present system can divide the cost calculated by the database by the number of rows, to generate a cost/row value for each of the SQL statements. Cost may also be determined by the number of logical reads required by the plan. After determining the cost, the present system sorts the SQL statements in ascending order (from the lowest to the highest cost).

In step (SI 8), the present system displays the original SQL statement and the SQL statement with the lowest cost in a Result dialog window 90 such as that shown in Fig. 6. The original SQL statement is displayed in a window portion 92 and the suggested alternative SQL statement(s) displayed in window portion 94. The user can then compare the suggested alternate SQL statement(s) with the original SQL statement, examine the changes, and then choose whether they want to replace the original SQL statement with the tuned SQL statement.

Clicking Replace SQL button 100 then replaces the original SQL statement with the suggested alternate SQL statement shown in window 94. The system adds comments and displays the tuned SQL statement in an edit or tuning window 104 as shown in Fig. 7. As shown in Fig. 7, the comments may include a date stamp, infoπnation indicating that the statement was tuned and identifying the optimization mode used (ALL- ROWS in this example). Clicking on Back button 96 returns to the previous window. Clicking on More Detail button 98 opens a Detailed Result Window 106 as shown in Fig. 8. The tuned SQL statement is displayed in window portion 108. Row 110 displays the Cost Statistics for the original SQL statement and one or more rows 112 display the cost statistics for the tuned alternate SQL statement(s). The statistics may include Elapsed Time 114 which is the amount of time required to execute the SQL (in seconds), CPU time 116 which is the number of CPU cycles per second required to execute the SQL, Log Reads 118 which is the number of logical reads required to execute the SQL, Physical reads 120 which is the number of physical reads required to execute the SQL, Rows 122 is the number of rows the SQL returns and Cost/Rows 124 is the database's cost divided by the number of rows.

The Detailed Results window 106 enables the user to view cost statistics for each statement to help the user choose the SQL best suited to their needs. The user can also choose to open the tuned statement in a system including an editor function for further editing.

If there is a bind variable in the SQL, clicking on the "Bind" button 126 displays a dialog where the user can specify a value for the variable before executing the SQL statement. Clicking on the "Execute" button 128 generates statistics for the selected statement. Clicking on the "Execute AU" button 130 generates statistics for all the SQL statements listed. The user can click "Print" button 132 to print the information currently displayed in the window. The user can click on the "Save" button 134 to save the information in this window as a text file for future reference. The user can select an alternate SQL entry in the table and click on the "PAFO" (Plan Analyzer for Oracle)

button 138 to display the code in a planner window.

Clicking on the "Replace SQL" button 140 replaces the original SQL with the tuned SQL.

Clicking on the "Back" button 136 returns to the Result window 90 (Fig. 6). Clicking on the

"Cancel" button 144 exits the present system without changing the original SQL.

The following descriptions list and describe the various terms used throughout the present

specification to reference parts or attributes of SQL statements as well as some of the concepts used

by the present system for tuning SQL statements. The terms are used throughout the specification

in describing the algorithms used by the present system to tune the SQL statements.

A. Branch and Level of Nesting

When a query has subqueries, the path from the top (main query) to the bottom is a "branch".

The "level of nesting" is the number of subqueries in the branch. The main query is considered at

"level" = 1. The subquery directly below it is at "level" = 2, and so on. Consider the following

SQL, line numbers for which have been added for ease of reference.

1. SELECT * FROM employees

2. WHERE emp_seq IN

3. (SELECT emp_seq

4. FROM time_sheets t

5. WHERE t.rpt_date = 'Ol-MAY-98'

6. AND t.proj_seq =

7. (SELECT proj_seq FROM proj ects

8. WHERE name - 'EXPLAIN SQL: DEVELOPMENT'))

9. AND hiredate > S YSDATE - 10 The branch extends from the main query, LEVEL= 1 (beginning line 1), to LEVEL = 2

(beginning line 3), to LEVEL = 3 the last subquery (beginning line 7). The "level of nesting" in this

example is 2, since the main query has a subquery and that subquery in turn has a subquery.

The following SQL statement contains 2 branches:

1. SELECT * FROM employees

2. WHERE emρ_seq LN

3. (SELECT emp_seq

4. FROM time_sheets t

5. WHERE t.rpt_date = 'Ol-MAY-98'

6. AND t.proj_seq =

7. (SELECT proj_seq FROM proj ects

8. WHERE name = 'EXPLAIN SQL: DEVELOPMENT'))

9. AND hiredate > S YSDATE - 10

10. AND emp_seq IN

11. (SELECT emp_seq FROM dependents

12. WHERE birthdate > 'Ol-JAN-67')

The first branch in this SQL statement is the same as in the previous example. However, this

SQL statement has a second branch. The main query (which begins on line 1) is level 1. Level 2

begins on line 11. This second branch has a level of nesting equal to 1.

B. Top-Most Parent Boolean Operator

When OR operators exist in a WHERE clause, care should be used before merging or moving

a subquery into a FROM clause. If the OR is a parent Boolean operator, the merge or move should

not take place. The following example illustrates this point. For the SQL statement:

SELECT e.emp_seq FROM employees e WHERE hiredate > SYSDATE - 155 OR emp_seq IN (SELECT emp_seq FROM assignments WHERE ρroj_seq = 1)

If we did not consider the OR operator, the subquery could be merged as follows:

SELECT e.emp_seq, x.emp_seq FROM employees e, (SELECT emp__seq FROM assignments

WHERE proj_seq = 1) x WHERE e.hiredate > SYSDATE - 155 OR e.emp_seq = x.emp_seq

However, the join will now become a Cartesian product when there is no joining row in the

subquery. For example, if the criterion "e.emp_seq = x.emp_seq" evaluates to FALSE, but the non-

join criterion evaluates to TRUE, the row from the subquery will be joined regardless. And for that

specific EMPLOYEES row, the non-join criterion will always evaluate to TRUE regardless of which

row from the subquery is joined. This creates the Cartesian product.

The following example illustrates that the problem may become even more difficult when

NOTs precede criteria.

For example, given the SQL statement:

SELECT e.emp_seq FROM employees e WHERE NOT(hiredate > SYSDATE - 155 OR emp_seq IN (SELECT emp_seq FROM assignments WHERE proj_seq = 1)) The NOT, using DeMorgan's Rule, reverses the enclosed Boolean and relational operators.

However, the present system may eliminate the NOT operators by using DeMorgan's rule, resulting

in the following SQL.

SELECT e.emp_seq FROM employees e WHERE hiredate <= SYSDATE - 155

AND emp_seq NOT IN (SELECT emρ_seq FROM assignments WHERE proj_seq - 1)

The reversed Boolean and relational operators are underlined. Making the reversals first may

simplify the process.

C. Correlated to Multiple SQL Modules

Correlated subqueries can be correlated to more than one SQL module as shown in the

following example:

SELECT * FROM employees e WHERE EXISTS

(SELECT null FROM assignments a WHERE a.emp_seq = e.emp_seq AND EXISTS

(SELECT null FROM time_sheets t WHERE t.emp_seq = e.emp_seq AND t.proj_seq = a.proj_seq))

The subqueiy on TIME_SHEETS (line 6) is correlated (via line 7) to both the surrounding

query (on ASSIGNMENTS) and the main query on EMPLOYEES (via line 8). D. Correlation Criteria

In the above SQL statement the criterion on lines 7 and 8 are called "correlation criteria".

These criteria are not "join criteria", but are more like "non-join criteria" in the sense that the

reference outside the query (such as "e.emp_seq" or "a.proj_seq") is like a constant when the

subquery is executed.

Correlation criteria can be checked in routines to ensure they utilize the equality operator.

However, it should be noted that correlation criteria that should be checked this way are those that

correspond to the columns comprising a unique index. For example, the following query has 2

correlation criteria in the subquery. However, the only one that matters is the one over PROJ_SEQ,

since that is the column having the unique index.

SELECT * FROM assignments a

WHERE EXISTS

(SELECT start_date FROM projects p WHERE a.proj_seq = p.proj_seq AND a.start_date < p.start_date)

E. Non- Join Criteria

In an SQL statement, when the code below references a non-join criterion, it is looking for

criterion where the relational operator is '=' and the other operand is either a constant or a subquery.

The assumption with the subquery is that since the user specified '=' as the relational operator, the

subquery must return at most a single row, which means a constant.

The following are examples of invalid WHERE clauses:

WHERE coll = 10 + col2 WHERE coll > 10

The following are examples of valid WHERE clauses:

WHERE coll = 10

WHERE coll = (select col2 from tabl)

WHERE coll = :Bind_Variable

F. Join Criteria

Join criteria are referenced below. In each case the relational operator should be '=', and

each side of the criterion should be a single column name and not an expression.

G. Subquery Interface

An interface between a surrounding query and a subquery typically contains at least a

"relational operator" such as "IN", "=", "NOT EXISTS", etc. Preceding all but the EXISTS and

NOT EXISTS relational operators is one or more columns. The columns preceding the operator are

called "interface columns". The columns in the corresponding SELECT list of the subquery are also

called "interface columns". These two types of "interface columns" can be distinguished by whether

they are interface columns for the surrounding query or the subquery.

It is possible that the columns do not belong to objects within the subquery and are not

correlated with any of the columns belonging to the tables in the subquery, as illustrated below:

1. SELECT * FROM employees e

2. WHERE emp_seq IN

3. (SELECT emp_seq FROM assignments a 4. WHERE (e.hiredate, a.proj_seq) IN

5. (SELECT rpt_date, proj_seq FROM time_sheets t))

The column "e.hiredate" (line 4) does not belong to the subquery (ASSIGNMENTS table)

and is not con-elated with any of the columns in ASSIGNMENTS. This document will refer to these

columns of the surrounding query as "foreign". If a column is foreign, the system will attempt to

merge it or convert it to a non-correlated subquery. However, the system will not attempt to move

that subquery.

H. Preferences

Null logic can present numerous problems when transforming SQL statements. If certain

columns are nullable in the surrounding query, different result sets can appear when transformed.

The transformation can be modified to ensure this does not happen, although there may be a small

cost in tenns of perfonnance. The modification is to AND a criterion, specifying that the column

can not be null (e.g., col IS NOT NULL). Since that criterion must be validated for each row

satisfying the original criteria, there will be a slight cost. On the other hand, in some cases the

perfonnance may improve because an optimizer such as ORACLE'S optimizer may now be able to

perfonn an anti-join. These preferences will be referenced in the algorithms below.

I. Aliases

When a subquery is merged or moved to a FROM clause, the criteria in the surrounding

query should be fully identified, via an alias, for example, as will be shown by reference to the

following example.

1. SELECT * FROM employees

2. WHERE emp_seq IN 3. (SELECT emp_seq

4. FROM time_sheets t, projects p

5. WHERE t.rpt_date = 'Ol-MAY-98'

6. AND t.proj_seq = p.proj_seq

7. AND p.stop_date IS NULL)

8. AND hiredate > SYSDATE - 10

In this example, an alias should be added for EMPLOYEES, and that alias should precede

all columns pertaining to that table, hi addition, the subquery should be given an alias. Adding an

"e" alias for EMPLOYEES, the transfonnation would then be:

SELECT e.* FROM employees e, (SELECT emp_seq FROM time_sheets t, projects p WHERE t.rpt_date = 'Ol-MAY-98' AND t.proj_seq = p.projjseq AND p.stop_date IS NULL) x WHERE e.emp_seq = x.emp_seq AND e.hiredate > SYSDATE - 10

If a subquery contains an aggregate, it should be given an alias before being moved into the

FROM clause. For example, the following SQL statement contains a "max" aggregate function in

the subqueiy (line 3).

1. SELECT * FROM sal iistory

2. WHERE (emp_seq, effective_date) IN

3. (SELECT emp_seq, max(effective_date)

4. FROM saljiistory

5. GROUP BY emp_seq) The SQL statement should thus be transformed to give the aggregate

'MAX(EFFECTIVEJDATE)" an alias as follows:

SELECT s.* FROM saljiistory s,

(SELECT emp_seq, max(effective_date) cl

FROM saljiistory

GROUP BY emρ_seq) x WHERE s.emp_seq = x.emp_seq AND s. effective date = x.cl

A question arises whether to allow the user to flag subqueries that are known to retum a

unique set via implicit knowledge.

For example,

SELECT * FROM employees e WHERE EXISTS

(SELECT null FROM projects p, assignments a

WHERE p.proj_seq = a.proj_seq

AND a.emp_seq = e.emp_seq

AND p.name = 'EXPLALN_SQL')

can be converted to:

SELECT * FROM employees e WHERE e.emp_seq IN

(SELECT a.emp_seq FROM projects p, assignments a

WHERE p.proj_seq = a.proj_seq AND p.name = 'EXPLALN_SQL')

or converted to:

SELECT e.* FROM employees e, projects p, assignments a WHERE e.emp_seq = a.emp_seq AND p.proj_seq = a.proj_seq AND p.name = 'EXPLALN_SQL'

Both cases are possible because the low level table in the subquery has the PK on PROJ_SEQ

and EMP_SEQ. Transitivity allows us to say there is an equality on PROJ 3EQ because of the

equality on NAME wliich also has a UNQ index. Therefore, the subquery is guaranteed to produce

one row per EMPLOYEES row, which means a merge is possible.

Now that some of the terminology and concepts used by the present system have been

explained, the following examples aid in describing algorithms that may be used by the present

system to transfomi SQL statements. Examples are shown for correlated and non-correlated

subqueries and are identified as such. A non-correlated subquery is a subquery that is not correlated

to the main query. That is, a non-correlated subquery may logically be executed before any row is

examined by the main query. In other words, the subquery is independent of the main query and

could exist as a query in its own right. A correlated subquery, on the other hand, depends on the

rows being examined by the main query.

1. The following SQL statement is a mix of correlated and non-correlated subqueries. In the

following SQL statement, the last subquery is a scalar subquery and since the surrounding subquery

has equality criteria on PROJ_SEQ and RPT_DATE, and the EMP_SEQ column is on the select list,

that subquery can be merged, even though the subquery on SALJIISTORY can not. This is a good

test of whether a nested query in the FROM clause was correctly identified as scalar.

SELECT * FROM employees WHERE emp_seq IN

(SELECT emp_seq FROM time_sheets t

WHERE proj_seq = (SELECT proj_seq FROM projects WHERE name = 'EXPLAIN SQL: DEVELOPMENT') AND rpt_date =

(SELECT MAX(effective_date) FROM sal_history s WHERE s.emp_seq = t.emp_seq))

should be transfonned to

SELECT e.* FROM employees e, time_sheets t, projects p,

(SELECT emp_seq, MAX(effective_date) coll FROM sal_history s GROUP BY emp_seq) x WHERE e.emp_seq = t.emp_seq AND t.proj_seq = p.proj_seq AND p.name = 'EXPLAIN SQL: DEVELOPMENT' AND t.emp_seq = x.emp_seq AND t.rpt_date = x.coll

2. The following example is non-con-elated and is slightly different from the immediately

previous example. The basic difference is the subquery on SALJHISTORY. In this case the

subquery initially is detennined to potentially return multiple rows. This means a run-time error may

be possible in the original query. This prevents merging or moving the subquery. That is, the

subsequery on SALJIISTORY should remain as is. However, since the criterion on RPT_DATE

has the equality operator, the merge of the subquery with T1ME SHEETS can still occur, and it will

potentially still encounter the same run-time error. In other words, the subquery on S ALJHISTORY

will not be moved since it can not be guaranteed to produce a scalar set. However, in case it does,

it will result in a single constant value for the criterion "rpt_date = ". The present system will

recognize the subquery as being equivalent to a constant, allowing the surrounding query to be

merged with it's surrounding query and carrying the nested select along. For example, the following SQL statement,

SELECT * FROM employees WHERE emp_seq IN

(SELECT emp_seq FROM time_sheets t WHERE proj_seq =

(SELECT ρroj_seq FROM projects WHERE name = 'EXPLAIN SQL: DEVELOPMENT') AND rpt_date =

(SELECT effective_date FROM saljiistory s WHERE sal > 100)

should be transformed to

SELECT e.* FROM employees e, time_sheets t, projects p

WHERE e.emp_seq = t.emp_seq

AND t.proj_seq = p.proj_seq

AND p.name = 'EXPLAIN SQL: DEVELOPMENT'

AND t.rpt_date =

(SELECT effective_date FROM saljiistory s

WHERE sal > 100)

3. The following example is correlated and deals with the ALL operator. If a subquery does

not retum any rows, the criterion may evaluate to TRUE. This means the subquery should be moved

only and done as an outer join where only the outer-join syntax is only applied to the initial

con-elation criteria. The original columns compared must account for no join match. The final

criterion added is evaluated after the outer-join takes place.

For example, the following SQL statement, SELECT emp_seq, birthdate FROM employees e WHERE birthdate < ALL

(SELECT birthdate FROM dependents d

WHERE e.emp_seq = d.emp_seq)

can be transfoπned to

SELECT e.emp_seq, e.birthdate FROM employees e,

(SELECT d.emp_seq, MLN(birthdate) coll, 1 col2 FROM dependents d

GROUP BY emp_seq) x WHERE e.emp_seq = x.emp_seq(+) AND (e.birthdate < x.coll OR x.col2 IS NULL)

The system should check to determine if any of the new outer-joined criteria are OR'ed to

anything, since that would produce a parse en-or. It may be desirable to have the database parse the

intennediate solution to see if the transfonnation should have been done. If it should not have been

done then nothing should be done with the subquery.

4. The following query is non-correlated and deals with subqueries that have GROUP BY

syntax. Even though the subquery has a GROUP BY clause, the SELECT list does not contain each

of the columns on the GROUP BY list. This means the potential does exist for duplicates unless the

DISTINCT keyword is added to the SELECT list of the subquery.

For example, the following SQL statement,

SELECT * FROM employees

WHERE emp_seq IN

(SELECT emp_seq FROM assignments GROUP BY emp_seq, proj_seq) can be transfonned to

SELECT e.* FROM employees e,

(SELECT DISTINCT a.emp_seq FROM assignments a GROUP BY a.emp_seq, a.proj_seq) x

WHERE e.emp_seq = x.emp_seq

5. A SQL statement may contain only aggregates, but no GROUP BY clause. Therefore it

does not require the DISTINCT keyword. This example is non-correlated.

The following SQL statement,

SELECT * FROM employees WHERE hiredate =

(SELECT MLN(effective_date) FROM saljiistory)

can be transfonned to

SELECT e.* FROM employees e,

(SELECT MLN(s.effective_date) coll FROM saljiistory s) x WHERE e.hiredate = x.coll

6. The following example is non-correlated and illustrates handling the NOT IN operator

when moving to the surrounding query. The basic points are that the system should convert to an

outer join and then look for rows that are the "outer-joined rows" (e.g., where a not nuUable column

on the select list of the subquery IS NULL (see the criterion "x.coll IS NULL" below)). A solution

is to add a constant to the Select list, such as 1, because that will work whether or not the subquery

has GROUP BY syntax. For example, the following SQL statement,

SELECT emp_seq FROM employees WHERE hiredate NOT IN

(SELECT effective_date FROM saljiistory)

can be transfonned to

SELECT e.emp_seq FROM employees e,

(SELECT DISTINCT effective_date, 1 coll FROM saljiistory) x WHERE e.hiredate = x.effective_date(+) AND x.coll IS NULL

7. The following example is similar to example 6 above and is also non-correlated.

However, in this example multiple columns are interfaced to the subquery. hi addition, the columns

represent multiple tables in the surrounding query. In this example, a problem may be encountered

since the transfonnation does not parse. An error message may be generated "ORA-01417: a table

may be outer joined to at most one other table". The reason is that X is outer-joined to

EMPLOYEES and DEPENDENTS. Accordingly, this transformation should not be made when the

relational operator requires an outer join but the columns interfacing the subquery are from multiple

tables.

For example, the following SQL statement,

SELECT e.emp_seq FROM employees e, dependents d WHERE (e.hiredate, d.birthdate) NOT IN

(SELECT hiredate, dl. birthdate FROM employees el, dependents dl

WHERE el.emp_seq = 1001

AND el .emp_seq = dl .emp_seq AND relation = 'SPOUSE') AND e.emp_seq = d.emp_seq AND d.relation = 'SPOUSE'

would be transfonned INCORRECTLY to

SELECT e.emp_seq FROM employees e, dependents d,

(SELECT DISTINCT el. hiredate, dl .birthdate, 1 coll FROM employees el, dependents dl

WHERE el.emp_seq = 1001

AND el.emp_seq = dl.emp_seq

AND dl. relation = 'SPOUSE') x WHERE e.emp_seq = d.emp_seq AND d.relation = 'SPOUSE' AND e.hiredate = x.hiredate(+) AND d.birthdate = x.birthdate(+) AND x.coll IS NULL

8. The following non-coιτelated example provides an interesting though strange situation.

This situation is seldom seen although, in theory, it can occur. The reason is that the subquery below

is non-correlated but interfaced to the surrounding query with the EXISTS operator. The EXISTS

operator is generally only used with a correlated subquery. In any event, the transformation is

simple; just move it to the surrounding query with no join clause, and add the criterion, "ROW NUM

= 1" as shown below, into the subquery.

For example, the following SQL statement,

SELECT * FROM employees WHERE EXISTS

(SELECT * FROM saljiistory

WHERE sal < 0) can be transformed to

SELECT e.* FROM employees e,

(SELECT * FROM saljiistory WHERE sal < 0 AND ROWNUM = 1)

The perfonnance will be tremendously better with the transformed SQL statement than with

the original SQL statement, since the subquery is now only executed once, instead of once per row

as would be perfoπned in the original SQL statement.

9. The following example is non-correlated and is listed because the subquery contains more

than one colunm name with the same name. Even though the column name is preceded by the table

alias, the column names should be given an alias. Also the asterisk on the SELECT list of the

surrounding query should be duplicated for each table in the surrounding query. Note the underlined

items in the transfonnation shown below.

The SQL statement,

SELECT * FROM employees e, dependents d WHERE e.emp_seq = d.emp_seq AND (e.birthdate, d.birthdate) IN

(SELECT el. birthdate, dl. birthdate FROM employees el, dependents dl

WHERE el.emp_seq = dl.emp_seq

AND el. hiredate = SYSDATE - 100)

can transfonned to

SELECT e.*. d.* FROM employees e, dependents d,

(SELECT DISTINCT el. birthdate coll, dl. birthdate col2 FROM employees el, dependents dl

WHERE el.emp_seq = dl.emp_seq

AND el. hiredate = SYSDATE - 100) x WHERE e.emp_seq = d.emp_seq AND e.birthdate = x.coll AND d.birthdate = x.co!2

10. The following example is non-correlated and illustrates a scalar subquery and the NOT

IN operator. Since the subquery is scalar, it can be merged with the surrounding as an outer-join. The

key is to retain only the "outer-joined" rows. This can be accomplished by ANDing a criterion with

a non-null column of the table in the subquery and the IS NULL operator. An even better solution

may be to use the ROWID (see the second transformation below).

For example,

SELECT emρ_seq FROM employees WHERE hiredate NOT IN

(SELECT hiredate FROM employees

WHERE emp_seq = 999)

can transformed to

SELECT e.emp_seq FROM employees e, employees el WHERE e.hiredate = el.hiredate(+) AND el.emρ_seq(+) = 999 AND el.emρ_seq IS NULL

or to SELECT e.emp_seq FROM employees e, employees el WHERE e.hiredate = el.hiredate(+) AND el.emp_seq(+) = 999 AND el.rowid lS NULL

11. The following con-elated example contains both correlation criteria and non-join criteria.

Since the subquery guarantees uniqueness the subqueiy can be merged.

For example, the SQL statement,

SELECT * FROM employees e WHERE EXISTS

(SELECT * FROM projects p, assignments a

WHERE ρ.proj_seq = a.proj_seq

AND e.emp_seq = a.emp_seq

AND p.name = 'EXPLAIN SQL: DEVELOPMENT')

can be transfonned to

SELECT e.* FROM employees e, projects p, assignments a

WHERE p.proj_seq = a.proj_seq

AND e.emp_seq = a.emp_seq

AND p.name = 'EXPLAIN SQL: DEVELOPMENT'

12. The following correlated example is different because the SELECT list contains columns

that are not returned to the surrounding query, (hi fact, if the subquery in the immediately preceding

example 11 had items on the select list, the transformation would still be exactly the same. Note that

in the iimnediately preceding example 11 the asterisk on the subqueries select list did not matter.)

For example, SELECT * FROM employees e WHERE EXISTS

(SELECT relation FROM dependents d

WHERE e.emp_seq = d.emp_seq)

can be transfonned to

SELECT e.* FROM employees e,

(SELECT DISTINCT emρ_seq FROM dependents d) x WHERE e.emp_seq = x.emp_seq

13. This correlated example is different from the immediately preceding example because

the SELECT list columns have corresponding columns in the surrounding query.

For example,

SELECT * FROM saljiistory si WHERE effective_date >

(SELECT MAX(effective_date) FROM saljiistory s2

WHERE sl.emp_seq = s2.emp_seq)

can transfonned to

SELECT si * FROM saljiistory si,

(SELECT emp_seq, MAX(effective_date) coll FROM saljiistory s2

GROUP BY emp_seq) x WHERE si .emp_seq = x.emp_seq AND si. effective date > x.coll 14. The following non-correlated example illustrates a non-correlated subquery interfaced

via the NOT EXISTS operator.

For example,

SELECT * FROM employees

WHERE NOT EXISTS (SELECT * FROM dependents)

can transfonned to

SELECT e.* FROM employees e,

(SELECT count(*) coll FROM dependents WHERE ROWNUM = 1) x WHERE x.coll = 0

Note that the SELECT list of the original subquery is changed to "COUNT(*)", and the

"ROWNUM = 1" criterion is added to the subquery.

15. The following non-correlated example is similar to the immediately preceding example

14, except that the subquery contains a GROUP BY. The transformation is similar to the

immediately preceding example 14.

For example,

SELECT * FROM employees WHERE NOT EXISTS

(SELECT null FROM saljiistory

GROUP BY emp_seq)

can transfonned to SELECT e.* FROM employees e,

(SELECT count(*) coll FROM sal_history

WHERE ROWNUM = 1

GROUP BY emρ_seq) x WHERE x.coll = 0

Whatever was on the SELECT list of the subquery is swapped with "COUNT(*)".

16. The following is a CORRELATED example using the NOT EXITS operator.

For example,

SELECT emp_seq FROM employees e

WHERE NOT EXISTS

(SELECT * FROM saljiistory s WHERE e.emp_seq = s.emp_seq AND effective ϊate = SYSDATE)

can be transfonned to

SELECT e.emp_seq FROM employees e, saljiistory s WHERE e.emp_seq = s.emp_seq(+) AND effective_date(+) = SYSDATE AND s.emp_seq IS NULL

17. The following correlated example illustrates a set subquery interfaced with the IN

operator.

For example, SELECT * FROM saljiistory si WHERE effective_date IN

(SELECT effective_date FROM saljiistory s2

WHERE sl.emp_seq = s2.emp_seq

AND sal > 100)

can be transfonned to

SELECT si.* FROM saljiistory si, saljiistory s2 WHERE sl.effective_date = s2.effective_date AND sl.emp_seq = s2.emp_seq AND s2.sal > 100

18. The following non-correlated example illustrates the < ALL operator.

For example,

SELECT * FROM employees

WHERE hiredate < ALL

(SELECT hiredate FROM employees WHERE birthdate > 'Ol-JAN-87')

can be transformed to

SELECT e.* FROM employees e,

(SELECT MIN(hiredate) coll FROM employees WHERE birthdate > 'Ol-JAN-87') x

WHERE (hiredate < x.coll OR x.coll IS NULL) Note the use of the "x.coll IS NULL" in the transformation. The point is that an aggregate

function will always retum a value even if no rows qualify the criteria. Since the ALL operator

always evaluates to TRUE when the subquery results in an empty set, the system should still return

the row in the surrounding query.

19. The following non-correlated example is similar to the immediately preceding example

18. However in this example, the > ALL operator is used instead of

< ALL.

For example,

SELECT * FROM employees

WHERE hiredate > ALL

(SELECT hiredate FROM employees WHERE birthdate > 'Ol-JAN-77')

can be transfonned to

SELECT e.* FROM employees e,

(SELECT MAX(hiredate) coll FROM employees WHERE birthdate > 'Ol-JAN-77') x WHERE (hiredate > x.coll OR x.coll IS NULL)

20. The following non-correlated example includes a set subquery that is merge-able.

For example,

SELECT * FROM employees WHERE emρ_seq NOT IN

(SELECT emp_seq FROM saljiistory

WHERE effective_date = '5-JUL-98') can be transfonned to

SELECT e.* FROM employees e, saljiistory s WHERE e.emp_seq = s.emp_seq (+) AND effective_date(+) = '5-JUL-98' AND s.emp_seq IS NULL

21. The following correlated example is similar to the immediately preceding example 20,

except correlated. The example illustrates how to add outer-join syntax to the correlation criteria.

For example,

SELECT * FROM assignments a

WHERE start_date NOT IN

(SELECT effective_date FROM saljiistory s WHERE a.emp_seq = s.emp_seq)

can transfonned to

SELECT a.* FROM assignments a, saljiistory s WHERE a.start_date = s.effective_date(+) AND a.emp_seq = s.emp_seq(+) AND s.emp_seq IS NULL

22. The following non-correlated example is similar to the immediately preceding example

21. However, in this example, the subquery has multiple tables. The example below illustrates how

to place the outer-join syntax on the join criteria within the subquery when merged to the

surrounding query. For example,

SELECT * FROM status Jist

WHERE status s NOT IN (SELECT p.status

FROM projects p, assignments a

WHERE a.proj_seq = p.proj_seq

AND p.name = 'EXPLAIN SQL: DEVELOPMENT')

can be transfonned to

SELECT s.* FROM status Jist s, projects p, assignments a

WHERE s. status = p.status(+)

AND p.name(+) = 'EXPLAIN SQL: DEVELOPMENT'

AND a.proj_seq(+) = p.proj_seq

AND a.ROWID IS NULL

In the subquery, uniqueness is guaranteed on the intermediately joined table (alias "p" in this

case since p.name is unique). At most 1 lowest level cliild in the join order can be non-unique, alias

"a" in this case. The join order is alias "s", then "p", then "a".

23. The following correlated example may be rather complicated. It should be an outer-join

because of the NOT LN operator. However, the problem is that both aliases "e" and "a" are

correlated with the subquery. The subquery should be mergable because the subquery contains only

1 table so we do not need to guarantee uniqueness.

For example,

SELECT * FROM employees e, assignments a WHERE e.emp_seq = a.emp_seq AND e.emp_seq NOT IN (SELECT t.emp_seq FROM time_sheets t WHERE a.proj_seq = t.proj_seq AND t.ιpt_date = '20-FEB-94')

The subquery can neither be merged nor moved because of the correlation criteria. If the

subquery' s con-elation criteria and interface columns reference more than one table in the

surrounding query, the subquery can be skipped. In this case the one correlation criterion references

alias "a." and the interface of the surrounding query references "e.emp_seq". If the system were to

transform the subquery, the transformation would end up with TIME_SHEETS being outer-joined

to more than 1 table which is not valid.

The above SQL statement would thus be INCORRECTLY merged to:

SELECT * FROM employees e, assignments a, time_sheets t

WHERE e.emp_seq = a.emp_seq

AND e.emp_seq = t.emp_seq(+)

AND a.proj_seq = t.proj_seq(+)

AND t.rpt_date(+) = '20-FEB-94'

AND t.emp_seq IS NULL

and is INCORRECTLY moved to

SELECT * FROM employees e, assignments a,

(SELECT t.emp_seq, t.ρroj_seq, 1 coll FROM time_sheets t

WHERE t.rpt_date = '20-FEB-94') x WHERE e.emp_seq = a.emp_seq AND e.emp_seq = x.emp_seq(+) AND a.proj_seq = x.proj_seq(+) AND x.coll IS NULL

Both of these cases result in an illegal outer-join and should be avoided. 24. This correlated example is similar to example 23 immediately above. Through

transitivity the system can transform the original statement as shown in example 23 to the following

original statement and then perfonns the transformation, thus eliminating the restriction encountered

in example 23.

For example, the original statement in example 23 can be transfonned to:

SELECT * FROM employees e, assignments a

WHERE e.emp_seq = a.emp_seq

AND a.emp_seq NOT IN

(SELECT t.emp_seq FROM time_sheets t WHERE a.proj_seq = t.proj_seq AND t.rpt_date = '20-FEB-94')

which in turn could be transfonned to

SELECT * FROM employees e, assignments a, timejsheets t

WHERE e.emp__seq = a.emp_seq

AND a.emp_seq = t.emp_seq(+)

AND a.proj_seq = t.proj_seq(+)

AND t.rpt_date(+) = '20-FEB-94'

AND t.emp_seq IS NULL

25. The following correlated example deals with the scalar operator, "=". This example

shows how transitivity in the subquery may be used to determine the subquery is scalar; e.g.

"e.emp_seq = t.emp_seq = a.emp_seq".

For example, SELECT * FROM time_sheets t, projects p WHERE t.proj_seq = p.proj_seq AND t.rpt_date =

(SELECT hiredate FROM employees e, assigmnents a

WHERE a.proj_seq = p.proj_seq

AND e.emp_seq = a.emp_seq

AND e.emp_seq = t.emp_seq)

could be transfonned to

SELECT * FROM time_sheets t, projects p, employees e, assignments a

WHERE t.proj_seq = p.proj_seq

AND t.rpt_date = e.hiredate

AND a.proj_seq = p.proj_seq

AND e.emp_seq = t.emp_seq

AND e.emp_seq = a.emp_seq

26. The following correlated example deals with the set operator, "IN'^'

For example,

SELECT * FROM employees e WHERE emρ_seq IN

(SELECT a.emp_seq FROM assignments a, time_sheets t

WHERE a.emp_seq = t.empjseq

AND a.proj_seq = t.proj_seq

AND a.proj_seq = 1

AND t.rpt_date = '20-FEB-94')

could be transfonned to SELECT * FROM employees e, assignments a, time_sheets t

WHERE e.emp_seq = a.emp_seq

AND a.emp_seq = t.emp_seq

AND a.proj_seq = t.ρroj_seq

AND a.proj_seq = 1

AND t.rpt_date = '20-FEB-94'

27. The following correlated example shows where the merge is possible although the

correlation criteria involves more than one of the tables in the subquery. In this case, 't." interfaces

"e." via the interface to the subquery; then "t." correlates with "a." via the PROJ_SEQ column.

Transitivity can be applied so "t" correlates with "a" via both the PROJ_SEQ and EMP_SEQ

columns. Moreover, uniqueness is guaranteed on "e" and "a" when joined.

For example,

SELECT * FROM time_sheets t WHERE t.rpt_date NOT IN

(SELECT hiredate FROM employees e, assignments a

WHERE a.proj_seq = t.proj_seq

AND e.emp_seq = a.emp_seq

AND e.emp_seq = t.emp_seq)

can be transfonned to

SELECT * FROM time_sheets t, employees e, assignments a WHERE t.rpt_date = e.hiredate(+) AND t.emp_seq = e.emp_seq(+) AND t.proj_seq = a.proj_seq(+) AND t.emp_seq = a.emp_seq(+) AND (a.emρ_seq IS NULL OR e.emp_seq IS NULL) Since "a" and "e" are joined to "t" separately, the system needs to add IS NULL to a non-

nullable column of each table and OR them together.

28. The following correlated example illustrates an aggregate on the SELECT list of a

subquery with a GROUP BY clause.

For example,

SELECT * FROM saljiistory s WHERE effective_date IN

(SELECT MAX(start_date) FROM assignments a

WHERE s.emp_seq = a.emp_seq

GROUP BY ρroj_seq)

can be transfonned to

SELECT * FROM saljiistory s,

(SELECT DISTINCT emp_seq, MAX(start_date) coll FROM assignments a

GROUP BY proj_seq, emp_seq) x WHERE s.emp_seq = x.emp_seq AND effective date = x.coll

29. The following non-correlated example illustrates transforming an aggregate into an

aggregate of an aggregate when the subquery is moved.

For example,

SELECT * FROM assignments a WHERE start_date < ANY

(SELECT MIN(rpt_date) FROM time_sheets GROUP BY projjseq)

can be transfonned to

SELECT * FROM assignments a,

(SELECT MAX(MLN(rpt_date)) coll FROM time_sheets

GROUP BY proj_seq) x WHERE a.start_date < x.coll

30. The following correlated example illustrates when a subquery interfaced through an

operator like "< ANY" cannot be transformed.

For example,

SELECT * FROM assignments a

WHERE start_date < ANY

(SELECT MLN(rpt_date) FROM time_sheets t WHERE a.emp_seq = t.emp_seq GROUP BY proj_seq)

would be INCORRECTLY transfonned to

SELECT * FROM assignments a,

(SELECT emp_seq, MIN(rpt_date) coll FROM time_sheets t

GROUP BY proj_seq, emp_seq) x WHERE a.start_date < x.coll AND a.emp_seq = x.emp_seq

This transfonnation is incorrect because the subquery is still a set, not a scalar as required,

e.g., per EMP_SEQ, there are still potentially many MTN( t_date) values due to the GROUPing by PROJ SEQ. Even if the subquery did not have the GROUP BY clause, the transfonnation would

still not be readily possible because the transformation would require putting the correlation columns

on the SELECT list and making the aggregate, and aggregate of an aggregate. That would end with

the syntax violation of a scalar and an aggregate of an aggregate.

Although the original GROUP BY column was not correlated nor part of a HAVING clause,

the system cannot get rid of it.

For example, the query could be incorrectly transformed to

SELECT e.* FROM employees e,

(SELECT emp_seq, MAX(rpt_date) coll FROM time_sheets t

GROUP BY t.emp_seq) x WHERE e.emp_seq = x.emp_seq AND e.hiredate < x.coll

The above transformation is incorrect because with the same PROJJSEQ, there could be a

rpt-date less than e.hiredate but another rpt_date greater than e.hiredate. In the original query, the

subquery with MLN(rpt_date) will retum the first data as on of the rows. However, in the above

query, the subquery will retum the maximum.

31. The following correlated example is an excellent example where the subquery can not

be moved because of the correlation criteria using the "<" operator, instead of the "=" operator.

For example,

SELECT * FROM saljiistory si, employees e WHERE e.emp_seq = sl.emp_seq AND sl.effective_date =

(SELECT MLN(s2.effective_date) FROM saljiistory s2

WHERE e.emp_seq = s2.emp_seq

AND e.hiredate < s2.effective_date)

would be INCORRECTLY transformed to

SELECT * FROM saljiistory si, employees e,

(SELECT s2.emp_seq, s2.effective_date, MLN(s2.effective_date) coll

FROM saljiistory s2

GROUP BY s2.emρ_seq, s2.effective_date) x WHERE e.emp_seq = x.emp_seq AND sl.effective__date = x.coll AND e.hiredate < x. effective date

One way to make a correct transfonnation would be to move the table that is involved in the

non-equality correlation, into the subquery as follows:

SELECT * FROM saljiistory si,

(SELECT e.emp_seq, MIN(s2.effective_date) coll

FROM saljiistory s2, employees e

WHERE e.emp_seq = s2.emp_seq

AND e.hiredate < s2.effective_date

GROUP BY e.emp_seq) x WHERE sl.emp_seq = x.emp_seq AND si .effective date = x.coll

The solution is to move the EMPLOYEES table into the subquery, and then move the

subquery.

32. The following correlated example is a scalar subquery without an aggregate.

For example, SELECT * FROM saljiistory s WHERE effective_date =

(SELECT hiredate FROM employees e

WHERE s.emp_seq = e.emp_seq)

can be transfonned to

SELECT s.* FROM saljiistory s, employees e WHERE s. effecti ve_date = e.hiredate AND s.emp_seq = e.emp_seq

33. The following correlated example is similar to the immediately preceding example 32.

However, in this example there is additional correlation criterion on column not part of a unique key.

Tins may be important when dealing with correlated subqueries; that is that the correlation criteria

over the PK or UNQ key of the subquery should utilize the "=" operator, and the operator used for

additional correlation criteria doesn't matter. In this case the correlation criterion

"t.rpt_date>=a.start_date" is not part of the unique key.

For example,

SELECT * FROM time_sheets t

WHERE rpt_date <

(SELECT stop_date FROM assignments a WHERE a.emp_seq = t.emp_seq AND a.projjseq = t.proj_seq AND t.rpt_date >= a.start_date)

can be transfonned to SELECT t* FROM time_sheets t, assignments a WHERE t.rpt_date < a.stop_date AND a.emp_seq = t.emp_seq AND a.proj_seq = t.proj_seq AND t.rpt_date >= a.start_date

34. The following correlated example utilizes a scalar operator to interface a scalar subquery,

where the subquery is scalar due to the aggregate and no GROUP BY clause.

For example,

SELECT * FROM employees e WHERE hiredate >

(SELECT MIN(effective_date) FROM saljiistory s

WHERE e.emp_seq = s.emp_seq)

can be transfonned to

SELECT e.* FROM employees e,

(SELECT emp_seq, MιN(effective_date) coll FROM saljiistory s

GROUP BY emp_seq) x WHERE e.emp_seq = x.emp_seq AND e.hiredate > x.coll

35. The following correlated example illustrates correlation in a HAVING clause. The

solution is that the HAVING clause becomes a criterion in the surrounding query, and the local item

compared in the correlated criterion should appear in the SELECT list of the subquery. It should be

noted that another syntax to check out will be where the subquery has no GROUP BY but does have

a HAVING clause. For example,

SELECT * FROM employees e WHERE NOT EXISTS

(SELECT a.proj_seq FROM assignments a, time_sheets t

WHERE a.proj_seq = t.proj_seq

AND a.emp_seq = t.emp_seq

AND e.emp_seq = a.emp_seq

GROUP BY a.proj_seq

HAVING MIN(t.rpt_date) < e.hiredate)

can be translated to

SELECT * FROM employees e,

(SELECT a.proj_seq, a.emp_seq, MLN(t.rpt_date) coll, 1 col2

FROM assignments a, time_sheets t

WHERE a.proj_seq = t.proj_seq

AND a.emp_seq = t.emp_seq

GROUP BY a.proj_seq, a.emp_seq) x WHERE e.emp_seq = x.emp_seq(+) AND x.coll < e.hiredate(+) AND x.col2 IS NULL

36. The following is a correlated example where the operator is the NOT IN operator and

there are non-equality correlation criteria. The system should notice that it doesn't have to move

another copy of the EMP_SEQ column onto the GROUP BY clause because it is already there.

For example,

SELECT * FROM saljiistory si WHERE sl.sal NOT LN (SELECT MLN(sal) FROM saljiistory s2 WHERE sl.emp_seq != s2.emp_seq AND sl.effective_date = s2.effective_date GROUP BY emp_seq)

can be transfonned to

SELECT si.* FROM saljiistory si,

(SELECT emp_seq, effective_date, MIN(sal) coll, 1 col2

FROM saljiistory s2

GROUP BY emp_seq, effective_date) x WHERE sl.sal = x.coll(+) AND sl.emp_seq != x.emp_seq(+) AND sl.effective_date = x.effective_date(+) AND x.col2 IS NULL order by sl.emp_seq, sl.effective_date

37. This correlated example is similar to example 30 above, except that in example 30 the

subquery does not have an aggregate. Though this query may not be practical, the example is used

to illustrate why a subquery with a GROUP BY should not be moved for any of the ANY, SOME

or ALL operators.

For example,

SELECT * FROM assignments a WHERE start_date < ALL

(SELECT rpt_date FROM time_sheets t

WHERE a.emp_seq = t.emp_seq

GROUP BY rpt_date

HAVING COUNT(*) > 1)

would be INCORRECTLY transformed to SELECT a.* FROM assignments a,

(SELECT MLN(rpt_date) coll, emp_seq, 1 col2 FROM time_sheets t

WHERE a.emp_seq = t.emp_seq

GROUP BY rpt_date, emp_seq

HAVING COUNTS) > 1) x WHERE (start_date < x.coll OR x.col2 IS NULL) AND a.emp_seq = x.emp_seq(+)

The problem is that there was already a grouping by RPTJDATE, and there is no join criteria

to the surrounding query. Thus we end up with a set per join over EMP_SEQ when we need a

scalar.

38. The following correlated example illustrates where a ANY, SOME or ALL subquery

cannot be moved due to inequality con-elation criteria. To be moved, the final transfonnation should

ensure that at most one row will be compared to the original surrounding query. However, with the

non-equality join criterion in the transformation, multiple rows may be joined.

For example,

SELECT * FROM saljiistory si

WHERE sal < ANY

(SELECT sal FROM saljiistory s2 WHERE sl.emp_^seq != s2.emp_seq AND sl.effective_date = s2.effective_date)

would be INCORRECTLY translated to

SELECT si.* FROM saljiistory si,

(SELECT emp_seq, effective_date, MAX(sal) coll FROM saljiistory s2) x WHERE si. sal < x.coll AND sl.emp_seq != x.emp_seq

AND si. effective date = x.effective date

39. The following is a correlated example with the ">ALL" operator. This example is

similar to example tliree above, except this example uses the MAX function instead of MIN function.

For example,

SELECT emp_seq, birthdate FROM employees e

WHERE birthdate >ALL

(SELECT birthdate FROM dependents d WHERE e.emp_seq = d.emp_seq)

can be transfonned to

SELECT e.emp_seq, e.birthdate FROM employees e,

(SELECT d.emp_seq, MAX(birthdate) coll, 1 col2 FROM dependents d

GROUP BY emp_seq) x WHERE e.emp_seq = x.emp_seq(+) AND (e.birthdate > x.coll OR x.col2 IS NULL)

40. The following correlated example illustrates the "<ANY" operator. The only difference

between the <ANY and the ">ANY" operator is that the MAX becomes a MIN.

For example,

SELECT * FROM saljiistory si WHERE emp_seq = 1001 AND sal <ANY (SELECT sal FROM saljiistory s2 WHERE si .emp_seq = s2.emp_seq AND sl.effective_date = s2.effective_date)

can be transfonned to

SELECT si.* FROM saljiistory si,

(SELECT emp_seq, effective_date, MAX(sal) coll FROM saljiistory s2

GROUP BY emp_seq, effective_date) x WHERE emp_seq = 1001 AND si. sal < x.coll AND sl.emp_seq = x.emp_seq AND si. effective date = x. effective date

41. The following correlated example illustrates the scalar operator, "=".

For example,

SELECT * FROM employees e WHERE hiredate =

(SELECT MLN(effective_date) FROM saljiistory s WHERE e.emp_seq = s.emp_seq)

can be transfonned to

SELECT e.* FROM employees e,

(SELECT emp_seq, MLN(effective_date) coll FROM sal_history s

GROUP BY emp_seq) x WHERE hiredate = x.coll AND e.emp_seq = x.emp_seq 42. The following correlated example illustrates the EXISTS operator where the subquery

contains a HAVING clause but no GROUP BY or aggregate. The point here is to add the GROUP

BY clause before the HAVING clause.

For example,

SELECT * FROM employees e

WHERE EXISTS

(SELECT null FROM saljiistory s WHERE e.emp_seq = s.emp_seq HAVING COUNT(*) > 1)

can be transfonned to

SELECT e.* FROM employees e,

(SELECT emp_seq FROM saljiistory s GROUP BY emp_seq HAVING COUNT(*) > 1) x

WHERE e.emp seq = x.emp_seq

43. This correlated example is similar to the immediately preceding example 42 where there

is a HAVING clause. However, in this example the GROUP exists as well.

For example,

SELECT * FROM employees e

WHERE EXISTS

(SELECT null FROM saljiistory s WHERE e.emp_seq = s.emp_seq GROUP BY effective date HAVING COUNT(*) > 1)

can be transfonned to

SELECT e.* FROM employees e,

(SELECT DISTINCT emp_seq FROM saljiistory s GROUP BY effective_date, emp_seq HAVING COUNT(*) > 1) x

WHERE e.emp_seq = x.emp_seq

44. The following non-correlated example illustrates when a subquery can not be moved for

the EXISTS operator. The reason for this is that the solution is to add the "ROWNUM = 1" criterion

to the subquery. However, the HAVING clause can disqualify the only row retrieved. In this

example there should be at least two rows qualifying else the subquery returns the empty set.

For example,

SELECT * FROM employees e WHERE EXISTS

(SELECT null FROM saljiistory s HAVING COUNT(*) > 1)

would be INCORRECTLY transfonned to

SELECT * FROM employees e,

(SELECT null FROM saljiistory s WHERE ROWNUM =1 HAVING COUNT(*) > 1) 45. The next correlated example illustrates subquery correlated to multiple levels above. This

example is included because in the section above, "Restriction", indicates that one should not

attempt to convert these type of correlated subqueries. The reason for this is that they are relatively

complicated, though they are possible as will be shown below.

For example,

SELECT * FROM employees e WHERE EXISTS

(SELECT null FROM assignments a WHERE a.emp_seq = e.emp_seq AND EXISTS

(SELECT null FROM time_sheets t WHERE t.emp_seq = e.emp_seq AND t.proj_seq = a.proj_seq))

can be transformed first to

SELECT e.* FROM employees e WHERE EXISTS

(SELECT null FROM assignments a,

(SELECT DISTINCT emρ_seq, proj_seq FROM time_sheets t) xl

WHERE a.emp_seq = e.emp_seq

AND xl .emp_seq = e.emp_seq

AND xl.proj_seq = a.proj_seq)

and next transfonned to

SELECT e.* FROM employees e,

(SELECT DISTINCT a.emp_seq coll, xl.emp_seq col2 FROM assignments a, (SELECT DISTINCT emρ_seq, proj_seq FROM time_sheets t) xl WHERE xl.proj_seq = a.proj_seq) x2 WHERE x2.coll = e.emp_seq AND x2.col2 = e.emp_seq

The first transfonnation is relatively simple. The subquery could not be merged, so it was

moved. This required the DISTINCT keyword in the Select list, plus the local columns that were

part of correlation criteria to be moved to the Select list. In addition, the correlated criteria were

moved to the surrounding query.

The second transfonnation is a bit more complex because now there are two columns from

different local tables of the subquery that had the same name but were correlated to the surrounding

queiy. That means unique aliases had to be given to the columns when they were moved to the Select

list. Then the system only had to move the correlated criteria to the surrounding query.

To better optimize this, the system should notice the transitivity of the criteria before the

move. For example, we had "a.emp_seq = e.emp_seq" and also "xl.emp_seq = e.emp_seq". This

is equivalent to saying "a.emp_seq = xl.emp seq". Recognizing this would have made the join in

the subquery more efficient and would have resulted in:

SELECT e.* FROM employees e,

(SELECT DISTINCT a.emρ_seq coll, xl.emp_seq col2 FROM assignments a, (SELECT DISTINCT emp_seq, projjseq FROM time_sheets t) xl

WHERE xl.proj_seq = a.proj_seq

AND a.emp_seq = xl.emp_seq) x2 WHERE x2.coll = e.emp_seq AND x2.co!2 = e.emp_seq

46. An OR operator is considered parent of a subquery. The following example illustrates

where a conversion to a join will produce a wrong answer. For example,

SELECT count(*) FROM employees WHERE hiredate > '01-jan-97'

OR emp_seq IN (SELECT emp_seq FROM saljiistory WHERE sal > 100)

If the OR is disregarded, the conversion will be to:

SELECT count(*)

FROM employees e, (SELECT DISTINCT emp_seq FROM saljiistory WHERE sal > 100) x

WHERE hiredate > 'Ol-jan-97'

OR e.emp_seq = x.emp_seq

One way to see the problem is to now convert the OR to a UNION ALL, as follows:

SELECT count(*)

FROM employees e, (SELECT DISTINCT emp_seq FROM saljiistory WHERE sal > 100) x

WHERE hiredate > '01-jan-97'

UNION ALL SELECT count(*)

FROM employees e, (SELECT DISTINCT emp_seq FROM saljiistory WHERE sal > 100) x WHERE hiredate > '01-jan-97'

47. The following example deals with OR'ed correlation criteria. If the correlation criteria

are OR'ed the conversion to an outer-join is impossible, and the conversion to a NOT IN is complex.

For example, SELECT * FROM employees e

WHERE NOT EXISTS (SELECT * FROM saljiistory s

WHERE s.effective_date = e.birthdate

OR s.effective_date = e.hiredate)

if the conversion to a NOT was performed, the system should do the following:

SELECT * FROM employees e

WHERE birthdate NOT IN (SELECT effective_^date FROM saljiistory s)

OR hiredate NOT IN (SELECT effective_date FROM saljiistory s)

if the conversion to an outer join is not impossible because it would produce un-parsable syntax:

SELECT e.* FROM employees e, (SELECT DISTINCT effective_date FROM saljiistory) x WHERE e.birthdate = x.effective__date(+) OR e.hiredate = x.effective_date(+)

48. The following example illustrates that a NOT IN to NOT EXISTS when original

subquery can produce empty set. The subquery was artificially forced to produce an empty set for

demonstration purposes.

For example,

SELECT * FROM employees WHERE hiredate NOT IN

(SELECT effective_date FROM saljiistory

WHERE 1=2)

should be transfonned as follows to ensure the same result set is produced: SELECT * FROM employees e WHERE NOT EXISTS

(SELECT * FROM sal history s

WHERE 1=2

AND e.hiredate = s.effective_date) AND (e.hiredate IS NOT NULL

OR

NOT EXISTS (SELECT * FROM saljiistory WHERE 1=2)

However, a problem with this solution is that we are back to the original again.

49. The following correlated example illustrates a HAVING clause with an aggregate

correlated to the surrounding query.

For example,

SELECT * FROM projects p WHERE EXISTS

(SELECT null FROM time_sheets t

WHERE t.proj_seq = p.proj_seq

GROUP BY emp_seq

HAVING MIN(t.rpt_date) = p.start_date)

This transfonnation may require moving the aggregate and correlation columns to the

SELECT list as follows:

SELECT * FROM projects p,

(SELECT proj_seq, MIN(rpt_date) coll FROM time_sheets

GROUP BY emp seq, proj_seq) x WHERE p.proj_seq = x.proj_seq AND p.start__date = x.coll 50. The following correlated example demonstrates a merge with NOT IN. Note that the

added criterion in the transformation contains the ROWID of the subquery table that is interfaced

to the surroimding query. It is important to use the ROWID of alias "t" rather than "a" because "t"

is the lowest level cliild in the join order, "e" is outer-joined to "a" and then "a" is outer-joined to

"t".

For example,

SELECT emρ_seq FROM employees e WHERE e.hiredate NOT IN

(SELECT start_date FROM assignments a, time_sheets t

WHERE e.emp_seq = a.emp_seq

AND a.emp_seq = t.emp_seq

AND a.proj_seq = t.proj_seq

AND t.rpt_date = 'Ol-jan-87'

AND a.proj_seq = 1)

can be transfonned to

SELECT e.emp_seq FROM employees e, assignments a, time_sheets t

WHERE e.hiredate = a.start_date(+)

AND e.emp_seq = a.emp_seq(+)

AND a.emp_seq = t.emp_seq(+)

AND a.proj_seq = t.proj_seq(+)

AND t.rpt_date(+) = 'Ol-jan-87'

AND a.projJseq(+) = 1

AND t.ROWID is null 51. The following correlated example is similar to the immediately preceding example 50,

except that the interface column of the subquery is not the same as the correlated table in the

subqueiy. For example START JD ATE belongs to ASSIGNMENTS but T.EMP_SEQ is correlated

but belongs to a different table. If transitivity is used we can switch T.EMP SEQ to A.EMP J3EQ

because of the join criteria between T and A.

For example,

SELECT emp_seq FROM employees e WHERE e.hiredate NOT IN

(SELECT start_date FROM assignments a, time_sheets t

WHERE e.empjseq = t.emp_seq

AND a.emp_seq = t.emp_seq

AND a.prcj_seq = t.proj_seq

AND t.rpt_date = '01-jan-87'

AND a.proj_seq = 1)

first add the transitivity to yield

SELECT emp_seq FROM employees e WHERE e.hiredate NOT IN

(SELECT start_date FROM assignments a, time_sheets t

WHERE e.emp_seq = a.emp_seq

AND a.emp_seq = t.emp_seq

AND a.proj_seq = t.proj_seq

AND t.rρt_date = 'Ol-jan-87'

AND a.proj_seq = 1)

we then have the same query as in the previous example 50, which means the same transformation

occurs. 52. This correlated example is similar to example 50 except now the table in the subquery

corresponding to the interface columns is TIME_SHEETS. The only difference is that

EMPLOYEES is outer-joined to TIME_^SHEETS and then TIME_SHEETS is outer-joined to

ASSIGNMENTS. Therefore, the added criterion uses ROWID of the ASSIGNMENTS table rather

than the TIME_SHEETS table as in example 50.

For example,

SELECT emp_seq FROM employees e WHERE e.hiredate NOT IN

(SELECT rpt_date FROM assignments a, time_sheets t

WHERE e.emp_seq = t.emp_seq

AND a.emp_seq = t.emp_seq

AND a.proj_seq = t.proj_seq

AND t.proj_seq = 1)

can be transfonned to

SELECT e.emp_seq FROM employees e, assignments a, time_sheets t

WHERE e.hiredate = t.rpt_date(+)

AND e.emp_seq = t.emp_seq(+)

AND t.emp_seq = a.emp_seq(+)

AND t.proj_seq = a.proj_seq(+)

AND t.proj_seq(+) = 1

AND a.ROWID is null

53. The following correlated example illustrates that an aggregate of aggregate is not a

possible translation.

For example, SELECT * FROM projects p WHERE stop_date IN

(SELECT MAX(MTN(ιpt_date)) FROM time_sheets t

WHERE t.proj_seq = p.proj_seq

GROUP BY emp_seq

HAVING MTN(t.rpt_date) = p.start_date)

would have to translated INCORRECTLY to

SELECT * FROM projects p,

(SELECT projjseq, MιN(rpt_date) coll, MAX(MTN(rpt_date)) col2 FROM time_sheets t GROUP BY emp_seq, proj_seq) x WHERE p.proj_seq = x.proj_seq AND p.start_date = x.coll AND p.stop_date = x.col2

However, such an SQL will not compile and should therefore be avoided.

54. The following correlated example illustrates that a non-equality con-elation criterion and

aggregate is not a possible translation.

For example,

SELECT * FROM assignments a

WHERE a.emp_seq IN

(SELECT emp_seq FROM time_sheets t WHERE t.proj_seq = a.proj_seq AND a.start_date > t.rpt_date GROUP BY emp_seq HAVING MAX(rpt_date) < a.stop_date)

would be INCORRECTLY transformed to SELECT * FROM assignments a,

(SELECT emp_seq, proj_seq, rpt_date, MAX(rpt_date) coll FROM timejsheets t

WHERE t.proj_seq = a.proj_seq

AND a.start_date > t.rpt_date

GROUP BY emp_seq, proj_seq, rpt__date,) x WHERE a.emp_seq = x.emp_seq AND a.proj_seq = x.proj_seq AND a.start_date > x.φt_date AND a.stop_date > x.coll

The problem here is that because of the non-equality criterion "a.start_date > t.φt_date", the

T.RPTJDATE will have to be moved to the SELECT list as well as the GROUP BY. However, the

aggregate can not also be moved to the SELECT list since it will contain the wrong group. In fact

the MAX will now be equal to RPT DATE.

55. The following correlated example also illustrates that an aggregate of aggregate is not

possible transfonnation.

For example,

SELECT * FROM assignments a WHERE a.stop_date IN

(SELECT MAX(MIN(ipt_date)) FROM time_sheets t

WHERE t.proj__seq = a.proj_seq

GROUP BY emp_seq)

would be INCORRECTLY transformed to

SELECT * FROM assignments a,

(SELECT proj_seq, MAX(MIN(φt_date)) coll FROM time_sheets t GROUP BY emp_seq, proj_seq) x WHERE a.stop_date = x.coll AND a.proj_seq = x.proj_seq

The problem is that you cannot have a scalar and an aggregate on the SELECT list.

Optimization

The following further lists the ways the present system uses to optimize an SQL statement

by generating altemative SQL statements, after the optimization mode has been selected by system

or user.. The hint for the optimization goal should be included in each SQL statement transformation.

The system will detennine all transitivity's in the non-join criteria, determine all join orders and

detennine all the various subquery transformations.

Non- Join Transitivity's

Briefly, looking at the following query,

SELECT * FROM employees e, assignments a, time_sheets t WHERE e.emp_seq = a.emp_seq AND a.emp_seq = t.emρ_seq AND a.proj_seq = t.proj_seq

AND e.emp_seq IN (SELECT emp_seq FROM dependents WHERE relation = 'SPOUSE')

In this example, the criterion on the subquery of DEPENDENTS can be interfaced to

"a.emp" or "t.emp_seq" just as well as the original which is "e.emp_seq". Other systems will not

recognize this. By switching to "a.emp_seq" the present system allows the database to utilize the

concatenated key index on ASSIGNMENTS. An altemative transfonnation would be to change "e.emp_seq" to "a.emp_seq" (line 5) as

follows:

1 SELECT * FROM employees e, assignments a, time_sheets t

2 WHERE e.emp_seq = a.emp_seq

3 AND a.emp_seq = t.emp_seq

4 AND a.proj_seq = t.proj_seq

5 AND a.emp_seq IN (SELECT emp_seq FROM dependents

6 WHERE relation = 'SPOUSE')

Since the operand, "e.emp_seq" can be replaced with two other values, the SQL can be

written tliree ways (one for the original and two for the number of different values that can replace

"e.emp_seq").

Multiple Non- Join Transitivities

The following describes how the system handles multiple non-join transitivities.

For example, given the SQL statement,

1 SELECT * FROM employees e, assignments a, time_sheets t

2 WHERE e.emp_seq = a.emp_seq

3 AND a.emp_seq = t.emp_seq

4 AND a.proj_seq = t.proj_seq

5 AND a.proj_seq = (SELECT proj_seq FROM projects

6 WHERE name = 'EXPLAIN SQL: DEVELOPMENT')

7 AND e.emp_seq IN (SELECT emp_seq FROM dependents

8 WHERE relation = 'SPOUSE')

The two operands "a.proj_seq" (line 5) and "e.emp_seq" (line 7) in the two different non-join criteria can be replaced with equivalent columns. Other systems will not recognize this. However,

the present system, by switching "e.emp-seq" (line 7) to "a.emp_seq" will allow the database to

utilize the concatenated key index on ASSIGNMENTS. The criterion with "e.emp_seq" is the same

as the last example. The present system can also replace "a.proj_seq" (line 5) with "t.proj_seq". This

means that there are a potentially 6 (2 X 3) different SQL statements that can be used (note that this

includes the original values).

Multi-Column Operands

What about operands that include multiple columns? For example:

1 SELECT * FROM employees e, saljiistory s

2 WHERE e.emp_seq = s.emp_seq

3 AND (s.emp_seq, s.effective_date) IN

4 (SELECT emp_seq, max(effective_date)

5 FROM saljiistory si

6 WHERE effective_date <= SYSDATE

7 GROUP BY emp_seq)

The criterion "s.emp_seq, s.effective_date" (line 3) represents the non-join criterion where

transitivity's can be applied. Since the join clause says "e.emp_seq" and "s.emp_seq" are

equivalent, there are two options for the non-join criterion operand. The first is the original and the

second is to replace "s.emp_seq" with "e.emp_seq".

The question is whether this should be done at all. The reason for the question is that both

columns in the operand reference the same table, and there is a concatenated index on both columns.

Even if there were only a single column indexes on both columns. The answer is not to worry and just add the other columns to the operand, and add a

con-esponding column to the select list as follows:

SELECT * FROM employees e, saljiistory s

WHERE e.emp_seq = s.emp_seq

AND "(e.emp_seq, s.emp_seq, s.effective_date) IN"

(SELECT "emp_seq", emp_seq, max(effective_date)

FROM saljiistory si

WHERE effective_date <= SYSDATE

GROUP BY emp_seq)

This simplifies the decision and lets the database decide which to use.

OR'ed Non- Jo in Criteria

What if OR'ed criteria exist in the query? If a non-join criterion is OR'ed to any of the join

criteria, the transitivities due to the OR'ed join criteria can not be used. The only transitivities that

can be used are those that are AND'ed.

The following SQL illustrates a query with an OR, but notice that all of the join criteria are

AND'ed to the non-join criterion.

SELECT * FROM employees e, assignments a, time__sheets t

WHERE e.emp_seq = a.emp_seq

AND a.emp_seq = t.emp_seq

AND a.ρroj_seq = t.proj_seq

AND (e.emp_seq IN (SELECT emp_seq FROM dependents

WHERE relation = 'SPOUSE') OR t.rpt_date = SYSDATE - 7) The following SQL does not use parenthesis to specify the precedence of the criteria. The

default, where ANDs are performed before ORs, is used.

1 SELECT * FROM employees e, assignments a, time_sheets t

2 WHERE e.emp_seq = a.emp_seq

3 OR e.emp_seq IN (SELECT emp_seq FROM dependents

4 WHERE relation = 'SPOUSE')

5 AND a.emp_seq = t.emp_seq

6 AND a.projjseq = t.proj_seq

In this case the only join criteria that are AND'ed to the non-join criterion are in lines 5 and

6. The transitivity where "e.emp_seq = a.emp_seq" can not be used. And since none of the join

criteria that are AND'ed to the non-join criteria reference "e.emp_seq", there are no transitive values.

Outer-Joins

What if a non-join criterion exists but there are outer-joins in the query? The next example

illustrates:

SELECT * FROM employees e, assignments a, time_sheets t WHERE e.emp_seq = a.emp_seq(+) AND a.emp_seq = t.emp_seq(+) AND a.proj_seq = t.prcj_seq(+)

AND e.emp__seq IN (SELECT emp_seq FROM dependents WHERE relation = 'SPOUSE')

Since the non-join criterion in question has an operand on "e.emp_seq", and there are no join

criteria where the equivalent column is from a non-outer-joined table, there are no transitive values. The following example is slightly different from the previous.

SELECT * FROM employees e, assignments a, time_sheets t WHERE e.emp_seq = a.emp_seq AND a.emp_seq = t.emp_seq(+) AND a.projjseq = t.proj_seq(+)

AND e.emp_seq IN (SELECT emp_seq FROM dependents WHERE relation = 'SPOUSE')

In this example, TIME_SHEETS is the only outer-joined table. Therefore in this example,

the transitivity "e.emp_seq = a.emp_seq" can be used but "e.emp_seq" can not be equated to

"t.emp_seq".

Transitivity of Simple Non- Join Criteria

The following query will be used to explain this concept.

1 SELECT * FROM employees e, dependents d

2 WHERE e.emp_seq = d.emp_seq

3 AND "d.emp_seq = 1001"

The non-join criterion is "d.emp_seq = 1001" (line 3). Due to the transitivity based on the

join criterion, the query could be rewritten as:

SELECT * FROM employees e, dependents d WHERE e.emp_seq = d.emp_seq AND "e.emp_seq = 1001"

Now the non-join criterion is on the EMPLOYEES table. This should be considered to be a single column operand with transitivity with two possible

values.

Transitivity of Non- Join Criteria and Nested Selects

This is one area of single column transitivity that can be very beneficial. If the nested select

does not have GROUP BY syntax or aggregates in the select list, and there is a non-join criterion

in the main query that references one of the join columns, the non-join criterion could be moved to

the nested select.

For instance, the non-join criterion is "e.emp_seq BETWEEN 1001 and 1010" (line 4).

1 SELECT * FROM employees e,

2 (SELECT DISTINCT emp_seq FROM dependents) x

3 WHERE e.emp_seq = x.emp_seq

4 AND "e. emp_seq BETWEEN 1001 and 1010"

Because of the join criterion and transitivity, the query can be rewritten as follows:

SELECT * FROM employees e,

(SELECT DISTINCT emp_seq FROM dependents WHERE "emp_seq BETWEEN 1001 and 1010") x

WHERE e.emp_seq = x.emp_seq

AND e.emp_seq BETWEEN 1001 and 1010

The point is not to consider this to be a single column operand that can take on multiple

values, but to simply duplicate the criterion in the nested select. Driving Tables

There can be queries with non-join criteria that are not indexed. For example, there is no way for an index search to start the query. There can be queries that do have indexed non-join criteria. And finally there can be queries with no non-join criteria. Another important consideration is whether the join criteria are indexed on both sides. All this must be considered along with what the optimization mode is.

Note the non-join criteria referenced below include those where one operand is a subquery.

Also note the potential driving tables should be detennined for all subqueries with more than one table as well as the main query. If Non- Join Criteria Exist

Determining a driving table is based on the optimization mode being used. FIRST ROWS

If the optimization mode is set to FIRST ROWS, then for table/view or nested select in From clause (that is not an outer-joined table), count the number of indexed non-join criteria and non- indexed non-join criteria. If the table has any indexed non-join criteria, then the table can be used to drive the query. If none of the tables have indexed non-join criteria, then the potential driving tables are those with any non-indexed non-join criteria. ALL ROWS

If the optimization mode is set to ALL ROWS, then for table/view or nested select in From clause (that is not an outer-joined table), just count the number of non-join criteria, whether indexed or not. The potential driving tables are those with any non-join criteria. (The count will be used later in determining the join order.) The following SQL statement will be used to illustrate both optimization mode

transformations.

SELECT * FROM employees e, dependents d, assignments a

WHERE e.emp_seq = d.emp_seq

AND e.emp_seq = a.emp_seq

AND Iname like 'SM%'

AND fname = 'ED'

AND relation = 'SPOUSE'

For the above SQL, if the optimization mode is FIRST ROWS, EMPLOYEES can be a

driving table because of the indexed non-join criteria on LNAME (we can forget about the fact that

FNAME is part of the concatenated index). The non-join criterion on RELATION is not indexed so

DEPENDENTS can not be a driving table. In addition, there are no non-join criteria on

ASSIGNMENTS. Accordingly, the only potential driving table is EMPLOYEES.

For the above SQL, if the optimization mode is ALL ROWS, both EMPLOYEES and

DEPENDENTS can be driving tables because they both have non-join criteria. Again,

ASSIGNMENTS does not have any non-join criteria and thus can not be a driving table.

If No Non- Join Criteria

If there are absolutely no non-join criteria, and we know the hierarchy either via referential

integrity, or though the user join (where the side using the unique index is the parent side, and the

side with the non-unique index is the child side), then the driving table should be the top-most

parent. And that should be for both FIRST ROW and ALL ROWS optimization modes.

For example, assume there is no referential integrity in the following example: SELECT * FROM employees e, dependents d, assignments a WHERE e.emp_seq = d.emp_seq AND e.emp_seq = a.emp_seq

The join between EMPLOYEES and DEPENDENTS uses a unique index on the

EMPLOYEES side and a non-unique on the DEPENDENTS side. Accordingly, EMPLOYEES is

considered the parent, and DEPENDENTS a child. The join between EMPLOYEES and

ASSIGNMENTS uses the unique index on the EMPLOYEES side. However, there is no index at

the ASSIGNMENTS side ( there is an index including EMP_SEQ but not as the leading column.)

The next section, "If Un-Indexed Join Criteria" will explain the significance of this.

Therefore for the example above, EMPLOYEES would be the only candidate for driving the

query (though you'll see in the next section that ASSIGNMENTS, due to the lack of join index, may

also be considered a potential driver.)

What about a query where more than one table can be considered the top-most table? The

following SQL illustrates.

SELECT * FROM employees e, dept Jiistory dh, departments d WHERE e.emp_seq = dh.emp_seq AND dh.dept_seq = d.dept_seq

In this case, both EMPLOYEES and DEPARTMENTS are parents of DEPT HISTORY.

Accordingly, both can be tested as driving tables. Un-Indexed Join Criteria

If a table's join is not indexed, then that table is a candidate for driving the query if the

optimization mode is ALL ROWS. Note that the table can drive an ALL ROWS query even if the

table doesn't have any non-join criteria whereas other tables do. If the optimization mode is FIRST

ROWS, a table like this is not considered unless there are non-join criteria on the table. For

example, use the previous rales of giving priority to tables with indexed criteria. The following

example illustrates:

SELECT * FROM employees e, assignments a, projects p WHERE a.prcj_seq = p.proj_seq AND e.emp_seq = a.emp_seq

ASSIGNMENTS in the above query does not have an index on the "a.proj_seq". Therefore

ASSIGNMENTS is also a candidate for driving the query in ALL ROWS optimization mode.

Outer Joins

If a table is outer-joined, it should not be used as a driving table. Accordingly, as illustrated

in the following example, even though there is a non-join criterion on DEPENDENTS,

DEPENDENTS still can not be used to drive the query.

SELECT * FROM employees e, dependents d WHERE e.emp_seq = d.emp_seq(+) AND d.relation(+) = 'SPOUSE' FROM Clause Nested Selects

The following query will illustrate the problem with outer joins. Since the nested select acts

as a subquery in a non-join criterion, the subquery should always be a driving table. In fact, all

nested SELECTS that are not outer-joined are joined first, whether or not they have non-join criteria.

In the following query, the best plan occurs when the subqueries are joined first, thereby

driving the selection of the historical records more efficiently. This is discussed more in the

ORDERED Hint section below.

SELECT e.emp_seq, e.lname, e.fhame, s.sal, d.dept_seq, j.job_seq FROM employees e, saljiistory s, jobjύstoryj, deptjiistory d,

(SELECT emp_seq, MAX(effective_date) coll FROM saljiistory

GROUP BY emp_seq) xl,

(SELECT emp_seq, MAX(effective_date) coll FROM job Jiistory

GROUP BY emp_seq) x2,

(SELECT emp_seq, MAX(effective_date) coll FROM deptjiistory

GROUP BY emp_seq) x3 WHERE e.emp_seq = s.emp_seq AND e.emp_seq = j.emp_seq AND e.emp_seq = d.emp_seq AND s.emp_seq = xl.emp_seq AND s.effective_date = xl.coll AND j.emp_seq = x2.emp_seq AND j.effective_date = x2.coll AND d.emp_seq = x3.emp_seq AND d. effective date = x3.coll

Views

Ideally, the system should check the view defimtion itself for non-join criteria, as well as the

query containing the view. If the view definition contains a view, you would not recursively check

the views. One level is enough and often that will be all that is needed. Also keep in mind that the view may be a join of multiple tables.

To simplify the process, consider the view as a potential driving table.

Join Orders

Before deciding on the join order, the system needs to know where to start. For example, what are the potential driving tables/views or nested selects in the From clause. (See Driving Tables above.)

The ORDERED hint will be used to specify the join order.

First, for the SQL statement that we will specify the join order for, test it without an ORDERED hint. For example, this will be the first pennutation (with no ORDERED hint) tested. The other permutations will be based on the various driving tables that were decided on. If a SQL statement has two tables that can drive the query, then there will be at least three pennutations; the original and two using the ORDERED hint with the different driving tables.

Pick one of the driving tables. Looking at the hierarchy between the objects in the FROM clause, check which join route will encounter the "most" non-join criteria (where FIRST ROWS looks for indexed non-join criteria if they exist, otherwise non-indexed non-join criteria; and ALL ROWS looks for just the most non-join criteria, whether indexed or not.) After all the objects are joined with non-join criteria, the other joins do not matter, so do not create variations of those tables.

The following examples will illustrate.

Example 1

The example below is for ALL ROWS optimization mode. SELECT e.* FROM employees e, time_sheets t, projects p,

(SELECT emp_seq, MAX(effective_date) coll

FROM saljiistory s

WHERE sal > 100

GROUP BY emp_seq) x WHERE e.emp_seq = t.emp_seq AND t.proj_seq = p.proj_seq AND p.name = 'PAFO' AND t.emp_seq = x.emp_seq AND t.rpt_date = x.coll

There is one non-join criterion on PROJECTS and one for the nested select. The hierarchy

for this example is shown in Fig. 10.

As shown in Fig. 10, the hierarchy depicts the nested select as a child of EMPLOYEES. This

works in this case because we know that SAL IISTORY is a child of EMPLOYEES. The dotted

line shows that the relationship is via transitivity, as will be further described later below.

Because of the transitivity, if PROJECTS is picked as the driving table, we would want to

join in the direction of the most non-join criteria. That means we want to get to the nested select as

quickly as possible. Therefore the join should be:

PROJECTS -> TIME__SHEETS -> nested select -> EMPLOYEES.

If the nested select is picked as the driver, the join should be:

nested select -> TIME_SHEETS -> PROJECTS -> EMPLOYEES

It turns out that all the non-join criteria in the previous SQL were also indexed, so the same join orders would be used.

Example 2

The following example is for FIRST ROW or ALL cOWS optimization modes:

SELECT * FROM employees e,

(SELECT DISTINCT emp_seq FROM dependents) x WHERE e.emp_seq = x.emp_seq AND e.emp_seq BETWEEN 1001 and 1010

The join criterion between EMPLOYEES and the nested select are indexed on both sides.

However, since the non-join criterion should be duplicated in the nested select (recall the section

above on transitivity), then both objects can drive the query.

Example 3

This query is intended to be used for FIRST ROWS optimization mode.

This is the standard historical data query used. The point here is that the non-join criteria

where one of the operands in each criterion is a subquery. Since the other operand in each case is

indexed, these are considered indexed non-join criteria.

SELECT e.emp_seq, e.lname, e.fname, s.sal, d.dept_seq, j.job_seq FROM employees e, saljiistory s, jobjiistoryj, deptjiistory d WHERE e.emp_seq = s.emp_seq AND e.emp_seq = j.emp_seq AND e.emp_seq = d.emp_seq AND s.effective_date =

(SELECT MAX(effective_date) FROM saljiistory si WHERE s.emp_seq = sl.emp_seq

AND effective_date <= SYSDATE) AND j .effective__date =

(SELECT MAX(effective_date) FROM job_history jl

WHERE j .emp_seq = j 1.emp_seq

AND effective_date <= SYSDATE) AND d.effective_date =

(SELECT MAX(effective_date) FROM dept_history dl

WHERE d.emp_seq = dl.emp_seq

AND effective_date <= SYSDATE)

The hierarchy for this example is shown in Fig. 11. The dotted lines represent the transitive

joins. Since each one of the history tables has an indexed non-join criterion, they can be joined

directly. Therefore any of the history tables can be a driving table. The different join orders, besides

the default, are as follows:

SAL HISTORY -> DEPT HISTORY -> JOB HISTORY -> EMPLOYEES

or

SAL HISTORY -> JOB HISTORY -> DEPT HISTORY -> EMPLOYEES

or

DEPT HISTORY -> SAL HISTORY -> JOB HISTORY -> EMPLOYEES or

DEPT HISTORY -> JOB HISTORY -> DEPT HISTORY -> EMPLOYEES

or

JOB HISTORY -> DEPT HISTORY -> SAL HISTORY -> EMPLOYEES

or

JOB JIISTORY -> SALJIISTORY -> DEPTJIISTORY -> EMPLOYEES

Example 4

Now, what if the subqueries in example 3 were moved to the FROM clause? Now the nested

selects can possibly drive the query. It is assumed that this is an ALL ROWS optimization mode.

SELECT e.emp_seq, e.lname, e.fname, s.sal, d.dept_seq, j.job_seq FROM employees e, saljiistory s, jobjiistoryj, deptjiistory d,

(SELECT emp_seq, MAX(effective_date) coll FROM saljiistory

WHERE effective_Jate <= SYSDATE

GROUP BY emp_seq) xl,

(SELECT emp_seq, MAX(effective_date) coll FROM jobjiistory

WHERE effective_^date <= SYSDATE

GROUP BY emp_seq) x2,

(SELECT emp_seq, MAX(effective__date) coll FROM deptjiistory

WHERE effective_date <= SYSDATE

GROUP BY emp_seq) x3 WHERE e.emp_seq = s.emp_seq AND e.emp_seq = j.emp_seq AND e.emp_seq = d.emp_seq AND s.emp_seq = xl.emp_seq AND s.effective_date = xl.coll AND j.emp_seq = x2.emp_seq AND j .effective_date = x2.coll AND d.emp_seq = x3.emp__seq AND d. effective date = x3.coll

The hierarchy for this example is shown in Fig. 12. Note that there are transitive join criteria

between the historical tables because of the transitivity of EMP_SEQ. For example, since

"s.emp_seq = xl.emp_seq", and "s.emp_seq = e.emp_seq", and "e.emp_seq = j.emp__seq", and

"j.emp_seq = x2.emp_seq", therefore "xl.emp_seq = x2.emp_seq".

In addition, note that there are no non-join criteria for the history tables, just the nested

selects. So none of the history tables can drive the query. In any case as mentioned above with

respect to the driving tables, the nested subqueries in the FROM clause can always be drivers.

Now we have even more possible join orders. To make it easier to view, aliases are given to

the nested selects; S.N. for the nested select on SALJIISTORY; DN for the nested select on

DEPTJIISTORY; etc. In addition, as noted previously, the system does not care about re-arranging

objects that have no non-join criteria.

S.N. -> DN -> JN -> SALJIISTORY -> DEPTJIISTORY -> JOB_HISTORY -> EMPLOYEES

S.N. -> JN -> DN -> SAL HISTORY -> DEPT HISTORY -> JOB HISTORY -> EMPLOYEES DN -> S.N. -> JN -> SALJIISTORY -> DEPTJIISTORY -> JOB ΪISTORY -> EMPLOYEES

DN -> JN -> S.N. -> SALJIISTORY -> DEPTJIISTORY -> JOB JΪISTORY -> EMPLOYEES

JN -> DN -> S.N. -> SALJIISTORY -> DEPTJIISTORY -> JOB JΪISTORY -> EMPLOYEES

JN -> S.N. -> DN -> SALJIISTORY -> DEPTJIISTORY -> JOB JIISTORY -> EMPLOYEES

Note again that the latter tables were constant in their order. The initial order could be

changed, but it would always remain constant for the various join orders.

Clustered Tables

If a table is clustered to one of the other tables in the query, those tables should be joined

together without intermediate joins of other tables. For example, ASSIGNMENTS, PROJECTS and

TIME SHEETS are in the same hash cluster. The following query illustrates:

SELECT * FROM employees e, assignments a, projects p, time_sheets t

WHERE e.emp_seq = a.emp_seq

AND a.emp_seq = t.emp_seq

AND a.proj_seq = p.proj_seq

AND a.proj_seq = t.proj_seq

AND t.rpt_date BETWEEN 'Ol-JAN-97' AND '30-JAN-97' The hierarchy for this example is shown in Fig. 13. The transitive join relationships are shown in Fig. 13 by dotted lines. Only ΗMEJSHEETS has a non-join criterion. So TIME_SHEETS will be the only driver. Because of the clustering and lack of non-join criteria anywhere else, all clustered tables should be joined next as follows:

TIME SHEETS -> ASSIGNMENTS -> PROJECTS -> EMPLOYEES

ORDERED Hint

The ORDERED hint always works. However, the system will be more efficient if the join criteria allow for the order of the join. Often this means adding join criteria that corresponds to the transitivity in the current join criteria.

When picking the join order, the system must check if the explicit join criteria exists. If not then the system should check via transitivity where the join criterion is implied. If it is, implied join criterion exists, that criterion should be added to the query prior to testing. Since parameters are already imbedded within the ORDERED hint, the implied join criteria should continue to be embedded in the ORDERED hint. The system handles the join order in this was when allowing the user to manually use the ORDERED hint, since it is not desirable to modify the SQL text for the other optimization modes.

For the example below, although there is no explicit join criterion between tables XI and X2, a join does exist via transitivity. The way to check is to follow the join paths from XI to X2. For example, from XI to S, the join is "xl.emp_seq = s.emp_seq AND xl.coll = s.effective_date". From S to E the join is over EMP SEQ also. This implicitly provides "xl .emp_seq = e.emρ_seq".

Continuing to follow the join path to X2, goes from E to D also via EMP_SEQ. It could thus be

said "xl.emp_seq = d.emp_seq". Finally the path goes from D to X2 via "d.emp_seq = x2.emp_seq

AND d.effective_date = x2.coll". Accordingly, the path is finally "xl .emp_seq = x2.emp_seq".

Since the ORDERED hint wants XI joined to X2, and the implicit join criterion, "xl.emp_seq =

x2.emp_seq" exists, that criterion should be added as a parameter to ORDERED.

The same process can be performed to get the transitive join criterion between X2 and X3.

For example, "x2.emp_seq = x3.emp_seq". Since X3 is not explicitly joined to E, the system

should include the join criterion "x3.emp_seq = e.emp_seq".

The suggested fonnat of the ORDERED hint with implicit join criteria is as follows:

SELECT/*+ ORDERED(xl,x2,x3,e,s,j,d, "xl.emp_seq=x2.emp_seq", "x2.emp_seq=x3.emρ_seq", "x3.emp_seq=e.emp_seq") */ e.emp_seq, e.lname, e.fname, s.sal, d.dept_seq, j.job_seq FROM employees e, saljiistory s, jobjvistoryj, dept_history d,

(SELECT emp_seq, MAX(effective_date) coll FROM saljiistory GROUP BY emp_seq) xl,

(SELECT emp_seq, MAX(effective_date) coll FROM jobjiistory GROUP BY emp_seq) x2,

(SELECT emρ_seq, MAX(effective_date) coll FROM deptjiistory GROUP BY emp_seq) x3 WHERE e.emp_seq = s.emp_seq AND e.emp_seq = j.emp_seq AND e.emp_seq = d.emp_seq AND s.emp_seq = xl.emp_seq AND s.effective_date = xl.coll AND j.emp_seq = x2.emp_seq AND j.effective_date = x2.coll AND d,emp_seq = x3.emp_seq AND d.effective date = x3.coll Nested Selects in FROM Clause

If a nested select is in the F ROM clause and non-join criteria reference it in the main SQL

module, the system will check out moving that criteria to the WHERE clause of the nested SELECT.

For example the SQL statement,

SELECT * FROM employees e, (SELECT DISTINCT emp_seq, relation

FROM dependents d WHERE birthdate > '01-jan-80') x

WHERE e.emp_seq = x.emp_seq

AND x.relation = 'SPOUSE'

can be transfonned to

SELECT * FROM employees e, (SELECT DISTINCT emp_seq, relation

FROM dependents d WHERE birthdate > 'Ol-jan-80' AND relation = 'SPOUSE') x

WHERE e.emp_seq = x.emp_seq

In this case, the non-join criterion is moved to the subquery to further reduce the number of

rows qualified prior to filtering later after the join.

It should be noted that although some databases may already account for this, moving the

criteria to the WHERE clause would be instractive, plus allows hints to be specified easier.

NOT'ed Logic

NOT'ed criterion are generally not considered herein in terms of optimizing because it

doesn't matter. That is, databases typically automatically handle the transposing of the NOT'ed

criterion. NVL Function

If the default parameter for the function does not equal the operand constant, then

WHERE NVL(cost,0) = 10 will evaluate to WHERE 0 = 10 when a NULL is returned.

Accordingly, since the default (0) is not equal to 10, the NVL function can be dropped, thus allowing

the criterion to use the index. If the criterion can not use an index do not do anything. If the column

is not NULLable, then do not, since there is no reason for the NVL function at all.

WHERE NVL(cost,0) = 0 should not be transposed unless other indexed criteria exist. For

example, if this did occur, the system would have to transpose to WHERE cost = 0 OR cost IS

NULL. That would still require a full table scan without other indexed criterion. For example, look

at the example below:

SELECT * FROM employees e, dependents d

WHERE e.emp_seq = d.emp_seq

AND e.hiredate = SYSDATE-7

AND NVL(d.birthdate,'01-JAN-80') = Ol-JAN-80'

can be transfonned to

SELECT * FROM employees e, dependents d

WHERE e.emp_seq = d.emp_seq

AND e.hiredate = SYSDATE-7

AND (d.birthdate = 'Ol-JAN-80' OR d.birthdate IS NULL)

now, this can be further transfonned to: SELECT * FROM employees e, dependents d WHERE e.emp_seq = d.emp_seq AND e.hiredate = SYSDATE-7 AND d.birthdate = 'Ol-JAN-80'

UNION ALL SELECT * FROM employees e, dependents d WHERE e.emp_seq = d.emp_seq AND e.hiredate = SYSDATE-7 AND d.birthdate IS NULL

It becomes more possible to optimize each SQL module (e.g., each nested query) separately.

In addition, note that the transfonnation does not include the standard inequality in the bottom

module. It does not have to, since the OR is with the same column. The OR'ed criterion will be

discussed in more detail below.

OR'ed Criteria

The following examples describe how the system handles OR'ed criteria.

For example, the SQL statement,

select * from tab where A=10 OR B=20

can be rewritten as

select * from tab where (A = 10)

UNION ALL

Select * from tab where (A != 10) AND (B = 20)

If columns A and B are nuUable, then the approach would be: select * from tab where (A = 10)

UNION ALL select * from tab where (A != 10 OR A IS NULL) AND (B = 20)

A WHERE clause written such as,

where A = 10 or (B = 20 AND C = 1)

can be rewritten as:

WHERE A = 10

UNION ALL

WHERE A != 10 AND ( B = 20 AND C = 1)

'IS NOT NULL' should be added if columns are nuUable.

New Plan Operations

BITMAP KEY ITERATION

Tins occurs when a subquery outputs multiple values that then goes to bitmap index. Check

if happens also when an IN, etc. Previously a BITMAP OR handled multiple values with IN

operator on bitmap index.

Other Hints

In data warehousing, a star schema can be used to depict one or more very large fact tables

that contain the primaiy information in the data warehouse and a number of much smaller dimension tables (e.g., lookup tables), each of which contains information about the entries for a particular attribute in the fact table.

A star query is a join between a fact table and a number of lookup tables. Each lookup table is joined to the fact table using a primary-key to foreign-key join. However, the lookup tables are not joined to each other.

The star transfonnation is a cost-based query transformation aimed at executing star queries efficiently. While a star optimization may work well for schemas with a small number of dimensions and dense fact tables, the star transformations may be considered as an altemative if any of the following is true. For example, star transformations may be useful if the number of dimensions is large or if the fact table is sparse, hi addition, the star transformations may be useful if there are queries where not all dimension tables have constraining predicates. The star transfonnation does not rely on computing a Cartesian product of the dimension tables, which makes it better suited for cases were fact table sparsity and/or large number of dimensions would lead to a large Cartesian product with few rows having actual matches in the fact table. In addition, rather than relying on concatenated indexes, the star transformation is based on combining bitmap indexes on individual fact table columns. The transfonnation can thus combine indexes corresponding precisely to the constrained dimensions. There is no need to create many concatenated indexes where the different column orders match different pattems of constrained dimensions in different queries.

"STARJTRANSFORMATION" - makes the present system use the best plan in which transfonnation was used. This occurs when there are bitmap indexes in the fact table and there are sufficient criteria on the dimension tables. Star transfonnation is not supported for tables with certain characteristics. For example,

tables with the following characteristics are not supported. Tables with a table hint that is

incompatible with a bitmap access path, tables with too few bitmap indexes (there should be a

bitmap index on a fact table column for the optimizer to consider generating a subquery for it),

remote 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 that have a good single-table access

path and tables that are too small for the transfonnation to be worthwhile.

The following example illustrates reducing the result of a nested select early in the query.

For example, if the user entered the query below with the nested select in the FROM clause.

SELECT e.* FROM employees e, time_sheets t, projects p,

(SELECT emp_seq, MAX(effective_date) coll FROM saljiistory s GROUP BY emp_seq) x

WHERE e.emp_seq = t.emp_seq

AND t.proj_seq = p.proj_seq

AND p.name = 'PAFO'

AND t.emp_seq = x.emp_seq

AND t.φt_date = x.coll

"AND x.coll < SYSDATE"

The user should have put the criterion with the nested select as shown below. That way, the

result of the nested select can be reduced early.

SELECT e.* FROM employees e, time_sheets t, projects p,

(SELECT emp_seq, MAX(effective_date) coll FROM saljiistory s GROUP BY emp_seq

"HAVING MAX(effective_date) < SYSDATE)" x WHERE e.emp_seq = t.emp_seq AND t.proj__seq = p.proj_seq AND p.name = 'PAFO' AND t.emp_^seq = x.emp_seq AND t.φt_date = x.coll

OBJECT JNSTANCE

If a table is not accessed during optimization, the system will be unable to discern which

tables are in the SQL statement. For example, in the following SQL statement, both

ASSIGNMENTS and PROJECTS are only represented in the ALL ROWS mode by their index

accesses. That is, the table does not need to be accessed. Accordingly, if a user clicks on the index

operation for ASSIGNMENTS nothing in the SQL would be highlighted to identify the relationship

between the plan and the SQL.

SELECT * FROM employees e,

(SELECT DISTINCT emp_seq FROM assignments a, projects p WHERE a.projjseq = p.proj_seq) x WHERE e.emp_seq = x.emp_seq

The system should at least check if the table is unique in the entire SQL statement. If it is,

then the system should highlight the table when the user clicks on the index operation.

OBJECT INSTANCE

When there is a nested select in the FROM clause, the present system does not account for

the OBJECT JNSTANCE value given to the nested select as well as the OBJECT JNSTANCE values given to the objects inside the nested select.

ALTERNATE SOL

If a user wants a FIRST ROWS mode it does not make sense in general to move a subquery

to the FROM clause. A merge is always good, but not necessarily for a move. Check the example

of historical data query for example. In fact, it might make sense to leave the correlated subqueries,

or even convert non-correlated to correlated.

When an ORDERED hint is added that requests a join order other than what the original SQL

specified, the system re-arranges the FROM clause and passes that to the database. However, when

the system re-arranges, it still shows the original ORDERED hint with the parameters. Since the

parameters are not part of the actual hint syntax, the system drops the parameters for the ORDERED

hint when displaying the HINTS mode for the SQL. For example, if the original SQL in the EDIT

window is:

SELECT/*+HINTl ORDERED(s,d,e,j) */ *

FROM employees e, saljiistory s, jobjiistoiyj, deptjiistory d

WHERE e.emp_seq = s.emp_seq

AND e.emp_seq = j.emp_seq

AND e.emp_seq = d.emp_seq

AND s.sal > 500

AND j.job_seq = 10

AND d.effective_date = j.effective_date

AND s.effective_date = j.effective_date the HINTS 1 window should display:

SELECT /*+ ORDERED */ *

FROM saljiistory s , deptjiistory d , employees e , jobjiistory j

WHERE e.emp_seq = s.emp_seq

AND e.emp_seq = j.emp_seq

AND e.emp_seq = d.emp_seq AND s.sal > 500

AND j.job_seq = 10

AND d.effective_date = j.effective_date

AND s.effective_date = j.effective_date

This would be consistent with the disappearance of the keyword HINTS 1, plus the FROM

clause change. That is, the SQL window for a specific mode should display exactly what would be

passed to the database. This should be done when join criteria is added to the ORDERED hint.

REPORTS

A useful report would be one that recommends which "indexed columns (single column

index)" might benefit from creating a histogram. Since a histogram should only be constructed on

columns that do not have a normal distribution, we need to find the columns that do not have normal

distribution.

"HINTS": For ALL ROWS, when the "joined to table" index is larger than the table, make FULL.

SOL Check

If a subquery is interfaced with the relational operator, NOT IN, or any form of ALL, the

system can display the following message and highlight, if possible, the subquery:

"If the subquery returns no row (empty set), the criterion including the subquery will evaluate to TRUE. I.e. it functions exactly the same as the NOT EXISTS when the subquery doesn't retum a row."

ALTERNATE SOL

When a subquery returns columns that will feed into criterion referencing the same table, the

select list of the subquery should be changed to contain the ROWTD instead. For example, look at the historical query below,

SELECT * FROM employees e, saljiistory si

WHERE e.emp_seq = sl.emp_seq

AND si . effective late =

(SELECT max(effective_date) FROM saljiistory s2 WHERE e.emp_seq = s2.emp_seq AND s2.effective_date <= SYSDATE)

could be transformed to

SELECT * FROM employees e, saljiistory si

WHERE e.emp_seq = sl.emp_seq

AND si. "ROWID" =

(SELECT "ROWID" FROM saljiistory s2 WHERE e.emp_seq = s.emp_seq AND s2.effective_date <= SYSDATE)

HINTS

When hints are applied, the user has to work at seeing if the hints took affect. The present

system can provide feedback to inform the user that the hints did have the desired affect. For

example, the system may infonn the user whether the ORDERED hint joined in the specified order.

For queries including NOT IN or NOT EXISTS subquery, transitivity can be applied so the

interface columns in the surrounding query belong to the same table. This merges NOT IN and NOT

EXISTS subqueries because a table cannot outer-join to more than one table.

For example the following SQL statement, SELECT *

FROM employees e, assignments a WHERE e.emp_seq = a.emp_seq AND (e.emp_seq, a.start_date) NOT IN (SELECT emρ_seq, φt_date

FROM time_sheets)

can be converted to

SELECT *

FROM employees e, assignments a WHERE e.emp_seq = a.emp_seq AND (a.emp_seq, a.start_date) NOT IN (SELECT emp_seq, φt_date

FROM time_sheets)

which can then be converted to

SELECT e.*, a.*

FROM employees e, assignments a, time_sheets tl WHERE e.emp_seq = a.emp_seq

AND a.emp_seq = tl.emp_seq (+)

AND a.start_date = tl.φt_date (+)

AND tl. ROWID IS NULL

The following example illustrates merging NOT IN and NOT EXISTS subqueries to join.

For example, the SQL statement,

SELECT * FROM employees e WHERE NOT EXISTS

(SELECT *

FROM assignments a

WHERE e.emp_seq = a.emp_seq ) will be converted to

SELECT /*+ ALL_ROWS */ e. * FROM employees e ,

(SELECT DISTINCT 1 col2 , a.emp_seq coll

FROM assignments a ) tl

WHERE e.emp_seq = tl.coll (+ ) AND tl.col2 IS NULL

however, the subquery should be merged as follows

SELECT e. *

FROM employees e , assignments a WHERE e.emp_seq = a.emp_seq (+ ) AND a.emp_seq IS NULL

The system can convert SQL 4 to SQL 5 and then to SQL 6, as follows, by moving the

constant in the surroiuiding interface column to the subquery.

For example, original SQL statement,

SELECT *

FROM employees e, assignments a WHERE e.emp_seq = a.emp_seq AND (e.emp_seq, '22-FEB-94') NOT IN (SELECT emρ_seq, φt_date

FROM time_sheets) SOL 4

can be transfonned to SELECT *

FROM employees e, assignments a WHERE e.emp_seq = a.emp_seq AND e.emp_seq NOT IN (SELECT emp_seq

FROM time_sheets WHERE ιpt_date = '22-FEB-94')

SOL 5

which can be transfonned to

SELECT e.*, a.*

FROM employees e, assigmnents a, time_sheets tl WHERE e.emp_seq = a.emp_seq

AND a.emp_seq = tl.emp_seq (+)

AND tl.φt_date (+) = '22-FEB-94'

AND tl .ROWID IS NULL SOL 6

The following query camiot be converted because the subquery table has to outer-join to a

table if merged. The subquery cannot be outer-joined to a constant alone.

SELECT *

FROM employees e, assignments a WHERE e.emp_seq = a.emp_seq AND '22-FEB-94' NOT IN (SELECT φt_date FROM time_sheets)

SOL 7

If a subquery interface column is a constant, merge can be perfonned if all tables are unique when joined. When merging the subquery, the constant Θ is set to NOT EQUALS to the

corresponding sun-ounding interface column and ORed with '<child>.ROWID IS NULL'

For example, the SQL statement,

SELECT * FROM employees e

WHERE (e.emρ_seq, lname) NOT IN (SELECT emp_seq, 'SMITH' FROM time_^sheets WHERE projjseq = 1 AND φt_date = '22-feb-94') SQL 8

can be transfonned to

SELECT e.* FROM employees e, time_sheets tl

WHERE e.emp_seq = tl.emp_seq (+) AND proj_seq (+) = 1 AND φt_date (+) = '22-feb-94' AND (tl. ROWID IS NULL

OR lname <> 'SMITH') SQL 9

Care must be exercised when attempting to transform SQL statements. For example, an

attempt can be made to transform the following SQL statement

SELECT * FROM employees e

WHERE emp_seq NOT IN (SELECT a.emp_seq FROM projects p, assignments a WHERE a.proj_seq = p.proj_seq AND p.name = 'EXPLAIN SQL: DEVELOPMENT') ORDER BY 1

into the following

SELECT/*+ ALL_ROWS */ e.*

FROM employees e, projects p, assignments a WHERE e.emp_seq = a.emp_seq (+)

AND a.proj_seq = p.proj_seq (+)

AND p.name (+) = 'EXPLAIN SQL: DEVELOPMENT'

AND p.ROWID IS NULL ORDER BY 1

However, this would not be correct, since it is not possible to merge the original SQL here

to the outer query. This would result in an incorrect transformation.

Another problem may occur if the system attempts to add the 'ROWID IS NULL' clause

with an arbitrary table. For example, sometimes the system might pick the projects table and

sometimes it might pick the assignment table, depending on the order in the FROM clause. This

would be incorrect. Care should be exercised to make certain that the lowest level cliild table is null

(e.g., the projects table). Otherwise, the join criteria to the child table (projects in this case), will not

be relevant.

The following rules and guidelines should be followed during transfonnation.

A) The present system should convert only if

i) All intennediate tables in the join order should be joined to guarantee uniqueness.

ii) There can be more than 1 branch. However, at most 1 lowest level child table can

be non-unique. All other tables should be unique when joined in order. In other words, when checking the above SQL, the assignments table is the table with the interfaced subquery column. Since it is not a lowest level child (projects table is joined to it), its columns in the select list or in the where clause join with a constant with an = operator must form a unique key. However, emp_seq is not a unique column. So this subquery cannot be merged. However, the Example SQL 10 below can be converted, when project name is a unique column. iii) If one of the subquery interface columns is a constant, all tables should be joined to guarantee uniqueness. iv) There is no aggregate function (GROUP BY, DISTINCT and HAVING is okay as long as there is no aggregate. When merged, ignore GROUP BY and DISTINCT and move HAVING criteria to WHERE clause) v) There is no CONNECT BY syntax vi) There is no set operation (e.g. UNION, MINUS) vii) Subquery contains no outer join.

B) For each lower level child, add '<child>.ROWID IS NULL' and OR them together.

C) If there is a constant in the surrounding interface column and there are other interface columns, change it to an equality criteria to the subquery where clause.

D) If there is a constant in a constant in the subquery interface column, set the constant to NOT EQUALS to the corresponding surrounding interface column and OR it with '<child>.ROWID IS NULL' SELECT * FROM status Jist WHERE status NOT IN (SELECT p.status

FROM projects p, assignments a WHERE a.proj_seq = p.proj_seq AND p.name = 'EXPLAIN SQL: DEVELOPMENT') SOL 10

Correlated subqueries may present a problem when determining whether or not an outer-join

can occur. Specifically, when the NOT IN operator is interfacing the correlated subquery, the

interface columns in the surrounding query must belong to the same table as the table correlated to

the subquery. For example, see the SQL statements immediately below. The interface columns in

the subquery should belong to the same table as the table correlated to the subquery. In the example

below, the HIREDATE column belongs to "e.", as well as one of the correlation criteria, but the

other correlation criterion references "a.". So the merge is not possible.

SELECT * FROM time_sheets t WHERE t.φt_date NOT IN

(SELECT hiredate FROM employees e, assigmnents a

WHERE a.proj_seq = t.proj_seq

AND e.emp_seq = a.emp_seq

AND e.emp_seq = t.emp_seq) SOL 11

It is not necessary that "the interface columns in the [NOT IN] subquery must belong to the

same table as the table correlated to the subquery", as long as rules/guidelines a) i) and a) ii) above

are true. However, the corresponding interface columns in the surrounding query should belong to

the same table because of the outer join. If converted, the assignment table needs to outer join to both time_sheets and employees. This is not legal. However, transitivity can be applied to convert

the SQL 11 statement shown above to SQL 12 shown below. Then SQL 12 can be merged to SQL

13 as shown below. The merge is possible because both employees and assignments have unique

indexes on the interface column(s).

SELECT * FROM trmejsheets t WHERE t.φt_date NOT IN

(SELECT hiredate FROM employees e, assignments a

WHERE a.proj_seq = t.projjseq

AND t.emp_seq = a.emp_seq

AND e.emp_seq = t.emp_seq) SQL 12

SELECT * FROM time_sheets t, employees e, assigmnents a WHERE e.hiredate(+) = t.rpt_date AND a.proj_seq(+) = t.proj_seq AND a.emp_seq(+) = t.emp_seq AND e.emp_seq(+) = t.emp_seq AND (a.rowid IS NULL or e.rowid IS NULL)

SQL 13

It should also be noted that you can not merge NOT EXISTS OR NOT IN subquery with

non-column (constant, function, etc.) in the subquery select clause. For example, an attempt may

be made to transfonn the following SQL 14 to SQL 15. However, SQL 15 does not produce the

same result set as SQL 14 because of t.rpt_date = '22-FEB-94'. The criteria could be changed to

t.rptjlate <> '22-FEB-94' . However, we would have to guarantee the original subquery returns

at least 1 row. This defeats the piu ose of the transfonnation, because we have to keep the subquery. SELECT * FROM time_sheets t

WHERE (t.emp seq, t.rpt_date) NOT IN (SELECT emp_seq, '22-FEB-94'

FROM employees) SOL 14

SELECT/*+ ALL_ROWS */ t*

FROM time_sheets t, employees tl WHERE t.emp_seq = tl.emp_seq (+)

AND t.rpt_date = '22-FEB-94'

AND tl. ROWID IS NULL

SQL 15

Set_NOT_rN_NOT_EXISTS Subquery_Can_Be_Ouerv

If an operator is NOT IN or NOT EXISTS, the following routine can be called attempt to

detennine whether it should attempt to convert the sub-query to a join.

«code begin»

IF subquery contains aggregate in Select list OR subquery contains aggregate in

HAVING clause OR subquery contains Connect By syntax OR subquery contains set operation THEN RETURN FALSE

LOOP over each table X in the FROM list, if there is more than 1 join criterion, apply transitivity so the table join to the same table, if possible.

Initialize Uniquejist to empty

NonJ nqiueJable := null

LOOP over each TABLE in the FROM list containing interface columns IF CheckTableUniquelndexCols retum TRUE

(CheclcTableUniquelndexCols routine checks for a given table, whether all the columns of its any one unique indexes are present in the where clause and have a '=' operation and the other operand is a constant or a correlated.) THEN add TABLE to the Uniquejist ELSE IF Non JJnqiue Jable is null THEN

Non_Unqiue Jable = TABLE ELSE

RETURN FALSE END LOOP over each TABLE in the FROM list containing interface coluimis LOOP over each TABLE in the FROM list containing interface columns

IF Check_Unique(TABLE, interface table of the surrounding query, Uniquejist, NonJJniqueJable) returns FALSE RETURN FALSE END LOOP over each table in the FROM list containing interface columns (Make sure all table is joined) LOOP over each TABLE

IF TABLE not in uniquejist and TABLE o Non_unique able THEN

RETURN FALSE END LOOP over each TABLE RETURN TRUE «code end»

CheclcJJnique

Input:

TABLE

PARENT JABLE

NONUNIQUEJTABLE

«code begin»

LOOP over each table Y joined to TABLE IF Y = PARENT ABLE continue loop IF Y is in Uniquejist or Y is Non_unique Jable (if Y is already joined, cannot join to another table) THEN RETURN FALSE LOOP over each index of Y

IF all indexed columns are joined to X with = operator OR index columns = constant

THEN add Y to Uniquejist IF CheckJJnique (Y, X, Uniquejist, Non_unique able) returns FALSE

THEN RETURN FALSE

BREAK (unique index found, no need to search anymore)

END LOOP over each index of Y IF Y not in Uniquejist

IF Non_uniqueJable is null Non_unique able = Y

(Since Y is not unique, make sure no other table is joined to Y) LOOP over each table Z joined to Y

IF Z not in Uniquejist and Z <> Non_uniqueJable THEN RETURN FALSE END LOOP over each table Z joined to Y ELSE

RETURN FALSE END LOOP over each table Y joined to TABLE RETURN TRUE «code end»

If a WHERE clause criteria only references table in the nested select in the FROM clause, move it

to the nested select.

For example, the following SQL

SELECT * FROM employees e,

(SELECT DISTINCT emp_seq, relation FROM dependents d WHERE birthdate > Ol-jan-80') x WHERE e.emp_seq = x.emp_seq AND x.relation = 'SPOUSE'

can be converted to

SELECT * FROM employees e,

(SELECT DISTINCT emp_seq, relation FROM dependents d WHERE birthdate > 'Ol-jan-80' AND x.relation = 'SPOUSE') x WHERE e.emp_seq = x.emp_seq

If a HAVING clause criteria does not involve group function, it can be moved to the

WHERE clause.

For example, the following SQL statement

SELECT LNAME , count (* ) FROM employees GROUP BY LNAME HAVING LNAME < 'D'

can be transformed to

SELECT LNAME , count (* ) FROM employees WHERE LNAME < 'D' GROUP BY LNAME

The present disclosure may be conveniently implemented using one or more conventional

general purpose digital computers and/or servers programmed according to the teachings of the

present specification. Appropriate software coding can readily be prepared based on the

teachings of the present disclosure.

In addition, although the above-description refers to specific database systems (e.g.,

Oracle), it should be understood that the present disclosure is not limited to any particular

database or type of database system.

Numerous additional modifications and variations of the present disclosure are possible

in view of the above-teachings. It is therefore to be understood that within the scope of the appended claims, the present disclosure may be practiced other than as specifically described herein.

Claims (33)

What is Claimed is:

1. A method of tuning a database query, comprising: selecting a database query; parsing the selected database query to determine relationships between portions of the selected database query; selecting an optimization mode from a plurality of available optimization modes; tuning the selected database query by modifying at least one portion of the selected database query based on the determined relationships and the selected optimization mode; and displaying the modified database query.

2. A method as recited in claim 1, wherein the parsing determines tokens withing the database query, tokens being words separated by delimiters.

3. A method as recited in claim 1 , wherein the plurality of available optimization modes include Cost based and Rule based modes.

4. A method as recited in claim 3, wherein the Cost based modes include First_Rows mode and AU_Rows mode.

5. A method as recited in claim 1, further comprising determining a cost associated with using the tuned database query.

6. A method as recited in claim 5, further comprising comparing a cost associated with using the selected database query to the cost associated with using the tuned database query.

7. A method as recited in claim 1, further comprising parsing the selected database query to detennine whether the database query includes at least one subquery joined by at least one of a NOT EXISTS, a NOT IN and an ALL clause.

8. A method as recited in claim 7, further comprising prompting a user to select preferences to be used during tuning based on whether the database query includes at least one of a NOT EXISTS, a NOT IN and an ALL clause.

9. A method as recited in claim 8, wherein the preferences include rewrite preferences for enabling a user to select at least one of a conversion of NOT EXISTS operators to a NOT IN operator and conversion of the selected database query to an outer join.

10. A method as recited in claim 8, wherein the preferences include rewrite preferences for enabling a user to select to convert subqueries joined by an ALL operator to a join or outer-join.

11. A method as recited in claim 8, wherein the preferences include rewrite preferences for enabling a user to select whether to use at least one of a NOT EXISTS operator and an outer-join to convert subqueries joined by a NOT In operator.

12. A computer storage medium including computer executable code for tuning a database query, comprising: computer executable code for allowing a user to select a database query; computer executable code for parsing the selected database query to determine relationships between portions of the selected database query; computer executable code for allowing a user to select an optimization mode from a plurality of available optimization modes; computer executable code for tuning the selected database query by modifying at least one portion of the selected database query based on the detennined relationships and the selected optimization mode; and computer executable code for displaying the modified database query.

13. A computer storage medium as recited in claim 12, wherein the parsing determines tokens withing the database query, tokens being words separated by delimiters.

14. A computer storage medium as recited in claim 12, wherein the plurality of available optimization modes include Cost based and Rule based modes.

15. A computer storage medium as recited in claim 14, wherein the Cost based modes include First Rows mode and All_Rows mode.

16. A computer storage medium as recited in claim 12, further comprising code for detennining a cost associated with using the tuned database query.

17. A computer storage medium as recited in claim 16, further comprising code for comparing a cost associated with using the selected database query to the cost associated with using the tuned database query.

18. A computer storage medium as recited in claim 12, further comprising code for parsing the selected database query to determine whether the database query includes at least one subquery joined by at least one of a NOT EXISTS, a NOT IN and an ALL clause.

19. A computer storage medium as recited in claim 18, further comprising code for prompting a user to select preferences to be used during tuning based on whether the database query includes at least one of a NOT EXISTS, a NOT IN and an ALL clause.

20. A computer storage medium as recited in claim 19, wherein the preferences include rewrite preferences for enabling a user to select at least one of a conversion of NOT EXISTS operators to a NOT IN operator and conversion of the selected database query to an outer join.

21. A computer storage medium as recited in claim 19, wherein the preferences include rewrite preferences for enabling a user to select to convert subqueries joined by an ALL operator to a join or outer-join.

22. A computer storage medium as recited in claim 19, wherein the preferences include rewrite preferences for enabling a user to select whether to use at least one of a NOT EXISTS operator and an outer-join to convert subqueries joined by a NOT In operator.

23. A programmed computer system for tuning a database query, comprising: a display for displaying at least one database queries to a user; a user input allowing the user to select a database query from among the displayed database queries and an optimization mode from a plurality of available optimization modes; and a processor for parsing the selected database query to detennine relationships between portions of the selected database query, and for tuning the selected database query by modifying at least one portion of the selected database query based on the determined relationships and the selected optimization mode, the modified database query being displayed to the user via the display.

24. A system as recited in claim 23, wherein the parsing determines tokens withing the database query, tokens being words separated by delimiters.

25. A system as recited in claim 23 , wherein the plurality of available optimization modes include Cost based and Rule based modes.

26. A system as recited in claim 25, wherein the Cost based modes include First_Rows mode and AUJRows mode.

27. A system as recited in claim 23, wherein the processor determines a cost associated with using the timed database query.

28. A system as recited in claim 27, wherein the processor compares a cost associated with using the selected database query to the cost associated with using the tuned database query.

29. A system as recited in claim 23, wherein the processor parses the selected database query to detennine whether the database query includes at least one subquery joined by at least one of a NOT EXISTS, a NOT IN and an ALL clause.

30. A system as recited in claim 29, wherein the processor prompts the user to select preferences to be used during tuning based on whether the database query includes at least one of a NOT EXISTS, a NOT IN and an ALL clause.

31. A system as recited in claim 30, wherein the preferences include rewrite preferences for enabling a user to select at least one of a conversion of NOT EXISTS operators to a NOT IN operator and conversion of the selected database query to an outer join.

32. A system as recited in claim 30, wherein the preferences include rewrite preferences for enabling a user to select to convert subqueries joined by an ALL operator to a join or outer-join.

33. A system as recited in claim 30, wherein the preferences include rewrite preferences for enabling a user to select whether to use at least one of a NOT EXISTS operator and an outer-join to convert subqueries joined by a NOT In operator.